proceedings of the land envsat workshop · discussions) during a one-day workshop convened in...

C S I R O L A N D a nd WAT E R

Proceedings of the Land EnvSat WorkshopAdelaide 25 August 2000

Edited by Tim R. McVicar

CSIRO Land and Water, Canberra

Technical Report 36/00, December 2000

Proceedings of the

Land EnvSat Workshop

10th Australasian Remote Sensing Photogrammetry Conference

Adelaide 25 August 2000

edited by

Tim R. McVicar

CSIRO Land and Water Canberra, ACT

December 200 0

© CSIRO 2000 All Rights Reserved This work is copyright. It may be reproduced in whole or in part for study or research subject to the inclusion of an acknowledgement of the source. Reproduction for commercial usage, training courses, or sale purposes requires written permission of the authors of each of the papers. Contact details for each lead author can be found in Appendix 1. ISBN 0 643 06085 5 Cover: The different spatial scales of data needed to develop and validate a model of understorey growth in woodlands. The photograph shows field measurement of biomass at a point scale, such values represent an area that is recognisable in LANDSAT TM imagery. The coloured image is a section of a LANDSAT TM scene showing (in yellow) the grazing experiment in which the photograph was taken. The black and white image is of the same area but at the spatial scale of the AVHRR instrument. AVHRR imagery has a high temporal density that allows unmixing of the tree and grass signals by time series analysis techniques, the results are shown in the graph. For details see Hume et al., this volume, p. 15. For bibliographic purposes, papers from this document may be cited as: Author, A.B., Author, C.D. and Author, E.F. (2000). Title of paper. In: Proceedings of the Land EnvSat Workshop, edited by T. R. McVicar, 10th Australasian Remote Sensing and Photogrammetry Conference, Adelaide, 25 August 2000, xx–yy. For additional copies of this publication please contact: Publication Officer CSIRO Land and Water PO Box 1666 Canberra, ACT 2601 Australia <[email protected]> This document is available for download in .PDF format from the WWW <http://www.clw.csiro.au/publications> Layout: Andrew Bell, Exact Editing. Phone (02) 6258 7276 or e-mail [email protected]

iii

Contents

Preface Regional ecohydrological indices: Monitoring moisture availability and drought in Australia and Papua New Guinea Tim R. McVicar, David L.B. Jupp and Phil N. Bierwirth 1 Estimating overstory and understory biomass in woodlands Iain H. Hume, Tim R. McVicar and Michael L. Roderick 15 Recent advances in the application of remote sensing for monitoring land surface fluxes William P. Kustas 29 Stitching the Australian 1-km archive Edward A. King 41 Interaction between surface and atmosphere in AVHRR shortwave channels Denis M. O’Brien, A. C. (Mac) Dilley and Mary Edwards 51 Pathfinder AVHRR Land NDVI data for Australia Jenny L. Lovell and R. Dean Graetz 61 Advances in CAPS Peter J. Turner and Harvey L. Davies 71 Summary of Discussions held during the EnvSat Workshop 81 Appendix 1 List of Attendees Address List

v

Preface These Proceedings include the papers presented (and a summary of associated discussions) during a one-day workshop convened in Adelaide on 25 August 2000. The workshop, which was attended by over 20 people (Appendix 1), brought together providers, operational users, and researchers of Environmental Satellites, abbreviated as EnvSats. The current EnvSats are AVHRR, the ATSR series, SeaWIFS, MODIS, and GMS (and other geostationary meteorological satellites). The unifying characteristic of these remote sensing systems is their ability to acquire data for large areas of the Earth’s surface at frequent intervals. The workshop was part of the 10th Australasian Remote Sensing and Photogrammetry Conference, and so there was a natural bias toward issues of the Australasian region. Nevertheless, we were pleased to have Dr Bill Kustas, from the United States Department of Agriculture – Agricultural Research Service, present his recent work and views. The visit to Australia by Dr Kustas was supported by the CSIRO Earth Observation Centre. Since the previous workshop in July 1998, we have seen the launch of MODIS, and increased interest in regional-scale applications of remote sensing. However, MODIS receives data in many more bands than AVHRR, and so the resulting image files are larger. While there are undoubted benefits in acquiring regional hyperspectral images, MODIS raises a problem: how do we, the individuals of the regional remote sensing community (and the organisations that employ us), deal with a decade-long time series of such images? Even now, archiving and curating 10 years of daily AVHRR data has raised difficulties that are only beginning to be resolved. Overcoming such problems will involve further development of an ‘end-to-end’ processing system, one that encompasses receiving station(s), through researcher(s), to operational user(s). Currently, the CSIRO Earth Observation Centre maintains a daily High Resolution Picture Transmission AVHRR data archive (nominally with a resolution of 1.1 km at nadir) for the whole of Australia and 2000 km surrounding it from 1992 until the present. The EOC also holds daily Global Area Coverage (GAC), nominally with an 8-km resolution, from 1981 until 1994. In sum, this means we have available 20 years of daily data for the entire Australian continent, a potentially valuable resource for many operational users and researchers. However, adequate institutional support must be provided to allow this data to be corrected, managed, and interpreted using algorithms and tools best suited for Australia’s climate, land surface, and institutional arrangements. The alternative, which would seem much less desirable, is to download subsets, covering Australia, of interpreted global images from US government WWW sites. In that case, US government agencies must be given credit as they have made a long time series of EnvSat data available. In Australia we need to put more energy into the development of an end-to-end processing system. Given the decreasing cost of terrabyte disk space, and the ever-increasing ability of Java-based interactive WWW interfaces, vast opportunities exist. In particular, I would like to see users being able to provide geographical information (either a point, line, or vector) to a WWW-based system and in return obtaining time

vi

series statistics for the last 20 years, or maps comparing conditions for some chosen period relative to the historical record. To be truly useful it would be ideal if this currently hypothetical system provided both corrected AVHRR data and geophysical properties derived from them. Examples of derived ecological variables could include percent greenness, percent vegetation cover, LAI, biomass, and moisture availability. I hope this vision will be shared by other members of the Australian EnvSat community, and that we can work together to achieve it. During the course of the workshop, 7 papers were presented: 1. McVicar, Jupp, and Bierwirth discussed the uses of time series of AVHRR data for monitoring regional hydrology. In Australia, regional moisture availability is determined by using daytime thermal AVHRR data in a “calculate then interpolate” approach. In Papua New Guinea, high amounts of cloud cover were overcome by combining both thermal and reflective data to monitor the 1997 drought. 2. Hume, McVicar, and Roderick reported on a project that aimed to quantify understorey biomass for regional woodland environments, a difficult proposition for remote sensing. To achieve this a plant-growth model was developed with remote sensing inputs at the LANDSAT TM scale. This will be validated using field-measured biomass. Evaluation of TM and AVHRR imagery suggests that geometric transformations in spectral space may be a reliable, effective means of linking LANDSAT TM and AVHRR data. These links overcome the “averaging problem” which precludes driving fine-scale models with coarse-scale data. 3. Kustas discussed recent advances in using the rate of change of surface temperature in conjunction with a Two-Source Time-Integrated model of atmospheric boundary layer growth. This approach reduces the need for absolute accuracy of surface temperature measurements, compared to methods that use remotely sensed data observed once during the day. The dual temperature-difference approach was tested with data from experimental sites: both ground and tower measurements and satellite imagery. 4. King outlined progress in stitching the 1-km continental daily AVHRR archive, which starts in 1992 and is still being acquired. Managing a 12-terrabyte (12288-gigabyte) data base has presented many difficult data-processing and IT issues for which solutions have been found. Since the same orbit is seen by several receiving stations, methods have been developed to use non-corrupted data in the final stitched image. Cross-checking of receiving station performance was also discussed. 5. O’Brien, Dilley, and Edwards discussed recent advances in simultaneously correcting AVHRR data for angular and atmospheric effects. Differences between corrected and uncorrected NDVI data were as large as 0.3 NDVI units. They showed that AVHRR data corrected for both effects simultaneously is closer to a 1:1 relationship than for either correction in isolation. 6. Lovell and Graetz reported methods to filter ‘10-day’ composite NDVI data sourced from the NASA/NOAA Pathfinder AVHRR Land (PAL) project. The

vii

filtering methods identify cloud-affected pixels which are used in the 10-day composite. The output of the filtering is a long time series, from July 1981 until September 1994, of NDVI data for the entire Australian continent free from cloud affects. 7. Turner and Davies presented recent advances made in the development of the Common AVHRR Processing Software (CAPS). CAPS promotes the production of standard calibrated and geolocated AVHRR data. Recent enhancements to CAPS included in the release of Version 2 include the ability to run in different operating systems and the provision of a FORTRAN and C subroutine interfaces. I wish to thank all speakers for taking the time to write about their research and to plan and present such good presentations. Given the historical context, I think it appropriate to paraphrase International Olympic Committee President Juan Antonio Samaranch and declare “this was the best EnvSat workshop ever”. After organising the last three AVHRR/EnvSat workshops, I look forward to someone from Brisbane, where the 11th Australasian Remote Sensing and Photogrammetry Conference will be held in 2002, volunteering to organise and host the next EnvSat workshop. Tim McVicar CSIRO Land and Water Canberra September 2000

1

Regional ecohydrological indices: Monitoring moisture availability and drought in Australia and Papua New Guinea Tim R. McVicar1, David L.B. Jupp2 and Phil N. Bierwirth3 1 CSIRO Land and Water, GPO Box 1666, Canberra, ACT 2601 2 CSIRO Earth Observation Centre, GPO Box 3023, Canberra, ACT 2601 3 Department of Geology, Australian National University, Canberra,

ACT 0200 Abstract: Regional water and land management agencies require spatial information to

make decisions to ensure that agri-ecological systems are sustainably managed. Regional

climate variability means that data which is updated often and is spatially complete offers

many advantages to extract ecohydrological indices. Regional remotely sensed data is an

ideal fundamental data source for providing ecohydrological near real-time information.

However, ground based data must be linked to the regional remotely sensed data to

provide validated interpretations. Recent examples using a time series of Advanced Very

High Resolution Radiometer (AVHRR) data are discussed in this paper. The two

examples are:

1. mapping moisture availability (ma), the ratio of actual to potential evapotrans-

piration (λλλλEa / λλλλEp), in the Murray–Darling Basin (MDB) using daytime land

surface temperatures (Ts); and

2. monitoring the 1997 drought in Papua New Guinea (PNG) using the ratio Ts/NDVI

derived from composite AVHRR data.

For these examples, data from most other EnvSats (if readily available) could have been

used; the main determinant governing the suitability of other EnvSats is that they acquire

both reflective and thermal data.

1. Introduction This paper briefly reports two recent applications that use time series AVHRR data to provide insight into regional ecohydrological processes. In the first example,

L a n d E n v S a t W o r k s h o p . . . . . . . . . . . . A d e l a i d e , A u g u s t 2 0 0 0

2

observations of Ts are linked to process understanding to map regional ma. This is achieved by using a Resistance Energy Balance Model (REBM, Monteith and Unsworth, 1990) at the isolated Australian Bureau of Meteorology (ABM) stations where at least daily air temperature extremes and rainfall are recorded. A ‘calculate then interpolate’ (CI) approach is then applied, using the remotely sensed data, which has high spatial density, as a covariate to interpolate ma away from the ABM stations. Different facets of this project are fully documented elsewhere (McVicar and Jupp, 1999a, Jupp et al., 1998, McVicar and Jupp, 1999b, McVicar and Jupp, 2000). In the second example, indices from composite AVHRR data are used to monitor drought conditions in tropical PNG. The ratio Ts/NDVI becomes higher during times of drought, due to the increase in Ts associated with more net radiation (Rn) being partitioned into the sensible heat flux (H) and the decrease in NDVI associated with decreasing amounts of vegetation cover (VegCov). We calculated the integral under the Ts/NDVI curve for 1997, denoted NDVI/T

Dec

Jan s∫ , which was plotted against the percentage area experiencing food supply problems so lives were at risk. A strong positive relationship (r2 = 0.82) is shown. This research is fully documented elsewhere (Bierwirth and McVicar, 1998, McVicar and Bierwirth, 2000).

2 Mapping Regional Moisture Availability Using Daytime Ts AVHRR Data

2.1 Introduction Many hydrologic and plant growth models have been developed at points. An approach frequently used to extend these models to regions is to interpolate all input parameters and driving variables, and then calculate at each location. Depending on the complexity of the model, and the spatial and temporal resolution and extent of the data, the interpolation can be a daunting task. This approach can be termed ‘interpolate then calculate’ (IC). Another less frequently used approach is to ‘calculate then interpolate’ (CI). The issue of IC or CI has received attention for spatially estimating moisture deficit (Stein et al., 1991), global solar radiation (Bechini et al., 2000), and soil properties (Heuvelink and Pebesma, 1999, Bosma et al., 1994). Here we summarise the use of the IC approach, which relies heavily on the inherent high spatial density of remote sensing, to monitor regional ma. This current study integrates data types with very different spatial and temporal scales. AVHRR data are spatially dense, with an at-nadir 1.1 km2 resolution, and are recorded over large areas in a matter of seconds at a specific time for specific wavelengths. This means for the extent of the image, remotely sensed data are a ‘census’. Depending on the amount of cloud coverage and the satellite repeat characteristics, optical remotely sensed data may only be available weekly or monthly. On the other hand, daily meteorological data are recorded sparsely with the points often separated by tens to hundreds of kilometres. The variables measured at these points represent a certain area. However, the exact area being represented by a given point measurement is

M c V i c a r e t a l . . . . . . . . . R e g i o n a l e c o h y d r o l o g i c a l i n d i c e s

3

unknown as the spatial autocorrelation is unknown. Thus, remotely sensed data is spatially dense but temporally sparse while meteorological data are spatially sparse but temporally dense. One aim of this research is to couple the high temporal density of meteorological data with the high spatial density of remotely sensed data.

2.2 Data Sets Used Remotely sensed data were acquired by the AVHRR sensor onboard the NOAA-9 and NOAA-11 satellites. The data archive, focussing on the MDB, consists of 97 AVHRR single overpass afternoon images from June 1986 until January 1994, recorded at approximately monthly time intervals. Our requirement for daily meteorological data are modest, only daily maximum (Tx) and minimum (Tn) air temperatures and daily rainfall (P) are essential. However in the 1.1 million km2 MDB there are only 63 ABM stations that have measured this data continuously from 1980 until the present (Fig. 1). To incorporate the influence of the north-west cloudband, a synoptic scale feature of the Australian region (Colls and Whitaker, 1990), 13 additional ABM stations to the north and west of MDB are included in the meteorological data base (Fig. 1).

SYDNEY

BRISBANE

MELBOURNE

CANBERRA

ADELAIDE

138o 142o 148o 150o 154o-39o

-36o

-30o

-23o

Q U E E N S L A N D

S O U T HA U S T R A L I A

N E W S O U T HW A L E S

A.C.T.

Murray River

Darling

River

MURRAY-DARLINGBASIN

AUSTRALIA

LEGENDBureau of Meteorology stationExtended NDTI areaFocused NDTI area

V I C T O R I A

Cobar

Figure 1. Location of the MDB, with sites of the 76 ABM stations are indicated. The focus and extended NDTI study areas are shown.


4

2.3 Development of the Normalised Difference Temperature Index (NDTI)

Ts is affected by a few environmental parameters. One controlling parameter is the partitioning of the available energy into the latent (λλλλE) and sensible (H) heat fluxes, which is in part determined by regional ma. The amount of energy partitioned into H is one determinant of the observed Ts. To remove the influence of time-of-day, day-of-year and geographic location from the Ts observation the NDTI has been developed, which has the form:

NDTI T TT T

s= −−

∞

∞ 0

where: T∞ is a surface temperature modelled as if there is an infinite surface resistance, that is, it is assumed that λλλλEa = 0 Wm-2; Ts is the AVHRR observed (in this case) surface temperature; and T0 is a surface temperature modelled as if there is zero surface resistance, that is, it is assumed that λλλλEa = λλλλEp. T∞ and T0 can be thought of as the physically-limited upper and lower temperatures respectively, for given meteorological conditions and surface resistances. They define an envelope within which meaningful AVHRR surface temperatures must fall. If Ts is close to the T0 value it is an indication of ‘wet’ conditions being. Whereas Ts close to the T∞ value suggests dryness (McVicar and Jupp, 1998). T∞ and T0 are calculated through the inversion of the REBM, and this is fully documented in Jupp et al. (1998). Meteorological variables required at the time of Ts acquisition include air temperature (Ta), incoming shortwave solar radiation (Rs), relative humidity (h) (or some other measure of vapor pressure (ea)) and wind speed (u). McVicar and Jupp (1999a) have tested and extended strategies to determine Ta, Rs, and h at the time of AVHRR data acquisition from only Tx, Tn and P. Daily wind run can be used to estimate u. Other surface variables including percent VegCov and surface albedo (α ) are obtained from reflective AVHRR data (McVicar and Jupp, 1999b). Vegetation height was measured when LAI was measured and can be obtained from simple GIS rules as a function of remotely sensed derived estimates of LAI.

2.4 Spatial Interpolation of the NDTI Using a CI Approach The NDTI is calculated at only those ABM stations, where data are recorded to support the calculation. Stations may be apart by up to 500 km. The NDTI surface was spatially interpolated using a commercially available spline interpolation package called ANU_SPLIN (Hutchinson, 1999). Three covariates were used to interpolate the NDTI: Ts - Ta; VegCov; and Rn (McVicar and Jupp, 1999b). Continuous grids which are used as covariates or in the modelling of Rn, were developed from:


5

(i) interpolated meteorological data (Ta, ea and effective beam transmittance at 0 m); (ii) remotely sensed data (Ts and α ); (iii) supervised classification of AVHRR reflective data used to stratify the time series of AVHRR reflective data (VegCov); and (iv) modelling Rn using all of the above six variables as inputs. The high spatial density of AVHRR data is conferred to the NDTI images. Figure 2 (a) shows the NDTI image for spring (22 September 1987) and figure 2 (b) shows the NDTI image for summer (25 December 1987). Changes in the NDTI between the two dates are shown in figure 2 (c), where blue is an increase, green fairly constant and red a decrease between the two dates. Figure 2 (a) shows the Menindee Lake system is the only area where the NDTI > 0.8, with the southern portion of wheat belt in the south east having a NDTI of about 0.5. Some areas around the Menindee Lakes, which are assumed to have received scattered rainfall, also have a NDTI of about 0.5. The larger extent of this pattern is confirmed by a decrease in Ts (McVicar and Jupp, 1999b). In figure 2 (b), some areas with a NDTI > 0.8 are lakes and some are associated with flood irrigation of crops and remnant River Red Gum forests (Barmah and Gulpa State Forests) along the Murray River. On the Cobar peneplain, there are scattered areas with NDTI values of approximately 0.5, which corresponds with areas where the NDTI increased between the two dates (Fig. 2 (c)). These areas have probably received scattered rainfall. The AVHRR derived estimates of VegCov are shown for spring (Fig. 2 (d)) and summer (Fig. 2 (e)). In figures 2 (d) and (e), the remnant deep rooted forests are identified by a stable VegCov between the two dates. This is especially seen for the remnant River Red Gum forests along the Murray River. In the woodlands of the Cobar peneplain, VegCov is mainly stable. However, there are some areas on the Cobar peneplain (the blue areas in Fig. 2 (f)) where VegCov has increased between the two dates. These areas are associated with an increase in the NDTI (Fig. 2 (c)). This increase in VegCov is most likely associated with an increase in the amount of grass cover. These grasses respond rapidly to rainfall and short term increases in available moisture and will be observable through the relatively open overstorey canopy (10 to 20%). The changes between the two dates in NDTI (Fig. 2 (c)) and VegCov (Fig. 2 (f)) are similar, though not identical. For example, along the north eastern border of both images there is a large decrease in VegCov due to harvesting of cereal crops, whereas in this area the NDTI remains fairly constant (it is less than 0.2 for both dates). To assist in regional agricultural management, including drought assessment, the resulting NDTI images and reflective based images, such as VegCov or the NDVI, and their interactions, need to be analysed in a spatio-temporal context. If the NDTI is viewed as an indicator of ma, and the NDVI is thought of as an indicator of moisture utilisation, then an opportunity exists to separate climatic induced variability from management induced variability (McVicar and Jupp, 1998). For example, in cropping or pasture agricultural systems, falling NDTI during a growing season in conjunction


6

with NDVI increases, would indicate that some of the decrease in ma was due to transpiration. However, there may be cases during the growing season where the NDTI declines while the NDVI fails to increase. Such a case could be the result of disease or insect damage early in the early growing season and the resulting decrease in ma may be solely due to soil evaporation. During crop growing seasons, these interactions may be best analysed by calculating the integrals of time series NDTI and NDVI images. The timing of the maximum NDTI and NDVI during crop growth periods would most likely have to be included.

1

0

100

0

(a) NDTI 22nd Sept 1987 (d) VegCov (%) 22nd Sept 1987

1 0

100 0

(b) NDTI 25th Dec 1987 (e) VegCov (%) 25th Dec 1987

0.5

-0.5

30

-30

(c) Change in NDTI (b) – (a) (f) Change in VegCov (%) (e) – (d) Figure 2. Images of the focus NDTI area shown in Figure 1 for: (a) NDTI for 22 September 1987; (b) NDTI for 25 December 1987 and (c) the difference of NDTI between these two dates, calculated as (b) – (a). (d) to (f) as for (a) to (c), except that VegCov is shown.


7

2.5 Validating a Water Balance by Co-Predicting a Surface Temperature In addition to using Ts to develop the NDTI, other opportunities exist to explicitly couple Ts with routines to optimise the parameters of water balance models. This is possible since λλλλEa is common to both water balance models and energy balance models. In water balance models, λλλλEa is defined in terms of a volume of water, usually expressed as a depth (millilitres per day). In energy balance models, λλλλEa is expressed in terms of energy (watts per m2). Water balance models based solely on meteorological data and soils properties, can produce estimates of soil moisture, λλλλEa, λλλλEp, and ma. Given suitable models for Rn and the ground heat flux (G), surface available energy (AE) can be calculated. This is be partitioned into either H or λλλλEa by simple energy balance (AE = Rn – G = H + λλλλEa). The water balance derived ma can be used to estimate the proportion of AE partitioned into λλλλEa. The remaining AE is partitioned toward H, which can then be physically inverted to model a surface temperature, denoted, Ts WB (McVicar and Jupp, 1999a). For the entire time series at each ABM station, the difference between Ts AVHRR and Ts WB was minimised using a global optimisation technique called simulated annealing. This alters some water balance model operating parameters. To validate this method, it must be shown that the post optimised water balance better fits field measured hydrologic data.

2.6 Conclusions The NDTI allows links to regional water balance modelling without recourse to spatial interpolation of daily rainfall. We are not advocating CI in preference to IC, each is useful depending on the underlying issues present. However, we are advocating, and have presented, a CI method, which inherently uses the high spatial density of AVHRR data as the backbone for the spatial interpolation. Assessing spatial and temporal interactions between the NDTI and VegCov or NDVI will provide useful information about regional ecohydrological processes, including agricultural management, within the context of Australia’s highly variable climate. Opportunities exist to alter water balance model parameters, by using remotely sensed data as ‘truth’.

3. Monitoring the 1997 Drought in Papua New Guinea (PNG)

3.1 Introduction During 1997, a strong El Niño event resulted in drought in PNG (Fig. 3) where both low plant water availability and frosts (restricted to areas above 1450 m) were experienced. In some cases, food production was severely curtailed and ceased. In extreme situations, particularly after severe frosts above 2200 m, food plants were


8

completely killed. This meant that populations relying on subsistence-based agriculture needed supplementary food supplies. Some 1.2 million rural villagers, almost 40 percent of the rural population, suffered severe, and in some cases fatal, food shortages. In PNG, this event is considered to be the ‘drought of the century’ and the associated food and water shortages strained emergency services beyond their limits.

Figure 3. Location of Papua New Guinea, its Provinces, and ground meteorological stations. The highland region is nominally comprised of five provinces: Southern Highlands, Enga, Western Highlands, Chimbu, and Eastern Highlands.

Much of PNG is covered by evergreen tropical rainforest, and shifting bush fallow cultivation practices are commonly used by the subsistence based agricultural communities that occupy 8% of PNG. This is where small gardens, commonly less than 0.1 ha, are used to grow plants which meet their food requirements. These gardens are surrounded by fallow land and land which is at some stage of vegetation regrowth. Commonly 5 to 15 years of regrowth elapse before the land is cultivated again (Saunders, 1993). For most of PNG, rainfall usually exceeds evapotranspiration and drought is a rare event (McAlpine et al., 1983). During the 1997 event, below average rainfall for up to 9 months was recorded at several locations. This caused a


9

decrease in the plant available water which, in turn, resulting in an increase in plant stress. During dry conditions, more Rs reaches the ground as there is less cloud cover, hence increasing Rn. More Rn is partitioned into H rather than λλλλEa, as there is less ma. This results in higher Ts, and increasing plant stress results in decreasing VegCov and a decrease in NDVI. It should be noted that Ts will increase slightly earlier than the corresponding NDVI decreases (McVicar and Jupp, 1998).

3.2 Data Sets Used Four-week composite AVHRR images were purchased from CSIRO Marine Research, Hobart, Australia, from January 1996 until December 1998. The original Local Area Coverage AVHRR data covering PNG was recorded at the Australian Centre for Remote Sensing (ACRES) satellite receiving station located at Alice Springs, Australia. The NDVI and Ts were calculated for pixels deemed to be clear of cloud, with Ts being calculated using a split window approach. Full processing details are documented in Bierwirth and McVicar (1998). To provide rapid assessment of the 1997 drought conditions in PNG, the negative correlation between Ts and NDVI (McVicar and Jupp, 1998, McVicar and Bierwirth, 2000) was utilised by dividing Ts by NDVI, denoted Ts/NDVI. Temporal variability of mean air temperature, daylength and evapotranspiration are small throughout the year in PNG (McAlpine et al., 1983). This means that much of the temporal variation exhibited in the Ts/NDVI can be attributed to changes in rainfall. Monthly precipitation data were available, however daily meteorological variables Tx, Tn and P were not.

3.3 Province-Level Analysis The mean Ts/NDVI value for each of the 14 mainland provinces has been plotted as a time series; six of these are presented here. Figure 4(a) illustrates that Western Province was the worst affected, as measured by the Ts/NDVI ratio, whereas the drought conditions experienced in the neighbouring Gulf Province (figure 4(b)) were not as severe. In addition to encountering a long period of stress during 1997, Milne Bay Province (figure 4(c)) also experienced a slight increase in stress in the middle of 1996 that is associated with rainfall seasonality (McAlpine, 1995). Figure 4(d) shows that West Sepik Province experienced little stress in 1997, as indicated by the Ts/NDVI response. Southern Highlands Province (SHP) was not as badly affected as Eastern Highlands Province (EHP), compare figures 4(e) and 4(f). It is interesting to note that during 1996 EHP has a higher base T /NDVI value than SHP, indicating EHP is generally s

drier than SHP. This is confirmed by spatial patterns of rainfall presented by (McAlpine et al., 1983); Bourke (1988) states “rainfall is lower in the eastern highlands and tends to increase in a westerly direction”. Chimbu Province also showed a similar elevated base Ts/NDVI value in 1996 (Bierwirth and McVicar, 1998).


10

Figure 4. Time series plots of mean T s/NDVI for: (a) Western Province; (b) Gulf Province; (c) Milne Bay Province; (d) West Sepik Province; (e) Southern Highlands Province; and (f) East Highlands Provinces. Refer to Figure 3 for the province locations. 8-week composites are used in 1996 and 4-week composites are used from January 1997.


11

3.4 Relating Ts/NDVI with Ground Observations of Subsistence-Based Food Supply

Two PNG-wide field-based assessments of subsistence based food supply and water supply were undertaken in 1997 (Allen et al., 1997, Allen and Bourke, 1997). These provided the basis for the emergency relief program. The assessment was, wherever possible, based on personal observations of the members of the teams. Almost all of the 520 Census Divisions in PNG were assessed. Phase 2 was completed in late November / early December. The subsistence based food supply categories during Phase 2 were defined by Allen et al. (1997) as follows: Table 1. Subsistence-based food supply categories (after Allen et al., 1997).

Category Definition

1 Unusually dry, but no real food supply problems.

2 Some inconvenience. Staple food is short but other food is available and health is okay.

3 Difficult, with food short and some famine food being eaten. Some babies and old people may be unwell. No lives at risk.

4 Very little food in gardens and it will not last for more than 3 weeks. Famine food being eaten by many people. There is said to be increasing sickness and there are signs of poor nutrition. The lives of small children and old people are at risk.

5 No food left in gardens. Only famine food available. An extreme situation with many people not eating, and small children and old people dying.

The percentage of each province experiencing subsistence-based food supply category

4 or 5 was calculated from the Geographic Information System (GIS) data base

collated after the Phase 2 survey. This is denoted % Cat4_5. For each of the 14

mainland provinces, the integral under the 4-week AVHRR composite Ts/NDVI curve

for 1997, denoted NDVI/TDec

Jan s∫ , was plotted against % Cat4_5. Figure 5 shows a

strong positive relationship between NDVI/TDec

Jan s∫ and % Cat4_5, where over 80%

agreement exists between the two variables. Additional validation of NDVI/TDec

Jan s∫ is

provided by plotting it against 1997 cumulative rainfall for meteorological stations that

recorded data continuously in 1997 (McVicar and Bierwirth, 2000).


12

Figure 5. Crossplot of NDVI/T

Dec

Jan s∫ and % Cat4_5. The numbers refer to the province

number, see Figure 3. The line of best fit is shown.

3.5 Conclusions The potential exists to use high frequency regional remote sensing in other tropical countries to track the onset of drought, especially for remote communities. Obviously such remotely sensed data and field based data are compatible and the advantages that each offer should be utilised in developing rapid operational drought monitoring systems in the tropics. For example, within PNG, remote sensing can be used to determine which provinces are experiencing drought conditions and field surveys could be used to assess which Census Divisions in the affected provinces require assistance. Provided access to the EnvSat data, with both thermal and reflective data, this style of analysis can be performed near real-time.

4. Overall Conclusions

In both examples, a time series of regional remotely sensed data, in these cases data from the AVHRR sensor, have been used to provide insights into regional ecohydrologic processes. Mapping regional ma relies on detailed process understanding at the points where ancillary meteorological data are available. Then, NDTI surfaces are generated using one strength of remotely sensed data: its high spatial density. Monitoring drought in PNG presented different problems: there were high amounts of cloud cover and no daily meteorological data was available. This


13

necessitated the development of the index Ts/NDVI, derived only from composite remotely sensed data. This was aggregated at the county level to provide information about 1997 drought conditions.

5. Acknowledgements The project to map ma in the MDB was supported by the Australian Centre for International Agricultural Research (ACIAR). The PNG Office of the United Nations Development Program (UNDP) supported the research evaluating the utility of AVHRR data to monitoring drought in PNG. Thanks to Iain Hume and Tom Van Niel for making comments on an earlier draft of this paper.

6. References Allen, B.J. and Bourke, R.M. (1997). Report of an Assessment of the Impacts of Frost and

Drought in Papua New Guinea. Papua New Guinea Department of Provincial and Local Government Affairs, Canberra, 20 pp.

Allen, B.J., Bourke, R.M., Burton, J., Flew, S., Gaupu, B., Heia, S., Igua, P., Ivahupa, S., Kanua, M., Kokoa, P., Lillicrap, S., Ling, G., Lowe, M., Lutulele, R., Nongkas, A., Poienou, M., Risimeri, J., Sheldon, R., Sowei, J., Ukegawa, K., Willson, N., Wissink, D. and Woruba, M. (1997). Report of an Assessment of the Impacts of Frost and Drought in Papua New Guinea – Phase 2. Papua New Guinea Department of Provincial and Local Government Affairs, Canberra, 27 pp.

Bechini, L., Ducco, G., Donatelli, M. and Stein, A. (2000). Modelling, interpolation and stochastic simulation in space and time of global solar radiation. Agriculture, Ecosystems and Environment, 81, 29-42.

Bierwirth, P.N. and McVicar, T.R. (1998). Rapid Monitoring and Assessment of Drought in Papua New Guinea using Satellite Imagery. Final Consultancy Report to United Nations Development Program, Port Moresby, Papua New Guinea. Australian Geological Survey Organisation, Canberra, 66 pp.

Bosma, W.J.P., Marinussen, M.P.J.C. and van der Zee, S.E.A.T.M. (1994). Simulation and areal interpolation of reactive solute transport. Geoderma, 62, 217-231.

Bourke, R.M. (1988). Taim hangre: variation in subsistence food supply in the Papua New Guinea highlands. In Department of Human Geography Australian National University, Canberra.

Colls, K. and Whitaker, R. (1990). The Australian weather book, Child & Associates, Sydney. Heuvelink, G.B.M. and Pebesma, E.J. (1999). Spatial aggregation and soil process modelling.

Geoderma, 89, 47-65. Hutchinson, M.F. (1999). ANUSPLIN Version 4.0 User Guide. ANU, Canberra. Jupp, D.L.B., Tian, G., McVicar, T.R., Qin, Y. and Fuqin, L. (1998). Soil Moisture and Drought

Monitoring Using Remote Sensing I: Theoretical Background and Methods. CSIRO Earth Observation Centre, Canberra, 96 pp.


14

McAlpine, J.R. (1995). Climate. In Papua New Guinea Inventory of Natural Resources, Population Distribution and Land Use Handbook, Vol. 6 (Eds, Bellamy, J. A. and McAlpine, J. R.) AusAID, Canberra, 162 pp.

McAlpine, J.R., Keig, G. and Falls, R. (1983). Climate of Papua New Guinea, ANU Press, Canberra.

McVicar, T.R. and Bierwirth, P.N. (2000). Rapidly Assessing the 1997 Drought in Papua New Guinea using Composite AVHRR Imagery. International Journal of Remote Sensing (in press).

McVicar, T.R. and Jupp, D.L.B. (1998). The current and potential operational uses of remote sensing to aid decisions on Drought Exceptional Circumstances in Australia: A Review. Agricultural Systems, 57, 399-468.

McVicar, T.R. and Jupp, D.L.B. (1999a). Estimating one-time-of-day meteorological data from standard daily data as inputs to thermal remote sensing based energy balance models. Agricultural and Forest Meteorology, 96, 219-238.

McVicar, T.R. and Jupp, D.L.B. (1999b). Using AVHRR data and meteorological surfaces to spatially interpolate moisture availability in the Murray-Darling Basin. CSIRO Land and Water, Canberra, 45 pp.

McVicar, T.R. and Jupp, D.L.B. (2000). Using covariates to spatially interpolate moisture availability in the Murray-Darling Basin: a novel use of remotely sensed data. Remote Sensing of Environment (in press).

Monteith, J.L. and Unsworth, M.H. (1990). Principles of Environmental Physics, Edward Arnold, London.

Saunders, J.C. (1993). Agricultural Land Use of Papua New Guinea: [1:1 000 000 map with explanatory notes]. PNGRIS Publication No. 1, AIDAB: Canberra, 12 pp.

Stein, A., Staritsky, I.G., Bouma, J., Van Eijsbergen, A.C. and Bgegt, A.K. (1991). Simulation of Moisture Deficits and Areal Interpolation by Universal Cokriging. Water Resources Research, 27, 1963-1973.

15

Estimating overstorey and understorey biomass in woodlands

Iain H. Hume1,2, Tim R. McVicar2, and Michael L. Roderick1

1 Ecosystems Dynamics Research Group, Research School of Biological Sciences, The Australian National University, Canberra, ACT 0200

2 Cooperative Research Center for Catchment Hydrology, CSIRO Land and Water, P.O. Box 1666, Canberra, ACT 2601

Abstract: An experimental approach is outlined to address the difficult problem of

measuring understorey growth in woodlands. The AVHRR GAC NDVI signal of grass

and trees are successfully unmixed by time series analysis. It is proposed that Landsat TM

imagery will be used to develop simple remote sensing models of pasture growth. These

models will be tested against field-measured growth in experiments identifiable at the TM

spatial scale. These models will use the unmixed grass NDVI to estimate the amount of

light absorbed by the grass and appropriate efficiency factors will be applied to estimate

growth.

Geometric transformation within image spectral space appears a promising technique for

translation between imagery of different spatial scales. These transformations have been

proposed using fine scale TM imagery (30 m pixel) which has been spatially degraded as

a surrogate of AVHRR imagery (1100 m pixel). These relationships need to be tested with

atmospherically corrected AVHRR imagery.

1. Background

A woodland contains trees that have a foliar projected cover (FPC) of less than 25% (Walker and Hopkins, 1984), much of Queensland is woodland (Figure 1). These woodlands are productive in that they support extensive cattle grazing between 2 and 50 cattle per 1000 ha (Carter et al., 1996). They are important from an agricultural and environmental perspective. Tothill and Gillies (1992) found widespread deterioration of Queensland's grazing lands and they noted that the most capable areas are most highly used and stressed. The work of Tothill (1992) stresses the need for resource monitoring as one key strategy in maintaining grazing lands in a sustainable condition.


16

Figure 1: Mean Foliar Projected Cover (FPC) for the calendar years of 1991, 1992 and 1993 calculated from the mean NDVI of those years using the regression equation FPC = –1.82 + 437.7 x NDVI (Danaher et al., 1992).

Condition assessments (for example Tothill and Gilles, 1992) are labor intensive, expensive, selective and slow. The information they provide is from a survey only and may not give timely warning of degradation trends. With appropriate choice of sensors remote sensing achieves high temporal density and wide extent of data acquisition and provides a complete census of the land surface.

Remotely sensed images are simply collections of systematically arranged measurements from the sensor. To measure individual scene elements the sensor's resolution cell size must be smaller than the size of those elements, an H resolution model (Strahler et al., 1986). Images formed from instruments with a resolution cell size larger than the scene elements belong to Strahler et al.’s (1986) L resolution case. These images contain pixels that are a spectral mix proportional to the spatial mix of elements in the sensor's resolution cell. Scene elements can be defined by image endmembers that are unique sets of image measurements.

There have been a number of attempts to use remote sensing to quantify pasture biomass in Australia. Filet et al. (1990) developed statistically significant regressions between the green biomass measured on the ground and the Normalised Difference Vegetation Index (NDVI) derived from the 1 km resolution local area coverage of the

0 1 0 00 100

H u m e e t a l . . . E s t i m a t i n g o v e r - a n d u n d e r s t o r e y b i o m a s s

17

National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Radiation Radiometer (AVHRR) instrument. The NDVI used in the regression was the monthly mean calculated from seven-day maximum NDVI value composites obtained during that month. Unique relationships were developed for each of the four main pasture communities of Queensland which occupy 62% (108 million ha) of the grazing area of that State. The overstorey cover of these communities ranged from zero in the Mitchell Grasslands and Aristida/Bothriochloa community thorough 3-5% in the black speargrass community to 25% in the mulga woodlands.

Carter et al. (1996) were unsuccessful in generalising Filet et al's (1990) relationships to develop Statewide estimates of green biomass. The Qld State Government have adopted an approach of mechanistic modelling of pasture growth to assess statewide pasture availability and identify exceptional circumstances. Remotely sensed NDVI was however retained as a test of the model output after its transformation into a 'synthetic NDVI'.

In Central Australia Hobbs (1995) found that maximum growing season NDVI and the NDVI amplitude were the best predictors of herbage biomass in exponential regressions. The integral beneath the NDVI time curve was a poor predictor of herbage. This was attributed to there being a high degree of shrub cover at the test sites and therefore in the NDVI signal. In Senagal Rasmussen (1998) measured as the sum of the annual increment of woody vegetation and herbaceous biomass which was called net primary production (NPP). The integral of NDVI over the growing season explained 71 % of the variation in above ground NPP, adding tree cover produced only a marginal improvement in the regression, explained only a further 7 % of the variation in NPP.

2. Unmixing overstorey and understorey biomass with low resolution satellite imagery

To unmix the growth and biomass of the tree and grass components of woodlands it is necessary to first "unmix" the remotely sensed signal recordeded by the satellite. At the spatial scale of remote sensing instruments woodlands are generally a mix of vegetation types, trees and grasses. This homogeneity makes unmixing by the application of linear mixture modelling problematic without the application of geo-optical modelling methods like those of Jupp and Walker (1996). In an alternative approach Roderick et al. (1999) used simple time series methods to unmix the NDVI signals of the perennial and annual vegetation. They interpreted the base of the time series trace as the signal from perennial vegetation and that annual vegetation causes departures from the base. Roderick et al. (1999) applied their methods to qualitative evaluation of changes in the long term perennial cover over the Australian continent. There has been no quantitative application of their or similar unmixing methods.

Roderick et al. (1999) calculated the fraction of photosynthetically active radiation absorbed by vegetation (F) was a linear transformation of the NDVI, this paper


18

considers NDVI (V) but the methods Roderick et al. (1999) hold because of the linearity between V and F and are reproduced here.

A time series of V is assumed to be composed of the sum of the evergreen (Ve) and raingreen (Vr) components: ( ) ( ) ( )e rV t V t V t= + (1)

An alternatively representation of the time series is as trend (T), seasonal (S), and random (R) components. ( ) ( ) ( ) ( )V t T t S t R t= + + (2)

The trend (T) in the time series (V) is estimated by the moving average of the raw time series. The seasonal component (S) was assumed to have a twelve month period and this is used for to estimate both the moving average of the time series and the seasonal component. T, S and R were estimated using the methods of Brockwell and Davis (1987). T is the moving average of the time series and S has both positive and negative values. The base of the trajectory in V is estimated by shifting T by a constant amount calculated as: ( )( )( ) ( ) minB t T t S t= − (3)

The base interpreted as the evergreen component, in this case trees, and the difference between V and B as the raingreen component due to understorey grass. A time series of NOAA Global Area Coverage NDVI data spanning the period 1981 to 1994 (Lovell and Graetz, 2000) were used to illustrate the unmixing in typical Queensland woodlands. The two locations chosen where QDPI have pasture measurements are Wambiana (20.5oS, 146.1oE) and Kielambete (23.4oS, 147.6oE).

Figure 2. Time series of Mixed, Tree, and Grass NDVI for the 8 km x 8 km NOAA GAC pixel

centered on the QDPI field experiment at Kielambete (23.4 oS, 147.6 oE).


19

Figure 3: Autocorrelation of the grass — and tree …… NDVI for the GAC pixel centered on Kielambete.

The raw, tree and grass NDVI at Kielambete are shown in Figure 2. The mean tree NDVI is consistently almost twice that of the mean grass NDVI and shows significantly less variation than the grass. Grass NDVI has a marked annual pattern with the highest positive autocorrelation at periods of 12, 24, and 36 months (Figure 3). Tree NDVI also showed no periodicity in the autocorrelation analysis (Figure 3). One major benefit of unmixing is that it allows comparisons of the different components of the woodland across the landscape and in time. There is clearly a larger NDVI signal at Kielambete than at Wambiana (Figure 4). The NDVI at both locations follow similar temporal trends suggesting that both locations experienced similar environmental and/or anthropogenic influences. Grass NDVI is much more dynamic than that of trees and differences in temporal trends are hard to identify in time series plots (Figure 4). Comparisons between grass NDVI are much easier on a seasonal basis, allowing inter-site comparisons in the same season (Figure 5a) and inter season comparisons at the same location (Figure 5b). Quantitative comparison of locations or times could be achieved by the by the application of metrics (DeFries et al., 1995) to the grass NDVI time series.

Figure 4. Time series at Kielambete — and Wambiana …… of the tree NDVI signal (top traces) and grass NDVI signal (bottom traces).


20

Figure 5. The grass component of NDVI (a) during 1991 at Kielambete — and Wambiana …

and (b) at Wambiana during 1985 — and 1991 ….

3. Modelling growth Monteith (1977) recognised that production of dry matter (DM) by agricultural crops in Britain was strongly correlated with the amount (f) of photosynthetically active radiation (PAR) they intercepted. The crops formed carbohydrate at a rate of about 1.4 g per MJ of intercepted energy, termed a conversion efficiency. Crop growth (DM) may be analysed simply in terms of fPAR during the growing season and the efficiency (ε) with which that light is used:

DM fPARε= ∑ (4)

The amount of interception depends on the seasonal distribution of leaf area which depends, in turn, on temperature and soil water supply. Kumar and Monteith (1981) showed that fPAR is proportional to the ratio between red and infrared radiation reflected from plant canopies. This forms the fundamental basis for the use of remotely sensed inputs to simple radiation absorption – conversion efficiency crop models, which are expressed in terms of some remotely sensed vegetation index (VI) and incoming photosynthetically active radiation (Rs) as:

( )DM a bVI Rsε= +∫ (5)

Where a and b are the constants of a linear relationship between the VI and f PAR. Myneni and Williams (1994) showed that a linear, scale invariant relationship exists between the fraction of photosynthetically active radiation absorbed by photosynthesising tissue in a vegetation canopy (fPAR) and the canopy normalised difference vegetation index (NDVITOC):

1.1638 0.1426TOCfPAR NDVI= − (6)

The reflectance of red radiation at the top of the atmosphere is higher than that at the earth's surface and that of near infrared is less (Kaufman, 1989). Thus the NDVI measured at the top of the atmosphere is less than that at the Earth's surface and this error will be transferred to any fPAR values it is used to calculate. There are many


21

ways of "correcting for the atmosphere" to calculate true reflectance on the ground from that measured at a satellite (see Kaufman, 1989). For the AVHRR instrument Dilley et al. (2000) have developed an operational system to correct red and near infrared reflectance measured at the top of the atmosphere to those at the earth's surface. The ground reflectance's have a standard, nadir, viewing geometry.

Simple "Kumar and Monteith" models have been applied at local, regional and global scales to predict both net and gross primary production. 3PG (Landsberg and Waring, 1997) is one such model used to estimate forest production. A variant which has been applied in a spatial context in Australia and New Zealand is 3PGS (Coops et al., 1998) which uses remotely sensed inputs of fPAR and GIS raster data for other climatic driving variables.

4. Temporal and Spatial Scaling Remote sensing can only infer characteristics of the Earth's surface through models or algorithms. These must be tested with accurate ground measurements of those inferred parameters or processes. The Queensland Department of Primary Industries (QDPI) has measured pasture growth under realistic grazing pressures over a number of years and locations. These experiments are also being used to calibrate a model of grass production GRASP (McKeon et al., 1990) developed by and the Queensland Department of Natural Resources (QDNR). Pasture production and tree basal area are measured on a regular basis in experimental plots large enough to be identified from space by Landsat TM (see Filet and Osten, 1996; O'Reagain et al., 1999). The difficulty in using this ground-based information is one of translating between the different spatial and temporal scale density and extent of the measurements.

4.1 Remote Sensing Model

An efficiency model will predict grass growth and fPAR is estimated from the NDVI derived from local area coverage (LAC) AVHRR data. The LAC images are acquired daily with an at nadir spatial resolution of 1.1 km. Ground truth data is measured on a scale of tens to hundreds of m2 in the QDPI experiments at Wambiana and Kielambete. These are controlled experiments in which different grazing pressures are applied to the land by stocking experimental paddocks at different densities. The amount of grass persisting under each treatment is measured only once per year at the perceived peak of the growing season, usually in April. Ground measurements provide a spatially accurate but temporally poor estimate of pasture amount that are too small to be identified as individual scene elements at LAC spatial scale.

In order to calibrate the GRASP model the growth of ungrazed pasture is also measured a number of times during the growing season. These provide insight into the temporal pattern of pasture production and will be used to constrain the temporal nature of modelled fPAR and growth.

Experimental treatments are clearly identifiable in Landsat TM imagery which has a spatial resolution of 30 m (Figure 6). This will allow the development of remote


22

sensing models of understorey biomass using the NDVI derived from Landsat imagery as the means of estimating fPAR. Biomass measured on the ground will allow the validation of remotely sensed biomass estimates and the remote sensing models to be tuned.

Figure 6. A portion of Landsat 7Enhanced Thematic Mapper scene acquired on 28 May 2000 and displayed as a false-colour image [R = band 4 (769–900nm), G = band 3 (630–690nm) and B = band 2 (520–600nm)] a gaussian stretch has been applied to each of the three bands. The external fence-lines of the Wambiana experiment are shown in yellow and the internal divisions between grazing-pressure treatments are clearly visible.

The paucity of cloud-free Landsat imagery during the summer growing season (Table 1) presents a challenge to the development of appropriate NDVI trajectories in time. More frequent estimates of ungrazed pasture biomass to calibrate the GASP model will serve to constrain the trajectory of NDVI, fPAR and growth within reasonable bounds.

Table 1. Dates (mm/dd/yy) when fine scale imagery was cloud free over the field experiments at Wambiana and Kielambete.

Season Kielambete Wambiana Season Kielambete Wambiana 1995/1996 10/31/95 10/29/95 1997/1998 10/20/97 11/3/97 01/19/96 03/21/96 03/13/98 11/15/97

03/23/96 06/25/96 06/01/98 03/11/98

06/27/96 07/11/96 06/17/98 06/15/98

07/29/96 08/12/96 08/18/98

1996/1997 11/18/96 09/13/96 1998/1999 12/26/98 11/22/98

02/22/97 04/25/97 05/03/99 05/1/99

05/12/97S 06/28/97 07/14/99 06/18/99

07/16/97 07/30/97 08/15/99 07/20/99

Note: All images are Landsat 5 TM except for that 1999/2000 12/05/99 11/9/99

marked S which is SPOT HRP and those marked 07/16/00ETM 05/28/00ETM

ETM which are Landsat 7 Enhanced Thematic 07/15/00ETM

Mapper scenes.


23

4.2 Image Scale and Resolution

It is therefore appears very feasible to develop fine scale remote sensing models of pasture growth using the field data gathered and imagery which have similar spatial and temporal scales. The high temporal density AVHRR data is spatially coarse and the fine scale experimental data and imagery become components of mixed pixels at the AVHRR spatial scale (Figure 7). The coarse scale mixed pixel is an average measure of the individual fine-scale scene elements. It has been shown that using coarse scale imagery to drive fine scale remote sensing models can cause large errors in the parameter being estimated (Chen, 1999). Linear mixture modelling is one strategy that can be employed to allow the use the coarse scale imagery. This can be applied in two ways. The first requires knowledge of the spectral properties of “pure” pixels which are then used to unmix the proportion of each pure land class within each mixed pixel. In its second application linear mixture modelling requires knowledge of the proportion of each fine scale scene elements within coarse pixels. The relative abundance of pixels of different types is then used to estimate the spectral properties of the individual scene elements. This second route dictates a classification of the landscape that must be valid over the modelling time frame. The data needed to pursue either of these linear mixture modelling avenues are not readily available.

A

B

Figure 7. NDVI images calculated from (A) the Landsat Enhanced Thematic Mapper (ETM+) image aquired on the 28 May 2000 and (B) as a result of spatial degradation of the ETM+ image to AVHRR spatial scale (1100 m pixels). In both figures the external boundary of the Wambiana experiment is shown in white. Both images are stretched linearly between 0.1 and 0.4 NDVI units.


24

Figure 8. The red and near infrared reflectance of the (ETM+) image acquired on the 28 May 2000 (+) and the red and near infrared reflectance of the same scene resampled to AVHRR spectral space (+).

The existence of spectral linkage between fine and coarse scale imagery would eliminate the need for complex mixture modelling. The relative size of scene elements and the sensor measurement cell size are a major factor influencing the success of spectral linkage across spatial scale. The strength of spectral linkage across image scale was tested using the ETM+ scene acquired on 28 May 2000. At the time of this analysis no atmospherically corrected AVHRR LAC data were available. As a surrogate the ETM+ image was spatially degraded to 1.1 km pixels. Exo-atmospheric reflectance of bands 3 (630–690 nm) and 4 (769–900 nm) of the ETM+ image were calculated from the quantized digital image using the methods of Markham and Barker (1987). An AVHRR pixel was assumed to enclose a window of 40 x 40 ETM+ pixels. The mean reflectance of 40 x 40 windows of band 3 and band 4 were calcu-lated and assumed to be the equivalent to band 1 and band 2 reflectance measured by the AVHRR instrument. This is recognised as a gross simplification since it ignores differences in the spectral bandwidths of the two instruments and the point spread function of the AVHRR instrument. The NDVI is calculated at both scales using band 3 for the red reflectance and band 4 for the near infrared reflectance.

Compared with the ETM NDVI image (Figure 6) there is marked loss of definition in the degraded image (Figure7), the Wambiana experiment is no longer visible as a defined scene element. Relatively small areas of extremely high or low fine scale data can markedly effect the value of the coarse scale. An example is the third pixel of the top row of the coarse scale image (pixel 1,3) which has an elevated NDVI because it is influenced by the high NIR reflectance of the vegetation flush associated with a creek. Pixel 1,5 has a surprisingly low value despite it’s proximity to a highly NIR reflective creek, the low NDVI is a consequence of dominating influence of larger area of relatively low NIR reflectance. This averaging phenomenon is the reason why coarse scale data do not work with fine scale models (Chen, 1999).

The coarse scale image occupied a smaller spectral space (Figure 8). This is an expected result since the averaging process will eliminate extreme values. It is


25

encouraging to see that the general shape of simplexs containing fine and coarse spectral data have similar shapes, and that some extreme data values are retained in the coarse data. These findings suggest that it is possible to translate between the two image domains by applying geometric corrections in spectral space. These translations will enable generation of pseudo fine scale imagery and for application in the remote sensing fine scale models of pasture growth. This concept must be tested using atmospherically and view angle corrected LAC AVHRR and TM data to ensure that image end-members at both scales represent the same scene end-member.

5. Conclusions and Future Directions Simple time series methods can separate the remotely sensed signals from grass and trees. Sufficient Landsat TM exists to build and test to simple light use efficiency models of grass growth. These use fine scale imagery to derive the radiation absorption variable of the model. The model will be validated with field measured grass growth measured in experimental treatments identifiable in Landsat imagery. Geometric transformations in spectral space may allow linkage between AVHRR and Landsat TM imagery. These linkages would allow the application of variables derived form coarse scale imagery to models developed at a finer scale.

6. Acknowledgements

This work is a subproject of research funded by the Land and Water Resources Research & Development Corporation, Enhanced framework for analysing climate variabilityand its impacts for policy purposes (BRR7). The authors are indebted to Dr Greg Laughlin,Bureau of Rural Sciences, who leads the parent project, for his continued enthusiasticsupport of this research. Special thanks to Ken Day, QDNR, who took time to show us the Queensland field experiments. A whistle stop tour of eastern Queensland with Ken allowed us to set the research in context. Without the huge efforts of Edward King, CSIRO Earth Observation Centre, in stitching the AVHRR archive, and Mac Dilley and Mary Edwards, CSIRO Atmospheric Research, for correcting the AVHRR data this work would not be possible. Their efforts will see that the potential of this valuable national resource is realised. Finally thanks to Paul Jones and Peter O'Reagain who run the field experiments at Kielambete and Wambiana, respectively. Their results and co-operation are the key to making the big step from qualification to quantification.


26

References Brockwell, P.J. and Davis, R.A., 1987. Time Series: theory and methods. Springer Verlag, New

York, 519 pp.

Carter, J., Flood, N., Danaher, T., Hugman, P., R.Young, Duncalfe, F., Beeston, G., Mlodawski, G., Hart, D., Green, D., Richards, R., Dudgeon, G., Dance, R., Brock, D. and D. P., 1996. Development of a national drought alert strategic information system: Volume 3, development of data rasters for model inputs. Final report on Project QDPI 20. Land and Water Resources Research and Development Corporation, Canberra.

Chen, J.M., 1999. Spatial scaling of a remotely sensed surface parameter by contexture. Remote Sensing of Enviroment, 69: 30-42.

Coops, N.C., Waring, R.H. and Landsberg, J.J., 1998. Assessing forest productivity in Australia and New Zealand using a physiologically-based model driven with averaged monthly weater data and satellite-derived estimates of canopy photosynthetic activity. Forest Ecology and Management, 104: 133-127.

Danaher, T.J., Carter, J.O., Brook, K.D., Peacock, A. and Dudgeon, G.S., 1992. Broad scale vegetation mapping using NOAA-AVHRR imagery, Sixth Australasian Remote Sensing Conference, Wellington, New Zealand, pp. 128-137.

DeFries, R., Hansen, M. and Townshend, J., 1995. Global discrimination of land cover types from metrics derived from AVHRR pathfinder data. Remote Sensing of Environment, 54(3): 209-222.

Dilley, A.C., Edwards, M., O'Brien, D.M. and Mitchell, R.M., 2000. Operational AVHRR processing modules: atmospheric correction, cloud masking and BRDF correction. EOC Project Final Report: CSIRO Atmospheric Research, Aspendale, Victoria, Australia.

Filet, P., Dudgeon, G., Scanlan, J., Elmes, D., Bushell, J., Quirk, M., Wilson, R. and Kelly, A., 1990. Rangeland vegetation monitoring using NOAA-AVHRR data: 2 Ground truthing NDVI data, Fifth Australasian Remote Sensing Conference, Perth, Australia, pp. 218-227.

Filet, P. and Osten, D., 1996. Ecology of a grazed woodland: Kielambete grazing trial methodology manual. DAQ.090, Queensland Departmant of Primary Industries, Emerald, Qld.

Hobbs, T.J., 1995. The use of NOAA-AVHRR NDVI data to assess herbage production in the arid rangelands of Central Australia. International Journal of Remote Sensing, 16(7): 1289-1302.

Jupp, D.L.B. and Walker, J., 1996. Detecting structural and growth changes in woodlands and forests: the challenge for remote sensing and the role of geometric-optical modelling. In: H.L. Gholz, K. Nakane and H. Shimoda (Editors), The use of remote sensing in the modelling of forest productivity. Kluwer Academic Publishers, Dortrecht, pp. 75-108.

Kaufman, Y.J., 1989. The atmospheric effect on remote sensing and its correction. In: G. Asrar (Editor), Theory and applications of optical remote sensing. John Wiley and Sons, New York, pp. 336-428.

Kumar, K. and Monteith, J.L., 1981. Remote sensing of crop growth. In: H. Smith (Editor), Plants in the daylight spectrum. Academic Press, London, pp. 133-144.

Landsberg, J.J. and Waring, R.H., 1997. A generalised model of forest productivity using simplified concepts of radiation-use efficiency, carbon balance and partioning. Forest Ecology and Management, 95: 209-228.


Lovell, J.L. and Graetz, R.D., 2000. Filtering Pathfinder AVHRR land data for Australia. CSIRO, Earth Observation Centre, Canberra.

Markham, B.L. and Barker, J.L., 1987. Thematic Mapper bandpass solar exoatmospheric irradiances. International Journal of Remote Sensing, 8(3): 517-523.

McKeon, G.M., Day, K.A., Howden, S.M., Mott, J.J., Orr, D.M., Scattini, W.J. and Weston, E.J., 1990. Nothern Australian savanas: management for pastoral production. Journal of Biogeography, 17: 355-372.

Monteith, J.L., 1977. Climate and the efficiency of crop production in Britain. Philosophical Transactions of the Royal Society London, 281: 277-294.

Myneni, R.B. and Williams, D.L., 1994. On the relationship between fAPAR and NDVI. Remote Sensing of Environment, 49: 200-211.

O'Reagain, P., Bushell, J. and C, B., 1999. Coping with rainfall variability: Grazing management strategies for seasonally variable tropical savannas. Annual Report, Queensland Department of Primary Industries, Charters Towers, Qld.

Rasmussen, M.S., 1998. Developing simple, operational, consistant NDVI-vegetation models by applying environmental and climatic information. Part I. Assessment of net primary production. International Journal of Remote Sensing, 19(1): 97-117.

Roderick, M.L., Nobel, I., R, and Cridland, S.W., 1999. Estimating woody and herbacious vegeation cover from time series satellite observations. Global Ecology and Biogeography Letters, 8: 501-508.

Strahler, A.H., Woodcock, C.E. and Smith, J.A., 1986. On the nature of models in remote sensing. Remote Sensing of Environment, 20: 121-139.

Tothill, J.C. and Gilles, C., 1992. The pasture lands of northern Australia: Their condition, productivity and sustainability. Occasional Paper No.5. Tropical Grassland Society of Australia.

Walker, J. and Hopkins, M.S., 1984. Vegetation, Australian soil and land survey field handbook. Inkarta Press, Melbourne, pp. 44-67.

27


28

29

Recent Advances in the Application of Remote Sensing for Monitoring Land Surface Fluxes William P. Kustas USDA-ARS Hydrology Lab Bldg. 007 BARC-WEST Beltsville, MD 20705, USA [email protected] Abstract: Satellite remote sensing data have provided regional and global coverage of landscape

properties relevant for monitoring land surface fluxes. This has prompted the development of

land–atmosphere exchange models that can use remotely sensed inputs to derive surface heat

fluxes. Many of these approaches use radiometric surface temperature as the key boundary

condition since surface air temperature differences reflect the overall balance of the various

surface energy flux components. The performance of each of these schemes varies widely due to

several factors. These include: 1) uncertainties in atmospheric correction, surface emissivity and

radiometer calibration; 2) non-uniqueness of the radiometric-aerodynamic temperature

relationship; 3) errors in defining meteorological variables for each satellite pixel from a sparse

network of weather station observations. A simple modeling approach will be described which

attempts to minimise the impact of these limiting factors. Examples from recent field studies will

illustrate its utility.

1. Introduction The temperature of the air near the surface and the temperature of the surface itself reflect the balance of the various surface energy flux components. The surface sensible and latent heat fluxes strongly interact with the overlying atmosphere and influence the characteristics of the planetary boundary layer, ultimately influencing local and regional weather patterns (Avissar, 1995). For these reasons, major efforts have focused on developing techniques to use surface and air temperatures to infer magnitudes of the components of the surface energy budget. In a recent review on monitoring the surface energy balance with remote sensing, Kustas and Norman (1996) conclude that obtaining reliable estimates of the heat fluxes using remotely

L a n d E n v S a t W o r k s h o p . . . . . . . . . . . . . . . . . . A d e l a i d e , A u g u s t 2 0 0 0

30

sensed radiometric temperature, TR(θ), at a viewing angle θ from remote brightness temperature measurements have been hampered by several factors. These include correcting the remotely sensed brightness temperatures for atmospheric and emissivity effects, calibration issues and the nonuniqueness of the aerodynamic-radiometric temperature relationship due to viewing angle and vegetation properties (e.g., Vining and Blad, 1992; Mahrt et al., 1997). Moreover, there is the issue of defining meteorological driving variables, particularly air temperature, for all satellite pixels for the region. In the best of cases, this has to be done using a sparse network of weather stations (i.e., where stations are typically spaced 100 km). Attempts at estimating spatial variability in air temperature, TA, at regional scales with remote sensing suggest an uncertainty of 3–4 K (Goward et al, 1994; Prince et al., 1998). An important conceptual step in improving the procedure for estimating surface fluxes came with the idea of using the time rate of change of radiometric temperature TR(θ) from a geostationary satellite such as GOES (Geosynchronous Orbiting Environmental Satellite) with an atmospheric boundary layer model (Wetzel et al., 1984). By using time rate of change of TR(θ), one reduces the need for absolute accuracy in satellite sensing and atmospheric corrections, both significant challenges. Diak (1990) improved this approach further with a method for partitioning the available energy into latent, E, and sensible heat, H, flux by using the rate of rise of TR(θ) from the GOES satellite and atmospheric boundary layer (ABL) growth. Diak and Whipple (1993) further refined the model by including a procedure to account for effects of horizontal and vertical temperature advection and vertical motions above the ABL. In a related approach, the Two-Source Time-Integrated model of Anderson et al. (1997) (presently called ALEXI: Mecikalski et al., 1999) provides a practical algorithm for using a combination of satellite data, synoptic weather data and ancillary information to map surface energy flux components on a continental scale (Mecikalski et al., 1999). The ALEXI approach builds on the earlier work with the Two-Source Model (TSM) scheme which accounts for radiometric-aerodynamic temperature differences using a physically-based approach instead of empirical techniques (Norman et al., 1995; Kustas and Norman, 1996). By using remote brightness temperature observations at two times in the morning hours and considering planetary boundary layer processes, the methodology removes the need for a measurement of near-surface air temperature and is relatively insensitive to uncertainties in surface thermal emissivity and atmospheric corrections on the GOES brightness temperature measurements. Anderson et al. (1997) and Mecikalski et al. (1999) have shown that surface fluxes retrieved from the ALEXI approach compare well with measurements. The ALEXI approach is a practical means to operational estimates of surface fluxes over continental scales with 5–10 km pixel resolution. This paper describes a relatively simple dual-temperature-difference (DTD) procedure for using time rate of change in TR( ) and TA to compute the surface heat fluxes (Norman et al., 2000). Data from several different field sites covering a wide range of environmental conditions are used to evaluate this new method. The impact of using weather station

K u s t a s . . . . . . . . . . . . . . . . . . . . . . . . . . . M o n i t o r i n g l a n d s u r f a c e f l u x e s

31

observations of TA and wind speed, u, 50–100 km from the field sites with local TR( ) observations is evaluated by contrasting heat flux predictions using the original TSM approach, which is sensitive to errors/bias in meteorological inputs (Anderson et al., 1997), with the DTD technique. The performance of the DTD approach is shown to be superior to TSM suggesting this technique is not significantly affected by errors in using non-local meteorological inputs. The performance of the DTD approach under conditions where TR( ) – TA < 0, even though observations indicate H > 0 from the surface, support the robustness of the technique. An example using the DTD method over central U.S. and comparison with the more complicated ALEXI model is included and suggests suitability for regional applications. The DTD approach reduces both the errors associated with deriving a radiometric temperature and defining meteorological quantities at large scales. The scheme does not require modeling boundary layer development or using a mesoscale atmospheric model linking surface temperature to ABL properties. Thus the potential exists that this method could be used operationally with the Japan Meteorological Agency-Geostationary Meteorological Satellite (JMA-GMS 5) or two of the National Oceanic & Atmospheric Administration-Advanced Very High Resolution Radiometer (NOAA-AVHRR) satellites. The DTD technique is most reliable and can be applied more frequently if the morning overpass time is approximately an hour or so after sunrise when fluxes are small and the second overpass is mid-morning (say ~1030 local time) minimising the likelihood of cloud cover.

2. Methodology The full derivation of the DTD method is given in Norman et al. (2000). Starting with the equations of Anderson et al. (1997), a double difference of radiometric and air temperatures can be formulated so that an estimate of sensible heat flux can be obtained from measurements of surface radiometric temperature, air temperature, wind speed, vegetation height, cover, type and approximate leaf size. The relationship between surface radiometric temperature and sensible heat flux can be obtained from Eq. (14) of Anderson et al. (1997):

where the i subscript refers to time, f( ) is the fraction of the radiometer’s field of view that is occupied by vegetative cover when the radiometer has a zenith viewing angle , is the density of air, cp is the specific heat of air, HC,i is the sensible heat flux from the vegetative canopy at time i, Hi is the total sensible heat flux above the canopy arising from both vegetation and soil, RA,i is the aerodynamic resistance to heat transport above the canopy, and RS,i is the resistance to heat transport arising from the air layer immediately above the soil. The quantity f( ) can be estimated from canopy architecture and view angle; for a random or clumped canopy with a spherical leaf angle distribution and leaf area index, F,

T T fH R

cf

H H R RcR i A i

C i A i

p

i C i A i S i

p, ,

, , , , ,( ) ( ) ( ( ) )( ) ( )

(1)θ θρ

θρ

− = + −− +

1

f F( ) exp . ( )cos

( )θ θθ

= − −

1 0 5 Ω 2


32

where ( ) is unity for random canopies and less than one for clumped canopies (Kucharik et al., 2000; Kustas and Norman, 1999). The sensible heat from the canopy, HC,i, is estimated from the net radiation divergence of the vegetative canopy (RnC,i) using

where PT is the Priestly–Taylor coefficient ~1.3 (Priestly and Taylor, 1972), fg is the fraction of the vegetative canopy that is green, s is the slope of the saturation vapor pressure versus temperature curve, and is the psychrometer constant. The net radiation divergence equation used by Norman et al. (1995) and modified by Anderson et al. (1997) for solar zenith angle assumes an exponential decay of Rn through the canopy layer (Ross, 1981). This is a reasonable parameterisation for moderate to high canopy covers (i.e., F 2) but not reliable for sparse canopies leading to significant underestimates of RnC,i because of inadequate treatment of thermal emission from the soil. If the equation presented by Norman et al. (1995) is used for the net solar radiation and net thermal radiation calculated by the approach of Kustas and Norman (1999, Eq. 2b), RnC,i estimates are more accurate. The equations for estimating the aerodynamic resistances above the canopy and above the soil surface are contained in Kustas and Norman (1999). The two terms on the right of Eq. (1) represent the contributions of vegetative canopy and soil to the difference between radiometric surface temperature and air temperature. Equation (1) is from the TSM scheme and accommodates the difference between radiometric and aerodynamic surface temperatures; it assumes the vegetative canopy and soil are in parallel, which usually is a good assumption (Norman et al., 1995). In the DTD method, Eq. (1) is applied at two times. The first time usually is chosen when all the fluxes are small and temperatures are similar; typically this occurs about one hour after sunrise so we choose i = 0 at this time. The second time can be any hour during the day. Applying Eq. (1) at two times (0, i) and subtracting the equations yields

The last two terms on the right side of Eq. (4) involving H0 and HC,0 are usually negligible, because the first time (i = 0) is chosen so that this occurs. Since H0 – HC,0 = HS,0, the sensible heat flux from the soil, this term clearly is negligible shortly after sunrise; this is fortunate because we do not have any means for estimating H0. The last term involving HC,0 usually is small but can be evaluated so is kept in the equation. The simplified equation for Hi is

,

( )i

5

H cT T T T

f R RH f

fR

R R

H HR RR R

H ff

RR R

i pR i R A i A

A i S iC i

A i

A i S i

CA S

A i S iC

A

A i S

=− − −

− +

+ −− +

+

+ −++

+

− +

ρθ θ

θθ

θ

θθ

, , , ,

, ,,

,

, ,

,, ,

, ,,

,

,

( ) ( )( ( ) ) ( )

( )( )

( )( )

0 0

0 00 0

00

11

1

1

( ) ( )

( )

H Rn f ssC i C i PT g, , (3)= −+

α

γ

(4)

H cT T T T

f R RH f

fR

R R

H ff

RR R

i pR i R A i A

A i S iC i

A i

A i S i

CA

A i S i

≅− − −

− +

+ −− +

+

+− +

ρθ θ

θθ

θ

θθ

, , , ,

, ,,

,

, ,

,,

, ,

( ) ( )( ( ) ) ( )

( )( )

( )( )

0 0

00

11

1

1

( ) ( )

(5)

K u s t a s . . . . . . . . . . . . . . . . . . . . . . . . . . . M o n i t o r i n g l a n d s u r f a c e f l u x e s

33

Equation (5) represents a relatively simple result with the advantage that any offset between measurements of TR,i( ) and TA,i cancels out of the temperature term. Given measurements or estimates of net radiation (Rni) and soil heat conduction flux (Gi) at time i, the latent heat flux, LEi, can be calculated as a residual in the energy balance equation, namely,

LEi = Rni – Gi – Hi . (6)

3. Results of Model Validation

3.1 Experimental Data Tower-based surface flux, meteorological and TR( ) data come from three experimental sites covering a wide range of climatic and vegetative cover conditions. For an overview of the study sites see Norman et al. (2000). One data set was collected over a semi-arid rangeland site containing ~ 25% shrub cover and the remaining ~ 75% predominantly bare soil. These data were collected as part of the Monsoon ’90 experiment (Kustas and Goodrich, 1994) and have been used to test several soil-vegetation-atmosphere modeling algorithms with details of the measurements described in several papers (Norman et al., 1995; Kustas et al., 1996; Flerchinger et al., 1998). Another data set was collected at four field sites in Oklahoma as part of the Southern Great Plains 1997 (SGP97) Hydrology Experiment. An overall summary of SGP97 experiment is given in Jackson et al. (1999) and further details can be found on the web (http://hydrolab.arsusda.gov/sgp97/ ). The measurement sites were located at one of the main study areas (El Reno facility) and consisted of native grasslands or rangeland, pasture and bare-soil fields. The micro-meteorological measurements are described in Twine et al. (2000) and have been used for validation of remote sensing energy balance modeling (French et al. 2000). The last data set was obtained over a riparian zone along the Rio Grande in northern New Mexico in the Bosque Del Apache, a nature preserve. The site contained a dense cover of non indigenous salt cedar vegetation, which have access to the shallow water table near the Rio Grande river. Basic vegetation/surface properties of the study sites are listed in Table 1.

To illustrate the application of the DTD approach to a large region, GOES and AVHRR satellite observations are combined with surface synoptic data for June 12, 1995 coverage over the central U.S. (U.S. Great Plains); the same data were used by Mecikalski et al. (1999) for testing ALEXI. The AVHRR data was used to calculate a NDVI (Normalized Difference Vegetation Index) and estimate a vegetative-cover fraction by the method of Carlson and Ripley (1997). GOES thermal images were obtained 1.5 and 5.5 hours after local sunrise and combined with near-surface air temperature and wind speed from the surface weather station nearest the pixel of interest. The output of latent heat flux from the more complex ALEXI model will be compared to the DTD approach.


34

Table 1. General site information and vegetation cover description

Site+ General Description Canopy Height (m) Leaf Area Index

ER01 Rangeland/grassland 0.6 4.2

ER05 Rangeland/grassland 0.5 2.6

ER09 Pasture 0.25 2.7

ER13 Bare soil – 0

LH01 Shrub land 0.5 0.5

BOSQ Riparian salt cedar 6.7 3.4

+ sites are described in the text; briefly, ER01, ER05, ER09, and ER13 are sites from the SGP97 experiment, LH01 is a site from the Monsoon ’90 experiment and the BOSQ site is from the experiment in the Bosque del Apache.

3.2 Results Using the Tower Data The impact of using weather station data 50–100 km away could be evaluated for the ER and LH sites (see Table 1) by comparing the original TSM formulation using absolute TR( ) – TA with the DTD method. For the LH site, a Tucson AZMET station (http://128.196.42.70/azmet/.html) was used which is ~ 120 km away and nearly 800 m lower in elevation. For the ER sites a MESONET station (http://okmesonet.ocs.ou.edu/) located in the Little Washita Watershed was used, which is ~75 km away and similar elevation. Differences in midday TA between LH01 and the Tucson weather station ranged between 4 to 8 K. For the ER sites differences in midday TA with the MESONET station were 0.5 to 2 K. In the application of the DTD technique, time differences in TR( ) and TA, namely TR( ) and

TA, were taken from approximately an hour after sunrise and at midday for LH01 and the four ER sites. Output of the fluxes using midday TR( ) – TA with the TSM model yields significant scatter (Figure 1). The root-mean-square-difference, RMSD, (Willmott, 1982) values for H and LE are ~ 90 W m-2. With the DTD method there is significantly less scatter in the midday fluxes (Figure 2) with RMSD ~ 60 W m-2 for H and LE . Comparing the results of Figures 1 and 2 indicates that the DTD method does account for a major part of the error/bias in TR( ) – TA observations caused by the use of non-local meteorological data. Observations from the BOSQ site are used to show how this technique can account for TR( ) measurement errors. TSM output using TR( ) – TA at each time step and the DTD method are compared using data for a clear day period. The comparisons between the two models and the observed heat fluxes are illustrated in Figure 3. Since TR( ) – TA was generally less than zero, the TSM computes H < 0, which is in total contradiction to the observations. The DTD method not only gets the sign correct but computes magnitudes similar to the observations of both H and LE over the whole course of the daytime period.

35

-200 -100 0 100 200 300 400 500 600

H Observed from Flux Towers (W/m^2)

-200

-100

0

100

200

300

400

500

600

H fr

om T

SM A

ppro

ach

(W/m

^2)

0 100 200 300 400 500 600 700 800

LE Observed from Flux Towers (W/m^2)

0

100

200

300

400

500

600

700

800

LE fr

om T

SM A

ppro

ach

(W/m

^2)

Figure 1. Comparison between observed and modeled midday sensible, H, and latent, LE, heat fluxes using absolute TR( ) – TA values in the TSM approach. Regional TA,i and u i observations are from weather stations ~ 50 to ~100 km away from study sites described in Table 1. Line represents perfect agreement with observations.

-200 -100 0 100 200 300 400 500 600

H Observed from Flux Towers (W/m^2)

-200

-100

0

100

200

300

400

500

600

H fr

om D

TD A

ppro

ach

(W/m

^2)

0 100 200 300 400 500 600 700 800

LE Observed from Flux Towers (W/m^2)

0

100

200

300

400

500

600

700

800

LE fr

om D

TD A

ppro

ach

(W/m

^2)

Figure 2. Comparison between observed and modeled midday sensible, H, and latent, LE, heat fluxes with new dual-temperature-difference (DTD) approach. Details as per Fig. 1 caption.

6 8 10 12 14 16 18

Local Time

-300

-200

-100

0

100

200

300

Sens

ible

Hea

t Flu

x (W

/m^2

) DDTTSMObs.

6 8 10 12 14 16 18

Local Time

-100

0

100

200

300

400

500

600

700

Late

nt H

eat F

lux

(W/m

^2) DDT

TSMObs.

Figure 3. Estimates of sensible, Hi, and latent, LEi , heat flux over the daytime period for a clear day at the BOSQ site using absolute TR( ) – TA values in TSM approach (diamonds), and the new dual-temperature-difference (DTD) method (squares) compared to the observations (triangles). Using local TA,i and ui observations in both modeling schemes.

K u s t a s . . . . . . . . . . . . . . . . . . . . . . . . . . . . M o n i t o r i n g l a n d s u r f a c e f l u x e s

36

3.3 Results Using Satellite Imagery For operational applications with the DTD method, synoptic weather station data would typically be used. However, these stations tend to be located near airports and populated areas; hence, air temperature differences at these locations may be systematically larger than at surrounding rural areas, which are generally more vegetated. Positive biases in air temperature differences from such measurements will result in an underestimate of sensible heat flux from the DTD method (see Eq. 6). This was observed at the field scale where DTD was applied using the weather station data from Tucson AZMET station. This station had 2 K larger TA value than the local LH station (LH0 ) around midday.

Kustas et al. (2000) investigated output from the ALEXI model since it diagnoses low-level air temperatures, their time differences and the relationship of air temperatures to surface fluxes to explore the relationship between TA and TR ( ). A systematic bias in TA from the synoptic weather station reports was identified and a simple correction was devised. Kustas et al. (2000) applied this first-order correction to the air temperature differences in the DTD method resulting in a 0% increase in the domain-average sensible heat flux (about 20 Wm-2), as well as an improved agreement between the DTD and ALEXI methods. Latent heating results for 5.5 hours after local sunrise for the domain are shown in Figure 4 from the DTD and ALEXI schemes. Areas that are white in this figure were those identified

Latent Heat Flux (W m ) 2 June 995-2

0 50 100 150 200 250 300 350 400

(a)

Latent Heat Flux (W m ) 2 June 995

0 50 100 150 200 250 300 350 400

-2

(b)

Figure 4. Regional map of latent heat flux for central U.S. (U.S. Great Plains) using (a) ALEXI and (b) DTD models. Both models use surface temperature differences from GOES at ~1.5 hr and 5.5 hr after local sunrise. The DTD approach uses air temperature differences from a network of synoptic weather stations, while ALEXI uses early morning soundings and simulates air temperature over the daytime period. Regions in white are cloudy or outside of the domain of analysis.


K u s t a s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M o n i t o r i n g l a n d s u r f a c e f l u x e s

37

as cloudy by screening procedures, and thus were not evaluated in either method. The DTD method displays very similar spatial features compared to the ALEXI output, although even with the adjustments to TA from Kustas et al. (2000) domain-average sensible heating results were still 20 W m-2 lower than with ALEXI. Regions of sparse cover in Texas and New Mexico (southwest) have been assigned low values of latent heating. Across the corn belt (midwest) soils are largely bare this early in the growing season and also show very low evaporation rates. Output from the DTD method is very encouraging. Its ability to duplicate the results from ALEXI, a much more complicated and data-intensive parameterization, suggests it can be applied operationally and provide output at near-real time. Computer processing time for the domain shown in Figure 4 for the ALEXI model was about 35 minutes, while the DTD scheme required only about one minute of processing time on the same UNIX workstation.

4. Conclusions The utility of the DTD method at the field scale is evaluated using ground-based data collected under a wide range of environmental conditions. Weather station data from ~100 km away caused large discrepancies with the TSM approach, which relies on absolute temperature differences, while the performance of the DTD method was satisfactory. Regional application of DTD using satellite data and synoptic weather stations gave satisfactory heat flux predictions compared to ALEXI, a much more complicated and data-intensive set of algorithms simulating atmospheric boundary layer growth for achieving energy balance closure (Mecikalski et al., 1999). Kustas et al. (2000) found that correction to weather station air temperature data may have to be applied in practice, however, because the DTD method cannot account for differences in time-rate-of-change in air temperature caused by local land cover/land use conditions. This error is due to the fact that synoptic weather stations tend to be located near airports and populated areas resulting in systematically larger temporal changes in air temperature than what would be observed in surrounding rural areas containing generally more vegetation. Positive biases in air temperature differences from such measurements will result in an underestimate of sensible heat flux from the DTD method; this is supported by both the field scale and regional scale studies. Fortunately, a relatively simple correction reduces the bias in heat fluxes to around 20 W m-2, which is well within model and measurement uncertainties. Although this correction is not entirely general, it was derived from observations over the central U.S. having a wide variety of climatic regimes ranging from the hot-dry desert southwest to the cool-wet northern coniferous forests. Therefore, this scheme for adjusting time-rate-of-change in air temperature has wider applicability than what is typically expected of empirically-based corrections. The utility of this simple correction needs further evaluation, particularly as affected by different seasonal conditions.

L a n d E n v S a t W o r k s h o p . . . . . . . . . . . . . . . . . . . . . . A d e l a i d e , A u g u s t 2 0 0 0

38

5. Acknowledgements The field data from Monsoon ’90 and SGP ’97 were made possible through the financial support of several government agencies, including National Aeronautics Space Agency (NASA) Land Surface Hydrology and Interdisciplinary Research Program in Earth Sciences and the United States Department of Agriculture-Agricultural Research Service (USDA-ARS). The data from the Bosque Del Apache was made possible through logistical support of the Bureau of Land Reclamation, the USDA Forest Service, and New Mexico State, Las Cruces NM.

6. References Anderson, M. C., Norman, J. M., Diak, G. R., Kustas, W. P., Mecikalski, J. R., A two-source time-

integrated model for estimating surface fluxes from thermal infrared satellite observations. Remote Sens. Environ., 60, 195-216, 1997.

Avissar, R., Scaling of land-atmosphere interactions: An atmospheric modelling perspective. Hydrol. Processes. 9, 679-695, 1995.

Carlson, T. N., Ripley, D. A., On the relation between NDVI, fractional vegetation cover, and leaf area index. Remote Sens. Environ. 62, 241-252, 1997.

Diak, G. R., Evaluation of heat flux, moisture flux and aerodynamic roughness at the land surface from knowledge of the PBL height and satellite-derived skin temperatures, Agric. For. Meteor., 52, 181-198, 1990.

Diak, G. R., Whipple, M. A., Improvements to models and methods for evaluating the land-surface energy balance and "effective" roughness using radiosonde reports and satellite-measured "skin" temperatures. Agric. For. Meteor., 63, 189-218, 1993.

Flerchinger, G. N., Kustas, W. P., Weltz, M. A., Simulating surface energy fluxes and radiometric surface temperatures for two arid vegetation communities using the SHAW model. J. Appl. Meteorol., 37, 449-460, 1998.

French, A. N., Schmugge, T. J., Kustas, W. P., Surface fluxes over the SGP site with remotely sensed data. Physics Chem. Earth (B), 25, 167-172, 2000.

Jackson, T. J., Le Vine, D. M., Hsu, A. Y., Oldak, A., Starks, P. J., Swift, C. T., Isham, J. D., and Haken, M., Soil moisture mapping at regional scales using microwave radiometry: The Southern Great Plains hydrology experiment. IEEE Trans. Geosci. Remote Sens., 37, 2136-2151, 1999.

Kucharik, C.J., Norman, J.M., Gower, S.T., Characterizing the radiation regime of nonrandom forest canopies: Theory, measurements, modeling and a simplified approach. Tree Physiol., 19, 695-706, 1999.

Kustas, W. P., Diak, G.R., Norman, J. M., Time difference methods for monitoring heat fluxes with remote sensing. In: (V. Lakshmi, Editor) AGU Water Science and Applications Series: Observations and Modeling of Land Surface Hydrological Processes (In Press) 2000.

Kustas, W. P., Goodrich, D.C., Preface, MONSOON'90 Multidisciplinary Experiment, Water Resour. Res., 30(5), 1211-1225, 1994.

Kustas, W. P., Norman, J. M., Use of remote sensing for evapotranspiration monitoring over land surfaces. Hydrol. Sci. J. 41, 495-516, 1996.

Goward, S. N., Waring, R, H., Dye, D. G., Yang, J., Ecological remote sensing at OTTER: Satellite macroscale observations. Ecol. Appl., 4, 322-343, 1994.

K u s t a s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M o n i t o r i n g l a n d s u r f a c e f l u x e s

39

Kustas, W. P., Humes, K.S., Norman J. M., Moran, M. S., Single-and dual-source modeling of surface energy fluxes with radiometric surface temperature. J. Appl. Meteorol., 35, 110-121, 1996.

Mahrt, L., Sun, J., MacPherson, J. I., Jensen, N. O., Desjardins, R. L., Formulation of surface heat flux: Application to BOREAS. J. Geophys. Res., 102(D24), 29,641-29,649, 1997.

Mecikalski, J. R., Diak, G. R., Anderson M. C., Norman, J. M., Estimating fluxes on continental scales using remotely-sensed data in an atmospheric-land exchange model. J. Appl. Meteorol., 38, 1352-1369, 1999.

Norman, J. M., Kustas, W. P., Humes, K. S., A two-source approach for estimating soil and vegetation energy fluxes from observations of directional radiometric surface temperature. Agric. For. Meteor., 77, 263-293, 1995.

Norman, J. M., Kustas, W. P., Prueger, J. H., Diak, G. R., Surface flux estimation using radiometric temperature: a dual-temperature-difference method to minimize measurement errors. Water Resour. Res., 36, 2263-2274, 2000.

Priestly, C. H. B., Taylor, R. J., On the assessment of surface heat flux and evaporation using large-scale parameters. Monthly Weath. Rev., 100, 81-92, 1972.

Prince, S. D., Goetz, S. J., Dubayah, R. O., Czajkowski, K. P., and Thawley, M., Inference of surface and air temperature, atmospheric precipitable water and vapor pressure deficit using Advanced Very High Resolution Radiometer satellite observations: comparison with field observations. J. Hydrol., 212-213, 230-249, 1998.

Ross, J., The radiation regime and architecture of plants, In Tasks for Vegetation Sciences 3 (H. Lieth, Series Ed.), Dr. W. Junk, The Hague, Netherlands, 1981.

Twine, T. E., Kustas, W. P., Norman, J. M., Cook, J. M., Houser, P. R., Meyers, T. P., Prueger, J. H., Starks, P. J., Wesely, M. L., Correcting eddy-covariance flux underestimates over a grassland. Agric. For. Meteorol., 103, 279-300, 2000.

Vining, R. C., Blad, B. L., Estimation of sensible heat flux from remotely sensed canopy temperatures. J. Geophys. Res., 97, D17, 18951-18954, 1992.

Wetzel, P. J., Atlas, D., Woodward, R., Determining soil moisture from geosynchronous satellite infrared data: A feasibility study. J. Clim. Appl. Meteorol., 23, 375-391, 1984.

Willmott, C. J., Some comments on the evaluation of model performance. Bull. Am. Meteorol. Soc., 11, 1309-1313, 1982.

Kustas, W. P., Norman, J. M., Evaluation of soil and vegetation heat flux predictions using a simple two-source model with radiometric temperatures for a partial canopy cover. Agric. For. Meteorol., 94, 13-29, 1999.

L a n d E n v S a t W o r k s h o p . . . . . . . . . . . . . . . . . . . . . . A d e l a i d e , A u g u s t 2 0 0 0

40

41

Stitching the Australian 1-km Archive Edward King CSIRO Earth Observation Centre, GPO Box 3023, Canberra, ACT 2601 [email protected]

1. Introduction

The Australian 1-km Archive is a collection of High Resolution Picture Transmission (HRPT) data from the NOAA polar orbiting spacecraft acquired in the Australian region since 1992. These data arise out of the Australian participation in the Global 1km Land Project1 and constitute the most complete coverage of the region in existence. The dataset began its life at the Australian Land Research Data Centre and continues to be maintained and grown today at the CSIRO Earth Observation Centre (EOC) in Canberra. At the time of writing (in late 2000) the archive is some 12 TB in size and consists of all acquisitions of NOAA HRPT taken at Darwin (NT), Perth (WA), Aspendale (Vic), Townsville (Qld) and Hobart (Tas) reception stations. The data is presently stored on some 3500 8mm and 4mm magnetic tapes. The daily coverage is of an area of about 50 million square km, including the entire Australian land surface, New Zealand, Papua New Guinea, most of Indonesia, and the surrounding ocean to at least 2000 km from the Australian Coast (Figure 1). The archive includes daylight and night passes from both the morning and afternoon spacecraft so the temporal coverage is also dense. During the eight years over which the data has been acquired, there have been numerous changes at the various contributing reception stations as a result of hardware replacement or operational changes. All these changes are reflected in the variety of data formats, media formats and media types present in the raw archive at the EOC. The tapes are inventoried only by approximate start/end dates and station of origin. This coarse cataloguing, coupled with the variety of formats/media, render the data practically inaccessible to anyone except those most familiar with its details. However this unique collection of data, being the densest and most complete record of the Australian region since 1992, is a potentially valuable National resource for climate and landcover research. Because it is the archive with the oldest data, its value increases, rather than decreases with the passage of time.

1 Eidenshink J.C. and Faundeen, J.L., The 1km AVHRR global land data set: first stages in implementation, Int. J. Remote Sensing, 1994, Vol 15, No 17, 3443-3462.

L a n d E n v S a t W o r k s h o p . . . . . . . . . . . . . . . A d e l a i d e , A u g u s t 2 0 0 0

42

Figure 1. Typical daily coverage for one satellite in one direction, in this case for NOAA-14 ascending (afternoon) passes on 4 Jan 1996. In fact, the equivalent area is covered at four times each day at approximately six hourly intervals by the two operational spacecraft both day and night.

2. Curation As custodians of these data, CSIRO/EOC is seeking to realise the potential of the archive by means of a pro-active management strategy comprised of the four following threads: 1. consolidation of the data by orbit stitching; 2. unification of the data to a common format; 3. cataloguing and documentation; and 4. protection by migration to a modern media.

The contribution of each of these to the goal of making the most value of the archive is described in turn below.

2.1 Consolidation by Stitching The AVHRR HRPT data are received at the Australian ground stations by direct broadcast from the NOAA spacecraft. When the spacecraft is within view of more than one station, multiple copies of the same data are acquired. These copies can differ markedly in quality depending on local conditions at each reception station. It is possible to exploit this inherent redundancy to select the best quality and most complete dataset, in the process reducing the total volume of storage required to hold

K i n g . . . . . . . . . . . . . . . . . . . S t i t c h i n g t h e A u s t r a l i a n 1 - k m A r c h i v e

43

the archive. In addition to eliminating the overlap, the parts of a pass from several different stations can be joined together (stitched) to form single continuous continent-spanning scenes, thereby reducing the total number of scenes that need to be managed and facilitating easier access to data for the whole of the continent. To process an archive of this size in this manner is a large undertaking, but precisely because it is such a comprehensive archive, the advantages make it worthwhile.

The major obstacles to be overcome include: 1. all simultaneously acquired scenes (existing in a variety of different media and file

formats) need to be located and read from the archive together; 2. the data need to be quality assessed and the line sequencing corrected; and 3. where data from more than one station is available, an algorithm to choose which

data is propagated to the output needs to be developed. A sophisticated system was developed to address these issues and is described in more detail below. Since the media from the reception stations are catalogued only by start and end date, the individual scenes can only be located by reading them from the tape. For efficiency it is more effective to read all the scenes from a single tape at once, and because tapes from different stations vary in both capacity and start/finish times, a considerable amount of disk space is required. The process which reads the tapes contains all the specific knowledge of both the media and file formats and therefore is able to output all the files to disk in a single common format. This dramatically simplifies the subsequent processing. As part of the ingestion procedure, each individual scene is pre-processed using specially developed software that identifies the spacecraft, sequences the lines correctly and quantifies the data quality on a line by line basis. Spacecraft identification is robustly achieved by examining the satellite address word present in every line and selecting the modal value. A novel system has been implemented to try and correctly sequence the lines in each file. This is necessary since (groups of) lines can be either dropped, or the timecode which tags each line, can be corrupted by poor quality reception. The procedure works by examining values present in each HRPT line, such as the minor frame counter, the day number, the TIP2 minor and major frame counters, as well as the timecode, and searching for consistent patterns to derive a model that predicts all of these for the entire file. This model is then used to identify the most probable location of any line with an ambiguous timecode. The data quality assessment exploits the fact that every HRPT line contains sequences of bits that are either fully or partially predictable. These include the synchronisation patterns and the TIP data which have parity bits. By comparing the expected bit

2 The Tiros Information Processor (TIP) multiplexes data from several other spacecraft instruments into the HRPT data stream at the rate of five frames per line. Each frame contains a pair of sequence numbers that can be used to determine where the line belongs.


44

patterns with those present in the file, and counting the number of TIP parity errors, an estimate of the bit error rate in each line can be made. The final processing step is the stitching of all the scenes from a particular pass into a single output file. This is done on a line by line basis and involves consideration of a number of different cases, depending on the number of input scenes in which each line in the output is present. The cases are: 1. only a single input line is available (only one receiving station recorded this data)

– it is copied directly to the output; 2. two input lines are available – the one with the lower measured bit error rate is

chosen; 3. three input lines are available – each bit in the output is determined by comparing

the three possible inputs and choosing the most common value (it is important to note that this assumes that the reception errors at the different stations are independent and not so frequent that the probability of the same bit being in error at two stations is significant); and

4. more than three inputs – the three with the lowest error rates are chosen and the process in 3 (above) is applied.

In each of cases 2, 3, and 4, the difference between each input line and the output line is stored in an auxiliary file. In most cases the differences are trivial (only one or two bits) so highly compressed storage is possible. However, the storage of these differences permits exact reconstruction of the input data, thus the stitching process amounts to a highly efficient compression of the input archive.

2.2 Unification As described above, the data are all converted to a common format as part of the ingestion process. This format is the LAS Archive Format3 in which a detailed header describing the data is created, and the sync words and TIP data are stripped from the HRPT. The TIP data and sync words are preserved locally by storing them in a separate file. The stitching process reads the input data and produces the output in the same format. Thus the entire archive of both input station-specific files and output files are in the one format. Both sets of files are archived to DLT4 tapes at this stage so that any future work with the data will require knowledge of only one format. The individual input scenes for each pass are archived as a group so that if the stitching itself ever needs to be repeated, there is no need to repeat the lengthy selection and grouping process.

3 The Land Analysis System (LAS) is a package of image analysis software developed by the US Geological Survey and incorporating an AVHRR processing suite. For further information refer to <http://edcwww.cr.usgs.gov/programs/sddm/lasdist>. 4 Digital Linear Tape (DLT) is a widely used modern recording standard capable of storing up to 35 GB per cartridge (uncompressed).


45

2.3 Cataloguing and Documentation As each scene is identified during the ingest process, it is possible to create a detailed and comprehensive catalogue of metadata. This catalogue is implemented as a database and contains details of all scenes and all media. Low resolution browse imagery is produced for all scenes, archived online, and similarly catalogued in the database. All ingest, pre-processing, stitching, and archiving is logged and the existence and location of the logs is also linked in the database. A simple web interface to this database allows users to explore the collection scenes held in the archive, view their browse imagery, locate the media upon which they are stored, and examine the log files that document the processing that was applied.

2.4 Protection The reading of the entire collection of station data serves to exercise the tape media in the input archive helping preserve the data on the tapes. In addition, the re-archival of the input files, grouped by pass, to DLT effectively migrates the whole archive to modern media. The output collection of stitched files is actually archived in duplicate, so that two identical sets of DLT media are produced. This provides some protection against the failure of individual media, as well as facilitating off-site storage and so permitting disaster recovery.

2.5 Summary The curation process described above overcomes all the difficulties mentioned in the introduction to produce a properly documented and catalogued dataset, migrated to modern high capacity media in a standard format. Selection and extraction of specific scenes is straightforward and processing of the entire dataset as a time-series is currently underway. The data is not only quality controlled, but by consolidating the duplicate input scenes, a best possible quality output dataset has been produced.

3. Export and Access A key step in enhancing utilisation of the archive is to ensure that it is available in a form that permits access by potential users. The Australian developed ASDA file format (<http://www.dar.csiro.au/publications/turner96/arcfrep.htm>; <http://www.bom.gov.au/climate/satellite/asda>) was chosen to export the data. While this format has the advantage of being a local standard, it is also both simple (in terms of physical format) and self-documenting. The data are stored in a single file as a sequence of standard HRPT lines, with a verbose header of readable text prepended. The header is extremely comprehensive containing, in addition to the physical properties of the file (lines, bytes per line etc), a description of the geographic location of the scene together with instrument calibration and navigation information. Such a format is ideal for long term storage or export of the dataset because it is both simple,


46

self-describing, and it entrains much of the auxiliary data that future users will need. However, creation of a complete ASDA header presents difficulties of its own since considerable auxiliary data must be collated and assembled in a suitable form. Furthermore, the complexity of the header provides many more opportunities toincorporate errors in the data description, necessitating a greater investment in quality control. However the long term benefits of properly documenting each scene in this manner outweigh the disadvantages. The metadata database, together with the browse images and the log files, is fundamental to enabling access by external users. As was already mentioned, a prototype web-based interface to these data is functional and will be available externally before the end of the year. Documentation describing the processing methods and providing details of the various auxiliary file formats is under development and will also be released to users. The final archive will be 6–8 TB in total volume. At present, the only practical means of maintaining such a large dataset online is in a tape library. We are presently experimenting with storage of the data in a StorageTek robotic tape store in Canberra, with a view to enabling efficient and automatic processing of the whole dataset as a time series. In addition, the stitched archive will be ingested into the robotic tape store at CSIRO Atmospheric Research in Aspendale. The tapes used to convey the data to Aspendale from Canberra will also be stored at CSIRO Marine Research in Hobart subsequently. Thus the entire archive will be replicated at three sites facilitating access by numerous potential users.

4. Progress At present, processing has been completed for all data in 1996 and 1997. A great deal of effort has been invested in extensive testing and debugging of the processing pipeline. This is because having to process the whole archive more than once if significant problems are discovered with the stitched product is such a large undertaking that every effort is being made to avoid it. Significant infrastructure has been assembled to support the processing effort. In terms of hardware this includes a dedicated network of 5 processing computers supported by approximately 300 GB of hard disk space and about 15 tape drives and libraries. Most of the software needed has had to be developed from scratch, including the programs to perform the noise assessment, line sequencing and stitching. A collection of TBUS navigation bulletins has been created to support the export in ASDA format. Processing will continue with 1995, and then with 1998–2000. It is anticipated that this stage will be complete by the end of January 2001. Processing each year takes about two weeks of continuous operation. The first three years of the archive (1992–1994) will be processed last.


47

5. Reception Network Performance The sophistication of the data quality assessment and stitching process, together with the large number of reception stations available, permits a novel analysis of the performance of the individual reception stations. This analysis is capable of detecting periods when (or directions from which) the station performance is degraded. It can also be used to discover where a station is not recording the data correctly. The technique of comparing nominally identical data from three or more stations allows the creation of virtually perfect data (subject to the assumption of independent errors and low error rates). The data from each station can then be compared individually with this “perfect” data to yield a very accurate measurement of the reception errors. Outside the areas where three or more stations are able to simultaneously receive the data, the data quality estimates based on the detection of errors in the predictable bits in each line can provide an indication of reception performance. Figure 2 shows that essentially the entire land area of Australia falls within the region where the highest quality data is obtained. Using these methods to identify which reception errors most probably occur at each station allows the creation of plots such as those in Figure 3 which show the average number of bit errors per line of data received at each station. Note that the extent of the region over which the data are plotted is effectively the horizon mask for each station, though the actual coverage area is considerably larger (in the East-West direction) since the scanner swath extends well beyond the sub-satellite point. In general the plots show that the quality of data received from every station is excellent (less than 5 bits in error per line) except for the occasional poor pass in Darwin (due to local interference?) and, as would be expected, at the limit of the field of view for each station when the satellite is close to the horizon. One obvious feature in Figure 3 which is not intuitively explained is the presence of two regions (light blue) of slightly higher error rate (5-10 bits per line) in the data for both Darwin and Perth. A hint as to the cause can be gained from observing that the regions are approximately the intersection with the areas covered by the East coast stations (but not the intersection of Perth with Darwin). A close scrutiny of the raw data files involved reveals that the very last pixel in every line from both Perth and Darwin are the same, but differ from the value of the same pixel common to all the other stations. Furthermore, examination of the adjacent pixels in the corresponding channel for the same line shows that the value obtained from the East coast stations is well correlated, while the value from both Perth and Darwin is essentially random. The Perth and Darwin stations are both run by the same institution and produce the data in the same format (presumably with the same software), a property shared by no other pair of stations. It thus seems reasonable to conclude that the error arises at the reception stations themselves subsequent to successful acquisition. Subsequent analysis shows that the problem was resolved later in the year. In any case it is not a serious obstacle for users since the extreme pixels in each scan are rarely used owing to the high view angles and low spatial resolution that they afford.

in sesending pascble on all asisi v14 wasAA-h NOom whic frationst of she numberTe 2. guriF


48

the first three months of 1996. This figure shows the number of stations recording the data at the sub-satellite point at every 100th line of the archived data (key at lower left corner).

Figure 3. Reception errors per HRPT line attributed to each of the Darwin, AIMS (Townsville), Perth and Hobart stations. All plots show the analysis averaged over 100 lines plotted at the sub-satellite point for the 50th line in each group. The spacecraft used is NOAA-14 with all ascending passes for the first three months of 1996 included (key in lower left corner, further explanation in the text).


49

Although the problem discussed amounts to only four or five bits, its presence is extremely clear from the analysis, illustrating the sensitivity of the technique. As the whole archive is processed, it will be possible to produce a performance history for each contributing station over the duration of the acquisition. More importantly, by collating the data and stitching it in near realtime (ie, soon after acquisition), it will be possible to detect any problems and provide feedback to the operators before too much data is affected.

6. Conclusions The curation of the 12 TB Australian 1km archive is a very large undertaking by CSIRO’s Earth Observation Centre. The primary benefits are:

1. making the largest possible dataset accessible for users through a process of cataloguing and consolidation to a common format;

2. protecting the data by migrating it to modern media and storing it in several locations; and

3. creation of a single best quality dataset with minimal corrupt data.

A secondary output is the development of a capacity to analyse reception station performance. The need to handle such a large volume of data is stimulating research into techniques for accessing and processing time-series datasets efficiently using large tape stores. A major outcome of this work is that it enables the longest possible comprehensive study of change in the Australian regional biosphere. As carbon accounting becomes more important, these data will be uniquely crucial in assessing quantitatively the evolution of Australia’s contribution to global change, both through landcover change and biomass burning. Investigations of the trends in ocean-related climate change will be similarly facilitated. As a library of continental-scale monitoring data for Australia the dataset is unsurpassed. As a long sequence of high quality raw data, this dataset will enable the refinement of many techniques for the consistent processing of such time-series. This will include navigation, calibration, cloud detection and compositing algorithms. As these techniques improve they will, in turn, enable the detection of more subtle trends within the dataset itself, and so improving our understanding of our environment.


50

51

Interaction between surface and atmosphere in AVHRR shortwave channels

D. M. O’Brien*, A. C. Dilley, and M. Edwards CSIRO Atmospheric Research, PMB1, Aspendale, Vic 3195 *Corresponding author. Email: [email protected]

1. Introduction Although new generation satellite sensors such as MODIS offer improved spectral coverage and radiometric accuracy over AVHRR, the long time series of AVHRR will be important for years to come in assessing changes in the environment, whether man-made or natural. However, the full potential of AVHRR (and for that matter MODIS) will not be realised unless the issues of calibration, directionality of surface reflectance and atmospheric absorption and scattering are addressed carefully and comprehensively. These are not easy tasks, and all too frequently they have been treated in isolation, even though the processes are coupled. Fortunately the situation is changing, and it is now possible to correct AVHRR for atmosphere and Bidirectional Reflectance Distribution Function simultaneously and expeditiously. The purpose of this short paper is to illustrate through examples the interaction between atmospheric and BRDF effects in AVHRR data and to outline a correction algorithm developed recently at CSIRO Atmospheric Research (Dilley et al., 2000).

2. Algorithms The coupling between atmosphere and surface was taken into account by Mitchell and O’Brien (1993), who showed how AVHRR data may be corrected accurately and efficiently if both the functional form of the indicatrix, defined to be the BRDF normalized by the surface albedo, and the optical properties of the atmosphere are known independently. In practice, the indicatrix may be deduced from multi-angle satellite data (such as POLDER, Deschamps et al. 1994), the water vapour field may be derived from either a meteorological forecast model (such as DARLAM, McGregor et al. 1993) or climatology (such as NVAP, Randel et al. 1995), and the aerosol properties may be obtained from either climatology or observations, if the latter are available.

L a n d E n v S a t W o r k s h o p . . . . . . . . . . . . . . . . . . . . A d e l a i d e , A u g u s t 2 0 0 0

52

In order to specify the process more precisely, we let ),,,( 00 φθφθdR denote the reflectance measured at the top of the atmosphere with the sun at zenith angle 0θ and azimuth 0φ and the satellite at ),( φθ , and we let )( 0θdA denote the corresponding albedo. The subscript d stands for ‘data’. We let lower case variables ),,,( 00 φθφθdr and )( 0θαd denote corresponding quantities measured at the surface. With these definitions, the data indicatrix gd is

=),,,( 00 φθφθdg /),,,( 00 φθφθdr )( 0θαd .

The purpose of the software is to estimate dr and dα , not only for the sun-satellite geometry that pertained at the time of the observation, but also for a standard geometrical configuration with the sun at ),( 00 φθ and satellite at ),( φθ where the overbar indicates an arbitrary reference angle. These estimated quantities we denote by a circumflex; thus, dr and dα are estimates of

dr and dα provided by the software.

In the course of obtaining the estimates, we will need, in addition to the optical properties of the atmosphere, a model for the surface BRDF and albedo, for which we use the notation

mr and mα , where the subscript m stands for ‘model’. The model indicatrix gm is then

=),,,( 00 φθφθmg /),,,( 00 φθφθmr )( 0θαm .

As an illustration, the model proposed by Roujean et al. (1992) has

22110 fkfkkrm ++=

where 1f and 2f are kernels representing geometric and volume scattering, and 0k , 1k and 2k are numerical parameters. Both 1f and 2f depend only on the incident direction ),( 00 φθ and

the exit direction ),( φθ of the photons. For this model, the surface albedo is

22110 IkIkkm ++=α ,

where 1I and 2I are simple trigonometric functions that depend only on the solar zenith angle 0θ . It is clear in this case that the indicatrix

22110

22110

IkIkkfkfkkgm

++++=

depends only on the ratios 01 / kk and 02 / kk .

The algorithm and its various products are represented by the flow chart of Figure 1. The circle labelled ATCO is the atmospheric correction algorithm developed by Mitchell and O’Brien (1993), the inputs to which are the reflectance ),,,( 00 φθφθdR at the top of the atmosphere and a model ),,,( 00 φθφθmg for the indicatrix. The output of ATCO is )(ˆ 0θαd the estimate of the surface albedo with the sun at zenith angle 0θ . The subsequent steps to estimate )(ˆ 0θαd and ),,,(ˆ 00 φθφθdr simply scale )(ˆ 0θαd from the actual geometry to the reference geometry, with the scaling factor determined by the model indicatrix.

A measure of the accuracy of the atmospheric correction is given in Figure 2. Each panel contains nine tiles, and each tile has 11 x 11 pixels. The tiles are labelled by the satellite zenith angleθ and the satellite azimuth φ relative to the sun at 00 =φ . The solar zenith angle

0θ is fixed at 45º. The horizontal dimension in each tile is aerosol optical thickness τ , increasing from 0 to 0.25 in steps of 0.025, while the vertical dimension is surface albedo, increasing

O ’ B r i e n e t a l . . . . . . . . . . . . C o r r e c t i n g A V H R R s h o r t w a v e c h a n n e l s

53

Figure 1: Flow chart for the atmospheric and BRDF correction module.

from 0 at the top of the tile to 0.5 at the bottom in steps of 0.05. For each combination of

),,,( Aτφθ , a multiple scattering radiative transfer code (Mitchell et al., 1992) was run to predict the radiance at the top of the atmosphere over a Lambertian surface. That radiance was used as data for the atmospheric correction algorithm, which returned its estimate of the surface albedo. The value shown in the pixel labelled by ),,,( Aτφθ is the error in the calculated albedo. Figure 2 shows that the errors generally are small, that they increase with increasing aerosol optical thickness, and that they depend only weakly on the scattering geometry.

Where Figure 2 used one set of retrieval tables, calculated with the correct water column and the aerosol optical depth at the centre of the range 125.0=τ , the operational code uses thirty two sets: one for each of the four optical depths 0.05, 0.15, 0.25 and 0.35, and, for each of these, eight values of integrated water column, 0, 10, 20, 30, 40, 50, 60 and 70 kg m-2. The algorithm selects the optical depth closest to the mean optical depth for the scene to be analyzed and reads into memory all eight sets of retrieval tables corresponding to this optical depth. The algorithm then uses interpolation between consecutive water column tables to perform a retrieval for the water column of the scene. This latest version of the module was tested, as was the earlier version, with calculated values of reflectance factors and nine sun-satellite configurations. However, in this instance, the test used one value of surface albedo (0.25), three values of optical depth and three values of integrated water column. Errors in retrieved albedos were small, being less than about 3% for approximately 90% of cases and never greater than 7%.

The algorithm is highly optimized so that it may be applied pixel-by-pixel to large images without incurring a prohibitive computational penalty. Table 1 shows a comparison of the execution times for ATCO and the 6S code (Vermote et al., 1997). Although ATCO and 6S give similar estimates of the surface albedo, ATCO is much faster. The superior performance of ATCO is attributed to the fact the algorithm on which it is based makes extensive use of precompiled tables.

φ)θ,φ0,θ0,θ0)gm(φ) = αd(θ,φ0,

Rd(θ0, φ0, θ, φ)

gm(θ0, φ0, θ, φ)

θ0,

αm(θ0)

rd(

θ0)θ0) = αd(θ0)

αm(

ATCO

αd(θ0)

αd(


54

Figure 2: Errors in surface albedo estimated by the atmospheric correction code (ATCO). The tiles are labelled by satellite zenith angle θ and the satellite azimuth φ relative to the sun at 00 =φ . The solar zenith angle is fixed at 450 =θ ° . The pixels within each 11 x 11 tile are labelled horizontally by aerosol optical thickness τ and vertically by surface albedo A . See the text for details.

0.10

0.05

0.00

-0.05

-0.10

Channel 2 φ = 0.00° φ = 90.00° φ = 180.00°

θ = 5.90°

θ = 28.63°

θ = 63.10°

0.10

0.05

0.00

-0.05

-0.10

Program ATCO test results - errors in calculated albedo

Channel 1 φ = 0.00° φ = 90.00° φ = 180.00°

θ = 5.90°

θ = 28.63°

θ = 63.10°


55

Table 1. Relative speeds of the atmospheric correction code ATCO and 6S (Vermote et al., 1997). The times quoted are seconds per pixel. The speed for 6S depends on whether aerosol is present, indicated by non-zero aerosol optical depth.

Channel Optical depth ATCO 6S Ratio

1 =0 0.000160 5.3 33,000

1 ≠0 0.000160 96.6 604,000

2 =0 0.000154 3.1 20,000

2 ≠0 0.000154 55.5 360,000

3. Impact of atmospheric and BRDF correction

The impact of simultaneous correction of AVHRR for atmosphere and surface is illustrated in the four panels of Figure 3, computed using the algorithm described above with aerosol optical depth fixed at 0.080 and 0.085 in channels 1 and 2, water vapour from the NVAP climatology (Randel et al., 1995) and BRDF data from POLDER interpreted according to the model of Roujean et al. (1992). Although not shown, almost identical results were obtained using aerosol optical depth interpolated from gridded seasonal mean values derived from the Global Aerosol Data Set (GADS) compiled by Köpke et al. (1997) and water vapour column interpolated from gridded values generated twice daily by the DARLAM 18-level regional model (McGregor et al., 1993).

Figure 3: Calibrated and navigated maps of the reflectance in AVHRR channel 2 synthesised from data on two successive overpasses of NOAA-14. Panel (a) has no additional processing, while panels (b)–(d) show the effects of correcting for the atmosphere, correcting for BRDF and correcting simultaneously for both atmosphere and BRDF.

-10

(a) Uncorrected

-15

-20

0

-25

500

-30

1000

-35

1500

-40

2000

-45

2500

3000

160

3500

150140130120110

NOAA-14 11368 11367 15/03/97 Refl62

5000

4000

3000

2000

1000

0

(d) Corrected for atmos and BRDF

NOAA-14 11368 11367 15/03/97

Refl12

-10

110

-15

120

-20

130

-25

140

-30

150

-35

160

-40

-45

160150140130120110

NOAA-14 11368 11367 15/03/97

Refl22

4000

3000

2000

1000

0

(c) Corrected for BRDF

-10

-15

-20

-25

-30

-35

-40

-45

160150140130120110

NOAA-14 11368 11367 15/03/97 Refl32

4000

3000

2000

1000

0

(b) Corrected for atmos-10

-15

-20

-25

-30

-35

-40

-45


56

Panel (a) of Figure 3 shows calibrated and navigated, but otherwise uncorrected, AVHRR reflectances in channel 2. The image is a composite of two successive overpasses by NOAA-14, separated by approximately 102 minutes, and the discontinuity through the image marks the mid-point of the overlap region of the overpasses. The eastern edge of the western pass is brighter than the western edge of the eastern pass, an effect caused principally by the BRDF of the surface. Panels (b)–(d) show the effect of correcting for atmosphere, correcting for BRDF and correcting simultaneously for both atmosphere and BRDF. Correction for BRDF appears to remove most of the discrepancy between the passes, but on closer inspection a line of discontinuity is still visible. In contrast, correction for both atmosphere and BRDF leads to near perfect matching across the boundary between the overpasses.

Although the effect of the atmospheric scattering and absorption appears to be secondary, nevertheless it is important because it produces a significant brightening of the image, as can be seen by comparing panels (c) and (d) of Figure 3. Furthermore, the brightening is highly variable, depending on both the surface and the atmospheric composition. The impact on normalized difference vegetation index (NDVI), shown in Figure 4, is both large in magnitude and spatially non-uniform. In particular, it is clear that the effect cannot be represented as either a shift or rescaling of NDVI, and that the effect of the atmosphere should not be neglected in large scale or multi-temporal studies of surface vegetation.

4. Sensitivity analysis

In order to separate the components of atmospheric correction caused by water vapour, aerosol and molecular (Rayleigh) scattering, we processed a representative image of south-eastern Australia (NOAA-14, afternoon ascending overpass of 1997-12-10, orbit number 15177) under three sets of conditions: Rayleigh scattering, no aerosol and no water vapour; Rayleigh scattering, no aerosol and a water column of 30 kg m-2; Rayleigh scattering, aerosol optical depth of 0.10 and no water vapour. We extracted reflectances from the image for nine sites that encompassed a range of reflectance signatures. The results, plotted in Figure 5, allow the following conclusions.

1. Correction for molecular scattering decreases channel 1 reflectance at all sites but has relatively little effect on channel 2, with channel 2 reflectance generally decreasing over dark targets and increasing over bright targets.

2. Correction for water vapour has little effect in channel 1, but leads to significant increases in channel 2 relectance at all sites.

3. Correction for aerosol decreases the reflectance in both channel 1 and 2, with the magnitude of the response in channel 1 generally exceeding that in channel 2.

4. Although correction for water vapour increases the NDVI at all sites, the effect is less pronounced for greener sites, where the changes caused by water vapour are roughly parallel to the lines of constant NDVI.

5. Similarly, correction for aerosol increases the NDVI at all sites, but the effect is smaller over sites that are bright in channel 1, where the changes caused by aerosol are roughly parallel to the lines of constant NDVI.


57

Figure 4: AVHRR composite images of Australia showing NDVI computed from (a) uncorrected reflectances at the top of the atmosphere, (b) data corrected for atmosphere and BRDF, (c) uncorrected data with NDVI scaled to the same values as in (b), and (d) the difference between NDVI estimates obtained with the corrected and uncorrected data. (Scale: 700 = 0.7 NDVI units)

Figure 5: Uncorrected top-of-atmosphere reflectances and variously corrected surface reflectances extracted for nine sites in an image of south-eastern Australia. The image data derive from an afternoon ascending overpass of NOAA-14 for 1997-12-10, orbit number 15177.

700

NOAA-14 11368 11367 15/03/97 NDVC-U

300

0

(d) Difference400

NOAA-14 11368 11367 15/03/97

NDVIR1

0

(b) Corrected

0

(c) Uncorrected but scaledNOAA-14 11367 0412Z 15/03/97 NDVIR6

700

0

(a) Uncorrected

NOAA-14 11368 11367 15/03/97

NDVIR1

0.8

NDVI =

0.70.60.50.40.30.2

0.1

98

7

6 54

3

21

surface corrected for Rayleigh and aerosol ( τ = 0.10 ) and to reference geometry surface corrected for Rayleigh and water ( W = 30 kg m-2 ) and to reference geometry surface corrected for Rayleigh and to reference illumination and observation geometry TOA uncorrected

Ban

d 1

refle

ctan

ce (

%)

20

15

10

5

00 10 20 30 40 50

Band 2 reflectance (%)


58

Figure 6: BRDF and atmospheric corrections for several sites in the Strzelecki Desert of South Australia.

corrected

uncorrected

GN2

(a) BRDF correction

0

GN1

10 20

LM2

30 40

LM1

50 60

RO2

70

RO1

LW2

LW1

TE2

TE1

TW2

TW1

TT2

TT1

Ref

lect

ance

(%)

- fo

rwar

d sc

atte

ring

case

60

50

40

30

20

10

0

Reflectance(%) - backward scattering case

706050


403020100

corrected

(c) Both corrections

uncorrected

GN2

GN1

LM2

LM1

RO2

RO1

LW2

LW1

TE2

TE1

TW2

TW1

TT2

TT1

Ref

lect

ance

(%)

- fo

rwar

d sc

atte

ring

case

60

50

0

40

10

30

20

20

30

10

40

0

50

60


706050403020100

corrected

uncorrected

GN2

GN1

LM2

LM1

RO2

RO1

LW2

LW1

TE2

Ref

lect

ance

(%)

- fo

rwar

d sc

atte

ring

case

TE1

TW2

TW1

TT2

TT1

(b) Atmospheric correction


59

It is clear from Figure 5 that corrections for a Rayleigh atmosphere, water vapour and aerosol loading under commonly encountered conditions are associated with relatively large changes in reflectances and NDVI values, so failure to make these corrections normally would introduce substantial error. The non-additive nature of the effects of corrections on NDVI noted above is also evident in panel (d) of Figure 4.

Correction for BRDF and correction for atmospheric effects are considered separately in Figure 6 for a number of targets in the Strzelecki Desert of South Australia. This area is generally bright, but there are numerous dry lakes whose surfaces of hard-packed clay are comparatively dark. Each panel of Figure 6 plots along the horizontal axis the reflectance measured on an orbit to the west (backward scattering case), while along the vertical axis is plotted the reflectance measured on an orbit to the east (forward scattering case). The open and closed dots represent the uncorrected and corrected data respectively. The labels denote the sites and the trailing numerals indicate the AVHRR channels. Panel (a) shows that initially the backward scattering case is brighter (as expected) and that the effect of BRDF correction is slight overcompensation. Panel (b) shows that atmospheric correction darkens the dark targets and brightens the bright targets in channel 1, but reflectances increase for all targets in channel 2. The composite of both corrections, shown in panel (c), effects a near perfect correlation between the forward and backward views of the targets.

5. Conclusion

Robust procedures for simultaneous correction of AVHRR for both atmosphere and BRDF are now available. They require data from both meteorological forecast models and climatology to fix the atmospheric composition and from other satellites such as POLDER, MODIS and AVHRR (O’Brien et al., 1998) to determine the surface BRDF. The effects are both large and important, particularly for multi-season and continental studies, such as those concerned with assessing changes in land cover and net primary production. Work is in progress to produce ancillary data sets to enable the full historical archive of AVHRR data to be processed.

6. Bibliography Deschamps, P.-Y., F.-M. Bréon, M. Leroy, A. Podaire, A. Bricaud, J.-C. Buriez, and G. Sèze, 1994: The

POLDER mission: Instrument characteristics and scientific objectives. IEEE Trans. Geosci. Remote Sens., 32, 598–615.

Dilley, A. C., M. Edwards, D. M. O’Brien, and R. M. Mitchell, 2000: Operational AVHRR processing modules: atmospheric correction, cloud masking and BRDF compensation. Internal paper 14, CSIRO Atmospheric Research, PMB 1, Aspendale Vic 3195, Australia. 24 pp.

Köpke, P., M. Hess, I. Schult, and E. P. Shettle, 1997: Global aerosol data set. Technical report 243, Max-Planck-Institut für Meteorologie, Hamburg.

McGregor, J. L., K. J. Walsh, and J. J. Katzfey, 1993: Nested modelling for regional climate studies. Modelling change in environmental systems, A. J. Jakeman, M. B. Beck, and M. J. McAleer, Eds. Wiley, 367–386.


60

Mitchell, R. M., and D. M. O’Brien, 1993: Correction of AVHRR shortwave channels for the effects of atmospheric scattering and absorption. Remote Sens. Environ., 46, 129–145.

Mitchell, R. M., D. M. O’Brien, and B. W. Forgan, 1992: Calibration of the NOAA AVHRR shortwave channels using split pass imagery: I. Pilot study. Remote Sens. Environ., 40, 57–65.

O’Brien, D. M., R. M. Mitchell, M. Edwards, and C. C. Elsum, 1998: Estimation of BRDF from AVHRR short-wave channels: tests over semiarid Australian sites. Remote Sens. Environ., 66, 71–86.

Randel, D. L., T. H. V. Harr, M. A. Ringerud, D. L. Reinke, G. L. Stephens, C. L. Combs, T. J. Greenwald, and I. L. Wittmeyer, 1995: An introduction to the NASA water vapor project data set (NVAP). Technical report Contract NASW-4715, National Aeronautics and Space Administration, Washington, D. C. USA.

Roujean, J. L., M. Leroy, and P. Y. Deschamps, 1992: A bidirectional reflectance model of the earth’s surface for the correction of remote sensing data. J. Geophys. Res., 97, 20455–20468.

Vermote, E. F., D. Tanré, J. L. Deuzé, M. Herman, and J. J. Morcrette, 1997: Second simulation of the satellite signal in the solar spectrum (6s): an overview. IEEE Trans. Geosci. Remote Sens., 35, 675–686.

61

Filtering Pathfinder AVHRR Land NDVI Data for Australia J.L. Lovell and R.D. Graetz

CSIRO Earth Observation Centre, GPO Box 3023, Canberra, ACT, 2601 [email protected]; [email protected] Abstract: The Pathfinder AVHRR Land (PAL) data set is a consistently processed 13-year

time series which can be used in modelling and monitoring of climate and land cover

change. The data are derived from AVHRR global area coverage data, resampled to 8km

resolution and 10-day composite images. The compositing reduces the effect of cloud, but

there is still significant cloud contamination and other noise present in the PAL dataset.

The best index slope extraction algorithm was tested and a modified algorithm applied to

the PAL dataset for the Australian continent. This paper describes the modifications and

presents some examples of the reduction in noise which is achieved by this method. A

comparison is also made with the PAL cloud flags. The filtered dataset is available by

contacting the authors.

1. Introduction The NASA/NOAA-sponsored Earth Observing System Pathfinder projects produced time series of global data sets for use in climate change research. The Pathfinder AVHRR Land (PAL) dataset comprises global, 10-day composite images of AVHRR data covering the period mid-July 1981 to September 1994 (James and Kalluri 1994). This paper describes an NDVI filtering process which was applied to the PAL data for the Australian continent, to minimise noise such as cloud contamination.

2. PAL processing The PAL dataset was derived from global area coverage data (GAC) acquired by the 5-channel AVHRR instruments, which is produced by onboard averaging and sampling to a resolution of 4km at nadir. The GAC data were navigated and corrections for sensor degradation, Rayleigh scattering and ozone absorption were applied. Cloud contamination was flagged by the CLAVR (Stowe et al. 1991) process based on 2X2 pixel arrays. NDVI was then calculated from the channel 1 and 2


62

reflectances. Re-sampling to 8km resolution was performed by forward, nearest neighbour mapping, selecting the highest NDVI value from all pixels within 42° of nadir. Small gaps were filled by copying values from adjacent pixels. Pixels for which the solar zenith angle was greater than 80° were not used. These re-sampled data were composited in quasi 10-day periods according to the maximum value of NDVI. The CLAVR cloud flags were not used in the re-sampling or compositing processes. The PAL data are mapped to the Goodes Interrupted Homolosine projection. The data layers and their units are listed in Table 1 (see also, Agbu and James 1994). Table 1. Pathfinder AVHRR Land data set layers.

Data Units NDVI ndvi CLAVR flag Numeric value from lookup tables QC flag Numeric value from lookup tables Scan angle radians Solar zenith angle radians Relative azimuth radians Ch 1 reflectance % reflectance Ch 2 reflectance % reflectance Ch 3 btemp K Ch 4 btemp K Ch 5 btemp K Day of year DDD.HH Compositing by maximum value of NDVI is used to minimise the effect of cloud in NDVI data, the assumption being that all contamination results in lower values of NDVI. This can be effective if the compositing period is quite long (2-4 weeks). However if the period is too long, short-term changes in vegetation condition may be lost. The PAL 10-day compositing period was not sufficient to remove all cloud affected pixels. Gutman and Ignatov (1996) found that some improvement could be made by taking the CLAVR flags into account and reducing the dataset to a coarser spatial or temporal resolution. However, they note that a large number of pixels are labelled as ‘mixed’. They maintain that some of these are in fact cloud-free and have been wrongly classified by the CLAVR algorithm. Prince and Goward (1996) also noted the large number of flagged pixels, commenting that in some cases as many as 90% of pixels in a 10-day composite image are flagged with some type of cloud cover.

3. NDVI noise removal techniques Maximum value composition (MVC) is often used to minimise the effects of cloud contamination in NDVI data but, as already noted, this technique has inherent problems. The length of the compositing period affects the success of the noise removal. A long period will be more effective in removing noise, but may result in

L o v e l l a n d G r a e t z . . . P a t h f i n d e r A V H R R L a n d N D V I d a t a

63

the loss of important short term changes. Spurious high values due to data transmission errors will not be removed by MVC. The best index slope extraction (BISE) method of Viovy et al. (1992) was proposed as an alternative to MVC. This method is based on the predictability of seasonal changes in vegetation and accounts for the fact that growth and senescence phases are often asymmetric. It can also accommodate sudden changes due to fire, deforestation or crop harvest. Sudden rises and falls in NDVI values are not compatible with the gradual process of regrowth, but are a feature of changing cloud conditions or viewing angles. This is the basis for the BISE algorithm, which applies the following rules. From the first date of the time series, the algorithm searches forward, accepting points if their value is higher than the preceeding one. If a decrease is found, the search is continued over a pre-defined sliding period (e.g. 30 days). The decrease is only accepted if there are no points within the sliding period with values greater than the previous high value minus 20% of the size of the decrease. That is, the decrease is only accepted if it is followed by a gradual increase consistent with regrowth. If an acceptable point is found within the sliding window, this point is chosen and the low point is rejected. Intermediate values are filled by interpolation and the search begins again from the last accepted point. The threshold of 20% of the decrease was chosen empirically for West African conditions. It should be adjusted to suit the typical regrowth rate of the area under consideration. Similarly, the length of the sliding period should be adjusted to suit the predominant cloud conditions. In addition to the tests applied within the sliding window, the BISE algorithm rejects any points with a random increase of greater than 0.1 as such fluctuations from one day to the next are unlikely from natural surfaces and are more probably the result of data transmission errors.

4. Modified BISE filtering of PAL NDVI The BISE algorithm is designed for use with daily NDVI data. The performance is not expected to be as good with the PAL 10-day composites as artifacts of the compositing process are already present in this dataset. However, the PAL data contains some noise of the type that could be removed by a BISE-type procedure. Adjustments have been made to the BISE rules so that the algorithm works more effectively with the PAL NDVI data for Australia. The test for spurious high points was modified to look for a spike (i.e. an increase immediately followed by a decrease). This refinement was necessary because the time between consecutive points is not equal in a composited dataset. The time may range from 1 to 20 days, depending on which dates within the compositing decads were chosen, so a simple threshold test is not appropriate. The BISE algorithm performed poorly when there was a long-term, gradual decrease in NDVI. The test within the sliding window was modified to take into account the local gradient of the data. The local gradient was calculated for each point in the time series, over the same time period as the sliding window length. The resulting time


64

series of gradients was then smoothed. When the gradient was negative, the test value for high points within a sliding window was calculated by extrapolating from the high point before the decrease according to the local gradient, then subtracting 20% of the decrease. Figures 1 and 2 show the results of the modified BISE filtering on four test sites using sliding windows of 3 decads (Figure 1) and 6 decads (Figure 2). The threshold for spurious spikes was set at 0.1. The sites were chosen for their different vegetation types and climatic zones. The first is an area of forest in Western Australia. This site may be expected to be cloudy quite often due to its proximity to the coast. The second site is an area of wheat crop in Western Australia and clearly shows the annual growth and harvest cycle. The third time series is from the Nullabor plain, an arid and largely cloud-free region. The final test site was in tropical forest in Queensland. This area experiences long periods of cloud cover, particularly during the wet season. The results in Figures 1 and 2 show that the modified BISE filtering greatly reduces the noise in the NDVI time series. The filtered data are realistic and maintain seasonal variation. The sliding window in the first example (3 compositing periods) was not long enough to remove all the cloud effects in the forest and tropics sample sites, but is reasonable for the less cloudy wheat crop and Nullabor sites. The 6 decad window is more effective on the cloudy sites and still retains most of the sharp changes at the other sites. Ideally, the length of the window should be varied in different climatic zones across the continent, but a window length of 6 compositing periods was judged to be a good compromise and was applied for all of Australia. Figures 3–7 are examples of NDVI images of Australia before and after the modified BISE filtering. The colours range from red (low NDVI, dry vegetation and bare soil) through yellow to green (high NDVI, green vegetation). The NDVI images were reprojected from Goodes Interrupted Homolosine to Plate Carrée prior to filtering. These images demonstrate a variety of noise effects which were removed or reduced by the modified BISE. Figure 3 is an area of northern Western Australia. A line of spurious values was removed by the filtering. Figure 4 shows evidence of cloud (large red patches) in northern Australia during the tropical wet season in February 1992. The result of the filtering is a much more spatially consistent image. Another example of cloud removal is shown in Figure 5. This image shows south-eastern Australia during May 1982. There are mottled red patches in the original data. The filtered image is more spatially consistent and accords with the expected greenness of this area. Other similar examples can be found throughout the dataset. Both Figures 6 and 7 show the southern part of Australia and illustrate large areas of missing data. This is due to rejection of data when the solar zenith angle was greater than 80°. Some of these missing areas were repaired (interpolated) by the filtering process. Sharp discontinuities between adjacent satellite passes evident in the south east of Figure 6 and in southern Western Australia in Figure 7 were also normalised by the filtering.


65

Figure 1. NDVI from Pathfinder AVHRR land dataset for four test pixels through the time series July 1981 to September 1994 (solid grey lines) filtered with a sliding window of 3 decads and spike threshold of 0.1 (dashed black lines).


66

Figure 2. NDVI from Pathfinder AVHRR land dataset for four test pixels through the time series July 1981 to September 1994 (solid grey lines) filtered with a sliding window of 6 decads and spike threshold of 0.1 (dashed black lines).


67

Figure 3. NDVI from Pathfinder AVHRR land dataset for the compositing period 21-31 August 1981, in northern Western Australia. Left original, right filtered.

Figure 4. NDVI from Pathfinder AVHRR land dataset for the compositing period 21-28 February 1992, in northern Australia. Left original, right filtered.

Figure 5. NDVI from Pathfinder AVHRR land dataset for the compositing period 21-31 May 1982, in south eastern Australia. Left original, right filtered.

Figure 6. NDVI from Pathfinder AVHRR land dataset for the compositing period 11-20 June 1988, in southern Australia. Left original, right filtered.

Figure 7. NDVI from Pathfinder AVHRR land dataset for the compositing period 21-30 June 1993, in southern Australia. Left original, right filtered.


68

These examples demonstrate that, although the noise removal was by temporal filtering performed on individual pixel time series, the effect is to improve the spatial uniformity of the images. Another test of the process is to look for seasonal changes expected in a series of images. Figure 8 shows images for the middle decad (days 11-20) of each month in 1987. The evergreen forest regions in southern Western Australia, Tasmania and along the eastern sea-board are green throughout this time series. Some small red patches can be seen in Tasmania during June and July. This is most likely due to residual cloud and suggests that the length of the filtering window was not sufficient to remove persistent winter cloud in this area. Southern grasslands, shrublands and croplands are greenest in spring. This is evident in the extensive green area in south-eastern Australia which is largest in September. The tropical regions are greenest in March and April following the wet season. A seasonal crop cycle is most clearly illustrated by the wheat-growing region of southern Western Australia. This is seen as a red patch from January to May and again in November and December. The red area becomes green in the late winter-spring months when the crop is growing vigorously. A similar trend can also be seen in smaller cropping areas of south-eastern Australia. The full time series of monthly NDVI images can be viewed at www.eoc.csiro.au/jlat.htm. All these observations of seasonal effects are realistic and show a smooth progression through the year. Thus the temporal filter has preserved seasonal changes.

5. Comparison with CLAVR flags Figure 9a shows an example of the CLAVR cloud flags. Pixels shown in black were flagged as cloudy or mixed (only 1% of those shown were flagged as cloudy). The image in Figure 9b shows (for the same PAL image) the pixels which were changed by more more than 5% in the modified BISE filtering. The patterns are similar, but the number of pixels flagged by the CLAVR process is greater than the number that changed due to the NDVI filtering. Table 2 gives the number of pixels with CLAVR flags indicating clear, mixed and cloudy for the whole dataset. Also shown is the percentage of those pixels which changed in the NDVI filtering and also the percentage that changed by more then 5%. The percentages are of the total number of clear, mixed and cloudy pixels as flagged by CLAVR. The majority of the pixels changed by the NDVI filtering were flagged as either mixed or cloudy by CLAVR, but the two methods of cloud detection do not agree entirely. Table 2. Pixel numbers for CLAVR flags related to changes in NDVI due to filtering.

CLAVR class Number Percent in class with changed NDVI

Percent in class with NDVI changed by >5%

Clear 3.8E7 50% 11% Mixed 1.2E7 70% 39% Cloudy 7.3E5 77% 56%


69

Figure 8. NDVI time series for 1987.

a.

Figure 9. (a) Pixels marked cloudy or mixed by CLAVR (black). (b) Pixels changed by more than 5% in the modified BISE filtering.

b.


70

6. Conclusion The product of the modified BISE filtering is useful but not definitive. This is an intermediate product and it is hoped that this dataset will be of value to its users. EOC staff are in the process of producing a consistent AVHRR time series (including NDVI) for the Australian continent of the highest quality. This will employ CSIRO best practice algorithms for calibration, navigation, atmospheric correction, cloud clearing and removal of the effects of different solar and viewing angles (BRDF). The filtered PAL dataset as described in this report can be obtained by contacting the authors.

7. Acknowledgements Data used by the authors in this study include data produced through funding from the Earth Observing System Pathfinder Program of NASA’s Mission to Planet Earth in cooperation with National Oceanic and Atmospheric Administration. The data were provided by the Earth Observing System Data and Information System, Distributed Active Archive Center at Goddard Space Flight Center which archives, manages, and distributes this data set. Thanks to Susan Campbell for valuable assistance in the reprojection of the PAL dataset.

8. References Agbu, P.A. and James, M.E., 1994, The NOAA/NASA Pathfinder AVHRR land data set users

manual. Goddard Distributed Active Archive Center, NASA, Goddard Space Flight Center, Greenbelt.

Gutman, G. and Ignatov, A., 1996, The relative merit of cloud/clear identification in the NOAA/NASA Pathfinder AVHRR Land 10-day composites. Int. J. Remote Sensing, 17, 3295–3304.

James, M.E. and Kalluri, S.N.V., 1994, The Pathfinder AVHRR land data set: An improved coarse resolution data set for terrestrial monitoring. Int. J. Remote Sensing, 15, 3347–3363.

Prince, S.D. and Goward, S.N., 1996, Evaluation of the NOAA/NASA Pathfinder AVHRR Land data set for global primary reduction modelling. Int. J. Remote Sensing, 17, 217–221.

Stowe, L.L., McClain, E.P., Carey, R., Pellegrino, P., Gutman, G.G., Davis, P., Long, C. and Hart, S., 1991, Global distribution of cloud cover derived from NOAA/AVHRR operational satellite data. Advances in Space Research, 3, 51–54.

Viovy, N., Arino, O. and Belward, A.S., 1992, The Best Index Slope Extraction (BISE): A method for reducing noise in NDVI time-series. Int. J. Remote Sensing, 13, 1585–1590.

71

Advances in CAPS

Peter Turner and Harvey Davies

CSIRO Atmospheric Research, Private Bag 1, Aspendale, VIC 3195, Australia email: [email protected], [email protected]

Abstract: The CAPS (Common AVHRR Processing System) software package has been

developed by CSIRO to support the production of standard AVHRR data products. We

give a brief outline of how the CAPS software works, and illustrate the functionality of

the package with two simple examples. We also describe some new additions to the

package including a new Fortran/C subroutine interface and a cloud detection scheme.

1. Introduction CAPS is a software package that was designed to promote the production of standard calibrated and geolocated Advanced Very High Resolution Radiometer (AVHRR) data products. CAPS is an extension to the Tcl scripting language designed by Ousterhout (1994, 2000). CAPS adds AVHRR-specific processing functions (Turner et al. 1998) and general-purpose numeric array-processing functions to Tcl. In addition, CAPS adds binary and Hierarchical Data Format (HDF, NCSA 2000) file input and output to Tcl. CAPS has been designed to be efficient and portable and runs under a number of versions of Unix, and under Linux, Windows 95/98 and NT 4.0. The current version of CAPS uses shared libraries that provide access to other extensions of Tcl. The next section of this paper provides a brief summary of the CAPS user interface. In the following section an extension to the user interface is described that allows users to easily include their own Fortran and C subroutines. In the final section the requirements for a reliable cloud detection scheme for AVHRR data is discussed and the implementation of a cloud detection algorithm is described.

2. CAPS Implementation The Tcl language has the basic syntax: command arg1 arg2 arg3 …


72

Tcl has been designed as an extendable scripting language. New commands can be defined within the Tcl language. It is also possible to define a new command using C to produce a shared library (dynamic load library, dll). CAPS has been implemented by extending Tcl with the addition of two shared libraries. The NAP (numeric array processor) library defines the nap command that supports the application of a range of mathematical operators and functions on numeric array objects (NAOs). An NAO is essentially a container for n-dimensional binary data. NAOs are object oriented and can operate as commands in the Tcl environment. The default command for an NAO is to print its contents. There are a number of methods that can be applied to an NAO to perform different functions including writing the NAO's data to an HDF file. A full list of NAO methods can be found in the NAP section of the CAPS web documentation (Turner and Davies 2000). The CAPS library defines the caps command that supports general satellite instrument and AVHRR-specific data processing functions. To illustrate how the nap and caps commands work an example of each is given in the following subsections.

2.1 The n a p command The nap command evaluates mathematical expressions involving NAOs. For example:

nap "x = ap(-3.5,3.5,0.1)" nap "sincx = (x != 0) ? sin(1p1*x)/(1p1*x) : 1" $sincx set coord x plot_nao sincx Here ap is a function that generates an arithmetic progression, in this case starting at -3.5 and ending at 3.5 in steps of 0.1. The next line is a NAP conditional expression of the form:

nap "result=<condition>?<true value>:<false value>"

where in the example the result is the sinc(x) function for x non-zero and 1 otherwise. The 1p1 term is a lexeme for π. The x-axis coordinates are assigned to sincx using an object-oriented command with the set coordinate method. The plot_nao command is a Tcl procedure providing an interface to the graph command in the BLT package (BLT 2000). This command plots the vector in sincx as a line graph and the result is shown in figure 1. BLT offers an extensive range of options to customise the look of the graph that is plotted. Elements of the graph are treated as objects and can be modified separately without the need to redraw the entire graph. The BLT shared library has been included as part of the version 2.2 release of CAPS.

T u r n e r a n d D a v i e s . . . . . . . . . . . . . . . . . . . . . . A d v a n c e s i n C A P S

73

Figure 1. Sinc(x) plot using BLT.

2.2 The c a p s command The caps command, through a series of sub-commands, provides satellite-specific functions. The general form of the caps command is:

caps sub-command arg2 arg3 …

A range of AVHRR-specific caps commands covers AVHRR calibration and navigation and has been described by Turner et al. (1998) and Turner and Davies (2000). An example that reads ATSR-2 (Along Track Scanning Radiometer) data into CAPS illustrates the general principles. set fileName "aus_d-9511230007-232669-960924-2av100.gbt-tvlx" nap "lineCV = pixelCV = 1..512" caps get_atsr_gbt nadir_v670 nadir_v870 nadir_v16 nap "nadir_v670 = abs(nadir_v670)"

nap "nadir_v870 = abs(nadir_v870)" nap "nadir_v670 = nadir_v670(,-(1..512))" nap "nadir_v870 = nadir_v870(,-(1..512))" nap "nadir_v16 = nadir_v16(,-(1..512))"

plot_nao "nadir_v16 /// nadir_v870 /// nadir_v670" In this example fileName is set to the name of an ATSR-2 Gridded Brightness Temperature file. A range of lines and pixels to be read from the file is set in lineCV and pixelCV. The ATSR-2 GBT file contains a number of 512x512 pixel scenes, so setting line and pixel coordinate variables to 1..512 selects the whole scene. The caps command has a sub-command get_atsr_gbt. The braces allow the arguments, in this case the ATSR-2 channel names, to occupy a number of lines of the script. Data from the three selected ATSR-2 channels are read into the variables named nadir_v670, nadir_v870 and nadir_v16. The nadir_v670 and nadir_v870


74

contain cosmetic fill values that are set negative and are made positive by the absolute value function (abs) for the purpose of display. The image arrays in this example are reversed in pixel order and this is corrected by subscript expressions. The final command is the plot_nao that plots the three-dimensional image created by the application of the "///" concatenation operator on the three ATSR-GBT channels. The three ATSR-2 channels nadir_v16, nadir_v870 and nadir_v670 are assigned to red, green, and blue respectively, and the result is shown in Figure 2.

Figure 2. ATSR-2 false colour scene drawn by BLT.

3. Enhancements

3.1 Structure Version 1 of CAPS ran only under Unix and was based on special CAPS shells. Version 2 of CAPS is available for Unix, Linux and Windows 95/98/NT and uses the standard tclsh and wish shells/executables to load. Version 2 uses shared libraries for Unix and Linux and dynamic load libraries for Windows. This new library structure allows CAPS to be used with other Tcl extensions.

3.2 Tcl Procedures Much recent development has been done using the Tcl scripting language rather than C. The Tcl procedure avhrr2hdf provides a flexible interface to the CAPS facilities for reading AVHRR data, calibrating, geolocating and writing results to HDF files. Tcl comes with the Tk extension, which provides windowing commands. These have been used to develop a GUI for CAPS. When CAPS is started using wish a small menu appears as shown in Figure 3.


75

Figure 3. CAPS Top Level Menu.

The figure shows the menu with the browse button selected. This exposes a sub-menu allowing the user to look at the list of Tcl variables currently defined, or browse AVHRR, HDF or ATSR data files.

3.3 Numeric Types The number of data types in NAP has been increased to support most of those defined for use in HDF files including those types used for data from the NASA Terra instruments. NAP now supports the data types shown in Table 1. Table 1. NAP Data Types.

Name Description u8 unsigned 8-bit integer u16 unsigned 16-bit integer u32 unsigned 32-bit integer i8 signed 8-bit integer i16 signed 16-bit integer i32 signed 32-bit integer f32 32-bit float f64 64-bit float character 8-bit text ragged compressed data, all basic types boxed pointer type

4. Fortran and C Subroutine Interface The nap command supports a number of mathematical functions including functions for type conversion, array manipulation and special functions such as morphological erosion and dilation. Tcl procedures can be used to define new functions for the nap command, but these functions are sometimes inefficient. To overcome this limitation, a special interface has been added to CAPS to make it easy for a user to add their own C or Fortran code defining a new Tcl command. These new commands are special in that their inputs and outputs are defined as NAOs, allowing close interaction with the nap command. The new interface requires a C or Fortran subroutine and a definition describing the inputs and outputs to the subroutine. To illustrate how this new feature works the following example shows how a partial


76

sum function written in C can be included. The C code for the psum subroutine is shown below:

void psum(int *n, float *x, float *result) int i; float sum = 0; for (i = 0; i < *n; i++) result[i] = sum = sum + x[i];

The intent of the C code is to take a vector x containing n elements and produce a new vector result, containing the n element partial sum. In version 2.2 of CAPS there is a procedure called make_dll whose arguments define the name of the Fortran or C function and details (name, data type, intent) of its arguments. For example:

make_dll psum n i32 in x f32 in result f32 inout

defines the parameters for the psum subroutine. Make_dll generates the C interface code to add psum as a new Tcl command, compiles it, and creates the new shared library. The new Tcl command can be accessed by loading the shared library and used as shown in the following example:

load ./libpsum.so nap "a = 2f 1.5f 0f -0.5f" nap "result = +a" psum nels(a) a result $result 2 3.5 3.5 3 The inclusion of this new facility in CAPS makes it easy to define new commands using Fortran and C subroutines that can read and write data from NAOs. As an example, the CLAVR algorithm described in the next section could be recoded in C and added as a built-in command using this interface. The interface also provides an efficient way of interfacing the atmospheric correction and BRDF compensation modules (Dilley et al. 2000) to CAPS.

5. Cloud Detection for AVHRR The detection of cloud in remotely sensed satellite imagery and in particular for AVHRR data is an essential precursor to the accurate retrieval of the properties of the Earth’s surface. Small amounts of undetected cloud can affect sea-surface temperature retrievals significantly (Saunders 1986). Normalised Difference Vegetation Index values can also be seriously affected by both cloud and cloud shadow. From another perspective, measurements of the amount and type of cloud is vital in understanding the Earth's radiation budget and hence the impact of cloud on global climate change.


77

Many cloud detection schemes have been developed during the last 10 years. Most algorithms are constructed of a battery of tests with thresholds based on our understanding of cloud physics. Saunders (1986), Saunders and Kriebel (1988) published the Apollo algorithm, Stowe et al. (1999) developed the CLAVR algorithm (Clouds from AVhRr) and more recently Ackerman et al. (1998) have developed an algorithm that more cleverly combines the result of the tests it uses. The requirements for a successful operational cloud detection algorithm are that the algorithm accurately detects cloud over the full spatial and temporal domain, and does the detection process efficiently. The CLAVR algorithm fits these criteria and has been implemented as a Tcl script in CAPS. The algorithm proposed by Stowe et al. divides into four separate groups of tests (see figure 4) for combinations of cloud detection over land and water during the day and night. The Stowe et al. algorithm includes a range of tests for high reflectance, low temperatures and some spatial coherence tests. The spatial coherence tests are designed for GAC 4-km resolution data and the thresholds should be adjusted for AVHRR 1-km data. The methodology adopted in coding the algorithm has centred on the use of the NAP conditional expression described in section 2.1. For example the daytime land cloud algorithm starts with the “Reflectance Gross Cloud Test” (RGCT).

nap mask =(mask == clear && ref1 > 44) ? mcrgct : mask

where mask is a byte mask covering the satellite scene, ref1 is reflectance derived from AVHRR channel 1 and clear and mcrgct are unique numbers between 0 and 255. Subsequent parts of the RGCT test to see if there is potential for restoring the pixels flagged as mcrgct to clear. For example:

nap mask=(mask == mcrgct && bt4 > 293) ? prestore : mask

where bt4 is brightness temperature (K) derived from AVHRR channel 4 and prestore is the potential restore flag with a unique value between 0 and 255. The tests employed all fall into this general form. If the expression is true then set a value otherwise leave the value alone. There is an inefficiency in the implementation using nap commands because there is an implicit loop through all the data for every operation. Combining some of the tests would improve efficiency, but reduces clarity. The full set of tests is described in Stowe et al. (1999). Figure 4 shows the calling structure for the CLAVR component algorithms. The day or night condition is determined from the solar zenith angle and the land sea mask is derived from a digital elevation map of the world (USGS 1999). The component CLAVR algorithms are called if any of the pixels in the scene meet the criteria shown in Figure 4. Thus the processing time for a scene will vary according to the number of combinations of land, sea, day and night.


78

Figure 4. CLAVR cloud detection top-level flowchart.

The efficiency could be increased dramatically by recoding in C and including the C subroutine using the new C/Fortran subroutine interface. An example of the application of CLAVR is shown in Figure 5.

Figure 5. The image at left is an AVHRR false colour (band 1 and 2) image of Spencer Gulf, South Australia at 0603Z on 6 May 1997. The image below is the CLAVR cloud mask derived for the

scene.


79

An AVHRR scene showing cloudy regions over South Australia has been processed using the CLAVR algorithm as implemented in CAPS. The results from each CLAVR test are shown as different colours in the cloud mask scene. A land-sea mask has been used to allow the algorithm to select the correct set of tests for land and water. The coastlines have not been masked from the spatial coherence test and have, as a result, been flagged as cloud contaminated. This error could be prevented by creating a coastline mask for the spatial coherence tests from the difference of the morphological erosion and dilation of the land mask. The particular scene for the test has been selected at random and the result is encouraging. The CLAVR algorithm, as presently implemented, is not included as an integral part of the CAPS AVHRR processing script. The script (avhrr2hdf.tcl) is somewhat inflexible and is scheduled for recoding to increase its flexibility and include the option of generating a cloud mask. Also planned is the inclusion of the Ackerman cloud algorithm, which offers a more sophisticated approach to cloud detection.

6. Conclusion Since its inception approximately three years ago CAPS has developed into a stable software platform for processing AVHRR data. There is extensive documentation available to assist new and existing users to access the functionality of the system. The addition of the CLAVR cloud detection scheme extends the AVHRR processing capabilities. The addition of the Fortran and C subroutine interface allows CAPS to take advantage of the intrinsic in-memory processing nature of the package. Extension of the supported numeric types will allow use of the software with other satellite instruments. Binary versions of the CAPS software can be downloaded from http://www.dar.csiro.au/rs/capshome.html.

7. References Ackerman, S.A., Strabala, K.I., Menzel, W.P., Frey, A., Moeller, C.C. and Gumley, L.E., 1998:

Discriminating clear sky from clouds with MODIS. Journal of Geophysical Research, 103 No D24 32,141-32,157.

BLT, 2000, Open Tcl Home page. http://216.167.121.111/blt/index.html Dilley, A.C., Edwards, M., O'Brien, D.M., and Mitchell, R.M., 2000, Operational AVHRR

Processing modules: atmospheric correction, cloud masking and BRDF compensation. CSIRO Atmospheric Research Internal Paper, 14.

NCSA, 2000, The NCSA HDF Home Page. National Center for Supercomputing Applications, http://hdf.ncsa.uiuc.edu

Ousterhout, J.K., 1994, Tcl and the Tk Toolkit. Addison-Wesley, Reading Massachusetts. Ousterhout, J.K., 2000, Tcl Developer Exchange. Ajuba Solutions, http://dev.scriptics.com Saunders, R.W., 1986, An automated scheme for the removal of cloud contamination from

AVHRR radiances over Western Europe. International Journal of Remote Sensing, 7, 867.


80

Saunders, R.W. and K.T. Kriebel, 1988, An improved method for detecting clear sky and cloudy radiances from AVHRR data. International Journal of Remote Sensing, 9, No 1, 123-150.

Stowe, L.L, Davis, P.A. and McClain, E.P., 1999, Scientific Basis and Initial Evaluation of the CLAVR-1 Global Clear/Cloud Classification Algorithm for the Advanced Very High Resolution Radiometer. Journal of Atmospheric and Oceanic Technology, 16, 656-681.

Turner, P.J., Davies, H.L., Tildesley, P.C. and Rathbone, C.R., 1998, Common AVHRR Processing Software (CAPS). Proceedings of the Land AVHRR Workshop, 24 July 1998, (Sydney: 9th Australasian Remote Sensing and Photogrammetry Conference), pp 51-58 (http://www.dar.csiro.au/publications/turner98/capspaper.html).

Turner, P.J., and Davies, H.L., 2000, CAPS Documentation. CSIRO Atmospheric Research, http://www.dar.csiro/rs/capshome.html

USGS, 1999, GTOPO30 World Digital Elevation Model, U.S. Geological Survey, http://edcdaac.usgs.gov/gtopo30/gtopo30.html.

81

Summary of Discussions Held During the EnvSat Workshop During the course of the meeting it became obvious that operational users and researchers have different timeliness requirements. Operational users need quick access to data, and interpreted products, to make near-real-time decisions. Such applications include fire management, drought assessment, and crop yield prediction. On the other hand, many researchers require a long time series that has been consistently processed and corrected. Many users would have higher confidence in the data set if the correction was validated. However, for most users, consistent correction is arguably better than the current situation of no (or incomplete) correction. Differences in delivery system, and hence data requirements, do exist. Rather than focussing on differences, discussions centred on meeting the needs of both operational and research groups from a single National Processing Facility. Researchers produce products that operational agencies may wish to use in near-real time. If these are based on the same fundamental data-processing system, this would greatly increase user confidence and encourage technology transfer. It is likely that enhancements to CAPS, including the provision of a FORTRAN and C subroutine interfaces, will greatly facilitate this. There was some discussion about the possibility of developing a National Processing System, where 20 years of corrected AVHRR data could be housed and made available, possibly on a user-pays system. However, resources would be required to establish such a system, and one major issue was how to obtain these resources. Several ideas were floated, and collaboration between Federal and State Government Departments, CSIRO, and other organisations needs to be negotiated. Twenty years of corrected data is obviously of interest to many researchers and operational land management agencies. The difference between market push (a bunch of scientists telling managers “You need this”) and market pull (a bunch of managers telling scientists “We need this”) were discussed. There is a need to evaluate the business opportunities that would flow from a unique continent-wide 20-year data base. The problems of marketing such a data base were also discussed, since it would require individuals with long-term commitment to generate enthusiasm about the benefits afforded by EnvSats. However, once interest was generated, the need would be for a capacity to deliver. Several participants recounted the increase in phone calls from users interested in using EnvSat data since the previous workshop in July 1998. There were several interesting scientific opportunities that became apparent during the course of the workshop: 1. Fusing data from different EnvSat sensors offers opportunities to cross-validate

radiometric corrections; 2. Underpinning the calibration of EnvSat sensors by using high-quality airborne data; 3. Time series land-based applications that allow better discrimination of tree and

grass components of the NDVI signal; and

82

4. Analysing both spatial and temporal interactions between the NDTI and NDVI. The NDTI measures moisture availability and the NDVI is a measure of moisture utilisation; combining the two would allow climatic and anthropogenic variability to be mapped.

Operational users of EnvSats need to enthusiatically tell research managers how useful improvements made by working scientists have been to them. In the current economic climate, especially considering the broad-scale non-market-specific nature of many land EnvSat applications, this message needs to be delivered with strength. As is the nature of workshops, there were many smaller in-depth discussions held during breaks, other scientific issues and organisational problems were identified (and possibly solved), and opportunities unearthed. To facilitate future discussions, full contact details of all who attended this workshop are listed in Appendix 1.

83

Appendix 1: List of Attendees, 10th Land EnvSat Workshop Gary Bastin CSIRO Wildlife and Ecology GPO Box 284 Canberra ACT 2601 Australia Tel: +61-2-6242-1600 Fax: +61-2-6242-1555 e-mail: [email protected] Rosalie Booth ACRES PO Box 2 Belconnen ACT 2616 Tel: +61-2-6201-4130 Fax: +61-2-6201-4240 e-mail: [email protected] Vanessa Chewings CSIRO Wildlife and Ecology Centre for Arid Zone Research PO Box 2111 Alice Springs NT 0871 Tel: +61-8-8950-7127 Fax: +61-8-8950-7187 e-mail: [email protected] Ron Craig Satellite Remote Sensing Services Department of Land Administration Leeuwin Centre for Earth Sensing Technologies PO Box 471 Wembley, W.A. 6014 Australia Tel: +61-8-9340-9346 Fax: +61-8-9383-7142 e-mail: [email protected] David Griersmith Satellite Section Bureau of Meteorology GPO Box 1289K Melbourne Vic 3001 Tel: +61-3-9669-4594 Fax: +61-3-9669-4736 e-mail: [email protected]

Zhang Guanglu Shijiazhuang Institute of Agricultural Modernization Chinese Academy of Sciences P.O. Box 185 Shijiazhuang 050021 P.R.China Tel: +86-311-587-1747 Fax: +86-311-581-5093 e-mail: [email protected] Lee Hong Satellite Section Bureau of Meteorology GPO Box 1289K Melbourne Vic 3001 Tel: +61-3-9669-4419 Fax: +61-3-9669-4736 e-mail: [email protected] Iain Hume CSIRO Land and Water GPO Box 1666 Canberra ACT 2601 Tel: +61-2-6246-5819 Fax: +61-2-6246-5800 e-mail: [email protected] David Jupp CSIRO Earth Observation Centre GPO Box 3023 Canberra ACT 2601 Australia Tel: +61-2-6216-7203 Fax: +61-2-6216-7222 e-mail: [email protected] Edward King CSIRO Earth Observation Centre GPO Box 3023 Canberra ACT 2601 Australia Tel: +61-2-6216-7197 Fax: +61-2-6216-7222 e-mail: [email protected]

84

Bill Kustas USDA – ARS Hydrology Lab Bldg 007, BARC-WEST Beltsville, MD 20705 USA Tel: +1-301-504-8498 Fax: +1-301-504-8931 e-mail: [email protected] Jenny Lovell CSIRO Earth Observation Centre GPO Box 3023 Canberra ACT 2601 Australia Tel: +61-2-6216-7195 Fax: +61-2-6216-7222 e-mail: [email protected] Tim McVicar CSIRO Land and Water GPO Box 1666 Canberra ACT 2601 Tel: +61-2-6246-5741 Fax: +61-2-6246-5800 e-mail: [email protected] Denis O’Brien CSIRO Atmospheric Research Private Bag 1 Aspendale Vic 3195 Tel: +61-3-9239-4400 Fax: +61-3-9239-4444 e-mail: [email protected] Sumith Pathirana Southern Cross University School of Resource Science PO Box 157 Lismore NSW 2480 Tel: +61-2-6620-3036 Fax: +61-2-6621-2669 e-mail: [email protected] Richard Roger NSW Agriculture Locked Bag 21 Orange NSW 3554 Tel: +61-2-6391-3195 Fax: +61-2-6391-3767 e-mail: [email protected]

Medhavy Thankappan ACRES PO Box 2 Belconnen ACT 2616 Tel: +61-2-6201-4130 Fax: +61-2-6201-4240 e-mail: [email protected] Mark Thomas PIRSA Land Information GPO Box 1671 Adelaide SA 5001 Tel: +61-8-8303-9641 Fax: +61-8-8303-9302 e-mail: [email protected] Paul Trezise ACRES PO Box 2 Belconnen ACT 2616 Tel: +61-2-6201-4130 Fax: +61-2-6201-4240 e-mail: [email protected] Guy Tuddenham Bureau of Meteorology (NMOC) GPO Box 1289K Melbourne Vic 3001 Tel: +61-3-9669-4689 Fax: +61-3-0669-4736 e-mail: [email protected] Peter Turner CSIRO Atmospheric Research Private Bag 1 Aspendale Vic 3195 Tel: +61-3-9239-4674 Fax: +61-3-9239-4444 e-mail: [email protected] Michael Willmott Bureau of Meteorology Satellite Section GPO Box 1289K Melbourne Vic 3001 Tel: +61-3-9669-4419 Fax: +61-3-0669-4736 e-mail: [email protected]

proceedings of the land envsat workshop · discussions) during a one-day workshop convened in...

Documents