This article was downloaded by: [Center for Earth Observation and Digital Earth ],[Zhaoming Zhang]On: 07 October 2011, At: 17:44Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of RemoteSensingPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/tres20

Leaf area index estimation of bambooforest in Fujian province based on IRSP6 LISS 3 imageryZhaoming Zhang a , Guojin He a , Xiaoqin Wang b & Hong Jiang ba Centre for Earth Observation and Digital Earth, ChineseAcademy of Sciences, Beijing, 100190, Chinab Spatial Information Research Centre, Fujian Province, Fuzhou,350002, China

Available online: 11 Jul 2011

To cite this article: Zhaoming Zhang, Guojin He, Xiaoqin Wang & Hong Jiang (2011): Leaf areaindex estimation of bamboo forest in Fujian province based on IRS P6 LISS 3 imagery, InternationalJournal of Remote Sensing, 32:19, 5365-5379

To link to this article: http://dx.doi.org/10.1080/01431161.2010.498454

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representationthat the contents will be complete or accurate or up to date. The accuracy of anyinstructions, formulae, and drug doses should be independently verified with primarysources. The publisher shall not be liable for any loss, actions, claims, proceedings,demand, or costs or damages whatsoever or howsoever caused arising directly orindirectly in connection with or arising out of the use of this material.

International Journal of Remote SensingVol. 32, No. 19, 10 October 2011, 5365–5379

Leaf area index estimation of bamboo forest in Fujian province basedon IRS P6 LISS 3 imagery

ZHAOMING ZHANG†, GUOJIN HE*†, XIAOQIN WANG‡and HONG JIANG‡

†Centre for Earth Observation and Digital Earth, Chinese Academy of Sciences, Beijing100190, China

‡Spatial Information Research Centre, Fujian Province, Fuzhou 350002, China

(Received 6 January 2010; in final form 2 June 2010)

Leaf area index (LAI) is an important surface biophysical parameter as an inputto many process-oriented ecosystem models. Much work has been reported inthe literature on LAI estimation in boreal forests using remotely sensed imagery.However, few if any explicit LAI retrieval studies on bamboo forests in Asian sub-tropical monsoon-climate regions based on remote sensing technology have beenperformed. Our goal is to carry out a comparative study on the LAI estimationmethods of bamboo forest in Fujian province, China, based on IRS P6 LISS 3imagery. Both the traditional empirical–statistical approach and the newly pro-posed normalized distance (ND) method were employed in this study, and a total of18 modelling parameters were regressed against ground-based LAI measurements.The results show that simple ratio (SR) is the best predictor for LAI estimationin this study area, with the highest R2 (coefficient of determination) value of 0.68;modified simple ratio (MSR) and normalized difference vegetation index (NDVI)ranked second and third, respectively. The good performance of these three veg-etation indices (VIs) can be explained by the ratioing principle. The overall goodmodelling performance of the ND method in our study area also indicates it is apromising method.

1. Introduction

Leaf area index (LAI) is defined as one-half the total green-leaf area (all sided) perunit ground surface area (Chen and Black 1992). In hilly and mountainous areas, theground surface area should be projected to a horizontal area perpendicular to the ver-tical view direction. Leaf area index is an important surface biophysical parameter formany agronomic, ecological and meteorological applications, which acts as a measureof vegetation growth and productivity, and as an input to many ecosystem processmodels.

Leaf area index varies both spatially and temporally, and is difficult and expensiveto derive with ground measurement methods. However, the remote sensing methodprovides a promising and practical way to estimate LAI over a large area with hightemporal coverage, and hence, considerable effort has been expended in developingremote sensing-based techniques to map LAI in the past two decades.

*Corresponding author. Email: gjhe@ceode.ac.cn

International Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online © 2011 Taylor & Francis

http://www.tandf.co.uk/journalsDOI: 10.1080/01431161.2010.498454

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

5366 Z. Zhang et al.

Traditionally, remote sensing-based techniques to estimate LAI have eitherbeen based on the empirical–statistical approach that relates ground-measuredLAI to the spectral vegetation indices (or spectral reflectance) derived fromsatellite-measured reflectance, or on the inversion of a physically based canopyreflectance model. Both approaches have their advantages and limitations (Qi et al.2000).

The empirical–statistical approach is commonly employed for estimating LAIacross landscapes, which establishes an empirical relationship between vegetationindices (VIs) and LAI by statistically fitting ground-measured LAI to the correspond-ing VI. The advantage of this approach is its simplicity and ease of computation,and it is highly desirable for large-scale LAI retrievals (Houborg and Boegh 2008).However, the major limitation of this approach is its lack of generality. Numerous VIshave been proposed in the literature, among which the normalized difference vegeta-tion index (NDVI) (Rouse et al. 1974) and the simple ratio (SR) (Jordan 1969) arethe most frequently used VIs to retrieve LAI for various ecosystem types, for exam-ple coniferous forests (Chen and Cihlar 1996), deciduous forests (Madugundu et al.2008) and grassland (Friedl et al. 1994). NDVI and SR are fundamentally the same:they can be obtained from each other without additional information. NDVI has afixed range of values between 0 and 1, while SR is sometimes preferred, because itis more sensitive and more linear with biophysical parameters, such as LAI (Chen1996, Chen and Cihlar 1996). In both NDVI and SR, the soil background effectis not considered, and the soil-adjusted vegetation index (SAVI) (Huete 1988) wasdeveloped to reduce or remove this effect by introducing a soil adjustment factor (L)into the calculation of NDVI. In SAVI, the L value is assumed to be 0.5. The per-pendicular vegetation index (PVI) (Richardson and Wiegand 1977) accounts for theeffect of soil background by computing an orthogonal greenness vector to a soil linewith arbitrary slopes and intercepts. For low density vegetated areas, PVI reducedsoil effects and held a near linear relationship with LAI (Curran 1983), but in a highdensity vegetated area, it caused much noise (Baret and Guyot 1991). The globalenvironment monitoring index (GEMI) (Pinty and Verstraete 1992) was designedto reduce the atmospheric effects on remotely sensed imagery at the global scale.It is inconvenient for algorithm development, considering the fact that the relation-ships between VIs and surface biophysical parameters are often not linear; thereforethe non-linear index (NLI) (Goel and Qin 1994) and the re-normalized differencevegetation index (RDVI) (Roujean and Breon 1995) were proposed, in an effort tolinearize the relationship with surface biophysical parameters such as LAI. Based onRDVI, Chen (1996) proposed the modified simple ratio (MSR) for the purpose ofproviding a more linear relationship with surface parameters. By including spectralresponse from the short-wave infrared, three-band VIs, such as the reduced simpleratio (RSR) (Brown et al. 2000) and the modified normalized difference vegetationindex (MNDVI) (Nemani et al. 1993), were formulated in order to minimize back-ground effects. The study conducted by Stenberg et al. (2004) showed RSR performswell and is better than NDVI for the estimation of LAI in pine and spruce stands inFinland.

An alternative to the empirical–statistical approach is a modelling approach basedon a set of radiative transfer models (Li and Strahler 1985, Jacquemoud et al. 1995,Schlerf and Atzberger 2006, Houborg and Boegh 2008). In order to derive LAI,the model should be run in a reverse mode. Once the model and the data to beused are determined, an appropriate inversion technique is of great importance. In

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

Leaf area index estimation of bamboo forest 5367

recent years, different strategies have been proposed for the inversion of these models,including numerical optimization methods (Jacquemoud et al. 1995, 2000), look-up tables (Combal et al. 2002), neural network techniques (Qi et al. 2000, Walthallet al. 2004, Fang and Liang 2005), support vector machines regression (Durbha et al.2007) and predictive equations (Zarco-Tejada et al. 2001, Haboudane et al. 2002). Amajor advantage of the inversion approach is that it is physically based and indepen-dent of vegetation type. However, two major limitations exist in operational use of thisapproach. First of all, this method suffers from the expensive computational require-ment involved in a large number of inversion processes, unless an appropriate inversiontechnique is employed. Another limitation is the ill-posed nature of model inversion(Combal et al. 2002, Atzberger 2004). Different combinations of model parametersmay yield almost identical reflectance spectra. When searching for the optimum com-bination during the inversion process, the solution may be found at quite differentplaces in the parameter space, yielding large uncertainties in the estimated LAI. Oneway to solve the ill-posed problem of model inversion is to use a priori knowledge toregularize the inversion problem (Atzberger, 2004); however, the a priori knowledgerequired by the inversion cannot be easily obtained for most places on the globe. Dueto the aforementioned limitations of the inversion approach, it is not employed in thisstudy.

Over the years, much effort has been made to propose new LAI estimationapproaches based on remote sensing imagery. These approaches include reducedmajor axis (RMA) regression analysis (Curran and Hay 1986, Cohen et al. 2003);the red edge inflection point (REIP) method based on hyperspectral remotely sensedimagery (Pu et al. 2003, Darvishzadeh et al. 2009); geostatistical techniques such ascokriging, kriging with an external drift (KED) and sequential Gaussian conditionalsimulation (SGCS) (Goovaerts 1997, Deutsch and Journel 1998, Berterretche et al.2005); linear spectral mixture analysis (LSMA) (Hall et al. 1995, Peddle et al. 1999,Hu et al. 2004); and the normalized distance (ND) method (McAllister and Valeo2007). Comparisons between different approaches have been carried out. Chen (1996)evaluated various VIs against field-data sets of LAI and fractional photosyntheti-cally active radiation (FPAR) in boreal forests. Based on Chen’s work, Peddle et al.(2001) made a comparison of spectral mixture analysis (SMA) and the VIs approachfor estimating boreal forest biophysical information from airborne data. Berterretcheet al. (2005) compared reduced major axis and two geostatistical techniques, KED andsequential Gaussian conditional simulation (SGCS), for mapping LAI with LandsatETM+ data in a boreal forest in Canada. Darvishzadeh et al. (2009) compared theperformance of various narrow-band VIs with the red-edge inflection point (REIP)method in estimating LAI. The ND method was proposed by McAllister and Valeoin 2007, and this method achieved better results for the remote estimation of LAIin montane and boreal forest in Canada than linear spectral mixture analysis andmodification of spectral vegetation indices. The advantages of the ND technique areits mitigation of background effects and residual noise by the normalized distancetechnique, which then produces meaningfully bounded values of LAI, and its mod-elling robustness for both coniferous and deciduous vegetation (McAllister and Valeo2007).

In this study, both the traditional empirical–statistical approach and the NDmethod were employed for the estimation of LAI for bamboo forest in Fujianprovince, China, and their performances were evaluated. In the following section, thestudy area and data used are described, after which the empirical–statistical approach

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

5368 Z. Zhang et al.

and the ND method employed in this study are briefly introduced. Results from bothmethods are then discussed and analysed.

2. Study area and study object

The study area is situated in Yong’an county, Fujian province, located in the south-eastof China (figure 1). Fujian has a subtropical oceanic monsoon climate, with annualtemperature averaging between 15.3◦C and 21.9◦C and annual average precipitationbetween 930 and 1843 mm. The forest coverage rate of Fujian province has reached62.96%, the highest in China. Yong’an, situated in central Fujian, is one of the 48key forestry districts and counties (cities) in the south of China, with forest coveragehaving reached 83.2%. Yong’an county, characterized by its bamboo forest, has beenhonoured as the ‘home town of bamboo in China’. Farmers have 2.7 ha of bambooforest per capita, ranking first in the country.

Due to its unique community structure and asexual reproduction, bamboo forest isquite different from other forest types. With the rapid expansion of the bamboo forestarea in Fujian province, bamboo forest is playing a more and more important role inthe regional forest ecosystem.

N

(a)

0 6km

Figure 1. Location map of the study area (the red dot is Yong’an county and the area aroundit is Fujian province). The image in the lower-right corner (a) is a subset of IRS P6 LISS 3composite image (RGB: bands 543), yellow–green indicating bamboo forest.

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

Leaf area index estimation of bamboo forest 5369

In the past two decades, much work has been done for LAI estimation in borealforest (Chen 1996, Brown et al. 2000, Eklundh et al. 2001, Peddle et al. 2001), in tem-perate climates (Peterson et al. 1987, Schlerf et al. 2005, Soudani et al. 2006) and inrainforests (Thenkabail et al. 2004) based on remote sensing imagery. However, few ifany explicit LAI estimation studies for bamboo forest in Asian subtropical monsoon-climate regions have been performed. Therefore, the present study has been conductedto address this problem.

3. Data set

3.1 Ground-based LAI measurements

Ground LAI measurements were made using the Plant Canopy Analyser (PCA), LAI-2000 (Welles and Norman 1991) (LI-COR Inc., Lincoln, NE, USA). The LAI-2000measures the gap fraction in five zenith angles, ranging from 0◦ to 75◦. The measuredgap fraction data are inverted to obtain the effective LAI, under the assumption of arandom spatial distribution of leaves. The assumption of random spatial distributionof the leaves is generally true for bamboo forest in this study area.

The effective LAI (denoted as Le) can be accurately calculated by equation (1) usingthe LAI-2000 instrument based on Miller (1967) theory:

Le = 2

π/2∫0

ln[

1P(θ )

]cos θ sin θ dθ (1)

where P(θ ) is the measured canopy gap fraction at zenith angle θ , which is best whenaveraged over the entire azimuthal angle range. In order to obtain Le accurately, hemi-spherical P(θ ) data is required, which can be provided by the LAI-2000 instrumentthrough sensing the diffuse radiation from the sky over the hemisphere.

LAIs were measured at each of 27 sample plots in July 2008. For this measurement,only one instrument was available and it was used in turn to measure above and belowcanopy irradiance. For each sample plot, above canopy irradiance was measured in anopen area, close to the plot. All plots were geo-located using global positioning system(GPS) measurements with an accuracy of <20 m in both the x and y directions. Inorder to prevent the effects of direct sunlight on the sensor, the instrument was onlyoperated near dusk or dawn or under overcast conditions.

3.2 Remotely sensed data preprocessing and reflectance estimation

A cloud-free IRS P6 LISS 3 image, acquired on 24 March 2008, was used in this study.The LISS 3 sensor observes with 24 m pixel resolution at green (0.52–0.59 µm), red(0.62–0.68 µm), near infrared (0.77–0.86 µm) and short-wave infrared (1.55–1.70 µm)bands. Digital numbers stored in the LISS 3 image were converted to land surfacereflectance before any subsequent analysis was conducted. Firstly, the digital numbersof the scene were transformed into spectral radiance by using the gains and offsetsobtained from the image header file. The atmospheric correction to convert the radi-ance values into land surface reflectance was accomplished with the practical darkobject subtraction (DOS) model-based atmospheric correction approach proposed byZhang et al. (2010).

It should be noted that there is a discrepancy between image acquisition time andLAI collection time, because bad weather conditions in southern China resulted in

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

5370 Z. Zhang et al.

cloudy optical imagery being received when synchronous ground measurement of LAIwas performed. However, considering that the time interval between image acquisitionand LAI collection is not large and LAI variation in bamboo forest can be consideredsmall during this period, any error caused by the discrepancy should be acceptable.

4. Methodology

4.1 Empirical–statistical approach

In the empirical–statistical approach, empirical relationships between LAI andreflectance in the four single spectral bands, and ten VIs (details provided in table 1)which were selected from the commonly used VIs in forestry applications, were inves-tigated. Among the ten VIs, eight are two-band indices and the last two are three-bandindices.

Table 1. Definition and sources of VIs used in this study. ρr, ρn and ρs are red, near-infraredand short-wave infrared (SWIR) reflectance, respectively. ρs min is the SWIR reflectance obtained

from a completely closed canopy and ρsmax is the SWIR reflectance from an open canopy.

Vegetation index Formulae Reference

Normalized differencevegetation index,NDVI

NDVI = ρn − ρr

ρn + ρrRouse et al.

(1974)

Simple ratio, SR SR = ρn

ρrJordan (1969)

Modified simple ratio,MSR

MSR =ρnρr

− 1√ρnρr

+ 1Chen (1996)

Soil adjusted vegetationindex, SAVI

SAVI = (ρn − ρr)(1 + L)(ρn + ρr + L)

, L = 0.5 Huete (1988)

Perpendicular vegetationindex, PVI

PVI = ρn − aρr − b√1 + a2

, a = 0.9, b = 0.1 Richardsonand Wiegand(1977)

Global environmentmonitoring index,GEMI

GEMI = η(1 − 0.25η) − (ρr − 0.125)1 − ρr

,

η = 2(ρ2n − ρ2

r ) + 1.5ρn + 0.5ρr

ρn + ρr + 0.5

Pinty andVerstraete(1992)

Re-normalizeddifference vegetationindex, RDVI

RDVI = ρn − ρr√ρn + ρr

Roujean andBreon (1995)

Non-linear index, NLI NLI = ρ2n − ρr

ρ2n + ρr

Goel and Qin(1994)

Reduced simple ratio,RSR

RSR = ρn

ρr

(1 − ρs − ρs min

ρs max − ρs min

)Brown et al.

(2000)

Modified normalizeddifference vegetationindex, MNDVI

MNDVI = ρn − ρr

ρn + ρr

(1 − ρs − ρs min

ρs max − ρs min

)Nemani et al.

(1993)

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

Leaf area index estimation of bamboo forest 5371

4.2 Normalized distance (ND) method

The advantages of the ND technique are its mitigation of background effects andresidual noise and its modelling robustness for both coniferous and deciduous vege-tation, and this method exhibits improved performance over the RSR and MNDVImethods (McAllister and Valeo 2007).

To estimate LAI, the ND method firstly computes the scaling factor, then calculatesthe normalized distance from the scaling factor, and finally transforms the normalizeddistance based on the observed relationship with LAI.

The scaling factors F are computed as in equation (2) and vary between 0 and 1:

Fi = 1 −[

Ri − Rimin

Rimax − Ri

min

](2)

where F i is the scaling factor for band i, Ri is the reflectance for an individual pixel inband i, Ri

min and Rimax are the observed minimum and maximum reflectances in band

i, respectively.The normalized distance for a given pixel is then calculated according to equation

(3):

ND =

√√√√√n∑

i=1(αi − Fi)2

n(3)

where ND is the normalized distance, F i is the scaling factor computed in bandi, n is the number of bands for which analysis is being conducted, and αi is 0 ifreflectances observed in band i exhibit a direct relationship with LAI, or αi is equal to1 if reflectances in band i exhibit an inverse relationship with LAI.

The measured LAI data were plotted against reflectances observed in each of thegreen, red, near infrared and short-wave infrared channels and these relationshipsare displayed in figure 2. As shown in figure 2, there is an obvious negative corre-lation between red reflectance and LAI, thus α is equal to 1; reflectances in the nearinfrared band exhibit an obvious direct relationship with LAI and so α is equal to 0.In the green or short-wave infrared band, there is a weak positive correlation betweenreflectance and LAI, so α is equal to 0.

Through a combination of different bands, different ND calculation equations canbe constructed.

When green, red and near infrared bands are employed, ND (denoted as NDG,R,NIR)is computed according to equation (4):

NDG,R,NIR =√

(Fgreen)2 + (1 − F red)2 + (FNIR)2

3(4)

Similarly, NDR,NIR,SWIR, NDR,NIR and NDG,R,NIR,SWIR are calculated based onequations (5)–(7):

NDR,NIR,SWIR =√

(1 − F red)2 + (FNIR)2 + (FSWIR)2

3(5)

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

5372 Z. Zhang et al.

y = 2.1719x + 2.2625

0

1

2

3

4

5

0 0.01 0.02 0.03 0.04 0.05

Reflectance

LA

I

y = –80.837x + 4.4578

0

1

2

3

4

5

0 0.01 0.02 0.03 0.04 0.05

Reflectance

LA

I

(a) (b)

y = 14.579x – 0.659

0

1

2

3

4

5

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35

Reflectance

LA

I

y = 6.8038x + 1.4703

0

1

2

3

4

5

0 0.05 0.1 0.15 0.2

ReflectanceL

AI

(c) (d)

Figure 2. LAI versus surface reflectance in (a) green, (b) red, (c) near infrared, and (d) short-wave infrared bands; y indicating leaf area index (LAI) and x indicating surface reflectance.

NDR,NIR =√

(1 − F red)2 + (FNIR)2

2(6)

and

NDG,R,NIR,SWIR =√

(Fgreen)2 + (1 − F red)2 + (FNIR)2 + (FSWIR)2

4(7)

5. Results and discussion

In this study, the performance of the above-mentioned modelling parameters, includ-ing reflectance in the four single bands, ten VIs and four NDs, were compared, and therelationships derived from the regression (both linear and non-linear) of ground-basedLAI measurements against these eighteen modelling parameters are shown in table 2.All the modelling parameters except for GEMI were based on the atmospherically cor-rected reflectances. When computing GEMI, the top-of-atmosphere reflectances wereused because the atmospheric effect is already considered in the GEMI calculationequation.

As shown in table 2, SR, MSR and NDVI were the best among all the modellingparameters in both linear and non-linear cases, which is similar to the results foundby Chen (1996) in the LAI retrieval conducted in the boreal Jack Pine and BlackSpruce stands in both spring and summer. Both NDVI and MSR can be expressed asa function of SR, and their good performance can be explained by the ratioing prin-ciple: measurement errors in remotely sensed imagery cause simultaneous increases ordecreases in red and near infrared reflectances in approximately the same proportion,and by taking the simple band ratio between the near infrared and red reflectances,these errors can be greatly reduced (Chen 1996). Figure 3 also demonstrates that SR

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

Leaf area index estimation of bamboo forest 5373

Table 2. Regression results for predicting LAI of bamboo forests using surface reflectances,vegetation indices and ND.

Linear regression relationship Non-linear regression relationship

Rank Relationship R2 Rank Relationship R2

SR y = 0.6014x − 2.4778 0.68 SR y = 0.2204e0.282x 0.67MSR y = 2.8433x − 4.2581 0.66 MSR y = 0.0922e1.3491x 0.66NDVI y = 20.749x − 13.688 0.60 NDVI y = 0.0009e10.074x 0.63NLI y = 6.1092x + 0.948 0.60 NLI y = 1.105e2.8385x 0.58NDR,NIR y =−6.6522x + 6.0184 0.49 NDR,NIR y = 12.731e−3.2505x 0.52RDVI y = 18.425x − 4.5022 0.45 RDVI y = 0.0978e8.2699x 0.40SAVI y = 14.09x − 2.8156 0.39 SAVI y = 0.2149e6.2425x 0.34RSR y = 0.3725x + 0.0636 0.26 RSR y = 0.6839e0.1843x 0.29GEMI y = 10.576x − 2.9566 0.26 GEMI y = 0.1945e4.7595x 0.23ρNIR y = 14.579x − 0.659 0.22 ρred y = 6.6121e−43.599x 0.21NDG,R,NIR y =−4.3483x + 4.872 0.22 NDG,R,NIR y = 6.6774e−1.9787x 0.20PVI y = 19.271x + 1.1625 0.19 NDR,NIR,SWIR y = 6.1342e−1.8978x 0.18NDR,NIR,SWIR y =−3.886x + 4.5254 0.17 ρNIR y = 0.6058e6.0655x 0.17ρred y =−80.837x + 4.4578 0.17 PVI y = 1.3024e7.8936x 0.15NDG,R,NIR,SWIR y =−3.0278x + 4.0735 0.14 NDG,R,NIR,SWIR y = 4.4581e−1.3069x 0.11MNDVI y = 2.0118x + 1.1561 0.03 MNDVI y = 1.0119e1.249x 0.06ρSWIR y = 6.8038x + 1.4703 0.01 ρgreen y = 2.4594e−4.3983x 0.00ρgreen y = 2.1719x + 2.2625 0.00 ρSWIR y = 1.7638e1.3981x 0.00

Note: R2, coefficient of determination; y, leaf area index; x, modelling parameters.

had a larger dynamic range and was more sensitive to the variations in LAI, whileNDVI suffered from the drawback of saturation. SR was also more linearly corre-lated to LAI than NDVI. NLI and RDVI, which have been developed to linearizetheir relationships with land surface parameters, achieved relatively good results. Asan improved version of RDVI, MSR is much superior to RDVI, and ranked sec-ond in all the eighteen modelling parameters. MSR not only retains the advantagesof RDVI but is also in accord with the ratioing principle. The performance of VIsdesigned mainly to minimize the background effect, such as SAVI, PVI and MNDVI,was unsatisfactory in this study, possibly due to the fact that in bamboo forest areasthe background influence is not obvious; what is more, these VIs are not in accordwith the ratioing principle and therefore may retain the noise or even amplify it.The poor performance of GEMI, designed for minimizing the atmospheric effects,may also be caused by noise problems owing to its complicated mathematical com-putation. The major advantages of RSR over SR are: (1) RSR has the potential tounify different cover types in LAI retrieval so that the accuracy of LAI retrieval formixed cover types can be improved and (2) the background influence is suppressedusing RSR because the SWIR band is sensitive to the background effects (Brownet al. 2000). However, in this study area, no mixed cover types existed in the sam-ple plots and the background influence on the signal received by the satellite sensorwas relatively small, therefore RSR performed worse than SR or NDVI with R2 valuesmaller than 0.3.

As for the single spectral band, reflectance in the near infrared or the red band wasbetter correlated to the measured LAI data than the green and the short-wave infraredbands (figure 2 and table 2). As the bands indicative of vegetation, the near infrared

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

5374 Z. Zhang et al.

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5

LAI

ND

VI

0

1

2

3

4

0 1 2 3 4 5

LAI

MS

R

0.0

0.1

0.2

0.3

0 1 2 3 4 5

LAI

PV

I

0

2

4

6

8

10

12

0 1 2 3 4 5

LAI

SR

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5

LAI

SA

VI

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5

LAI

GE

MI

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0 1 2 3 4 5

LAI

RD

VI

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5

LAI

NLI

0

2

4

6

8

10

0 1 2 3 4 5

LAI

RS

R

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5

LAI

MN

DV

I

Figure 3. LAI versus the ten vegetation indices.

and the red bands are more sensitive to the vegetative information, while the green andthe short-wave infrared bands contained much noise and performed poorly for LAIestimation in this study area, which is different from the results found by other studies(Stenberg et al. 2004, Darvishzadeh et al. 2008).

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

Leaf area index estimation of bamboo forest 5375

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5

LAI

ND

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5

LAI

ND

(a) (b)

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5

LAI

ND

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5

LAI

ND

(c) (d)

Figure 4. LAI versus the normalized distances (ND) constructed with (a) the green, red andnear infrared bands; (b) the red and near infrared bands; (c) the red, near infrared and short-wave infrared bands; and (d) with all the four bands of IRS P6 LISS 3 imagery.

As a new LAI estimation method, the ND method ranked a respectable fifthamongst all eighteen modelling parameters. Of the four NDs, NDR,NIR performedthe best with the highest R2 value of 0.52 (figure 4 and table 2). The red and nearinfrared bands were most informative in our study area, therefore normalized distanceconstructed with these two bands is most effective. The overall good modelling perfor-mance of the ND method in this study area indicates it is a promising method. In areaswhere the background exerts serious influence on the satellite signals and substantialinformation is contained in the short-wave infrared band, the ND method would per-form better when this useful short-wave infrared band is added in the construction ofthe normalized distance. The major limitation of the ND method is that it is not inaccord with the ratioing principle, therefore, unlike the NDVI and SR spectral indices,it does not have the advantage provided by the band ratioing principle. Furthermore,special attention should be paid to band selection when constructing the normalizeddistance.

5.1 Cross-validation

The SR technique has proven to be the best in its modelling of LAI of bambooforest compared to the other seventeen modelling parameters. Validation of theSR technique was performed with ‘leave-one-out’ cross-validation in this section.Comparisons between model-estimated LAI and field-measured LAI are demon-strated in figure 5. As can be seen from figure 5, bamboo forest LAI estimates based onthe SR technique are slightly underestimated at higher LAI values and slightly over-estimated at lower LAI values, with the slope and intercept not significantly differentfrom 1 and 0, respectively. The root mean squared error (RMSE) is 0.61; errors may

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

5376 Z. Zhang et al.

y = 0.6624x + 0.7809

R2 = 0.63, RMSE = 0.61

0

1

2

3

4

5

0 1 2 3 4 5

Field-measured LAI

Mod

el-e

stim

ated

LA

I

Figure 5. ‘Leave-one-out’ cross-validation for bamboo forest LAI estimated based on the SRtechnique, y indicating model-estimated LAI and x indicating field-measured LAI.

have been caused by several factors including errors in LAI measurement, variation ofcanopy architecture, and errors in co-location of ground plots in the remotely sensedimagery.

6. Conclusions

In this study, LAI estimation for bamboo forest in the Fujian subtropical monsoon-climate region based on remote sensing imagery was performed. Eighteen modellingparameters, including reflectance in the four single spectral bands, eight two-band VIs,two three-band VIs and four normalized distances, were evaluated against ground-based LAI measurements. Several conclusions can be drawn from this study.

One conclusion that is evident from the results of this study is that, of the eighteenmodelling parameters, SR performed the best, followed by MSR and NDVI, both ofwhich can be expressed as a function of SR. Their good performance can be explainedby the ratioing principle: taking the simple band ratio between the near infrared andred reflectances can greatly reduce measurement errors in remotely sensed imagery.Cross-validation results show the SR technique can achieve a good LAI estimationresult.

A second conclusion drawn from this study is that the ND method, a new LAIretrieval method, ranked a respectable fifth amongst all eighteen modelling parame-ters, with the normalized distance constructed with the red and near infrared bands,which indicates it is a promising method. This method would perform better if more

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

Leaf area index estimation of bamboo forest 5377

informative bands were employed in constructing the normalized distance, whichrequires further validation in other study areas.

A third conclusion is that the background effect in the bamboo forest of thisstudy area was relatively small, therefore modelling parameters designed to reducethe background effect (such as SAVI, PVI and RSR) did not perform well in thisstudy. However, these parameters may have advantages in other study areas wherebackground effects exert significant influence on the satellite signals.

AcknowledgementThis work was funded by the Ministry of Science and Technology of the People’sRepublic of China under grant number 2005DFA20420.

ReferencesATZBERGER, C., 2004, Object-based retrieval of biophysical canopy variables using artificial

neural nets and radiative transfer models. Remote Sensing of Environment, 93, pp. 53–67.BARET, F. and GUYOT, G., 1991, Potentials and limits of vegetation indices for LAI and APAR

assessment. Remote Sensing of Environment, 35, pp. 161–173.BERTERRETCHE, M., HUDAK, A.T., COHEN, W.B., MAIERSPERGER, T.K., GOWER, S.T. and

DUNGAN, J., 2005, Comparison of regression and geostatistical methods for mappingleaf area index (LAI) with Landsat ETM+ over a boreal forest. Remote Sensing ofEnvironment, 96, pp. 49–61.

BROWN, L., CHEN, J.M., LEBLANC, S.G. and CIHLAR, J., 2000, A shortwave infrared modifica-tion to the simple ratio for LAI retrieval in boreal forests: an image and model analysis.Remote Sensing of Environment, 71, pp. 16–25.

CHEN, J.M., 1996, Evaluation of vegetation indices and a modified simple ratio for borealapplications. Canadian Journal of Remote Sensing, 22, pp. 229–242.

CHEN, J.M. and BLACK, T.A., 1992, Defining leaf area index for non-flat leaves. Plant, Cell andEnvironment, 15, pp. 421– 429.

CHEN, J.M. and CIHLAR, J., 1996, Retrieving leaf area index for boreal conifer forests usingLandsat TM images. Remote Sensing of Environment, 55, pp. 153–162.

COHEN, W.B., MAIERSPERGER, T.K., GOWER, S.T. and TURNER, D.P., 2003, An improved strat-egy for regression of biophysical variables and Landsat ETM+ data. Remote Sensing ofEnvironment, 84, pp. 561–571.

COMBAL, B., BARET, F., WEISS, M., TRUBUIL, A., MACÉ, D., PRAGNÈRE, A., MYNENI, R.,KNYAZIKHIN, Y. and WANG, L., 2002, Retrieval of canopy biophysical variables frombidirectional reflectance: using prior information to solve the ill-posed inverse problem.Remote Sensing of Environment, 84, pp. 1–15.

CURRAN, P.J., 1983, Multispectral remote sensing for the estimation of green leaf area index.Philosophical Transactions of the Royal Society A, 309, pp. 257–270.

CURRAN, P.J. and HAY, A.M., 1986, The importance of measurement error for certain pro-cedures in remote sensing at optical wavelengths. Photogrammetric Engineering andRemote Sensing, 52, pp. 229–241.

DARVISHZADEH, R., ATZBERGER, C., SKIDMORE, A.K. and ABKAR, A.A., 2009, Leaf area indexderivation from hyperspectral vegetation indices and the red edge position. InternationalJournal of Remote Sensing, 30, pp. 6199–6218.

DARVISHZADEH, R., SKIDMORE, A., ATZBERGER, C. and WIEREN, S.V., 2008, Estimation ofvegetation LAI from hyperspectral reflectance data: effects of soil type and plant archi-tecture. International Journal of Applied Earth Observation and Geoinformation, 10, pp.358–373.

DEUTSCH, C. and JOURNEL, A., 1998, GSLIB: Geostatistical Software Library and User’s Guide,(2nd edn) (New York: Oxford University Press).

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

5378 Z. Zhang et al.

DURBHA, S.S., KING, R.L. and YOUNAN, N.H., 2007, Support vector machines regression forretrieval of leaf area index from multiangle imaging spectroradiometer. Remote Sensingof Environment, 107, pp. 348–361.

EKLUNDH, L., HARRIE, L. and KUUSK, A., 2001, Investigating relationships between LandsatETM+ sensor data and leaf area index in a boreal conifer forest. Remote Sensing ofEnvironment, 78, pp. 239–251.

FANG, H. and LIANG, S., 2005, A hybrid inversion method for mapping leaf area indexfrom MODIS data: experiments and application to broadleaf and needleleaf canopies.Remote Sensing of Environment, 94, pp. 405–424.

FRIEDL, M.A., MICHAELSON, J., DAVIS, F.W., WALKER, H. and SCHIMEL, D.S., 1994,Estimating grassland biomass and leaf area index using ground and satellite data.International Journal of Remote Sensing, 15, pp. 1401–1420.

GOEL, N.S. and QIN, W., 1994, Influences of canopy architecture on relationships between var-ious vegetation indices and LAI and FPAR: a computer simulation. Remote SensingReviews, 10, pp. 309–347.

GOOVAERTS, P., 1997, Geostatistics for Natural Resources Evaluation (New York: OxfordUniversity Press).

HABOUDANE, D., MILLER, J.R., TREMBLAY, N., ZARCO-TEJADA, P.J. and DEXTRAZE, L.,2002, Integrated narrow-band vegetation indices for prediction of crop chlorophyllcontent for application to precision agriculture. Remote Sensing of Environment, 81,pp. 416–426.

HALL, F.G., SHIMABUKURO, Y.E. and HUEMMERICH, K.F., 1995, Remote sensing of forestbiophysical structure using mixture decomposition and geometric reflectance models.Ecological Applications, 5, pp. 993–1013.

HOUBORG, R. and BOEGH, E., 2008, Mapping leaf chlorophyll and leaf area index using inverseand forward canopy reflectance modelling and SPOT reflectance data. Remote Sensingof Environment, 112, pp. 186–202.

HU, B., MILLER, J.R., CHEN, J.M. and HOLLINGER, A., 2004, Retrieval of canopy leaf areaindex in the BOREAS flux tower sites using linear spectral mixture analysis. RemoteSensing of Environment, 89, pp. 176–188.

HUETE, A.R., 1988, A soil-adjusted vegetation index (SAVI). Remote Sensing of Environment,25, pp. 295–309.

JACQUEMOUD, S., BACOUR, C., POILVÉ, H. and FRANGI, J.-P., 2000, Comparison of four radia-tive transfer models to simulate plant canopies reflectance: direct and inverse mode.Remote Sensing of Environment, 74, pp. 471–481.

JACQUEMOUD, S., BARET, F., ANDRIEU, B., DANSON, F.M. and JAGGARD, K., 1995, Extractionof vegetation biophysical parameters by inversion of the PROSPECT+SAIL models onsugar beet canopy reflectance data: application to TM and AVIRIS sensors. RemoteSensing of Environment, 52, pp. 163–172.

JORDAN, C.F., 1969, Derivation of leaf area index from quality of light on the forest floor.Ecology, 50, pp. 663–666.

LI, X. and STRAHLER, A.H., 1985, Geometric-optical modelling of a conifer forest canopy.IEEE Transactions on Geoscience and Remote Sensing, 23, pp. 705–721.

MADUGUNDU, R., NIZALAPUR, V. and JHA, C.S., 2008, Estimation of LAI and above-groundbiomass in deciduous forests: Western Ghats of Karnataka, India. International Journalof Applied Earth Observation and Geoinformation, 10, pp. 211–219.

MCALLISTER, D.M. and VALEO, C., 2007, A robust new method for the remote estimationof LAI in montane and boreal forests. International Journal of Remote Sensing, 8,pp. 1891–1905.

MILLER, J.B., 1967, A formula for average foliage density. Australian Journal of Botany, 15, pp.141–144.

NEMANI, R.R., PIERCE, L.L., RUNNING, S.W. and BAND, L., 1993, Forest ecosystem pro-cesses at the watershed scale: sensitivity to remotely sensed leaf area index estimates.International Journal of Remote Sensing, 14, pp. 2519–2534.

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

Leaf area index estimation of bamboo forest 5379

PEDDLE, D.R., BRUNKE, S.P. and HALL, F.G., 2001, A comparison of spectral mixture analysisand ten vegetation indices for estimating boreal forest biophysical information fromairborne data. Canadian Journal of Remote Sensing, 27, pp. 627–635.

PEDDLE, D.R., HALL, F.G. and LEDREW, E.F., 1999, Spectral mixture analysis and geometric–optical reflectance modelling of boreal forest biophysical structure. Remote Sensing ofEnvironment, 67, pp. 288–297.

PETERSON, D.L., SPANNER, M.A., RUNNING, S.W. and TEUBER, K.B., 1987, Relationship ofthematic mapper simulator data to leaf area index of temperate coniferous forests.Remote Sensing of Environment, 22, pp. 323–341.

PINTY, B. and VERSTRAETE, M.M., 1992, GEMI: a nonlinear index to monitor globalvegetation from satellites. Vegetation, 10, pp. 15–20.

PU, R., GONG, P., BIGING, G.S. and LARRIEU, M.R., 2003, Extraction of red edge opti-cal parameters from Hyperion data for estimation of forest leaf area index. IEEETransactions on Geoscience and Remote Sensing, 41, pp. 916–921.

QI, J., KERR, Y.H., MORAN, M.S., WELTZ, M., HUETE, A.R., SOROOSHIAN, S. and BRYANT,R., 2000, Leaf area index estimates using remotely sensed data and BRDF models in asemiarid region. Remote Sensing of Environment, 73, pp. 18–30.

RICHARDSON, A.J. and WIEGAND, C.L., 1977, Distinguishing vegetation from soil backgroundinformation. Photogrammetric Engineering and Remote Sensing, 43, pp. 1541–1552.

ROUJEAN, J.L. and BREON, F.M., 1995, Estimating PAR absorbed by vegetation from bidirec-tional reflectance measurements. Remote Sensing of Environment, 51, pp. 375–384.

ROUSE, J.W., HAAS, R.H., DEERING, D.W., SCHELL, J.A. and HARLAN, J.C., 1974, Monitoringvegetation systems in the Great Plains with ERTS. In Proceedings, 3rd Earth ResourceTechnology Satellite (ERTS) Symposium, Washington, DC (Washington, DC: NASA,Scientific and Technical Information Office), pp. 48–62.

SCHLERF, M. and ATZBERGER, C., 2006, Inversion of a forest reflectance model to estimatestructural canopy variables from hyperspectral remote sensing data. Remote Sensing ofEnvironment, 100, pp. 281–294.

SCHLERF, M., ATZBERGER, C. and HILL, J., 2005, Remote sensing of forest biophysical vari-ables using HyMap imaging spectrometer data. Remote Sensing of Environment, 95,pp. 177–194.

SOUDANI, K., FRANCOIS, C., MAIRE, G.L., DANTEC, V.L. and DUFRENE, E., 2006,Comparative analysis of IKONOS, SPOT, and ETM+ data for leaf area index esti-mation in temperate coniferous and deciduous forest stands. Remote Sensing ofEnvironment, 102, pp. 161–175.

STENBERG, P., RAUTIAINEN, M., MANNINEN, T., VOIPIO, P. and SMOLANDER, H., 2004,Reduced simple ratio better than NDVI for estimating LAI in Finnish pine and sprucestands. Silva Fennica, 38, pp. 3–14.

THENKABAIL, P.S., ENCLONA, E.A., ASHTON, M.S., LEGG, C. and DIEU, M.J.D., 2004,Hyperion, IKONOS, ALI and ETM+ sensors in the study of African rainforests.Remote Sensing of Environment, 90, pp. 23–43.

WALTHALL, C., DULANEY, W., ANDERSON, M., NORMAN, J., FANG, H. and LIANG, S., 2004,A comparison of empirical and neural network approaches for estimating corn andsoybean leaf area index from Landsat ETM+ imagery. Remote Sensing of Environment,92, pp. 465–474.

WELLES, J.M. and NORMAN, J.M., 1991, Instrument for indirect measurement of canopyarchitecture. Agronomy Journal, 83, pp. 818–825.

ZARCO-TEJADA, P.J., MILLER, J.R., NOLAND, T.L., MOHAMMED, G.H. and SAMPSON, P.H.,2001, Scaling-up and model inversion methods with narrowband optical indices forchlorophyll content estimation in closed forest canopies with hyperspectral data. IEEETransactions on Geoscience and Remote Sensing, 39, pp. 1491–1506.

ZHANG, Z., HE, G. and WANG, X., 2010, A practical DOS model-based atmospheric correctionalgorithm. International Journal of Remote Sensing, 31, pp. 2837–2852.

Dow

nloa

ded

by [

Cen

ter

for

Ear

th O

bser

vatio

n an

d D

igita

l Ear

th ]

, [Z

haom

ing

Zha

ng]

at 1

7:44

07

Oct

ober

201

1

View publication statsView publication stats