Dynamics of Land Surface Temperature in Response to Land-Use/Cover Change

Download Dynamics of Land Surface Temperature in Response to Land-Use/Cover Change

Post on 21-Jul-2016

214 views

Category:

Documents

1 download

TRANSCRIPT

Dynamics of Land Surface Temperature inResponse to Land-Use/Cover Changegeor_686 23..36XIAOLU ZHOU1* and YI-CHEN WANG21Department of Geography, National University of Singapore, 1 Arts Link, AS2 #03-01, Singapore,117570.2Department of Geography, National University of Singapore, 1 Arts Link, AS2 #03-05, Singapore,117570.*Corresponding author. Email: g0800627@nus.edu.sg; xiaoluzhou86@gmail.comReceived 21 July 2010; Revised 10 October 2010; Accepted 14 November 2010AbstractIn this study, we employed Geographical Information Systems and remotesensing techniques to investigate the impact of land-use/cover change on landsurface temperature (LST) in a rapidly urbanisation city, Kunming in south-west China. Spatial patterns of LST and land use for 1992 and 2006 werederived from Landsat images to examine how LST responded to urban growth.Remote sensing indices were used to quantify land-use types and employed asexplanatory variables in LST modelling. The geographically weighted regres-sion (GWR), a location dependent model, was performed to explore the influ-ences of the spatially varied land-use conditions on the LST patterns. Resultsrevealed that rapid urbanisation in Kunming altered the local thermal environ-ment, particularly in increasing the LST in the zone surrounding the urban core.Remote sensing indices demonstrated that water and vegetation played animportant role in mitigating the urban heat island effect, while built-up andbarren land accounted for the increase in LST. The GWR improved thegoodness-of-fit for LST modelling and provided insights into the spatiallyvaried relationship between LST and land-use conditions.KEY WORDS land-use/cover change; urban heat island; remote sensing; geo-graphical information systems; geographically weighted regression; KunmingIntroductionRapid population growth and continuous exploi-tation of natural resources during the pastcentury have caused land-use/cover change(LUCC) worldwide. Urbanisation is occurring inmany countries, increasing the concentration ofpopulation in cities. The rate of urbanisation insome developing countries is remarkable overrecent decades because of the pursuit of fast eco-nomic development (Li et al., 2009). In China,the launch of economic reforms since the late1970s has largely accelerated the urbanisationprocess (Luo and Wei, 2009).The process of urbanisation often replacesnatural vegetation and agricultural land withimpervious surfaces, such as buildings and roads(Thi Van and Xuan Bao, 2010). This trend hasproduced a series of environmental impacts onbiodiversity, local climate, hydrologic processes,and so forth (Streutker, 2002; Roy et al., 2009).Of these impacts, the urban heat island (UHI),the phenomenon where temperatures in urbancores are higher than they are in surroundingrural areas (Voogt and Oke, 2003), is commonlyassociated with cities. Because each land-use/cover type has its unique thermal, moisture, and23Geographical Research February 2011 49(1):2336doi: 10.1111/j.1745-5871.2010.00686.xoptical spectral properties, LUCC will affectthe local thermal environment (Oke, 1982). Forexample, the expansion of the impervioussurface area alters the heat capacity and radiativeproperties of the land surface, largely reducingthe evapotranspiration in urban areas (Streutker,2002).Prior studies on urban thermal environmentshave been based on air temperature obtainedfrom weather stations (Li et al., 2009). Althoughthese in situ data can provide accurate local tem-perature, they are normally costly for large-scaleanalyses and subject to private or governmentalrestrictions (Owen et al., 1998). Furthermore,data from weather stations are discrete points inspace, which hardly reflect the spatial variationof temperature caused by different land-use/cover types. An alternative to measuring urbanthermal environment is using land surface tem-perature (LST) because it is able to modulate theair temperature of the layer immediately abovethe earth surface and is a major parameterassociated with surface radiation and energyexchange (Voogt and Oke, 1998; Weng, 2009). Avariety of satellite sensors that capture thermalinfrared information have been available, such asLandsat Thematic Mapper (TM) and EnhancedThematic Mapper Plus, and Moderate-resolutionImaging Spectroradiometer, facilitating the in-vestigation of LST in high spatial and temporalresolutions.Studies have investigated the relationshipbetween LST and LUCC (Chen et al., 2006;Xiao and Weng, 2007; Thi Van and Xuan Bao,2010) with two main foci. The first focus hasbeen on the comparison of the LST of differentland-use conditions. Multispectral techniqueshave been used because thermal and land-useinformation can be obtained simultaneously froma single sensor (Voogt and Oke, 2003). Analy-tical functions of Geographical InformationSystems (GIS), such as spatial overlay and imagedifferencing, coupled with biophysical para-meters of surface temperature and emissivityderived from remotely sensed images, haveeffectively unveiled the impacts of LUCCon thermal environment change (Chen et al.,2006).The second focus has been on modelling LSTbased on remote sensing indices. Instead ofdirectly using land-use types that are categoricaldata, remote sensing indices are employed toquantitatively represent land-use/cover types.For example, the Normalized Difference Vegeta-tion Index (NDVI) has been used to validate therole of green space in mitigating UHI (Yuanand Bauer, 2007). The Normalized DifferenceBuilt-up Index (NDBI) (Zha et al., 2003) and theNormalized Difference Water Index (NDWI)(Gao, 1996) have been employed to representurban and water areas quantitatively. Althoughprior studies have adopted these indices to modelLST (Chen et al., 2006), only a few of them havecompared the modelling results of differentyears.When modelling the LST, global regressionmodels, such as ordinary least squares regression(OLS), are commonly developed to estimate LSTbased on explanatory variables (Weng et al.,2004; Chen et al., 2006). These global modelsassume that the relationships between LSTand explanatory variables are spatially constant(Bagheri et al., 2009). However, the explanatoryvariables, such as land-use/cover types and theirchanges, often have various effects on LSTacross the space. Research is thus needed thatemploys localised statistical models in analysingthe relationship between LST and LUCC.In this study, the effect of LUCC on LST in arapidly urbanisation city, Kunming, China, wasinvestigated. LST was derived from satelliteimages, and the change of the LST pattern inresponse to urbanisation was explored using GISand remote sensing techniques. Three remotesensing indices were compared in terms oftheir effectiveness in LST estimation. Effects ofspatial variation of land-use/cover distribution onLST were examined by comparing the OLS, aglobal model, with the geographically weightedregression (GWR), a location dependent method.Specifically, the study addressed three questions.First, how has urbanisation affected the localthermal environment? Second, how effective aredifferent remote sensing indices in LST estima-tion? Third, does the location dependent GWRmethod perform better in LST modelling than theOLS method?Study areaThe study area, Kunming, is located from10210 to 10340E and 2423 to 2633Nin the north-central Yunnan province, China(Figure 1). It is situated on a plateau with eleva-tion ranging from 1500 m to 2800 m. The met-ropolitan area of Kunming is located in theDianchi basin with an elevation of 1890 m, sur-rounded by mountains to the north, east, andwest. The city is characterised by a subtropicalhighland climate with warm and humid summersand cold and dry winters. Mean annual precipi-24 Geographical Research February 2011 49(1):2336 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographerstation is around 1000 mm; mean annual sunexposure is about 2250 hours.Kunming is the capital ofYunnan province andhas experienced rapid development in the lasttwo decades. The gross domestic product (GDP)increased from 16.9 billion ChineseYuan in 1992to 56.2 billion in 1998. From 1998 to 2008, theGDP nearly tripled, attaining 160.5 billionChinese Yuan (KBS, 2009). The fast economicgrowth is accompanied by rapid urbanisation,causing a great loss of green space in the subur-ban areas and an increase of the urbanised areafrom 184.4 km2 in 1992 to 257.8 km2 in 2005(Cai, 2007). The study area is within metropoli-tan Kunming, an area of 538.9 km2, consisting ofurban area (dominated by impervious land),urban fringe (combination of impervious landand agriculture land), and suburban area (domi-nated by forest land).Materials and methodsDataRemotely sensed images and field survey datawere used in this study. The remotely sensedimages used were Landsat 5 TM images obtainedfrom the US Geological Survey website withstandard systematic corrections. Because of dataquality, availability, and comparability, twoscenes of images, dated 16 August 1992 and 19May 2006, were used in this study. The opticalbands of TM images (bands 15 and 7) havea spatial resolution of 30 m and the thermalband (band 6) has a spatial resolution of 120 m.Field survey data consisted of ground controlpoints collected using Global PositioningSystems (GPS) for geo-referencing the two TMimages and photographs for assisting in imageinterpretation.Image preparationFigure 2 is the flow chart of the analysis. Forimage preparation, the GPS ground controlpoints were used to geo-reference and rectifythe two TM images to the Universal TransverseMercator coordinate system zone 48, using theWGS84 datum. The nearest neighbour methodwas employed to resample all the bands into30 m. The root mean square error was controlledwithin 0.5 pixels for these two images. Toremove the effect of cloud on land-use classifi-cation and LST derivation, cloud cover, whichFigure 1 Location of the study area, the Kunming metropolitan area, in Yunnan Province, China.X.L. Zhou and Y.-C. Wang: Dynamics of Land Surface Temperature 25 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographersaccounted for less than 1% of the images, wasmasked out from the analysis. Atmospheric cor-rection was carried out in ENVI 4.6 using theFast Line-of-sight Atmospheric Analysis ofSpectral Hypercubes (FLAASH) atmosphericcorrection module to convert digital numbers(DN) into reflectance. FLAASH is an atmo-spheric correction tool based on the MODTRANradiation transfer codes and algorithms by whicha unique model can be computed for each image(ENVI, 2009). The process filters the interfer-ence from the path radiation, such as aerosol,dust particle, and water vapour.Image classificationThe two TM images were classified into fiveland-use types, including built-up land, water,barren land, forest, and agriculture land, for theassessment of LST change in response to urbani-sation. This classification system was feasibleconsidering the medium resolution satelliteimages used. In addition, it conformed to theChinese land-use system (SAC, 2007), and aclassification system used in prior researchwhose study area shared the same characteristicsas Kunming city (Xiao and Weng, 2007).The maximum likelihood classifier wasemployed. Statistics of training samples wereplotted to ensure the separability of differentclasses and comparability between two images(Jensen, 2005). Accuracy assessment was per-formed using the stratified random samplingmethod. Field survey assisted by photographsand knowledge of the area were used as refer-ences in the assessment. The overall accuracy forthe images of 1992 and 2006 was 83.7% and80.8%, respectively. The Kappa coefficients forboth images were above 0.78.Derivation of LSTThe DN values of both Landsat images wereconverted into spectral radiance based on the fol-lowing formula:Radiance gain DN offset= + (1)Where, gain and offset were derived from thehead files of images and the Landsat currentradiometric calibration coefficients (Chanderet al., 2009).The retrieved radiance was converted intoat-satellite brightness temperature (BT) basedLandsatimage in1992Image preparationOptical bands (15, 7) Thermal band (6)Landsatimage in2006Classified imagesChange detectionSpatial overlay (a)Spatial overlay (b) Correlation andregressionUnivariate regressionMultivariate regressionDerivation of remotesensing indicesDerivation of BTDerivation of LSTOLS GWRFigure 2 Flow chart of the analysis. Spatial overlay (a) is conducted using the land-use layer and the LST layer, while spatialoverlay (b) uses the land-use change layer and the LST layer. BT, brightness temperature; GWR, geographically weightedregression; LST, land surface temperature; OLS, ordinary least squares regression.26 Geographical Research February 2011 49(1):2336 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographerson the following equation, assuming that landcover had the same emissivity (Weng and Lu,2008):T KK Lb=+( )211ln (2)Where Tb is the at-satellite BT in Kelvin (K);Ll is the spectral radiance in Wm-2 sr-1 mm-1;K1 and K2 are the calibration constantswith K1 = 607.76 Wm-2 sr-1 mm-1 and K2 =1260.56 K.Because the focus of the study, like most of theprior work (e.g. Weng et al., 2004; Chen et al.,2006; Li et al., 2009), was on the spatialvariation of surface temperature, the followingmethods were employed. To retrieve the LST, theat-satellite BT (Tb) needs to be corrected accord-ing to the real object properties. Therefore, theemissivity corrected surface temperature wascomputed following Artis and Carnahan (1982):T TTsbb=+ ( )1 ln (3)Where Ts is the LST in K; l is the wavelength ofemitted radiance (l = 11.5 mm) (Markham andBarker, 1985); a equals 1.438 10-2 mK, calcu-lated as a = hc/s, with h as the Planck constant(6.626 10-34 Js), c as the velocity of light (2.998 108 m s-1), and s as the Boltzmann constant(1.38 10-23 J/K); e is the surface emissivityderived from NDVI (Artis and Carnahan, 1982;Roerink et al., 2000).Derivation of remote sensing indicesThree remote sensing indices, NDVI, NDBI, andModified Normalized Difference Water Index(MNDWI) (Xu, 2006), were computed to char-acterise the land-use types of this study. TheNDVI is an index widely used to describe thegreenness of an area (Chen and Brutsaert, 1998),and is used to represent the extent of forest andagricultural land. The NDBI is an indicator ofurban areas, which can reveal the built-up andbarren land (Zha et al., 2003; Chen et al., 2006).The MNDWI is selected to represent water areasbecause water areas often reveal remarkabledifferences in thermal characteristics when mod-elling the urban thermal environment, and com-pared with NDWI, MNDWI can remove thebuilt-up land noise on water in the urban areas(Xu, 2008). These three indices are calculated asfollows:NDVI R RR RNIR REDNIR RED=+(4)NDBI R RR RMIR NIRMIR NIR=+(5)MNDWI R RR RGREEN MIRGREEN MIR=+(6)Where RNIR is the reflectance in the near infraredband; RRED and RGREEN, respectively, stand for thereflectance in red and green bands; RMIR denotesthe reflectance in the middle infrared band.Analysis of the impacts of LUCC on LSTThe classified land-use maps and the derivedLST layers of 1992 and 2006 were incorporatedinto ArcGIS 9.3 to analyse the relationshipbetween the spatial patterns of LST and land-usetypes. Because of spatial autocorrelation, a largeportion of the data was redundant for analysis.Spatial sampling was thus carried out to extract5929 points evenly distributed over the studyarea for analysis. The LST layers were thenimposed onto the land-use layers to explore theaverage temperature for each land-use type(Spatial overlay (a) in Figure 2). The LSTs ofdifferent land-use types in 1992 and 2006 werecompared to examine the thermal environmentchange. To examine how LUCC might haveinfluenced the LST pattern, change detection wasperformed in ArcGIS to identify the areas whereLUCC took place between 1992 and 2006. Thedetected changed areas were then overlaid withLST layers to calculate the LST differencesduring this time period (Spatial overlay (b) inFigure 2).Variation in LST may be subject to three majorfactors: seasonality, time of day, and land-usecondition. To investigate the LST variation asso-ciated only with changes in land-use condition,the effects of the other two factors need to beminimised. Since the Landsat satellite follows asun-synchronous orbit, it ensures that imagestaken for the same area have a similar local time.The two images used in this study were bothcaptured at around 10:30 am local time, therebyexcluding the influence of the time of day factor.As for seasonality, temperature difference for thesame land-use type between the two images inAugust 1992 and May 2006 was assumed to becaused by different seasons. To exclude the sea-sonal influence, LST difference for the sameX.L. Zhou and Y.-C. Wang: Dynamics of Land Surface Temperature 27 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographersland-use type was normalised against all otherLST differences:dT T Tij j i= ( ) ( )2006 1992 (7)T T Ti i i= ( ) ( )2006 1992 (8)dT dT Tn ij i= (9)Where dTij is the temperature difference betweenland-use type j in 2006 and land-use type i in1992; DTi is the temperature difference for thesame land-use type i in 2006 and 1992; dTn isthe normalised temperature by subtracting DTifrom dTij.Univariate and multivariate regression of LSTon remote sensing indicesThe land-use types of the study area were repre-sented using remote sensing indices of NDVI,NDBI, and MNDWI. Areas with more greennesswill exhibit higher NDVI values. High NDBIvalues generally signify areas with intensiveurban development. The MNDWI is sensitive towater bodies, so it will exhibit high values forareas covered by water.The strength of the correlation between LSTand individual explanatory variables (i.e. each ofthe three remote sensing indices) was first calcu-lated using univariate regression. Multivariateregression was next conducted to model the LSTbased on the combination of the explanatoryvariables. The multivariate regressions were per-formed based on OLS and GWR. As opposed toOLS which may mask out local variation, GWRis a localised regression that examines spatiallynon-stationary phenomena (Fotheringham et al.,2002; Lloyd and Shuttleworth, 2005) and isexpressed as:y u v u v Xi i k i i kiki= ( ) + ( ) + 0 , , (10)Where b0 is the constant that depends on thespecific location i; bk is the coefficient of inde-pendent variable Xk at the location i; ei is theresidual term at the location i; (ui, vi) indicatesthe coordinates of the point i. GWR allowslocal parameters to be estimated based on adja-cent points. Coefficient bk is controlled by theweight which is assigned according to thespatial proximity of point i to its adjacentpoints. Weight can be obtained through twotypes of functions, fixed or adaptive (Pez et al.,2002a; 2002b). In this study, an adaptive func-tion is used which takes on the form below(Fotheringham et al., 2002):wdbif j Sw otherwiseijijij= { }=102 2(11)Where dij is the distance from point i to j; S is theset that indicates the specified N nearest neigh-bour points; b is the bandwidth, defined as thedistance from the Nth nearest neighbor point toi. The corrected Akaile Information Criterion(AICc), an indicator of the distance between theunknown true model and the derived model(Fotheringham et al., 2002), was used to deter-mine the bandwidth. A lower value of AICc gen-erally reflects a better simulation; thus, followingCharlton and Fotheringham (2009), a computa-tional process was done for each point to searchfor the lowest AICc to determine the correspond-ing bandwidth.To evaluate the robustness of the spatialregressions, the spatial distribution of residualsfrom OLS and GWR were analysed. If theresiduals cluster together and display an obviousspatial pattern, the constructed model may be inquestion. Morans I, an index that measures thecorrelation degree of the adjacent observations,was used. The index ranges from -1 (perfectnegative autocorrelation) to 1 (perfect positiveautocorrelation) with values close to zero sug-gesting no obvious spatial autocorrelation.ResultsImpacts of LUCC on LSTThe LST maps of 1992 and 2006 (Figure 3) notonly exhibited the magnitude and spatial varia-tion of surface temperature, but also showed theeffects of urbanisation on LST. The central urba-nised area with a light tone indicated a warmersurface temperature. The area with light tonespread dramatically from 1992 to 2006, suggest-ing the expansion of UHI.The spatial pattern change of the remotesensing indices also revealed the effect of LUCCon LST (Figure 4). Areas with high NDVI shranktremendously from 1992 to 2006, suggestinga green space loss (Figure 4a). Light tones inthe NDBI maps signified the urbanised areas(Figure 4b), and they showed a marked urbangrowth from 1992 to 2006. These results echoed28 Geographical Research February 2011 49(1):2336 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographersthe increased LST in the urban fringe (Figure 3).The MNDWI maps reflected the water areas, butdid not display any major pattern change duringthe time period (Figure 4c).The LST maps of the two years were superim-posed onto land-use maps to retrieve the surfacetemperature for each land-use type. Of all the fiveland-use classes, water exhibited the lowest tem-Figure 3 Land surface temperature (LST) in (a) 1992 and (b) 2006. K, Kelvin degrees.Figure 4 Remote sensing indices of (a) Normalized Difference Vegetation Index; (b) Normalized Difference Built-up Index;and (c) Modified Normalized Difference Water Index in 1992 and 2006.X.L. Zhou and Y.-C. Wang: Dynamics of Land Surface Temperature 29 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographersperature with the minimum variation between thetwo years because its relatively high thermalinertia reduced the heat difference. Built-up andbarren land showed the highest temperatures in1992 and 2006. The temperature of forest wasbetween that of water area and built-up land, witha 1.3 K difference between 1992 and 2006, indi-cating a relatively stable thermal environment forforested land. Agricultural land exhibited a greatthermal difference of 5 K between the two years,probably due to the different crop growingstages. In August, crops in Kunming are in theirprime time so agricultural land is covered by fullgreen vegetation. While in May, crops are justplanted, with much of the land still covered bysoil. This resulted in a large temperature differ-ence between the two dates for agricultural land(Table 1).To investigate the effect of LUCC on LSTchange, differences in LST were calculated bysubtracting LST in 1992 from LST in 2006 foreach land-use change category (Table 2). LSTincreased about 3.4 K and 1.9 K, respectively,for forest and agricultural land that were con-verted into built-up areas. The LST dropped3.1 K when agricultural land changed toforest, but rose 1.5 K when forest was turnedTable 1 Derived land surface temperature in Kelvin degreesfor each land-use type with the standard deviation inparentheses.Land-UseTypes16 August199219 May2006dTAgriculture 293.8 (1.5) 298.9 (1.7) 5.1Water 292.4 (1.1) 291.7 (1.0) -0.7Barren 297.6 (2.1) 300.4 (2.0) 2.8Built-up 298.3 (2.2) 301.0 (2.0) 2.7Forest 295.1 (2.1) 296.4 (2.0) 1.3dT is the temperature difference between 1992 and 2006.Table 2 Effects of land-use/cover change on land surface temperature.Land-Use Type dT (s.d.) Adjusted dT16 August 1992 19 May 2006Agriculture Agriculture 4.7 (2.2) 0Built-up 6.6 (2.7) 1.9Barren 7.1 (2.4) 2.4Water -2.8 (1.7) -7.5Forest 1.6 (2.2) -3.1Forest Forest 1.7 (2.2) 0Built-up 5.1 (2.8) 3.4Barren 4.8 (3.0) 3.1Water -0.7 (1.7) -2.4Agriculture 3.2 (1.8) 1.5Built-up Built-up 3.0 (1.9) 0Agriculture 2.8 (2.2) -0.2Barren 3.5 (2.7) 0.5Water -1.4 (1.5) -4.4Forest 1.7 (1.0) -1.3Barren Barren 2.0 (2.4) 0Built-up 3.7 (2.3) 1.7Agriculture 1.7 (2.4) -0.3Water -1.5 (1.4) -3.5Forest -0.6 (2.2) -2.6Water Water -0.9 (0.9) 0Barren 3.3 (4.6) 4.2Forest 1.2 (1.8) 2.1Built-up 4.3 (4.6) 5.2Agriculture 4.9 (3.3) 5.8dT is the temperature difference between two years with standard deviation (s.d.) in parentheses. Adjusted dT is calculated basedon equations (7)(9).30 Geographical Research February 2011 49(1):2336 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographersinto agricultural land. The conversion of barrenland had two main pathways. For barren landconverted into built-up land, an average of1.7 K was gained, while for barren landchanged into green space, a temperature declinewas observed. A small fraction of the built-upland in 1992 changed into green space in 2006,accompanied by a decreased LST.To highlight the spatial correspondencebetween LST and LUCC, the land-use types ofbuilt-up land and barren land were combined asnon-vegetated areas, and forest and agricultureland were combined as vegetated areas. Theresults exhibited a noticeable correspondencebetween LUCC and LST change. It was evidentthat urban sprawl took place on the fringes of theKunming city from 1992 to 2006, convertingvegetated areas into built-up or barren lands,while some non-vegetated areas were turned intovegetated areas in eastern Kunming (Figure 5a).The thermal environment changed accordingly,as seen in the increased LST in the surroundingsof the urban core and the decreased LST ineastern Kunming.Univariate modelling of LST based on remotesensing indicesTo better understand how LST dynamics wereassociated with LUCC, correlation strengthbetween LST and remote sensing indices wasexamined. Scatter plots of LST and remotesensing indices assessed the effectiveness of eachindex in modelling LST (Figure 6). For NDVIvalues above zero, a negative relation with LSTwas found, suggesting that densely vegetatedareas were associated with lower LST. For NDVIvalues below zero, there were some scatteredpoints deviating from the trend line, probably dueto the presence of water that had a lower LST thanother land-use types (cf., Table 1). The LST dis-played a positive relationship with NDBI. Thecorrelation coefficient increased from 0.55 to 0.59from 1992 to 2006, suggesting that the NDBIperformed better when urban areas became domi-nant. MNDWI showed a negative relation withLST, which confirmed that water areas had anotably cooling effect. However, the robustness ofthe regression was reduced by the large variationof LST where MNDWI was below zero.Figure 5 Spatial correspondence between (a) changes in land use and (b) differences in land surface temperature (LST)between 1992 and 2006. Each dot represents one of the 5929 sampled points. K, Kelvin degrees.X.L. Zhou and Y.-C. Wang: Dynamics of Land Surface Temperature 31 2010 The AuthorsGeographical Research 2010 Institute of Australian GeographersComparison of the OLS and the GWRmodelling on LSTMultivariate regression was carried out toexamine whether it would improve the LSTmodelling. When the multivariate regressionoperated based on the OLS method, resultsdid not reveal a significant improvement. TheMNDWI was not statistically significant atthe 0.05 level in 1992 (Table 3). Because of theheterogeneity of the land-use mosaic and thespatial variation in LST, the spatially variedrelationship between LST and remote sensingindices cannot be reflected by the global mod-elling method. Conversely, the results of thelocalised regression method GWR showed amuch higher adjusted r2 value than the OLS(Table 4). The performance of OLS and GWRwere further compared using the AICc indica-tor; the AICc values for GWR were much lowerthan those for OLS, suggesting that GWR per-formed better than OLS (Table 4). Results fromMorans I did not show any spatial autocorrela-tion among the residuals of the OLS and GWRmodels, indicating the robustness of the models(Table 4).While the global regression of OLS uni-formly showed that high NDVI values wereassociated with low LST (Table 3), GWRFigure 6 Relations between (a) NDVI and LST; (b) NDBI and LST; and (c) MNDWI and LST. K, Kelvin degrees; LST, landsurface temperature; MNDWI, Modified Normalized Difference Water Index; NDBI, Normalized Difference Built-up Index;NDVI, Normalized Difference Vegetation Index.Table 3 Parameter coefficients of the multivariate regression models based on the OLS and GWR methods.Coefficients 1992 2006OLS GWR (Median) OLS GWR (Median)Intercept 299.131 298.512 298.907 299.767NDVI 0.141 -5.150 -2.827 -4.036NDBI 8.431 7.984 7.557 7.603MNDWI -0.4221 -6.083 -5.778 -4.3151 Parameter is not significant based on the t statistic (at the 0.05 level).GWR, geographically weighted regression; MNDWI, Modified Normalized Difference Water Index; NDBI, Normalized Dif-ference Built-up Index; NDVI, Normalized Difference Vegetation Index; OLS, ordinary least squares regression.32 Geographical Research February 2011 49(1):2336 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographersallowed the spatially varied relationshipbetween NDVI and LST to be revealed. In mostplaces, coefficients of NDVI displayed as nega-tive values, suggesting that vegetated areas(often high NDVI) relieved the urban heat (thuslow LST) (Figure 7a). However, the coefficientsof GWR revealed that there were areas withpositive coefficients between NDVI and LST,particularly in areas covered by water (cf., Fig-ures 1 and 7a). This illustrated that GWR effec-tively modelled the relationship between LSTand NDVI in areas with different land surfacecharacteristics.Coefficients of NDBI showed a positiverelationship with LST across the study area(Figure 7b). The coefficients increased from1992 to 2006 in southern Kunming where largeareas of agricultural land was converted intourban areas, indicating the effect of urban expan-sion on enhancing the LST. Apart from repre-senting water area, MNDWI can be used torepresent surface moisture (Zha et al., 2003;Chen et al., 2006). The general negative coeffi-cient of MNDWI (Figure 7c) thus suggested thatwater bodies and surface moisture can moderatethe surface temperature. The comparison of theMNDWI coefficients between 1992 and 2006showed that areas with positive coefficients ofMNDWI expanded in the south-east, inferringa loss of surface moisture because of theTable 4 Comparison of the diagnostics for OLS and GWRmodelling.Year Method Adjust r2 AICc Morans I1992 OLS 0.531 28 128.7 -0.018GWR 0.721 23 046.1 -0.0032006 OLS 0.615 28 322.6 -0.009GWR 0.746 22 668.1 -0.004AICc, corrected Akaile Information Criterion; GWR, geo-graphically weighted regression; OLS, ordinary least squaresregression.Figure 7 Local coefficients of the geographically weighted regression for (a) Normalized Difference Vegetation Index;(b) Normalized Difference Built-up Index; and (c) Modified Normalized Difference Water Index.X.L. Zhou and Y.-C. Wang: Dynamics of Land Surface Temperature 33 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographersconversion of vegetated land to impervioussurface (Figure 7c).DiscussionProcess of thermal environment changeThe LUCC has extensively modified thethermal environment of the Kunming city. TheLST became higher as the land use changedfrom vegetated areas to barren and built-up land(Table 2, Figure 5) in the following process. Atthe beginning stage, the removal of vegetatedareas changed the surface energy balance,which increased the sensible heat flux at thecost of the latent heat flux (Owen et al., 1998).The evapotranspiration of vegetation was sub-stantially reduced. Following this stage was aconversion of bare soil into impervious surface,which altered the moisture balance. The reduc-tion of the water storage of land furtherstrengthened the sensible heat flux. In addition,the land surface conversion largely modified thethermal property of the land surface (e.g. heatcapacity and thermal conductivity), creating awarmer thermal environment (Li et al., 2009).At the last stage, a great deal of impervioussurface was developed. Built-up areas mightfacilitate sunlight absorption because of mul-tiple times of reflections. This urban topogra-phy, known as the canyon effect, furtherintensified the UHI (Oke, 1982). Apart fromland-use change, urbanisation with increasedhuman population also contributed to the urbanthermal environment change with the risinganthropogenic heat discharge.Effects of green policies on thethermal environmentAlthough areas with high LST have increasedfrom 1992 to 2006 (Figure 3), several spots withdark tone indicating a low LST can be detected inthe urban area in 2006 (Figure 3b). These areaswere associated with urban parks and riparianbelts with vegetation cover. In recent years,Kunming implemented several policies to intro-duce green space into the city. The campaigns toallot more green patches and corridors into thecity ameliorated the urban environment becauseof the role of vegetation in mitigating UHI, asseen that the temperature of vegetated areas wasaround 34 K lower than the built-up land(Tables 1 and 2). As a result, more cooling spotsin the urban core appeared on the 2006 LST map(Figure 3b). The dark line going through theurban centre in Figure 3b showing the position ofthe Panlong Canal also became noticeable in2006, suggesting a cooling trend along thePanlong Canal (Figures 1 and 3). This was prob-ably because of the greening initiatives along thebanks of the Panlong Canal, which expanded thegreen strips and thereafter reduced the LST. Also,the greening initiatives implemented in the eastof Kunming created or enlarged several urbanparks. Accordingly, some barren lands wereconverted into vegetated areas and noticeablyreduced the surface temperature in easternKunming (Figure 5).Localised statistics in LST modellingPrior studies pointed out that a single remotesensing index could not reflect the combinedimpacts of vegetation, buildings, water bodies,and other land cover on LST (Buyantuyev andWu, 2010). Multivariate regression conducted inthis study confirmed the importance of combin-ing remote sensing indices in LST modelling.The global models based on the OLS methodfailed to capture the spatial variation of LSTbecause all the parameters were uniformlyapplied to the data points regardless of their loca-tions. In contrast, GWR provided local coeffi-cients and residuals, allowing comparison of themodelling results in different places. The localcoefficients could be visualised on maps to illus-trate the spatially varied relationship betweeneach of the remote sensing indices and LST(Figure 7). The increased values of AICc and theadjusted r2 both suggested the effectiveness ofGWR in LST modelling.Conclusion and future workThis study contributed to the understandingof LUCC on the thermal environment. Rapidurbanisation notably caused the increase ofLST in Kunming, China, particularly in the areasurrounding the urban core. Among the threeremote sensing indices examined in the study,NDBI revealed a strong relation with LST.Multivariate regression results showed that thelocal regression model based on GWR signi-ficantly improved the goodness-of-fit andprovided insights into the spatially varied rela-tionship between LST and land-use conditions.The findings also suggested that sustainableurban green policies would be important inmoderating the urban thermal environment.Because different seasons of remote sensingimages may have caused the LST difference in34 Geographical Research February 2011 49(1):2336 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographersagricultural land in this study, future researchcan be conducted to examine how LST varieswith different land-use/cover types in differentseasons, particularly in the seasonal change ofthe UHI intensity and the effect of green spacein mitigating the UHI.ACKNOWLEDGMENTThe authors wish to thank the National University of Sin-gapore for providing funding support through the GraduateResearch Support Scheme and the research grant R-109-000-070-101 and two anonymous reviewers for their insightfulcomments.REFERENCESArtis, D.A. and Carnahan, W.H., 1982: Survey of emissivityvariability in thermography of urban areas. RemoteSensing of Environment 12, 313329.Bagheri, N., Holt, A. and Benwell, G.L., 2009: Using geo-graphically weighted regression to validate approaches formodelling accessibility to primary health care. AppliedSpatial Analysis 2, 177194.Buyantuyev, A. and Wu, J., 2010: Urban heat islands andlandscape heterogeneity: linking spatiotemporal variationsin surface temperatures to land-cover and socioeconomicpatterns. Landscape Ecology 25, 1733.Cai, H., 2007: Land use change comparative study ofKunming City based on remote sensing. Guizhou Science25, 178184. (In Chinese).Chander, G., Markham, B.L. and Helder, D.L., 2009:Summary of current radiometric calibration coefficientsfor Landsat MSS, TM, ETM+, and EO-1 ALI sensors.Remote Sensing of Environment 113, 893903.Charlton, M. and Fotheringham, A.S., 2009: Geographicallyweighted regression white paper. National Centre for Geo-computation, National University of Ireland MaynoothMaynooth, Co Kildare, Ireland, 117.Chen, D. and Brutsaert, W., 1998: Satellite-sensed distribu-tion and spatial patterns of vegetation parameters over atallgrass prairie. Journal of the Atmospheric Sciences 55,12251238.Chen, X., Zhao, H., Li, P. and Yin, Z., 2006: Remote sensingimage-based analysis of the relationship between urbanheat island and land use/cover changes. Remote Sensing ofEnvironment 104, 133146.ENVI, 2009: Atmospheric Correction Module: QUAC andFLAASH Users Guide.Fotheringham, A.S., Brunsdon, C. and Charlton, M.E., 2002:Geographically Weighted Regression: The Analysis ofSpatially Varying Relationships. Wiley, Chichester.Gao, B.-C., 1996: NDWI A normalized differencewater index for remote sensing of vegetation liquid waterfrom space. Remote Sensing of Environment 58, 257266.Jensen, J.R., 2005: Introductory Digital Image Processing ARemote Sensing Perspective, 3th edn. Prentice Hall, UpperSaddle River, NJ.KBS (Kunming Bureau of Statistics), 2009: KunmingStatistical Year Book. China Statistics Press, Beijing. (InChinese).Li, J.-J., Wang, X.-R., Wang, X.-J., Ma, W.-C. and Zhang, H.,2009: Remote sensing evaluation of urban heat island andits spatial pattern of the Shanghai metropolitan area, China.Ecological Complexity 6, 413420.Lloyd, C. and Shuttleworth, I., 2005: Analysing commutingusing local regression techniques: Scale, sensitivity, andgeographical patterning. Environment and Planning A 37,81103.Luo, J. and Wei, Y.H.D., 2009: Modeling spatial variationsof urban growth patterns in Chinese cities: The case ofNanjing. Landscape and Urban Planning 91, 5164.Markham, B.L. and Barker, J.K., 1985: Spectral characteris-tics of the LANDSAT Thematic Mapper Sensors. Interna-tional Journal of Remote Sensing 6, 697716.Oke, T.R., 1982: The energetic basis of the urban heat island.Quarterly Journal of the Royal Meteorological Society108, 124.Owen, T.W., Carlson, T.N. and Gillies, R.R., 1998: Anassessment of satellite remotely-sensed land cover param-eters in quantitatively describing the climatic effect ofurbanization. International Journal of Remote Sensing 19,16631681.Pez, A., Uchida, T. and Miyamoto, K., 2002a: A generalframework for estimation and inference of geographical-lyweighted regression models 1, Location-specific kernelbandwidths and a test for locational heterogeneity. Envi-ronment and Planning A 34, 733754.Pez, A., Uchida, T. and Miyamoto, K., 2002b: A generalframework for estimation and inference of geographicallyweighted regression models. 2. Spatial association andmodel specification tests. Environment and Planning A 34,883904.Roerink, G.J., Su, Z. and Meneti, M., 2000: S-SEBI: A simpleremote sensing Algorithm to estimate the surface energybalance. Physics and Chemistry of the Earth 25, 147157.Roy, A.H., Dybas, A.L., Fritz, K.M. and Lubbers, H.R.,2009: Urbanization affects the extent and hydrologic per-manence of headwater streams in a midwestern US metro-politan area. Journal of the North American BenthologicalSociety 28, 911928.SAC (Standardization administration of the peoples republicof China), 2007: Chinese Current Land Use Classification,GB/T 21010_2007. Chinese Standard Press, Beijing. (InChinese).Streutker, D.R., 2002: A remote sensing study of the urbanheat island of Houston, Texas. International Journal ofRemote Sensing 23, 25952608.Thi Van, T. and Xuan Bao, H.D., 2010: Study of the impact ofurban development on surface temperature using remotesensing in Ho Chi Minh City, northern Vietnam. Geo-graphical Research 48, 8696.Voogt, J.A. and Oke, T.R., 1998: Effects of urban surfacegeometry on remotely sensed surface temperature. Inter-national Journal of Remote Sensing 19, 895920.Voogt, J.A. and Oke, T.R., 2003: Thermal remote sensingof urban areas. Remote Sensing of Environment 86, 370384.Weng, Q., 2009: Thermal infrared remote sensing for urbanclimate and environmental studies: Methods, applications,and trends. ISPRS Journal of Photogrammetry and RemoteSensing 64, 335344.Weng, Q. and Lu, D., 2008: A sub-pixel analysis of urban-ization effect on land surface temperature and its interplaywith impervious surface and vegetation coverage in India-napolis, United States. International Journal of AppliedEarth Observation and Geoinformation 10, 6883.Weng, Q., Lu, D. and Schubring, J., 2004: Estimation of landsurface temperaturevegetation abundance relationship forurban heat island studies. Remote Sensing of Environment89, 467483.Xiao, H. and Weng, Q., 2007: The impact of land use andland cover changes on land surface temperature in a karstX.L. Zhou and Y.-C. Wang: Dynamics of Land Surface Temperature 35 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographersarea of China. Journal of Environmental Management 85,245257.Xu, H., 2006: Modification of normalized difference waterindex (NDWI) to enhance open water features in remotelysensed imagery. International Journal of Remote Sensing27, 30253033.Xu, H., 2008: A new index for delineating built-up landfeatures in satellite imagery. International Journal ofRemote Sensing 29, 42694276.Yuan, F. and Bauer, M.E., 2007: Comparison of impervioussurface area and normalized difference vegetation indexas indicators of surface urban heat island effects inLandsat imagery. Remote Sensing of Environment 106,375386.Zha, Y., Gao, J. and Ni, S., 2003: Use of normalized differ-ence built-up index in automatically mapping urban areasfrom TM imagery. International Journal of RemoteSensing 24, 583594.36 Geographical Research February 2011 49(1):2336 2010 The AuthorsGeographical Research 2010 Institute of Australian Geographers

Recommended

View more >