studyforpredictinglandsurfacetemperature(lst)using...

Research ArticleStudy for Predicting Land Surface Temperature (LST) UsingLandsat Data A Comparison of Four Algorithms

Elhadi K Mustafa 1 Yungang Co12 Guoxiang Liu12 Mosbeh R Kaloop 3

Ashraf A Beshr3 Fawzi Zarzoura3 and Mohammed Sadek 4

1Department of Surveying and Geo-Informatics Faculty of Geosciences and Environmental EngineeringSouthwest Jiaotong University Chengdu 610031 Sichuan China2State-Province Joint Engineering Laboratory Spatial Information Technology for High-Speed Railway SafetySouthwest Jiaotong University Chengdu 610031 Sichuan China3Public Works and Civil Engineering Department Mansoura University Mansoura Egypt4School of Human Settlements and Civil Engineering Institute of Global Environmental ChangeDepartment of Earth amp Environmental Science Xirsquoan Jiaotong University Xirsquoan 710049 China

Correspondence should be addressed to Elhadi K Mustafa elhadik30yahoocom

Received 11 September 2019 Revised 30 January 2020 Accepted 6 February 2020 Published 31 March 2020

Academic Editor Dong Zhao

Copyright copy 2020 Elhadi K Mustafa et al is is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

e soft computing models used for predicting land surface temperature (LST) changes are very useful to evaluate andforecast the rapidly changing climate of the world In this study four soft computing techniques namely multivariateadaptive regression splines (MARS) wavelet neural network (WNN) adaptive neurofuzzy inference system (ANFIS) anddynamic evolving neurofuzzy inference system (DENFIS) are applied and compared to find the best model that can be usedto predict the LST changes of Beijing area e topographic change is considered in this study to accurately predict LSTfurthermore Landsat 45 TM and Landsat 8OLI_TIRS images for four years (1995 2004 2010 and 2015) are used to studythe LSTchanges of the research areae four models are assessed using statistical analysis coefficient of determination (R2)root mean square error (RMSE) and mean absolute error (MAE) in the training and testing stages and MARS is used toestimate the important variables that should be considered in the design models e results show that the LST for thestudied area increases by 028degCyear due to the urban changes in the study area In addition the topographic changes andpreviously recorded temperature changes have a significant influence on the LSTprediction of the study area Moreover theresults of the models show that the MARS ANFIS and DENFIS models can be used to predict the LSTof the study area eANFIS model showed the highest performances in the training (R2 099 RMSE 078degC MAE 055degC) and testing(R2 099 RMSE 036degC MAE 016degC) stages therefore the ANFIS model can be used to predict the LST changes in theBeijing areae predicted LSTshows that the change in climate and urban area will affect the LSTchanges of the Beijing areain the future

1 Introduction

Climate change is one of the most critical challenges that theworld faces Previous studies have reported that climatechange has a significant effect on the land surface temper-ature and its other parameters [1 2] e expansion of urbanareas is considered a significant factor for the change in landuse and land surface temperature [3 4] Previous studies

show that in China urban growth and climate changeincluding temperature change are serious problems resultingfrom economic and social development [5 6]erefore it isvital to investigate ways to predict change in land surfacetemperature (LST) change

LST is defined as the temperature felt when there is anexchange of long-wave radiation and turbulent heat fluxeswithin the surface-atmosphere interface [7ndash9] e LST is

HindawiAdvances in Civil EngineeringVolume 2020 Article ID 7363546 16 pageshttpsdoiorg10115520207363546

being increasingly utilized to evaluate the climate changes inurban zones Satellite images are used as sources of infor-mation for surface temperature trends and variability Liet al [8] summarized the basics and theoretical backgroundfor the models deployed for LSTdetermination from satelliteimages Avdan and Jovanovska (2016) presented somemethods for calculating LST from Landsat images and theyconcluded that the methods used were successful in re-trieving the LST Zhang et al [10] evaluated the changes inLST of the Ebinur lake between 1998 and 2011 by applyingLandsat image processing and found that Landsat imageprocessing is a good tool to estimate the relationship be-tween LST and land cover factors Vlassova et al [11]assessed three different techniques for Landsat image pro-cessing ey applied radiative transfer equation (RTE)inversion using the atmospheric correction parameterscalculator (ACPC) tool and two algorithms based on theapproximations of RTE namely the single-channel (SC)method and the mono-window (MW) method and foundthat the three methods can be utilized for the estimation ofLST from Landsat thermal images

Recently soft computing techniques are widely used forstudying environmental changes Valipou [12] examinedneural network (NN)methods for detecting drought and wetyears and he found that NN methods are suitable for es-timating the environmental changes Tran et al [13] applieda regression model to estimate the LST patterns in Hanoicity Vietnam and found that this method can be used inother cities that possess a similar urban growth Ahmed et al[14] used an NN model to simulate land cover changesbetween 2019 and 2029 in Dhaka Bangladesh and theirresults showed that the change in temperature would begreater than or equal to 30degC Ranjan et al [15] utilized NNin conjunction with geoinformatics technology to estimateLST and they found that this model could precisely predictLST Lee and Jung (2014) used an NN to approximate theLST in Korea and found that it can be applied to predict LSTchanges with high accuracy Kestens et al [16] utilized ageneral linear model based on Landsat images to model theLST change at Quebec Canada and found that the dataestimated from satellite images were useful for detecting theLST changes Yang et al [17] applied four interpolationmethods to predict the LST changes in England namelyinverse distance weighting (IDW) spline kriging andcokriging and recommended using kriging method whenapproximating LST from image processing Li et al [18]studied the effectiveness of image clearance on LST esti-mation and observed that the resolution of the Landsatimages is important for accurately estimating the LSTchanges

More recently integrated soft computing models werefound to be better to detect different environmental changes[19 20] Advanced regression and integrated NN modelshave shown good performance as prediction models fornonlinear performances changes of LST water level changesmonthly temperature changes and evaporation[11 19ndash21]For instance the wavelet neural network (WNN)was developed to forecast water level changes while adaptiveneurofuzzy inference system (ANFIS) is used to forecast

both water level changes and LST changes [21 22]eANFIS provides better performances for the dynamic be-haviors [23 24] Kisi et al [19] investigated different caseswith the ANFIS model to predict temperature changes inTurkey ey established that the ANFIS with grid partitionis the optimal model for change prediction In addition Kisiet al [20] developed a dynamic evolving neurofuzzy in-ference system (DENFIS) which is a fuzzy integrated modelto predict the evaporation of water surface and found thatthe performance of DENFIS is superior to that of ANFIS as aprediction model for water evaporation Wang et al [25 26]utilized the soft computing to model the solar radiation inChina and they found that the Multilayer Perceptron (MLP)and radial basis NN models are more accurate in detectingsolar radiation at different climatic zones in China fur-thermore the ANFIS and M5 model tree are found to bebetter in predicting solar radiation at some stations in ChinaMeanwhile the multivariate adaptive regression splines(MARS) technique is considered to be an effective tool forprediction and classification of input data for modelingespecially when the input data is limited [27] MARStechnique was found to be useful for land cover classificationand prediction of solar radiation [28 29]

Here the input variables affect the accuracy of pre-dictionmodels e input variables for the prediction of LSTcan be classified into urban changes and uses and landgeometry [14 15] e correlation between LST and nor-malized difference vegetation index (NDVI) has beendocumented in Weng [30] and Weng et al [31] Rinner andHussain [32] and Chen et al [33] showed a correlationbetween LST and the normalized difference built-up index(NDBI) e normalized difference water index (NDWI)and normalized difference bareness index (NDBaI) wereanalyzed for delineating the water content in vegetation andidentifying the bareness of soil respectively [33 34] Fenget al [35] found a strong correlation between LSTand NDBIfollowed by NDWI and NVDI in Suzhou city Furthermorethe urban index (UI) was found to have a high correlationwith surface temperature in Harare Zimbabwe [36] eintensity of LST is directly related to the rate of urbanizationland use patterns and building density [14] LST is related tothe patterns of land usecover changes for example thecomposition of the built-up area vegetation and waterbodies [33] Xiao et al [37] reported that the impervioussurface is positively correlated with LST in Beijing Chinaerefore the digital elevation should be considered in thisstudy Other studies utilized geodetic coordinates for de-signing the LST prediction models For instance Ranjanet al [15] used real-time LST measurements and geodeticcoordinates to estimate the LST

In this study WNN DENFIS ANFIS and MARSmodels are applied to accurately calculate the LST fromLandsat images in order to predict the temperature changesin Beijing city In addition the input variables (NDVINDWI NDBI UI and topographic changes) are evaluatedand assessed for LST prediction Here it should be men-tioned that the WNN DENFIS and MARS models have notbeen used so far to predict the LST changes from satellitedata in addition the ANFIS model is limited to the

2 Advances in Civil Engineering

prediction of LST changes [22] erefore the main aim ofthis paper is to investigate LSTmodeling using novel designof MARS ANFIS DENFIS andWNNmodels for predictingLST changes from visible Landsat near-infrared and TIRimagery obtained from Landsat e four models aredesigned and compared to obtain an optimal model fordetecting nonlinear LST changes of Beijing area

2 Study Area and Data Resources

21 Study Area e study area Beijing is the capital ofChina and it is one of the biggest cities in the world covering16 districts and two counties with a land area of 16410 km2e administrative region of Beijing is divided into threeparts the central city suburbs and outer suburbs ecentral city includes two districts Xicheng and Dongchenge suburbs aremade up of four districts Haidian ChaoyangFengtai and Shijingshan and the outer suburbs compriseeight districts and two counties namely Tongzhou ShunyiFangshan Daxing Changping Mentougou Huairou andPinggu districts and Miyun and Yanqing counties as pre-sented in Figure 1(a) Beijing has been identified as one of therapidly developing cities in China both economically and inarea e population distribution depends on the elevation ofthe land as presented in Figure 1(b) e increasing pop-ulation requires additional space which necessitates newresidential areas thereby encouraging rapid expansion of theregions urbanized us the rapid population growth andexpansion of built-up areas present a daunting challengewhile estimating the cultivate land LST

22 Data Resources In this research Landsat data is used asthe main source of data Landsat satellite images takenduring the dry seasons of 1995 2004 2010 and 2015 wereobtained from the US Geological Survey (USGS) website(Table 1) e years were considered as an indicator for thetime that the images were taken Path 123 and row 32 coverthe entire research area e map projection of the collectedsatellite images is Universal Transverse Mercator (UTM)and the images are of type Zone 50 N-Datum WorldGeodetic System (WGS) 84 with pixel size of 30m A total of32 ground truth points were directly observed using a high-resolution handheld GPS (Garmin GPSMAP 62s) in fieldsurvey conducted during 1995ndash2015 to make the accuracyassessment highly accurate e ancillary data used in thisresearch include land surface parameters derived from the90m grid spacing DEM data of the Shuttle Radar Topo-graphic Mission (SRTM) (Figure 1(b))

3 Methodology

is section describes the general procedures such as imageprocessing LSTestimation and prediction of LST which aredescribed in the flowchart shown in Figure 2

31 Land UseLand Cover Maps e acquired satellite im-ages (1995 2004 and 2015) were classified into seven landuse and land cover (LULC) classes within a five-classification

scheme urban (high and low residential areas) vegetation(high-density vegetation and low-density vegetation)bareland waterbodies and forestland A supervised imageclassification with the support vector machine (SVM)classification method was applied to classify the LULCclasses [38] In this method prior knowledge of the studyregion is used to select the training sites and the spectralinformation contained in individual pixels is utilized togenerate the LULC classes Finally the summed up image isrenamed to create the final form of land cover maps forvarious years (Figure 3) Satisfactory results were obtainedwith regard to classification accuracy and Kappa coefficientse lowest accuracy was obtained for the built-up areas in1995 which reached 881 this is expected because built-upareas have mixed pixels with trees home gardens and actualbuildings e overall accuracies were 9515 for 19959093 for 2004 and 9428 for 2015 (Table 2)

32 Retrieval of Land SurfaceTemperature fromLandsat-5TMand Landsat-8

321 Retrieval of Satellite Brightness Temperature (BT)

(1) Retrieval of Satellite Brightness Temperature (BT) from theLandsat-5TM Images Based on Chen et al [39] a two-stepprocess was applied to obtain the brightness temperaturefrom the Landsat-5TM images in this research In the firststep the digital numbers (DNs) of band 6 were transformedinto radiation luminance (RTM6) using the followingformula

RTM6 V

255Rmax minus Rmin( 1113857 + Rmin (1)

where V represents the DN of band 6 and

Rmax 1896 mWlowast cmminus 2lowast srminus 11113872 1113873 (2)

Rmin 01534 mWlowast cmminus 2lowast srminus 11113872 1113873 (3)

In the second step the radiation luminance was trans-formed into at-satellite brightness temperature (BT) inCelsius (degC) using the following equation

BT K1

ln K2 RTM6b( 1113857 + 1( 1113857( 1113857minus 27315 (4)

where K1 126056 K and K2 60766 (mWlowast cmminus2lowastsrminus1 μmminus1) which are prelaunch calibration constants underan assumption of unity emissivity and b represents theeffective spectral range when the sensorrsquos response is con-siderably higher than 50 b 1239 (μm) [29]

(2) Retrieval of Satellite Brightness Temperature (BT) fromthe Landsat-8 Images e first step is to transform the DN ofthe thermal infrared band into spectral radiance (Lλ) byusing the following equation obtained from the Landsatuserrsquos handbook [40]

Lλ MLQcal + AL (5)

Advances in Civil Engineering 3

where Lλ is the TOA spectral radiation in watts(m2lowast sterlowast μm) ML is the band specific multiplicativerescaling factor from the metadata AL is the band specificadditive rescaling factor and Qcal represents the symbolizedvalues of quantized and calibrated standard product pixels(D)

e second step is to convert the band radiance into BTin Celsius using the following conversion formula

BT K2

ln K1Lλ + 1( 1113857minus 27315 (6)

where BT is the satellite brightness temperature in Celsiusand K1 and K2 represent thermal conversion from themetadata

322 Surface Emissivity (ε) Retrieval e land surfaceemissivity (ε) values are obtained using equation (7) [41]

ε m middot PV + n (7)

In this study we used m 0004 and n 0986 PV is theproportion of vegetation extracted from equation (9)

PV NDVI minus NDVImin

NDVImax minus NDVImin1113890 1113891

2

(8)

NDVI NIR minus REDNIR + RED

1113876 1113877 (9)

where NDVI is the normalized difference vegetation in-dex NDVImin and NDVImax are the minimum andmaximum values of the NDVI respectively e emis-sivity-corrected LST values were then retrieved usingequation (10)

Brightness temperatures assume that the Earth is ablackbody which it is not and this can lead to errors insurface temperature In order to minimize these errorsemissivity correction is important and is done to retrieve theLST from Tb using equation (9) [42]

LST BT

1 + (λ(BT)lowast ln(ε)ρ) (10)

where LST is Celsius BT is the at-sensor brightness tem-perature in Celsius λ (115 μm) is the wavelength of the

117deg45prime0PrimeE117deg0prime0PrimeE116deg15prime0PrimeE115deg30prime0PrimeE

41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N39

deg0prime0Prime

N

41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N39

deg0prime0Prime

N


Ji

Yi

Chicheng

Luanping

Ba

Xinglong

Zhuolu

Baodi

Huailai

Fengning Manchu

Chongli

Jinghai

Laishui

Wuqing

Ninghe

Yutian

Dagang

Li

Longhua

Renqiu

Zhuo

WenanAnxin

Guan

Qing

Xuanhua

Xushui

Longfang

Zunhua

Tanggu

Sanhe

Xiong

Qing Yuan

Xiqing

Yongqing

Dingxing

Daicheng

Guyuan

Dongli

Xincheng

Chengde Xiagraven

Beichen

Mancheng

Xianghe

Zhangbei

Gaoyang

Hejian

Hangu

Fengrun

Wan

RongchengTianjin

Huanghua

Chengde Shigrave

Wangdu

Baoding

Dachang Hui

Fengnan

Chengde Xiagraven

BoyeDing

YingshouyingzikuangMiyun

Huairou

Yanqing

Fangshan

Beijing

Shunyi

Daxing

Pinggu

Mentougou

Changping

Tongzhou

Beijing

0 10 20 30 405Miles

Study area

District

N

(a)

Digital elevation model for Beijing

0 8 16 24 324Miles

High 2283

Low 5

117deg0prime0PrimeE116deg15prime0PrimeE115deg30prime0PrimeE


41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N

41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N

N

(b)

Figure 1 Beijing cityrsquos (a) location and (b) digital elevation model (DEM)

Table 1 Image information will be used in the study

Slno Sensor Date of

acquisitionPathrow

Resolution(m)

1 Landsat45TM April 1995 12332 30m2 Landsat45TM April 2004 12332 30m3 Landsat45TM April 2010 12332 30m4 Landsat8OLI_TIR April 2015 12332 30m


emitted radiance ρ hlowast cσ 1438lowast 10minus 2 mK σ is theStefanndashBoltzmann constant h is Planckrsquos constant c is thevelocity of light and ε is the land surface emissivity (LSE)

e calculation of LST for the month of April is presentedin Figure 4 and Table 3 for the years 1995 2004 2010 and2015

Imageprocessing

Opticalbands

Atmosphericcorrection

Correctedsurface

reflectanceimage

Clipped tostudy area

Input Processing Intermediateoutput

Result

Select the best model to predict the LST

Compare the modelsresult for both

training and testingphases

DENFIS

ANFIS

WNN

MARS

Land surfacetemperature

Brightnesstemperature

Radiance

Thermalbands

Surfaceemissivity

PV

NDVI

NIR and REDbands

Reprojection

LandsatTMOLImultispectral images

LSTestimation LST prediction

Figure 2 Flowchart of the downscaling algorithms used in this study

0 30 60 90 12015Miles

UrbanVegetationBareland

WaterbodiesForests

N

(a)

0 30 60 90 12015Miles


WaterbodiesForests

N

(b)

0 30 60 90 12015Miles


WaterbodiesForests

N

(c)

Figure 3 Land use pattern by year in Beijing (a) 1995 (b) 2004 and (c) 2015


33 Design and Evaluation of Prediction Models In thissection four methods are applied and compared for pre-dicting temperature changes e regression and integratedNN models are used to detect the LST e MARS WNNANFIS and DENFIS models are applied and investigatede summary of the four methods is presented below In thisstudy the LST for 2010 and 2015 are utilized to design andevaluate the models To predict the LST for 2010 and 2015the data for the past two years and topographic changes forthe estimated image nodes are used

331 Multivariate Adaptive Regression Splines (MARS)Algorithm e MARS model is a nonlinear and nonpara-metric regression model that can be used for predictingnonlinear LST measurements [28 29] A spline is a utilitydefined piecewise by polynomials and it is applied to a classof functions used in input-output data interpolation [28] Acommon spline function is the cubic spline or knote [43]e knote numbers and placement are fixed for regressionsplines e forward and backward processes are applied toestablish the knote Basic functions (BFs) are used to searchfor knotes A linear combination of BFs for theMARSmodelis developed as a set of functions representing the rela-tionship between the prediction (y) and target variables forthe LST as follows

y β0 + 1113944

M

m1βmBFm (11)

whereβm m 0 1 M are unknown coefficients that canbe estimated by the least square method andBFm are m-thbasis functions that denote reflected pairs [28] ese arelinear BFs of the forms (xminus t)+ and (tminus x)minus with t being theknote [28]

erefore theMARS process is performed in three stepsFirst the forward algorithm chooses all possible funda-mental functions and their related knots Second thebackward algorithm gets rid of all basic functions in order toproduce the best combinations of existing knots and finallythe smoothing operation is performed to obtain continuouspartition borders [43]

Quiros et al [29] summarized the classification methodsby MARS e first one relates to pairwise classification inwhich the output can be coded as 0 or 1 thereby treating theclassification as a regression e second possibility involvesmore than two classes with the classification serving as ahybrid of MARS called Poly MARS [29] is study adoptsthe first technique

332 Wavelet Neural Network (WNN) Algorithm eWNN was introduced by Zhang and Benveniste [44] ealgorithm of WNN connects the NN and wavelet decom-position with a highly nonlinear wavelet function [21] eLST prediction (yk) based on WNN can be obtained asfollows

yk 1113944m

i1Ckminusiψ ai xkminusi minus bi( 1113857( 1113857 + w (12)

where Ckminusi are coefficient variables ai are dilation variablesbi are translation variables andψ is a wavelet function

e WNN contains three layers namely input waveletor hidden as well as output respectively In this study weuse five input parameters 25 hidden layers withMexican hatwavelet function and a single output layer for predictingLSTvaluese choice of the wavelet function is based on theapplication and many wavelet functions can be applied inthe WNN prediction models In this study the Mexican hatwavelet function is used to perform nonlinear LSTmeasurements

333 Adaptive Neurofuzzy Interference System (ANFIS)Algorithm In the ANFIS the FIS parameters are calculatedusing NN back-propagation algorithmse system dependson FIS and has fuzzy IF-THEN rules e ANFIS modelintroduced by Jang [45] is capable of approximating any realcontinuous function on a compact set to any level of pre-cision In other words ANFIS is defined as a system that usesfuzzy rules and the associated fuzzy inference method forlearning and rule optimization purposes [46] In this studywe apply the combined learning rule which uses conven-tional back-propagation and least square methods to esti-mate the parameters of the model accurately [47]emodelis made up of five layers e process of each layer is pre-sented in [20] Kisis et al [19] summarized that the input andoutput neurons can be fuzzified by a fuzzy quantizationapproach

334 Dynamic Evolving Neurofuzzy Inference System(DENFIS) Algorithm e DENFIS model is also an inte-grated model like the ANFIS model [19 48] In DANFISthe FIS parameters are calculated using NN back-prop-agation algorithms by applying evolving fuzzy neuralnetworks (EFuNN) in the model [20]erefore the majordifference between ANFIS and DENFIS is the applicationof EFuNN [20] e EFuNN employs Mamdani-type fuzzy

Table 2 Accuracy of multitemporal LULC classifications maps

Land cover1995 2004 2015

User () Producer () User () Producer () User () Producer ()Urban 901 881 891 901 911 901Vegetation 933 932 915 912 943 936Barren land 933 905 923 935 942 941Waterbodies 981 911 911 969 961 952Forest 952 945 934 905 912 932Overall 9515 9093 9428Kappa 093 089 092


0 10 20 30 405Miles

High 27

Low 3

N

Distribution of land surface temperature of Beijing

(a)

0 10 20 30 405Miles

High 27

Low 3

N


(b)

0 10 20 30 405Miles

High 27

Low 3

N


(c)

0 10 20 30 405Miles

High 27

Low 3

N


(d)

Figure 4 LST 1995 (a) 2004 (b) 2010 (c) and 2015 (d) distributions

Table 3 Statistical analysis for the extraction LST (degC)

Year Mean Max Min STD1995 15887 24564 3118 plusmn42152004 15968 24980 3631 plusmn42982010 19702 30569 0457 plusmn46592015 21190 27855 8778 plusmn4053


rules [48]erefore the input variables are directlyassigned to rules through fuzzy membership functionswithout weights layer [48] In addition the evolvingconnectionist system is a distance-based fast one-passalgorithm that performs dynamic approximation of thenumber of clusters in dataset and allocates the position ofcluster centers in the input data space [20] In thisclustering method the maximum distance between thecluster center and data point must be less than a user-defined threshold value In this model the describedmodel architecture is carried out in an evolving approachbased on cross-linking to Takagi-Sugeno fuzzy systems[19 20]

335 Input Variables and Performance Evaluation of theModels In this study the NDVI NDWI NDBI UI andtopographic changes of the selected points are used andassessed e MARS method is used to classify the inputvariables for predicting the LST of Beijing city

e performance analysis of the four applied models isconducted based on three statistical analyses methodsnamely coefficient of determination (R2) root mean squareerror (RMSE) and mean absolute error (MAE) given byequations (13)ndash(15)

R2

1 minus1113936

Ni1 yi minus fi( 1113857

2

1113936Ni1 yi minus y( 1113857

2 (13)

RMSE

051113944N

i1yi minus fi( 1113857

2

11139741113972

(14)

MAE 1N

1113944

N

i1yi minus fi

11138681113868111386811138681113868111386811138681113868 (15)

where y and f are the actual and predicted LST values re-spectively y is the mean value of the actual values and N isthe number of measurements

4 Results and Discussions

41 Observed LULC and Transitions from 1995 to 2015e spatial patterns of LULC in the study area for 1995 2004and 2015 are illustrated in Figure 3 e forestlands andvegetation lands dominated the entire area however theydiminished with the continuous increment of urban areasIn 1995 urban region vegetation barren land and forestwere the predominant land-use types with rates of 14351514 1318 and 5608 respectively It was found thatthere were dense built-up surfaces in the urban area whichwere surrounded by farmlands andmountains In 2004 highresidential area vegetation barren land and forests were thepredominant land-use types with rates of 1813 13741579 and 5137 respectively In 2015 built-up surfaceregions were the dominant land-use type with a percentageof 2302 followed by barren land and vegetation (1430and 1007 respectively) e statistical postclassificationcomparison results are presented in Tables 4ndash6 is

comparison shows a noteworthy increment in the region ofbuilt-up surfaces there was a significant change between1995 and 2004 and dense vegetation zones changed intourban regions or bareland On analyzing the changes in landuse and land cover between 1995 and 2015 it was observedthat there was an increase in the built-up region by 868and a decrease in vegetation by 507 Here the funda-mental change is that the vegetation zones and barelandschanged into developed areas

42 LST in 1995 2004 2010 and 2015 e LST which arepresented in Figure 5 were extracted using ArcGIS103According to Figures 1(b) and 4 the geometry of the cityaffects the LST and the results cited in Xiao et al [37]confirm this fact Mushore et al [36] evaluated the signif-icance of the predicted changes in temperature over a studyarea using selection points Herein we tested 250 randomdistribution points which were chosen to accurately reflectthe distribution of the LSTof the study area as presented inFigures 4 and 6 e geodetic coordinates (latitude longi-tude and height) and LST of the research area for each yearare extracted and presented in Figures 5 and 6 e existingurban areas are selected from points one to 180 the watersurface and areas near water are selected from points 180 to200 and the forest and farm lands which are suitable forurban activities are selected from points 200 to 250

Table 3 lists the statistical parameters (mean maximumminimum and standard deviation (STD)) of the LST for thefour years of study From Table 3 it is observed that therewere LST changes from 1995 to 2015 in the Beijing area Inaddition it is shown that the standard deviations for all casesare close and the difference is small Moreover the tem-perature changes for the years 1995 2004 2010 and 2015were 2145 2135 3011 and 1901degC respectively e re-sults show that there was significant temperature change inthe research area and that the topography of the study areahas a considerable effect on the forecast of temperaturechanges In addition the measurement values show that theLST increased rapidly with a slope of change of 028degCyearbased on the changes in mean value from 1995 to 2015Meanwhile from Figure 1 it can be observed that the to-pographic change is high erefore the geodetic coordi-nates (latitude (Lat) longitude (Long) and orthometricheight (H)) are used as input factors in this study

43 Predictionof LandSurfaceTemperature Firstly the inputvariables should be evaluated and assessed e MARS is abest classification algorithm for the prediction models[27 29] According to the previous studies the NDVI NDWINDBI UI and topographic changes are the main factors thataffect the LSTchanges From the literature it can be seen thatthe influence of these variables depends on the case underconsideration Xiao et al [37] showed a high correlationbetween LSTand the change in land geometry of Beijing cityHere NDVI NDWI NDBI UI topographic changes andprevious temperatures records are evaluated and assessedusing MARS classification us the NDVI NDWI NDBIand UI of the study area for 2010 are calculated and used to


model the LST for 2010 In addition the LST for 1995 and2005 and the geodetic coordinates are used for the selectionpoints e important factor (IF) is calculated from 0 to 100for each variable Figure 7 shows the IF of the variables FromFigure 7 it can be seen that the impact of topographic changesis high and this confirms the conclusion made by Xiao et al[37] for Beijing area In addition the influence of previoustemperature records is greater than 70 with regard to theprediction of LST It means that the previous LSTrecords can

be used to estimate the future LST Table 7 presents thecoefficient of determination values of different input variablesused to predict the LSTfor 2010 using theMARSmodel FromTable 7 it can be seen that model 4 is suitable to use in ourcase Secondly the selected input parameters are used todesign the models as presented below

431 Training and Design Stage In this study the selecteddata are utilized to study the performance of MARS WNNANFIS and DENFIS in LST prediction here the modelsare built using Matlab software e temperature at theareas in the vicinity of Beijing increased during the periodfrom 1995 to 2015 as shown in Figure 4 and Table 3Moreover from Figure 4 it can be observed that the urbanarea increased by approximately 20 during the periodfrom 1995 to 2015 e distribution of LST changesdepending on land use and geography In addition thetemperatures of the urban areas are higher than those ofother land-use types erefore the prediction models canbe used for detecting future LSTchanges e novel modelsdesigned in this research are applied to estimate thetemperature change e models are obtained using thegeodetic coordinates of several selected points and previousLSTchangese data for 2010 is deployed as the target LSTfor the training stage and the LST for 2015 is utilized as thetarget for the testing stage

(1) MARS Model To design the MARS model five parametersare used as input latitude longitude orthometric height the

Table 4 A real change of LULC categories in different time spans between 1995 and 2004

1995 (sq km) 2004 (sq km) ∆ (sq km) 1995 () 2004 () ∆ ()Urban 23498 29701 6203 1435 1813 379Vegetation 24798 22509 minus2289 1514 1374 minus140Barren land 21636 25862 4225 1318 1579 258Water 1997 1578 minus419 122 096 minus026Forest 91853 84131 minus7721 5608 5137 minus471


2004 (sq km) 2015 (sq km) ∆ (sq km) 2004 () 2015 () ∆ ()Urban 29701 37708 8007 1813 2302 489Vegetation 22509 16495 minus6014 1374 1007 minus367Barren land 25862 23426 minus2435 1579 1430 minus149Water 1578 2012 434 096 123 026Forest 84131 84140 09 5137 5137 001


1995 (sq km) 2015 (sq km) ∆ (sq km) 1995 () 2015 () ∆ ()Urban 23498 37708 14210 1435 2302 868Vegetation 24798 16495 minus8303 1514 1007 minus507Barren land 21636 23426 1790 1321 1430 109Water 1997 2012 15 122 123 001Forest 91853 84140 minus7713 5608 5137 minus471

0 50 100 150 200 250Sample no

0

5

10

15

20

25

30

35

LST

(degC)

19952004

20102015

Figure 5 LST for the month of April at the specified four years


LST for the period from 1995 to 2004 and the LST for 2010(which is used as the target value) e number of BFs can becalculated using generalized cross-validation (GCV) [28]Figure 8(a) shows the GCV for 199 BFs From this figure it isobserved that the minimum GCV is observed at 151 BFsHowever the model designed in this instance used a constantparameter 07414 and 151 BFs e performance of the

designedmodel is demonstrated in Figure 8(b) In addition thestatistical analysis for the prediction model is carried out andR2 for the model is found to be 097 while the RMSE andMAEare 078 and 055degC respectively From the statistical analysis itis found that the correlation between LSTmeasurements andpredictions is high and the model error is small erefore themodel can be used for detecting LST changes over the selectedpoints of the study area

(2) WNN Model For the WNN design many wavelet net-work models have tried to optimize the structure of thewavelet network for the training stage It was found that theoptimummodel that can be applied in this study is presentedin [21] is structure enables the use of statistical analysisR2 RMSE and MAE for WNN are 067 257degC and 186degCrespectivelyerefore the input year has five values selectedfor the geodetic coordinates and previous observations ofLST in addition the hidden layer of 25 wavelet neurons isemployed to forecast the future values of the LST mea-surements Figure 9 illustrates the measurements and pre-dicted LST for year 2010 From Figure 9 and the statisticalanalysis it observed that the performance of the WNNmodel is lower than that of the MARS model in the trainingstage

115deg30prime0PrimeE 116deg15prime0PrimeE

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE

115deg30prime0PrimeE 116deg15prime0PrimeE 117deg0prime0PrimeE

N

0 55 19 285 38479Miles

Figure 6 Distribution of selection points for predicting LST

0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004

Figure 7 e important factor of input variables


(3) ANFIS Model For the ANFIS model the number andtype of membership functions (MFs) for the model are themain factors that determine the accuracy of the model Eightdifferent MFs can be applied in the ANFIS model e eightMFs are evaluated with two MFs and ten epochs to reducethe CPU time the RMSE for eight MFs are also calculatede RMSE for the Gaussian trapezoidal triangular d sig-moid p sigmoid pi-shaped two Gaussian and generalized

bell are 171 164 238 165 152 161 145 and 162degCrespectively However the two-Gaussian function isdeployed to predict the LST In addition two and three MFsare assessed to limit CPU and memory usage e RMSE fortwo and three MFs are 145 and 142degC respectively In thiswork three MFs are selected to predict the LST Finally theANFIS model based on three MFs and two Gaussianfunctions with 50 epochs was applied in this study e

25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)

Figure 8 MARS model performance (a) GCM (b) measurement and prediction of LST-2010

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

Figure 9 Measurements and WNN prediction for training stage of LST

Table 7 MARS models evaluation for different inputs

Model Variables R2

1 Lat Long H NDVI NDWI NDBI UI LST_95 LST_2004 0612 Lat Long H NDVI NDWI LST_95 LST_2004 0743 Lat Long H NDVI LST_95 LST_2004 0794 Lat Long H LST_95 LST_2004 097


designed model is shown in Figure 10(a) e performanceof the ANFIS model is evaluated using equations (13) to (15)R2 RMSE andMAE for the designed model are 099 046degCand 026degC respectively Figure 10(b) shows the measure-ments and prediction values for the training data of LSTFrom Figure 10(b) it is seen that there is no loses of in-formation during the LST measurements Also the corre-lation between the measurements and the predicted LST ishigh and minimum RMSE and MAE are observed Fromthese results it can be concluded that the ANFIS model isbetter than MARS and WNN for predicting the LST fromtraining data and can therefore be used to predict the BeijingLST

(4) DENFISModel In the training stage the major differencebetween ANFIS and DENFIS is the application of EFuNN as

mentioned previously Moreover the maximum distancebetween the cluster center and data point must be less than auser-defined threshold value [20] In this study the distancethreshold is 01 erefore the EFuNN is applied to estimatethe next epochs Figure 11(a) illustrates the mean squareerror (MSE) for the training epochs of EFuNN From thisfigure it can be observed that 16 epochs are required toobtain the best performance for the DENFIS model Oth-erwise the same parameters that were used for the MFs ofthe ANFIS model are used in DENFIS model that is threeMFS and a Gaussian function e performance of thedesigned DENFIS design model is shown in Figure 11(b)e statistical evaluation of the model error was conductedand the values of R2 RMSE and MAE were found to be 091142degC and 090degC respectively Furthermore Akaikersquos in-formation criterion (AIC) [49] of the models was calculated

ANDORNOT

Logical operations

Input OutputOutputmfInputmf Rule

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)

Figure 10 ANFIS model design and results (a) ANFIS model diagram (b) LST measurements and ANFIS prediction results

0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)

Figure 11 (a) e epochs model performance and (b) DENFIS model performance for the training data


and found to be minus1172 and 423 for ANFIS and WNNmodels respectively e results for the training stage showthat the ANFIS model is the optimal model for predictingthe LST values while the WNN is the worst model

432 Testing and Comparison Stage e performance ofthe four models was also evaluated in the testing stageFigure 12 and Table 8 show a scatter plot for the observationsand prediction values of LSTand the statistical analysis of themodel errors e linear fitting is also shown in the figureFrom Figure 12 and Table 8 it is can be seen that thecorrelation between the observation and prediction is high

for the ANFIS and DENFIS models Moreover the worstmodel during the testing stage is the WNN model eminimum model error is observed with the ANFIS modeland its RMSE and MAE are 036 and 016degC respectivelye slope of linear fitting is found to be 098 093 087 and070 for the ANFIS DENFIS MARS and WNN modelsrespectively It means that the ANFIS model optimallydetects the LST in the testing stage

Based on the performances in the training and testingstages it is safe to argue that the ANFIS model is the optimalmodel to predict the LST for the Beijing area e ANFISmodel is applied to detect the LST for the years 2025 and2035 and the results are presented in Figure 13(a) In

Table 8 Testing stage performance analysis

Models MARS WNN ANFIS DENFISR 2 087 053 099 093RMSE (degC) 143 278 036 105MAE (degC) 111 196 016 069

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)

282624222018161412108Measured LST (degC)

y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 098lowastx + 044

(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)

Figure 12 Testing performance for (a) MARS (b) WNN (c) ANFIS and (d) DENFIS


addition the linear fitting is calculated and presented for themean values of the observations and prediction years asdemonstrated in Figure 13(b) e prediction values show thatthe LSTwill increase rapidly and the slope of change is 033degCyear It is also observed that R2 for the linear fitting is 094 andthe prediction trend is similar to the LSTmeasurements for theyears 2010 and 2015 In addition it is seen that the predictionvalues for the LSTof water forest and farm areas do not showany significant changes compared to the urban areas It meansthat the changes in climate and urban area will affect the LSTchanges of the Beijing area in the future

5 Conclusions

Although the land surface temperature (LST) is widelyevaluated in global temperature variation as a result ofclimate change the application of integrated soft computingmodels is still limited to use for detecting LST is studyaimed to design an integrated soft computing model todetect the LST of Beijing area Four models were developedand compared namely MARSWNN ANFIS and DENFISIn addition the MARS method was used to classify the inputvariables Landsat images and different software were uti-lized to mine the LST changes from April 1995 to 2015 Inaddition a statistical analysis was conducted to assess theperformance of the developed models From the results thefollowing can be concluded

First the comparison of temperature values for thestudied periods through imagersquos analyses shows that therehas been a significant change in temperature in Beijing citythe temperature is increased by 028degCyear In addition thevegetation cover decreased due to the population explosionand economic activities In addition the classification ofNDVI NDWI NDBI UI and geographic changes of Beijingand previous temperature records show that the surface andprevious LST changes have a significant impact on the LSTprediction of the study area

Second the evaluation of the prediction models showsthat the integrated fuzzy models ANFIS and DENFIS arethe optimum models for detecting the LST changes in theBeijing area In addition it is shown that the MARS modelexhibits good performance with limited input parameterse WNN model showed the weakest performance inpredicting the LST changes e performance of the ANFISmodel for the training and testing stages is satisfactorybecause it has low minimum errors the RMSE for thetraining and testing stages were 046 and 036degC respectively

Finally the predicted LST values show that the changesin climate and urban area will affect the LST changes of theBeijing area in the future Moreover this approach could beused to evaluate and predict the LST of other areas fromsatellite images

Data Availability

Landsat data were download from EarthExplorer httpsearthexplorerusgsgov e ancillary data included landsurface parameters derived from the Shuttle Radar Topo-graphic Mission (SRTM) 90m grid spacing DEM data

Conflicts of Interest

e authors declare no conflicts of interest

Authorsrsquo Contributions

All the authors have contributed equally to this work

Acknowledgments

is work was supported by the National Key Research andDevelopment Program of China under Grant no2016YFC0803105 and the National Natural Science Foun-dation of China under Grant no 41771451

LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)

Figure 13 (a) 2025 and 2035 prediction LST using ANFIS model (b) linear fitting of mean values for the LS


References

[1] P Brohan J J Kennedy I Harris S F Tett and P D JonesldquoUncertainty estimates in regional and global observedtemperature changes a new data set from 1850rdquo Journal ofGeophysical Research Atmospheres vol 111 no D12 2006

[2] J Hansen R Ruedy M Sato and K Lo ldquoGlobal surfacetemperature changerdquo Reviews of Geophysics vol 48 no 42010

[3] D Argueso J P Evans A J Pitman and A Di Luca ldquoEffectsof city expansion on heat stress under climate change con-ditionsrdquo PLoS One vol 10 no 2 Article ID e0117066 2015

[4] I R Hegazy and M R Kaloop ldquoMonitoring urban growthand land use change detection with GIS and remote sensingtechniques in Daqahlia Governorate Egyptrdquo InternationalJournal of Sustainable Built Environment vol 4 no 1pp 117ndash124 2015

[5] Y Tang S Xi X Chen and Y Lian ldquoQuantification ofmultiple climate change and human activity impact factors onflood regimes in the Pearl River Delta of Chinardquo Advances inMeteorology vol 2016 Article ID 3928920 11 pages 2016

[6] L Zhou R E Dickinson Y Tian et al ldquoEvidence for asignificant urbanization effect on climate in Chinardquo Pro-ceedings of the National Academy of Sciences vol 101 no 26pp 9540ndash9544 2004

[7] C J Myneni L Chapman J E ornes and C BakerldquoRemote sensing land surface temperature for meteorologyand climatology a reviewrdquo Meteorological Applicationsvol 18 no 3 pp 296ndash306 2011

[8] Z-L Li B-H Tang H Wu et al ldquoSatellite-derived landsurface temperature current status and perspectivesrdquo RemoteSensing of Environment vol 131 pp 14ndash37 2013

[9] U Sobrino and G Jovanovska ldquoAlgorithm for automatedmapping of land surface temperature using LANDSAT 8satellite datardquo Journal of Sensors vol 2016 Article ID1480307 8 pages 2016

[10] F Zhang H Kung V C Johnson B I LaGrone and J WangldquoChange detection of land surface temperature (LST) andsome related parameters using Landsat image a case study ofthe Ebinur lake watershed Xinjiang Chinardquo Wetlandsvol 38 no 1 pp 65ndash80 2018

[11] L Vlassova F Perez-Cabello H Nieto P Martın D Riantildeoand J de la Riva ldquoAssessment of methods for land surfacetemperature retrieval from Landsat-5 TM images applicableto multiscale tree-grass ecosystemmodelingrdquo Remote Sensingvol 6 no 5 pp 4345ndash4368 2014

[12] M Valipour ldquoOptimization of neural networks for precipi-tation analysis in a humid region to detect drought and wetyear alarmsrdquo Meteorological Applications vol 23 no 1pp 91ndash100 2016

[13] D X Tran F Pla P Latorre-Carmona S W MyintM Caetano and H V Kieu ldquoCharacterizing the relationshipbetween land use land cover change and land surface tem-peraturerdquo ISPRS Journal of Photogrammetry and RemoteSensing vol 124 pp 119ndash132 2017

[14] B Ahmed M Kamruzzaman X Zhu M Rahman andK Choi ldquoSimulating land cover changes and their impacts onland surface temperature in Dhaka Bangladeshrdquo RemoteSensing vol 5 no 11 pp 5969ndash5998 2013

[15] A Ranjan A Anand P Kumar S Verma and L MurmuldquoPrediction of land surface temperature using artificial neuralnetwork in conjunction with geoinformatics technologywithin Sun City Jodhpur (Rajasthan) Indiardquo Asian Journal ofGeoinformatics vol 17 no 3 pp 14ndash23 2017

[16] Y Kestens A Brand M Fournier et al ldquoModelling thevariation of land surface temperature as determinant of risk ofheat-related health eventsrdquo International Journal of HealthGeographics vol 10 no 1 p 7 2011

[17] J Yang Y Wang and P August ldquoEstimation of land surfacetemperature using spatial interpolation and satellite-derivedsurface emissivityrdquo Journal of Environmental Informaticsvol 4 no 1 pp 37ndash44 2004

[18] F Li T J Jackson W P Kustas et al ldquoDeriving land surfacetemperature from Landsat 5 and 7 during SMEX02SMA-CEXrdquo Remote Sensing of Environment vol 92 no 4pp 521ndash534 2004

[19] O Kisi V Demir and S Kim ldquoEstimation of long-termmonthly temperatures by three different adaptive neuro-fuzzyapproaches using geographical inputsrdquo Journal of Irrigationand Drainage Engineering vol 143 no 12 Article ID04017052 2017

[20] O Kisi I Mansouri and J W Hu ldquoA new method forevaporation modeling dynamic evolving neural-fuzzy in-ference systemrdquo Advances in Meteorology vol 2017 ArticleID 5356324 9 pages 2017

[21] M El-Diasty and S Al-Harbi ldquoDevelopment of waveletnetwork model for accurate water levels prediction withmeteorological effectsrdquo Applied Ocean Research vol 53pp 228ndash235 2015

[22] K Bisht and S Dodamani Estimation and Modelling of LandSurface Temperature Using Landsat 7 ETM+ Images and FuzzySystem Techniques American Geophysical Union Wash-ington DC USA 2016

[23] Y Liu W Zhang and Y Zhang ldquoDynamic neuro-fuzzy-based human intelligence modeling and control in GTAWrdquoIEEE Transactions on Automation Science and Engineeringvol 12 no 1 pp 324ndash335 2015

[24] Y Liu and Y Zhang ldquoIterative local ANFIS-based humanwelder intelligence modeling and control in pipe GTAWprocess a data-driven approachrdquo IEEEASME Transactionson Mechatronics vol 20 no 3 pp 1079ndash1088 2015

[25] L Wang O Kisi M Zounemat-Kermani G A SalazarZ Zhu and W Gong ldquoSolar radiation prediction usingdifferent techniques model evaluation and comparisonrdquoRenewable and Sustainable Energy Reviews vol 61 pp 384ndash397 2016a

[26] L Wang O Kisi M Zounemat-Kermani et al ldquoPrediction ofsolar radiation in China using different adaptive neuro-fuzzymethods and M5 model treerdquo International Journal of Cli-matology vol 37 no 3 pp 1141ndash1155 2016b

[27] V N Sharda S O Prasher R M Patel P R Ojasvi andC Prakash ldquoPerformance of multivariate adaptive regressionsplines (MARS) in predicting runoff in mid-Himalayan mi-cro-watersheds with limited datardquo Hydrological SciencesJournal vol 53 no 6 pp 1165ndash1175 2008

[28] B Keshtegar C Mert and O Kisi ldquoComparison of fourheuristic regression techniques in solar radiation modelingkriging method vs RSM MARS and M5 model treerdquo Re-newable and Sustainable Energy Reviews vol 81 pp 330ndash3412018

[29] E Quiros A Felicısimo and A Cuartero ldquoTesting multi-variate adaptive regression splines (MARS) as a method ofland cover classification of TERRA-ASTER satellite imagesrdquoSensors vol 9 no 11 pp 9011ndash9028 2009

[30] Q Weng ldquoA remote sensing GIS evaluation of urban ex-pansion and its impact on surface temperature in the ZhujiangDelta Chinardquo International Journal of Remote Sensingvol 22 no 10 pp 1999ndash2014 2001


[31] Q Weng D Lu and B Liang ldquoUrban surface biophysicaldescriptors and land surface temperature variationsrdquo Pho-togrammetric Engineering amp Remote Sensing vol 72 no 11pp 1275ndash1286 2006

[32] C Rinner and M Hussain ldquoTorontorsquos urban heat island-exploring the relationship between land use and surfacetemperaturerdquo Remote Sensing vol 3 no 6 pp 1251ndash12652011

[33] X-L Chen H-M Zhao P-X Li and Z-Y Yin ldquoRemotesensing image-based analysis of the relationship betweenurban heat island and land usecover changesrdquo RemoteSensing of Environment vol 104 no 2 pp 133ndash146 2006

[34] H M Zhao and X L Chen ldquoUse of normalized differencebareness index in quickly mapping bare areas from TMETM+rdquo in Proceedings of the 2005 IEEE International Geo-science and Remote Sensing Symposium 2005 IGARSSrsquo05Seoul South Korea pp 1666ndash1668 July 2005

[35] Y Feng C Gao X Tong S Chen Z Lei and JWang ldquoSpatialpatterns of land surface temperature and their influencingfactors a case study in Suzhou Chinardquo Remote Sensingvol 11 no 2 p 182 2019

[36] T Mushore J Odindi T Dube and O Mutanga ldquoPredictionof future urban surface temperatures using medium resolu-tion satellite data in Harare Metropolitan City ZimbabwerdquoBuilding and Environment vol 122 p 397e410 2017

[37] R-B Xiao Z-Y Ouyang H Zheng W-F Li E W Schienkeand X-K Wang ldquoSpatial pattern of impervious surfaces andtheir impacts on land surface temperature in Beijing ChinardquoJournal of Environmental Sciences vol 19 no 2 pp 250ndash2562007

[38] J R Jensen Remote Sensing of the Environment An EarthResource Perspective Pearson Education India BengaluruIndia 2009

[39] Y-H Chen J Wang and X-B Li ldquoA study on urban thermalfield in summer based on satellite remote sensingrdquo RemoteSensing for Land and Resources vol 4 no 1 2002

[40] USGS Landsat 8 Data Users Handbook USGS Reston VAUSA 2016

[41] J A Sobrino J C Jimenez-Muntildeoz and L Paolini ldquoLandsurface temperature retrieval from LANDSAT TM 5rdquo RemoteSensing of Environment vol 90 no 4 pp 434ndash440 2004

[42] Q Weng D Lu and J Schubring ldquoEstimation of land surfacetemperature-vegetation abundance relationship for urbanheat island studiesrdquo Remote Sensing of Environment vol 89no 4 pp 467ndash483 2004

[43] K M Turpin D R Lapen E G Gregorich et al ldquoUsingmultivariate adaptive regression splines (MARS) to identifyrelationships between soil and corn (Zea mays L) productionpropertiesrdquo Canadian Journal of Soil Science vol 85 no 5pp 625ndash636 2005

[44] Q Zhang and A Benveniste ldquoWavelet networksrdquo IEEETransactions on Neural Networks vol 3 no 6 pp 889ndash8981992

[45] J-S R Jang ldquoANFIS adaptive-network-based fuzzy inferencesystemrdquo IEEE Transactions on Systems Man and Cyberneticsvol 23 no 3 pp 665ndash685 1993

[46] H Sanikhani O Kisi E Maroufpoor and Z M YaseenldquoTemperature-based modeling of reference evapotranspira-tion using several artificial intelligence models application ofdifferent modeling scenariosrdquo Leoretical and Applied Cli-matology vol 135 no 1-2 pp 449ndash462 2019

[47] M R Kaloop M El-Diasty and J W Hu ldquoReal-time pre-diction of water level change using adaptive neuro-fuzzy

inference systemrdquo Geomatics Natural Hazards and Riskvol 8 no 2 pp 1320ndash1332 2017

[48] N Kasabov and Q Song ldquoDENFIS dynamic evolving neural-fuzzy inference system and its application for time seriespredictionrdquo IEEE Transactions on Fuzzy Systems vol 10no 2 2002

[49] R Moghadam M Izadbakhsh F Yosefvand andS Shabanlou ldquoOptimization of ANFIS network using fireflyalgorithm for simulating discharge coefficient of side orificesrdquoApplied Water Science vol 9 p 84 2019


being increasingly utilized to evaluate the climate changes inurban zones Satellite images are used as sources of infor-mation for surface temperature trends and variability Liet al [8] summarized the basics and theoretical backgroundfor the models deployed for LSTdetermination from satelliteimages Avdan and Jovanovska (2016) presented somemethods for calculating LST from Landsat images and theyconcluded that the methods used were successful in re-trieving the LST Zhang et al [10] evaluated the changes inLST of the Ebinur lake between 1998 and 2011 by applyingLandsat image processing and found that Landsat imageprocessing is a good tool to estimate the relationship be-tween LST and land cover factors Vlassova et al [11]assessed three different techniques for Landsat image pro-cessing ey applied radiative transfer equation (RTE)inversion using the atmospheric correction parameterscalculator (ACPC) tool and two algorithms based on theapproximations of RTE namely the single-channel (SC)method and the mono-window (MW) method and foundthat the three methods can be utilized for the estimation ofLST from Landsat thermal images

Recently soft computing techniques are widely used forstudying environmental changes Valipou [12] examinedneural network (NN)methods for detecting drought and wetyears and he found that NN methods are suitable for es-timating the environmental changes Tran et al [13] applieda regression model to estimate the LST patterns in Hanoicity Vietnam and found that this method can be used inother cities that possess a similar urban growth Ahmed et al[14] used an NN model to simulate land cover changesbetween 2019 and 2029 in Dhaka Bangladesh and theirresults showed that the change in temperature would begreater than or equal to 30degC Ranjan et al [15] utilized NNin conjunction with geoinformatics technology to estimateLST and they found that this model could precisely predictLST Lee and Jung (2014) used an NN to approximate theLST in Korea and found that it can be applied to predict LSTchanges with high accuracy Kestens et al [16] utilized ageneral linear model based on Landsat images to model theLST change at Quebec Canada and found that the dataestimated from satellite images were useful for detecting theLST changes Yang et al [17] applied four interpolationmethods to predict the LST changes in England namelyinverse distance weighting (IDW) spline kriging andcokriging and recommended using kriging method whenapproximating LST from image processing Li et al [18]studied the effectiveness of image clearance on LST esti-mation and observed that the resolution of the Landsatimages is important for accurately estimating the LSTchanges

More recently integrated soft computing models werefound to be better to detect different environmental changes[19 20] Advanced regression and integrated NN modelshave shown good performance as prediction models fornonlinear performances changes of LST water level changesmonthly temperature changes and evaporation[11 19ndash21]For instance the wavelet neural network (WNN)was developed to forecast water level changes while adaptiveneurofuzzy inference system (ANFIS) is used to forecast

both water level changes and LST changes [21 22]eANFIS provides better performances for the dynamic be-haviors [23 24] Kisi et al [19] investigated different caseswith the ANFIS model to predict temperature changes inTurkey ey established that the ANFIS with grid partitionis the optimal model for change prediction In addition Kisiet al [20] developed a dynamic evolving neurofuzzy in-ference system (DENFIS) which is a fuzzy integrated modelto predict the evaporation of water surface and found thatthe performance of DENFIS is superior to that of ANFIS as aprediction model for water evaporation Wang et al [25 26]utilized the soft computing to model the solar radiation inChina and they found that the Multilayer Perceptron (MLP)and radial basis NN models are more accurate in detectingsolar radiation at different climatic zones in China fur-thermore the ANFIS and M5 model tree are found to bebetter in predicting solar radiation at some stations in ChinaMeanwhile the multivariate adaptive regression splines(MARS) technique is considered to be an effective tool forprediction and classification of input data for modelingespecially when the input data is limited [27] MARStechnique was found to be useful for land cover classificationand prediction of solar radiation [28 29]

Here the input variables affect the accuracy of pre-dictionmodels e input variables for the prediction of LSTcan be classified into urban changes and uses and landgeometry [14 15] e correlation between LST and nor-malized difference vegetation index (NDVI) has beendocumented in Weng [30] and Weng et al [31] Rinner andHussain [32] and Chen et al [33] showed a correlationbetween LST and the normalized difference built-up index(NDBI) e normalized difference water index (NDWI)and normalized difference bareness index (NDBaI) wereanalyzed for delineating the water content in vegetation andidentifying the bareness of soil respectively [33 34] Fenget al [35] found a strong correlation between LSTand NDBIfollowed by NDWI and NVDI in Suzhou city Furthermorethe urban index (UI) was found to have a high correlationwith surface temperature in Harare Zimbabwe [36] eintensity of LST is directly related to the rate of urbanizationland use patterns and building density [14] LST is related tothe patterns of land usecover changes for example thecomposition of the built-up area vegetation and waterbodies [33] Xiao et al [37] reported that the impervioussurface is positively correlated with LST in Beijing Chinaerefore the digital elevation should be considered in thisstudy Other studies utilized geodetic coordinates for de-signing the LST prediction models For instance Ranjanet al [15] used real-time LST measurements and geodeticcoordinates to estimate the LST

In this study WNN DENFIS ANFIS and MARSmodels are applied to accurately calculate the LST fromLandsat images in order to predict the temperature changesin Beijing city In addition the input variables (NDVINDWI NDBI UI and topographic changes) are evaluatedand assessed for LST prediction Here it should be men-tioned that the WNN DENFIS and MARS models have notbeen used so far to predict the LST changes from satellitedata in addition the ANFIS model is limited to the






3 Methodology







RTM6 V






BT K1

ln K2 RTM6b( 1113857 + 1( 1113857( 1113857minus 27315 (4)



Lλ MLQcal + AL (5)




BT K2

ln K1Lλ + 1( 1113857minus 27315 (6)







2

(8)


1113876 1113877 (9)



LST BT




41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N39

deg0prime0Prime

N

41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N39

deg0prime0Prime

N


Ji

Yi

Chicheng

Luanping

Ba

Xinglong

Zhuolu

Baodi

Huailai

Fengning Manchu

Chongli

Jinghai

Laishui

Wuqing

Ninghe

Yutian

Dagang

Li

Longhua

Renqiu

Zhuo

WenanAnxin

Guan

Qing

Xuanhua

Xushui

Longfang

Zunhua

Tanggu

Sanhe

Xiong

Qing Yuan

Xiqing

Yongqing

Dingxing

Daicheng

Guyuan

Dongli

Xincheng

Chengde Xiagraven

Beichen

Mancheng

Xianghe

Zhangbei

Gaoyang

Hejian

Hangu

Fengrun

Wan

RongchengTianjin

Huanghua

Chengde Shigrave

Wangdu

Baoding

Dachang Hui

Fengnan

Chengde Xiagraven

BoyeDing


Huairou

Yanqing

Fangshan

Beijing

Shunyi

Daxing

Pinggu

Mentougou

Changping

Tongzhou

Beijing

0 10 20 30 405Miles

Study area

District

N

(a)


0 8 16 24 324Miles

High 2283

Low 5



41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N

41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N

N

(b)



Slno Sensor Date of

acquisitionPathrow

Resolution(m)





Imageprocessing

Opticalbands


Correctedsurface

reflectanceimage



Result




DENFIS

ANFIS

WNN

MARS



Radiance

Thermalbands

Surfaceemissivity

PV

NDVI

NIR and REDbands

Reprojection




0 30 60 90 12015Miles


WaterbodiesForests

N

(a)

0 30 60 90 12015Miles


WaterbodiesForests

N

(b)

0 30 60 90 12015Miles


WaterbodiesForests

N

(c)





y β0 + 1113944

M

m1βmBFm (11)





yk 1113944m







Land cover1995 2004 2015



0 10 20 30 405Miles

High 27

Low 3

N


(a)

0 10 20 30 405Miles

High 27

Low 3

N


(b)

0 10 20 30 405Miles

High 27

Low 3

N


(c)

0 10 20 30 405Miles

High 27

Low 3

N


(d)








R2

1 minus1113936


2

1113936Ni1 yi minus y( 1113857

2 (13)

RMSE

051113944N


2

11139741113972

(14)

MAE 1N

1113944

N

i1yi minus fi

11138681113868111386811138681113868111386811138681113868 (15)



















0 50 100 150 200 250Sample no

0

5

10

15

20

25

30

35

LST

(degC)

19952004

20102015







39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE


N

0 55 19 285 38479Miles


0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004





25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References

























































3 Methodology







RTM6 V






BT K1

ln K2 RTM6b( 1113857 + 1( 1113857( 1113857minus 27315 (4)



Lλ MLQcal + AL (5)




BT K2

ln K1Lλ + 1( 1113857minus 27315 (6)







2

(8)


1113876 1113877 (9)



LST BT




41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N39

deg0prime0Prime

N

41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N39

deg0prime0Prime

N


Ji

Yi

Chicheng

Luanping

Ba

Xinglong

Zhuolu

Baodi

Huailai

Fengning Manchu

Chongli

Jinghai

Laishui

Wuqing

Ninghe

Yutian

Dagang

Li

Longhua

Renqiu

Zhuo

WenanAnxin

Guan

Qing

Xuanhua

Xushui

Longfang

Zunhua

Tanggu

Sanhe

Xiong

Qing Yuan

Xiqing

Yongqing

Dingxing

Daicheng

Guyuan

Dongli

Xincheng

Chengde Xiagraven

Beichen

Mancheng

Xianghe

Zhangbei

Gaoyang

Hejian

Hangu

Fengrun

Wan

RongchengTianjin

Huanghua

Chengde Shigrave

Wangdu

Baoding

Dachang Hui

Fengnan

Chengde Xiagraven

BoyeDing


Huairou

Yanqing

Fangshan

Beijing

Shunyi

Daxing

Pinggu

Mentougou

Changping

Tongzhou

Beijing

0 10 20 30 405Miles

Study area

District

N

(a)


0 8 16 24 324Miles

High 2283

Low 5



41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N

41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N

N

(b)



Slno Sensor Date of

acquisitionPathrow

Resolution(m)





Imageprocessing

Opticalbands


Correctedsurface

reflectanceimage



Result




DENFIS

ANFIS

WNN

MARS



Radiance

Thermalbands

Surfaceemissivity

PV

NDVI

NIR and REDbands

Reprojection




0 30 60 90 12015Miles


WaterbodiesForests

N

(a)

0 30 60 90 12015Miles


WaterbodiesForests

N

(b)

0 30 60 90 12015Miles


WaterbodiesForests

N

(c)





y β0 + 1113944

M

m1βmBFm (11)





yk 1113944m







Land cover1995 2004 2015



0 10 20 30 405Miles

High 27

Low 3

N


(a)

0 10 20 30 405Miles

High 27

Low 3

N


(b)

0 10 20 30 405Miles

High 27

Low 3

N


(c)

0 10 20 30 405Miles

High 27

Low 3

N


(d)








R2

1 minus1113936


2

1113936Ni1 yi minus y( 1113857

2 (13)

RMSE

051113944N


2

11139741113972

(14)

MAE 1N

1113944

N

i1yi minus fi

11138681113868111386811138681113868111386811138681113868 (15)



















0 50 100 150 200 250Sample no

0

5

10

15

20

25

30

35

LST

(degC)

19952004

20102015







39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE


N

0 55 19 285 38479Miles


0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004





25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References























































BT K2

ln K1Lλ + 1( 1113857minus 27315 (6)







2

(8)


1113876 1113877 (9)



LST BT




41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N39

deg0prime0Prime

N

41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N39

deg0prime0Prime

N


Ji

Yi

Chicheng

Luanping

Ba

Xinglong

Zhuolu

Baodi

Huailai

Fengning Manchu

Chongli

Jinghai

Laishui

Wuqing

Ninghe

Yutian

Dagang

Li

Longhua

Renqiu

Zhuo

WenanAnxin

Guan

Qing

Xuanhua

Xushui

Longfang

Zunhua

Tanggu

Sanhe

Xiong

Qing Yuan

Xiqing

Yongqing

Dingxing

Daicheng

Guyuan

Dongli

Xincheng

Chengde Xiagraven

Beichen

Mancheng

Xianghe

Zhangbei

Gaoyang

Hejian

Hangu

Fengrun

Wan

RongchengTianjin

Huanghua

Chengde Shigrave

Wangdu

Baoding

Dachang Hui

Fengnan

Chengde Xiagraven

BoyeDing


Huairou

Yanqing

Fangshan

Beijing

Shunyi

Daxing

Pinggu

Mentougou

Changping

Tongzhou

Beijing

0 10 20 30 405Miles

Study area

District

N

(a)


0 8 16 24 324Miles

High 2283

Low 5



41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N

41deg1

5prime0Prime

N40

deg30prime

0PrimeN

39deg4

5prime0Prime

N

N

(b)



Slno Sensor Date of

acquisitionPathrow

Resolution(m)





Imageprocessing

Opticalbands


Correctedsurface

reflectanceimage



Result




DENFIS

ANFIS

WNN

MARS



Radiance

Thermalbands

Surfaceemissivity

PV

NDVI

NIR and REDbands

Reprojection




0 30 60 90 12015Miles


WaterbodiesForests

N

(a)

0 30 60 90 12015Miles


WaterbodiesForests

N

(b)

0 30 60 90 12015Miles


WaterbodiesForests

N

(c)





y β0 + 1113944

M

m1βmBFm (11)





yk 1113944m







Land cover1995 2004 2015



0 10 20 30 405Miles

High 27

Low 3

N


(a)

0 10 20 30 405Miles

High 27

Low 3

N


(b)

0 10 20 30 405Miles

High 27

Low 3

N


(c)

0 10 20 30 405Miles

High 27

Low 3

N


(d)








R2

1 minus1113936


2

1113936Ni1 yi minus y( 1113857

2 (13)

RMSE

051113944N


2

11139741113972

(14)

MAE 1N

1113944

N

i1yi minus fi

11138681113868111386811138681113868111386811138681113868 (15)



















0 50 100 150 200 250Sample no

0

5

10

15

20

25

30

35

LST

(degC)

19952004

20102015







39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE


N

0 55 19 285 38479Miles


0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004





25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References























































Imageprocessing

Opticalbands


Correctedsurface

reflectanceimage



Result




DENFIS

ANFIS

WNN

MARS



Radiance

Thermalbands

Surfaceemissivity

PV

NDVI

NIR and REDbands

Reprojection




0 30 60 90 12015Miles


WaterbodiesForests

N

(a)

0 30 60 90 12015Miles


WaterbodiesForests

N

(b)

0 30 60 90 12015Miles


WaterbodiesForests

N

(c)





y β0 + 1113944

M

m1βmBFm (11)





yk 1113944m







Land cover1995 2004 2015



0 10 20 30 405Miles

High 27

Low 3

N


(a)

0 10 20 30 405Miles

High 27

Low 3

N


(b)

0 10 20 30 405Miles

High 27

Low 3

N


(c)

0 10 20 30 405Miles

High 27

Low 3

N


(d)








R2

1 minus1113936


2

1113936Ni1 yi minus y( 1113857

2 (13)

RMSE

051113944N


2

11139741113972

(14)

MAE 1N

1113944

N

i1yi minus fi

11138681113868111386811138681113868111386811138681113868 (15)



















0 50 100 150 200 250Sample no

0

5

10

15

20

25

30

35

LST

(degC)

19952004

20102015







39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE


N

0 55 19 285 38479Miles


0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004





25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References























































y β0 + 1113944

M

m1βmBFm (11)





yk 1113944m







Land cover1995 2004 2015



0 10 20 30 405Miles

High 27

Low 3

N


(a)

0 10 20 30 405Miles

High 27

Low 3

N


(b)

0 10 20 30 405Miles

High 27

Low 3

N


(c)

0 10 20 30 405Miles

High 27

Low 3

N


(d)








R2

1 minus1113936


2

1113936Ni1 yi minus y( 1113857

2 (13)

RMSE

051113944N


2

11139741113972

(14)

MAE 1N

1113944

N

i1yi minus fi

11138681113868111386811138681113868111386811138681113868 (15)



















0 50 100 150 200 250Sample no

0

5

10

15

20

25

30

35

LST

(degC)

19952004

20102015







39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE


N

0 55 19 285 38479Miles


0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004





25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References





















































0 10 20 30 405Miles

High 27

Low 3

N


(a)

0 10 20 30 405Miles

High 27

Low 3

N


(b)

0 10 20 30 405Miles

High 27

Low 3

N


(c)

0 10 20 30 405Miles

High 27

Low 3

N


(d)








R2

1 minus1113936


2

1113936Ni1 yi minus y( 1113857

2 (13)

RMSE

051113944N


2

11139741113972

(14)

MAE 1N

1113944

N

i1yi minus fi

11138681113868111386811138681113868111386811138681113868 (15)



















0 50 100 150 200 250Sample no

0

5

10

15

20

25

30

35

LST

(degC)

19952004

20102015







39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE


N

0 55 19 285 38479Miles


0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004





25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References
























































R2

1 minus1113936


2

1113936Ni1 yi minus y( 1113857

2 (13)

RMSE

051113944N


2

11139741113972

(14)

MAE 1N

1113944

N

i1yi minus fi

11138681113868111386811138681113868111386811138681113868 (15)



















0 50 100 150 200 250Sample no

0

5

10

15

20

25

30

35

LST

(degC)

19952004

20102015







39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE


N

0 55 19 285 38479Miles


0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004





25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References































































0 50 100 150 200 250Sample no

0

5

10

15

20

25

30

35

LST

(degC)

19952004

20102015







39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE


N

0 55 19 285 38479Miles


0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004





25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References

























































39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

39deg4

5prime0Prime

N40

deg30prime

0PrimeN

41deg1

5prime0Prime

N

117deg0prime0PrimeE


N

0 55 19 285 38479Miles


0102030405060708090

100

IF

Variables

Latit

ude

Long

itude

Hei

ght

ND

VI

ND

WI

ND

BI UI

LST_

1995

LST_

2004





25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References























































25

20

15

10

5

0

GCV

0 20 40 60 80 100 120 140 160 180 200BFs no

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted



Model Variables R2






ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References
























































ANDORNOT

Logical operations


(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sample no

MeasuredPredicted

(b)


0 2 4 6 8 10 12 14 16Epochs

MSE 10ndash4

10ndash6

10ndash8

10ndash2

100

TrainBestGoal

(a)

25

20

15

10

5

0

35

30

LST

(degC)

0 50 100 150 200 250Sampling no

MeasuredPredicted

(b)









28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References



























































28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)

y = 087lowastx + 27

(a)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 07lowastx + 6

(b)

28

26

24

22

20

18

16

14

12

10

8282624222018161412108

Pred

icte

d LS

T (deg

C)

Measured LST (degC)


(c)

28

26

24

22

20

18

16

14

12

10

8

Pred

icte

d LS

T (deg

C)


y = 093lowastx + 14

(d)




5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References






















































5 Conclusions





Data Availability






Acknowledgments


LST

(degC)

40

35

30

25

20

150 50 100 150 200 250

Sample no

20252035

(a)

y = 03297x ndash 64307R2 = 09445

10

12

14

16

18

20

22

24

26

28

30

1990 2000 2010 2020 2030 2040

LST

(degC)

Year

(b)



References





















































References





















































studyforpredictinglandsurfacetemperature(lst)using...

Documents