1007-4619 (2010) 06-1074-16 Joural of Remote Sesig 遥感学报 Spatial scalig of et primary productivity model based o remote sesig WANG Liwe 1, WEI Yaxig 1, NIU Zheg 2 1. College of Urba ad Eviromet Sciece, Liaoig Normal Uiversity, Liaoig Dalia 116029, Chia; 2. The State Key Laboratory of Remote Sesig Sciece, Istitute of Remote Sesig Applicatios, Chiese Academy of Scieces, Beijig 100101, Chia Abstract: Spatial scalig for et primary productivity (NPP) refers to the trasferrig process of establishig quatitative correlatio betwee simulated NPP derived from data at differet spatial resolutios. How to trasfer NPP at oe scale by the algorithm with smaller error to at aother is the urget problem. Noliearity ad effects from lad cover type are two mai problems i NPP scalig. I this paper, the cotextural approach based o mixed pixels ad support vector machie (SVM) algorithm are used to make the scalig model from the fie resolutio (TM) to the coarse resolutio (MODIS). Spatial scalig from NPP retrieved from fie resolutio data to NPP derived from coarse resolutio images is performed, ad the correctio of scale effect to NPP retrieved from coarse resolutio data of MODIS is accomplished. The result shows that the correlatio betwee R j_corrected of the correctio factor for scale effect ad 1-F middle desity grasslad estimated by SVM regressio model is higher (R 2 =0.81). Before the correctio for scale effect, the correlatio betwee NPP MODIS ad NPP TM is lower (R 2 =0.69; RMSE=3.47), while the correlatio betwee NPP TM ad corrected NPP MODIS_corrected is higher (R 2 =0.84; RMSE=1.87). Therefore, NPP corrected for scale effect has bee greatly improved i both correlatio ad error. Key words: et primary productivity, light use efficiecy model, remote sesig, scalig, support vector machie CLC umber: TP761 Documet code: A Citatio format: Wag L W, Wei Y X ad Niu Z. 2010. Spatial scalig of et primary productivity model based o remote sesig. Joural of Remote Sesig. 14(6): 1074 1089 1 INTRODUCTION The advatage of remote sesig is to provide lad surface iformatio frequetly ad loger. A importat reaso is that problems i remote sesig are complex ad the icosistet scale data are derived from the measured data, data from differet methods of remote sesig ad data for applicatios. How to scale betwee differet scale data is the core of scale problem (Li et al., 2001; Turer et al., 1996). Spatial heterogeeity of the earth system limits the trasferrig from oe scale to aother. That is the major cause that data at differet scale retrieved from satellite sesors eed to be uified (Su et al., 2001; Pierce et al., 1995). Spatial scalig for et primary productivity (NPP) refers to the trasferrig process of establishig quatitative correlatio betwee simulated NPP derived from data at differet spatial resolutios. I the process of scalig, simulated NPP at oe resolutio ca be deduced from simulated NPP at aother resolutio (Ehleriger & Field 1993). Errors i lad surface parameters at the coarse resolutio are derived from the sythesis of various kids of spectral sigals from lad surface objects received by sesors. I additio, the quatitative correlatio betwee mixed spectral sigals ad lad surface parameters is difficult to be precisely defied. Therefore, the urget problem is how to trasfer NPP at oe scale by the algorithm with smaller error to at aother. I scalig, oliearity ad lad cover types that have effects o scalig are two mai problems (Che et al., 1999; Nemai et al., 2003). I this paper, a cotextural approach based o mixed pixels was used to make the scalig model from the fie resolutio (TM) to the coarse resolutio (MODIS). The algorithm of support vector machie (SVM) was adopted to establish the regressio model betwee correctio factor (R j ) of scale effect for a pixel labeled as cover type j ad F ij of the odomiat cover type i i this pixel. That is, F ij was the fractio of a odomiat cover type i i a pixel labeled as cover type j i fie resolutio data of TM. Afterwards Spatial scalig from NPP retrieved at the fie resolutio to NPP derived from the coarse resolutio was performed, ad the correctio of scale effect to NPP retrieved at the coarse resolutio was accomplished. Received: 2009-12-14; Accepted: 2010-06-07 Foudatio: Chiese Liaoig Provice Educatio Bureau Geeral Sciece Research Project (No. L2010226); Chiese Educatio Miistry Humaities ad Social Scieces Key Research Base Project (No.08JJD790142); Chiese Liaoig Provice Educatio Bureau Iovatio Team Project (No. 2007T095); Chiese Special Fuds for Major State Basic Research Project (No. 2007CB714406). First author biography: WANG Liwe (1971 ), female, lecturer. She graduated from Istitute of Remote Sesig Applicatios, Chiese Academy of Scieces i 2009, majors i Cartography ad Geographic Iformatio System. She is focusig o the research of remote sesig applicatios. She has published more tha 10 sciece papers. E-mail: wlw9585@163.com
WANG Liwe et al.: Spatial scalig of et primary productivity model based o remote sesig 1075 2 STUDY AREA This study is coducted i Dari Couty, located i the southeaster of Qighai Provice i West Chia (Fig.1). Dari Couty is geographically situated betwee 98 15 29 ad 100 32 41 East logitude ad betwee 32 36 42 ad 34 15 20 North latitude. The legth of Dari Couty from east to west is about 162 km, ad the width from orth to south is almost 126 km. The total area of Dari Couty is 14600 km 2, ad the average altitude is 4426 m. The climate of Dari Couty is semi-humid plateau type without obvious seaso variatio i additio to warm ad cold seasos. The average temperature of Dari Couty rages from 0.1 to 3.5. The aual precipitatio is 569 mm usually takig place from Jue to September. The aual evaporatio is 1119.07 mm, ad the mothly highest is 150.1 mm occurrig i May. The mai vegetatio i Dari Couty is alpie meadow domiated by Kobresia Pygmaea, K. Humilis ad K. Tibetica, which are widely distributed o suy sides of hills ad wide valley. Alpie shrub is maily domiated by Dasiphora Bsuticosa, Salix Cupularis, which are ofte located o shade sides of hills. 3 DATA PREPARATION Satellite data used i this paper iclude MODIS data products ad TM data. I the process of makig the NPP model, MODIS data products were used to calculatio ad validatio icludig MOD09A1 (surface reflectace 8-day L3 global 500 m), MOD11A2 (lad surface temperature/emissivity 8-day L3 global 1 km), MOD12Q1 (lad cover type yearly L3 global 1 km), MOD13A2 (vegetatio idices 16-day L3 global 1 km), MOD15A2 (leaf area idex/fpar 8-day L4 global 1 km), MOD17A2 (gross primary productivity 8-day L4 global 1 km) ad MOD43B3 (albedo 16-day L3 global 1 km), which were supplied by LPDAAC (Lad Process Distributed Active Archive Ceter, U.S.A.). Geometric rectificatio, cloud elimiatio processsig ad modifyig abormal pixel were made to MODIS data i advace. The study used 5 scees of TM images derived from Ladsat-5, of which path/row umbers were 132/37, 133/36, 133/37, 134/36 ad 134/37. Image extet covered the whole Dari Couty. Geometric rectificatio ad atmospheric correctio was performed to TM data. The received date of MODIS ad TM was i July 2006, which represeted the period that vegetatio grew better i a year. The meteorology data icludig average temperature, precipitatio, average atmospheric relative humidity, sushie duratio, air pressure ad mea wid speed i the paper was derived from Chiese atioal meteorological ceter. The method of vegetatio cover classificatio i Qighai Provice (Wag et al., 2008) was used i the research. DEM data were dowloaded from Cetro Iteratioal de Agriculture Tropical (CIAT, http://srtm.csi.cgiar.org/). Measured data i the paper was from experimet i Dari Couty i July 2006. Experimetal istrumet icluded LI- 6400 portable photosythesis system, LAI-2000 vegetatio caopy aalysis system, SUNSCAN caopy aalysis system ad so o. 4 METHODS 4.1 NPP simulatio model Fig. 1 Locatio map of the study area Remote sesig data ca be directly used i the NPP simulatio model based o light use efficiecy (LUE). Thus spatio-temporal resolutio of the model ca be greatly improved. It makes it possible to simulate ad dyamic moitor NPP for large scale. The NPP simulatio model i the paper is based o LUE ad lears from advatages of MODIS-PSN, CASA, GLO-PEM, VPM models. At the same time, the NPP model fully cosiders characteristics of vegetatio LUE ad eviromet i the study area. The NPP simulatio model based o remote sesig ca be expressed as follows. NPP = PAR FPAR mi ( Ts, Ws) εmax (1) where PAR is photosythetic active radiatio (MJ m 2 ), FPAR is fractioal photosythetic active radiatio (uitless), T s is temperature stress factor (uitless), W s is water stress factor (uitless), ad ε max is maximum LUE (gc MJ 1 ). 4.1.1 Maximum LUE (ε max ) Values of maximum LUE for various plats are differet. They have great effect o fial NPP simulatio results. Optimal values of maximum LUE for differet vegetatio are simulated by maximum likelihood estimate based o literature ad field experimetal data i the study area (Wag et al., 2006; Zhu et
1076 Joural of Remote Sesig 遥感学报 2010, 14(6) al., 2006; Li et al., 2004; Yag et al., 1987; Pu et al., 2005). xiyi xy ˆ ε i 1 max = = 2 2 xi x i= 1 where max ˆε is the simulated value of maximum LUE, x i is the calculated value of PAR FPAR mi ( Ts, Ws) for the plat, y i is the NPP value for the plat, i is the sample umber for the plat, is maximum sample umber for the plat, ad 1 yi i = 1 y =. (2) 1 xi i = 1 x =, After Eq.(2) is used to simulate maximum LUE value of mai type of grasslad or shrub i the study area, Eq. (3) is adopted to calculate maximum LUE value of every kid of grasslad or shrub remote sesig-based classificatio i order to improve simulated precisio of maximum LUE. ε max i i= 1 ( ) = c ε (3) where c i is area fractio of grasslad or shrub labeled as cover type i, ad (ε max ) i is maximum LUE of grasslad or shrub labeled as cover type i. Calculated results of maximum LUE for mai vegetatio i the study area are 0.908 gc MJ 1 for broad-leaved forest, 0.725 gc MJ 1 for eedle-leaved ad broad-leaved mixed forest, 0.645 gc MJ 1 for eedle-leaved forest, 0.312 gc MJ 1 for dese grasslad, 0.538 gc MJ 1 for dese shrub, 0.234 gc MJ 1 max i for middle desity grasslad, 0.208 gc MJ 1 for grasslad mixed with farmlad, 0.175 gc MJ 1 for sparse grasslad, ad 0.114 gc MJ 1 for sparse shrub. 4.1.2 FPAR The correlatio equatio betwee NDVI ad FPAR is made based o data i the study area. FPAR i o-vegetatio cover regios is assumed as a miimum value of 0, ad FPAR i complete vegetatio cover areas is assiged to a maximum value of 0.95 (Che et al., 2007): SR SR mi FPAR = mi,0.95 SR max SR mi SR max SR (4) mi 1+ NDVI SR = (5) 1 NDVI where SR mi of all kids of vegetatio is assiged to a same value of 1.06, values of SR max are 6.14 for broad-leaved forest, 3.88 for eedle-leaved ad broad-leaved mixed forest, 3.44 for eedle-leaved forest, 3 for dese grasslad, 3.27 for dese shrub, 2.17 for middle desity grasslad, 1.94 for grasslad mixed with farmlad, 1.78 for sparse grasslad, ad 1.70 for sparse shrub. 4.1.3 Water stress factor (W s ) I this paper, evaporative fractio (EF) which has bee prove i may researches is used to calculate water stress factor (W s ) (Suleima & Crago, 2004; Veturii et al., 2004; We et al., 2006; Zhag et al., 2004): LE Ws = EF = (6) LE + H where LE is latet heat flux (W m 2 ), ad H is sesible heat flux (W m 2 ). Fig. 2 shows the flowchart of calculatig EF. Fig. 2 Flowchart of EF
WANG Liwe et al.: Spatial scalig of et primary productivity model based o remote sesig 1077 4.1.4 Temperature stress factor (T s ) Temperature stress factor is estimated by the method from the Terrestrial Ecosystem Model (TEM) of Raich (1991). ( T Tmi )( T Tmax ) Ts = (7) 2 ( T Tmi )( T T max ) ( T Topt ) where T is air temperature, T mi, T max ad T opt are miimum, maximum ad optimum air temperature for photosythetic activity, respectively. If air temperature is below T mi or beyod T max, T s is set to 0. I this paper, T mi ad T max are set to 0 ad 36, respectively. Because vegetatio have adapted well to the temperature of growth eviromet, T opt may be set to the mea temperature of growig seaso. 4.2 Spatial scalig of NPP 4.2.1 Cotextural approach based o mixed pixels Accordig to primary problems affectig scalig, cotextural approach based o mixed pixels of Aita et al. (2004) is used to perform spatial scalig of NPP. The theory is as follows. Mixed pixels are commo pheomea i remote sesig images. The presece of mixed pixels is the mai reaso that precisios of lad surface parameters ca ot meet requiremets. I order to improve precisios of remote sesig applicatios, umixig of pixels must be solved ad remote sesig applicatios eed to trasfer from pixels level to sub-pixels level. A pixel i coarse resolutio images ca be deemed as a mixed pixel composed of may pixels at correspodig coordiates i fie resolutio images, ad these pixels i fie resolutio images ca be cosidered as sub-pixels of this pixel i coarse resolutio images (Che et al., 1999; Tog et al., 2006). Spatial scalig based o cotextural approach adopts the theory of aalyzig area fractios for lad cover types to compute lad surface parameters at various resolutios. A pixel i remote sesig classificatios is ofte defied as the pixel of domiat cover type whose area is the largest i the pixel, while other odomiat cover types i this pixel is igored durig calculatig lad surface parameters. This leads to every pixel oly to be labeled as a kid of domiat cover type. Therefore, fial calculatio results cotai much ucertaity (Che et al., 1999). The ideal method is that lad surface parameters i coarse resolutio images should be calculated as mea values of parameters at correspodig coordiates i fie resolutio images. I this paper, NPP TM retrieved from fie resolutio data of TM is assumed to represet the real value of NPP for terrestrial vegetatio (Fig.3(a)), ad NPP MODIS derived from coarse resolutio data of MODIS is the approximate value which eeds to be corrected for scale effect (Fig.3(b)). A pixel i MODIS (1 1 km 2 ) is supposed to cotai 1089 pixels of TM (30 30 m 2 ). Due to the presece of mixed pixels, 1089 pixels i TM cotai may kids of lad cover types. While a pixel at the correspodig coordiate i MODIS is labeled as a kid of domiat cover type which covers the largest area i this pixel, ad other odomiat cover types i the same pixel are eglected. TM ad MODIS data which all covers the study area are respectively iput as parameters of the NPP model based o LUE, ad NPP TM ad NPP MODIS ca be achieved respectively (Fig.3). It should be oticed that estimated NPP results for two kids of scales eed to be coverted to the same resolutio i order to cotrastively aalyze NPP results. That is, the resolutio of NPP TM retrieved from fie resolutio data of TM eeds to be switched to 1 km. whe NPP TM of 30 m resolutio is coverted to 1 km NPP TM, arithmetical mea value of NPP for may pixels at the correspodig coordiate i NPP TM images is cosidered as estimated NPP for a pixel i NPP TM images. Due to the resolutio differece betwee TM ad MODIS which are all iput parameters of the model, much error appears betwee NPP TM ad NPP MODIS. The reaso that calculatig results for two scales are differet is that large amout of mixed pixels are i coarse resolutio data of MODIS comparig with fie resolutio data of TM ad odomiat cover types i mixed pixels are igored. The NPP MODIS is corrected for scale effect based o NPP TM to retrieve NPP MODIS j_correcred so that the precisio of estimated NPP derived from coarse resolutio data of remote sesig ca be improved. Fig. 3 Flowchart of calculatig NPP by fie or coarse resolutio data of remote sesig (a) Flowchart of NPP scalig retrieved from fie resolutio data of TM; (b) Flowchart of NPP retrieved from coarse resolutio data of MODIS NPP MODIS retrieved from coarse resolutio data of MODIS ca be corrected as follows. NPP = NPP R (8) TM j MODIS j j where NPP TM j is estimated NPP result for a pixel labeled as cover type j derived from fie resolutio data of TM, NPP MODIS j is estimated NPP result for a pixel labeled as cover type j retrieved from coarse resolutio data of MODIS, ad R j is a correctio factor for scale effect. So R j is give by: NPPTM j R j = (9) NPP MODIS j
1078 Joural of Remote Sesig 遥感学报 2010, 14(6) The relatioship is theoretically as follows. R j 1 CijFij i= 1 = (10) where F ij is the fractio of odomiat cover type i (30 m resolutio TM) i a pixel labeled as cover type j (1 km resolutio TM), C ij is the regressio coefficiet betwee R j for a pixel labeled as cover type j ad F ij of odomiat cover type i i this pixel, ad is the amout of odomiat cover types i a pixel labeled as cover type j. Based o above formulas ad area fractios of odomiat cover types i every pixel, NPP MODIS j_correcred which has bee corrected for scale effect ca be achieved as follows. NPPMODIS j_corrected = NPPMODIS j 1 CijF ij (11) i= 1 4.2.2 Support Vector Machie (SVM) algorithm Whe NPP is performed scalig i this paper, regressio coefficiet C ij betwee R j for a pixel labeled as cover type j ad F ij of odomiat cover type i i this pixel eeds to be calculated as i Eq. (10) ad Eq. (11). The precisio of estimated C ij has direct effect o successful accomplishmet of spatial scalig method for the NPP model based o remote sesig. Therefore, SVM algorithm is used to make the regressio model betwee R j ad F ij i the paper. Comparig with the traditioal statistical studyig theories, it is foud that fewer adjustmet parameters, faster studyig speed, optimum process ad better spread capability from techical performace make SVM predomiat. Studyig results of SVM are ofte better tha that of other patter idetificatios ad regressio predictio methods. SVM has uique advatages i solvig small sample, oliearity ad higher dimesioal patter idetificatio (Li et al., 2006). I the paper, Libsvm is adopted as the SVM algorithm. Libsvm is a software package which has simple structure, easy usage ad fast efficiet support vector classificatio ad regressio. It ca solve classificatio (e.g. C-SVM ad u-svm), regressio (e.g. e-svm ad u-svm), distributed estimatio (oe-class-svm) ad so o. Liearity, polyomial, RBF ad S-shaped fuctios are four kids of commo kerel fuctios to be selected to solve multi-classificatio problems, choose parameters of cross validatio, weight ubalace sample, estimate probability of multi-classificatio problems ad so o (Li, 2008). RBF kerel fuctio is selected to ru Libsvm program i this paper. acquired by calculatig arithmetic mea value of NPP i may correspodig pixels at 30 m resolutio. Durig the process, the priciple of vegetatio cover type with domiat area is used to defie vegetatio cover type of the pixel. I additio, various vegetatio types have differet cotributio to NPP i scalig, ad they have ot liear or oliear relatioship. So some errors are produced i scalig of vegetatio classificatio data, ad they are passed i NPP scalig. Through above aalysis, estimated NPP results greatly rely o lad cover type, classificatio precisio eeds to be ehaced to esure successful accomplishmet of NPP scalig. I this paper, based o literature of Dari Couty ad field ivestigatio, preprocessed MODIS at 1 km resolutio ad TM data at 30 m resolutio of the study area i 2006 were respectively classified by highprecisio sythetic classificatio method which ca reflect the spatial distributio of vegetatio. Fig.4 ad Fig.5 show lad cover classificatio maps retrieved from MODIS ad TM data i the study area, respectively. Lad cover classificatio results i Dari Couty were statistically aalyzed. The result shows that the area of middle desity grasslad is the largest (60.96% of the total area of the whole couty). The secod area is sparse grasslad (16.81%). Sparse shrub covers 6.27% of the total area. Gobi ad dese grasslad cover approximate area (5.67% ad 5.65%). Dese shrub oly covers 3.25%. The area of forest i Dari Couty is sparse. Needle-leaved forest, eedle-leaved ad broad-leaved mixed forest ad broad-leaved forest covers 0.46%, 0.40% ad 0.53%, respectively. Fig. 4 Lad cover classificatio map of the study area at 1 km resolutio 5 RESULTS 5.1 Lad cover classificatio i the study area Vegetatio type ca make great differece of estimatig LUE ad FPAR i the NPP model based o LUE, so it affects the precisio of fial NPP results. Durig NPP scalig, NPP at 30 m resolutio is retrieved from 30 m resolutio TM, ad estimated NPP of a correspodig pixel at 1 km resolutio ca be Fig. 5 Lad cover classificatio map of the study area at 30 m resolutio
WANG Liwe et al.: Spatial scalig of et primary productivity model based o remote sesig 1079 5.2 Cotrastive aalysis of NPP TM ad NPP MODIS map NPP TM (Fig.6) ad NPP MODIS (Fig.7) maps of the study area i July 2006 were respectively achieved based o the NPP model made i the paper. Fig.8 shows NPP TM map at 1 km resolutio which was coverted from NPP TM at 30 m resolutio. Fig.7 (NPP MODIS ) ad Fig.8 (NPP TM ) at 1 km resolutio were cotrastively aalyzed, ad the differece betwee them was foud. 30m resolutio data ad 1 km resolutio data as iput parameters of the NPP model are differet, ad durig the course of the model calculatig, the differece is elarged due to oliearity of algorithm. Fig. 6 NPP TM map of the study area at 30 m resolutio Accordig to statistics, the mea value of NPP TM i the study area i July 2006 is 55 gcm 2 moth 1 (Fig.8). Mea values of NPP TM for various vegetatio types are 28 gcm 2 moth 1 for sparse shrub, 40 gcm 2 moth 1 for sparse grasslad, 63 gcm 2 moth 1 for middle desity grasslad which is the mai vegetatio cover type i the study area, 67 gcm 2 moth 1 for dese shrub, 69 gcm 2 moth 1 for dese grasslad, 72 gcm 2 moth 1 for eedle-leaved forest, 75 gcm 2 moth 1 for eedle-leaved ad broad-leaved mixed forest, 78 gcm 2 moth 1 (the largest value) for broad-leaved forest, respectively. Based o statistics aalyzig to Fig.7, the average value of NPP MODIS i the study area i July 2006 is 62 gcm 2 moth 1. Mea values of NPP MODIS for various vegetatio types are 34 gcm 2 moth 1 for sparse shrub, 45 gcm 2 moth 1 for sparse grasslad, 65 gcm 2 moth 1 for middle desity grasslad which is the mai vegetatio cover type i the study area, 68 gcm 2 moth 1 for dese shrub, 71 gcm 2 moth 1 for dese grasslad, 74 gcm 2 moth 1 for eedle-leaved forest, 77 gcm 2 moth 1 for eedle-leaved ad broad-leaved mixed forest, 79 gcm 2 moth 1 (the largest value) for broad-leaved forest, respectively. From the above calculatig results, NPP MODIS is bigger tha NPP TM, ad variable rage of the former is also bigger tha that of the latter. Spatial heterogeeity ad mixed pixels may be reasos. Nodomiat cover types i the pixel are igored i estimatig NPP from MODIS data at the coarse resolutio, while other cover types i may pixels are cosidered i estimatig NPP from TM data at the fie resolutio. The distributio characteristic of lad cover type i the study area has a great effect o estimated NPP results, ad it will lead that NPP MODIS is bigger or smaller tha NPP TM. I this paper, the ratio of vegetatio cover i Dari Couty is higher (94%). The primary vegetatio cover is middle desity grasslad (60.96%), ad the secod is sparse grasslad (16.81%). This distributio characteristic of vegetatio leads that NPP MODIS is a little bigger tha NPP TM i the whole study area. Fig. 7 NPP MODIS map of the study area at 1 km resolutio 5.3 Regressio aalysis betwee R j ad F ij Fig. 8 NPP TM map of the study area at 1 km resolutio coverted from NPP TM by NPP scalig Durig NPP scalig, the regressio relatioship betwee R j for a pixel labeled as cover type j ad F ij of odomiat cover type i i this pixel eeds to be quatified. I this paper, SVM algorithm is used to make the regressio model betwee R j ad F ij. I the paper, RMSE is adopted to estimate the error betwee the calculated value ad the real value. I the regressio model of SVM, various model parameters determie differet regressio results, ad model parameters iclude pealty coefficiet C ad parameter γ of RBF kerel fuctio. Pealty coefficiet C ad parameter γ are defied by experimets. Selectig the value i the bigger extet is to acquire approximate optimum regio, the the value is selected i the smaller extet. Cosiderig sythetic performace ad traiig time, values of parameter C
1080 Joural of Remote Sesig 遥感学报 2010, 14(6) ad γ are selected by the experimet while the error is less. Fig.9 shows the scatter plot of R j_corrected ad 1-F middle desity grasslad estimated by the regressio model based o SVM (for pixels labeled as middle desity grasslad). From Fig.9, the correlatio betwee R j_corrected ad 1-F middle desity grasslad estimated by SVM regressio model is higher (R 2 =0.81). NPP MODIS is lower (R 2 =0.69; RMSE=3.47), while the correlatio betwee NPP TM ad corrected NPP MODIS_corrected is higher (R 2 =0.84; RMSE=1.87).Therefore, NPP which has bee corrected for scale effect has bee greatly improved i both correlatio ad error. 6 CONCLUSION AND DISCUSSION Fig. 9 Correlatio aalysis betwee R j_corrected ad 1-F middle desity grasslad for pixels labeled as middle desity grasslad 5.4 Correlatio compariso betwee NPP TM ad NPP which has bee ucorrected or corrected for scale effect I NPP scalig, to multiply NPP MODIS by R j_corrected estimated from SVM regressio model is used to calculate NPP MODIS_corrected which has bee corrected for scale effect. Fig.10 ad Fig.11 are aalysis maps of correlatio compariso betwee NPP TM ad NPP which has bee ucorrected or corrected for scale effect (samples used i the figures is differet from samples adopted to make the regressio model betwee R j ad F ij ). From the figures, the correlatio betwee NPP TM ad ucorrected Fig. 10 Compariso betwee ucorrected NPP (NPP MODIS ) ad NPP TM Fig. 11 Compariso betwee corrected NPP (NPP MODIS_corrected ) ad NPP TM Spatial scalig of NPP model based o remote sesig is the mai research cotet i this paper. The primary coclusios are draw as follows. (1) Cotextural approach based o mixed pixels is ideal i coductig spatial scalig of NPP. The approach cosiders the effect of odomiat cover types i mixed pixels, ad it used the theory of area fractios covered by various cover types to calculate lad surface parameters at various resolutios. It realizes spatial scalig from NPP at the fie resolutio to NPP at the coarse resolutio based o remote sesig data. The study shows Cotextural approach is a better method of scalig. (2) The regressio model betwee correctio factor (R j ) of scale effect for a pixel labeled as cover type j ad F ij of odomiat cover type i i this pixel is made by support vector machie (SVM). Results show that SVM ca efficietly covert realistic problems to high dimesio characteristic space by oliear trasform i order to costruct the liear fuctio to acquire optimum process result for limited samples. The correlatio betwee R j_corrected ad 1-F middle desity grasslad estimated by SVM regressio model is higher (R 2 =0.81). (3) The scalig model from fie resolutio data of TM to coarse resolutio data of MODIS is made, ad NPP results retrieved from coarse resolutio data of MODIS is corrected for scale effect. The correlatio betwee NPP TM ad ucorrected NPP MODIS is lower (R 2 =0.69; RMSE=3.47), while the correlatio betwee NPP TM ad corrected NPP MODIS_corrected is higher (R 2 =0.84; RMSE=1.87).Therefore, NPP which has bee corrected for scale effect has bee greatly improved i both correlatio ad error. With the developmet of remote sesig techology, various kids of spatial ad temporal resolutio sesors appear. This makes space scalig more importat i quatified aalysis of remote sesig. Lad cover types are so stable for a period that they suit to be acquired by high spatial resolutio images of remote sesig (e.g. Ladsat). However, lad surface variatio by atural disaster (e.g. forest fire) eeds to be foud by high temporal resolutio images of which occurrig frequecy is the same with that of atural disaster. Therefore, spatial scalig amog remote sesig data at various resolutios i applicatio eeds to be solved. Due to the ifluece of sub-pixels heterogeeity, NPP retrieved from remote sesig data at coarse resolutios have cosiderable errors. I the study, the relative error of NPP derived from MODIS is about 13%. At preset, this kid of error is commo i regioal ad global NPP maps, but spatial scalig based o mixed pixels ca efficietly reduce it. Cotextural approach based o mixed pixels adopted i this
WANG Liwe et al.: Spatial scalig of et primary productivity model based o remote sesig 1081 paper ca also be used i spatial scalig amog NPP at other resolutios. REFERENCES Aita S, Che J M ad Liu J. 2004. Spatial scalig of et primary productivity usig subpixel iformatio. Remote Sesig of Eviromet, 93(1 2): 246 258 Che J M, Liu J ad Cihlar J. 1999. Daily caopy photosythesis model through temporal ad spatial scalig for remote sesig applicatios. Ecological Modellig, 124(2 3): 99 119 Che L F, Gao Y H, Li L, Liu Q H ad Gu X F. 2007. Estimatio of daily et primary productivity for forest based o white sky data of MODIS. Sciece i Chia Ser. D, 37(11): 1515 1521 Ehleriger J R ad Field C B. 1993. Scalig Physiological Processes: Leaf to Globe. Bosto: Academic Li G Z, Wag M ad Zeg H J. 2006. A Itroductio to Support Vector Machies. Beijig: Publishig House of Electroics Idustry Li X W, Wag J F, Wag J D ad Liu Q H. 2001. Multi-Agle ad Thermal Iferred Remote Sesig. Beijig: Sciece Press Li Y N, Zhao X Q ad Cao G M. 2004. Aalyses o climates ad vegetatio productivity backgroud at Haibei Alpie Meadow ecosystem research statio. Plateau Meteorology, 23(4): 558 567 Li H D. 2008. A simple itroductio to LIBSVM. Iteret: www.csie.tu.edu.tw/~cjli/libsvm Pierce L L ad Ruig S W. 1995. The effects of aggregatig sub-grid lad surface variatio o large-scale estimates of et primary productio. Ladscape Ecology, 10(4): 239 253 Nemai R R, Keelig C D ad Hashimoto H. 2003. Climate-drive icreases i global terrestrial et primary productio from 1982 to 1999. Sciece, 300(5625): 1560 1563 Pu J Y, Li Y N, Zhao L ad Yag S H. 2005. Seasoal chages of Kobresia Humilis Meadow biomass with climate factor. Acta Agrestia Siica, 13(3): 238 241 Raich J W, Rastetter E B ad Melillo J M. 1991. Potetial et primary productivity i South America: applicatio of a global model. Ecological Applicatios, 1(4): 399 429 Su L H, Li X W ad Huag Y X. 2001. A review o scale i remote sesig. Advace i Earth Scieces, 16(4): 544 548 Suleima A ad Crago R. 2004. Hourly ad daytime evapotraspiratio from grasslad usig radiometric surface temperatures. Agroomy Joural, 96: 384 390 Tog Q X, Zhag B ad Zheg L F. 2006. Hyperspectral Remote Sesig: Theory, Techology ad Applicatio. Beijig: Higher Educatio Press Turer D P, Dodso R D ad Marks D. 1996. Compariso of alterative spatial resolutios i the applicatio of a spatially distributed biogeochemical model over complex terrai. Ecological Modelig, 90(1): 53 67 Veturii V, Bisht G ad Islam S. 2004. Compariso of evaporative fractios estimated from AVHRR ad MODIS sesors over South Florida. Remote Sesig of Eviromet, 93(1-2): 77 86 Wag L W, Wei Y X ad Niu Z. 2008. Spatial ad temporal variatios of vegetatio i Qighai Provice based o satellite data. Joural of Geographical Scieces, 18(1): 73 84 Wag X Z, Tao B Z ad Qiu W N. 2006. Advaced Surveyig Adjustmet. Beijig: Surveyig ad Mappig Press We P Y, Shu G L ad Guag S Z. 2007. Derivig a light use efficiecy model from eddy covariace flux data for predictig daily gross primary productio across biomes. Agricultural ad Forest Meteorology, 143(3 4): 189 207 Yag F T, Wag Q J ad Shi S H. 1987. The allocatio of the biomass ad eergy i Kobresia Humilis Meadow, Haibei district, Qighai provice. Acta Phytoecologica Et Geobotaica Siica, 11(2): 106 112 Zhag Y Q, Liu C M ad Yu Q. 2004. Eergy fluxes ad the Priestley-Taylor parameter over witer wheat ad maize i the North Chia Plai. Hydrol. Process, 18: 2235 2246 Zhu W Q, Pa Y Z ad He H. 2006. Simulatio of maximum light use efficiecy for mai vegetatio i Chia. Chiese Sciece Bulleti, 51(6): 700 706
1082 Joural of Remote Sesig 遥感学报 2010, 14(6) 王莉雯 1, 卫亚星 1 2, 牛铮 1., 116029; 2., 100101 摘要 : (SVM), (TM) (MODIS), NPP NPP (MODIS) NPP : SVM R j _ corrected 1 F, R 2 0.81 NPP MODIS NPP TM, R 2 0.69, RMSE 3.47; NPP MODIS_corrected NPP TM, R 2 0.84, RMSE 1.87, NPP 关键词 :,,,, SVM 中图分类号 : TP761 文献标志码 : A :,,. 2010.., 14(6): 1074 1089 Wag L W, Wei Y X ad Niu Z. 2010. Spatial scalig of et primary productivity model based o remote sesig. Joural of Remote Sesig. 14(6): 1074 1089 1 (, 2001; Turer, 1996), (, 2001; Pierce, 1995) (NPP) NPP, NPP NPP (Ehleriger & Field 1993),,, NPP, NPP (Che, 1999; Nemai, 2003), (TM) (MODIS) (SVM), j R j i F ij ( TM i j ) NPP NPP, (MODIS) NPP 收稿日期 : 2009-12-14; 修订日期 : 2010-06-07 基金项目 : ( : L2010226); ( : 08JJD790142); ( : 2007T095) (973) ( : 2007CB714406) 第一作者简介 : (1971 ),,, GIS 10, 4 SCI E-mail: wlw9585@163.com
: 1083 2 ( 1), 98 15 29 100 32 41, 32 36 42 34 15 20 162 km, 126km, 14600 km 2 4426 m,, 0.1 3.5 569 mm, 6 9 1119.07 mm, 5, 150.1 mm,,, 1km L4 ) MOD43B3( 16 1km L3 ), LP DAAC(Lad Process Distributed Active Archive Ceter, U.S.A) MODIS TM Ladsat-5 5 TM, 132/37 133/36 133/37 134/36 134/37, TM MODIS TM 2006 7,,, 41 (Wag, 2008) DEM (Cetro Iteratioal de Agriculture Tropical, CIAT, http://srtm. csi. cgiar. org/) 2006 7, LI-6400 LAI-2000 SUNSCAN 4 4.1 NPP 3 1 MODIS TM NPP MODIS MOD09A1( 8 500m L3 ) MOD11A2 ( 8 1km L3 ) MOD12Q1( 1km L3 ) MOD13A2( 16 1km L3 ) MOD15A2( 8 1km LAI FPAR L4 ) MOD17A2( 8 NPP,, NPP NPP, MODIS-PSN CASA GLO-PEM VPM NPP,, NPP NPP : ( ) NPP=PAR FPAR mi T, W ε (1) s s max, PAR, MJ m 2 ; FPAR (PAR), ; T s, ; W s, ; ε max, gc MJ 1 4.1.1 ε max ε max
1084 Joural of Remote Sesig 遥感学报 2010, 14(6), NPP (, 2006;, 2006), (, 2004;, 1987;, 2005), (2) ε max xiyi xy ˆ i= 1 ε max = 2 2 xi x i= 1 (2), max ˆε ε max, x i PAR FPAR mi(t s, W s ) ; y i NPP ; i ; ; 1 yi i = 1 y = 1 xi i = 1 x = ; ε max, (2), (3) ε max max i i= 1 ( ) ε = c ε (3) max, c i i ( ), ( εmax ) i i ( ), i : 0.908 gc MJ 1, 0.725 gc MJ 1, 0.645 gc MJ 1, 0.312 gc MJ 1, 0.538 gc MJ 1, 0.234 gc MJ 1, 0.208 gc MJ 1, 0.175 gc MJ 1, 0.114 gc MJ 1 4.1.2 FPAR NDVI FPAR,, FPAR, 0;, FPAR, 0.95(, 2007): SR SR FPAR = mi, 0.95 mi SR max SR mi SR max SR m i (4) 1+ NDVI SR = (5) 1 NDVI, SR mi 1.06, SR max, : 6.14; 3.88; 3.44; 3; 3.27; 2.17; 1.94; 1.78; 1.70 4.1.3 W s (EF)(Suleima & Crago, 2004; Veturii, 2004; We, 2006; Zhag, 2004), W s, : LE Ws = EF = (6) LE + H, LE W m 2 ; H, W m 2 2 EF 2 EF
: 1085 4.1.4 T s Raich (1991) (the terrestrial ecosystem model, TEM) T s : ( T Tmi )( T Tmax ) Ts = (7) 2 ( T T )( T T ) ( T T ) mi max opt, T, T mi T opt T max T mi T max, T s 0, T mi T max 0 36,, T opt 4.2 NPP 4.2.1, Aita (2004) NPP, :,,,,, (Che, 1999;, 2006),, (Che, 1999), NPP, TM NPP TM NPP ( 3(a)); MODIS NPP MODIS NPP ( 3(b)) 1 MODIS (1 km 1 km) 1089 TM (30 m 30 m), 1089 TM 3 NPP (a) TM NPP ; (b) MODIS NPP 1 MODIS,, TM MODIS, NPP, NPP NPP TM NPP MODIS ( 3), NPP, NPP, TM NPP TM 1km 30m NPP TM 1km NPP TM, NPP TM NPP, NPP TM NPP TM MODIS, NPP TM NPP MODIS, TM, MODIS, TM NPP TM, MODIS NPP MODIS, NPP MODIS j_corrected, NPP MODIS NPP MODIS : NPP = NPP R (8) TM j MODIS j j, NPP TM j j TM NPP, NPP MODIS j j MODIS NPP, R j, R j (a) : NPPTM j R j = (9) NPP MODIS j
1086 Joural of Remote Sesig 遥感学报 2010, 14(6) : R = 1 C F (10) j ij ij i= 1, F ij TM( 30 m) i j ( TM 1km ), C ij j R j i F ij, j, (11) NPP MODIS j_corrected : NPPMODIS j_corrected = NPPMODIS j 1 CF ij ij (11) i= 1 4.2.2 (SVM) NPP, j R j i F ij C ij, (10) (11) C ij C ij, NPP (SVM) R j F ij, :,,,,, (, 2006) SVM Libsvm Libsvm, ( C-SVM u-svm) ( e-svm u-svm) (oe-class-svm), S 4, (, 2008) Libsvm RBF NPP, 30 m TM 30 m NPP, 30 m NPP, 1 km NPP,,, NPP, NPP,, NPP, NPP,, NPP,, 2006 1 km MODIS 30 m TM MODIS TM 4 5,, 4 1km 5 5.1 NPP FPAR, NPP 5 30m
王莉雯等: 净初级生产力遥感估算模型空间尺度转换 1087 总面积占 60.96%; 其次为低覆盖度草地, 占 16.81%; 将分辨率 1km 的图 8(NPPTM)和图 7(NPPMODIS) 稀疏灌丛占 6.27%; 戈壁和高覆盖度草地的面积近 对比分析, 发现他们二者存在差异 这是由于作为 似, 分别占 5.67%和 5.65%; 郁闭灌丛仅占 3.25%; NPP 遥感估算模型输入参数的分辨率 30 m 和 1 km 该县森林面积稀少, 针叶林 针阔混交林和阔叶林 本身存在差异, 在模型的运算过程中, 由于算法的 分别占 0.46% 0.40%和 0.53% 非线性, 加剧了这种差异 NPPTM 和 NPPMODIS 分布图的对比分析 5.2 首先, 采用本文所构建的基于光能利用率的 NPP 遥感估算模型模拟, 分别得到该区域 2006 年 7 月的 NPP 模拟结果 NPPTM (图 6)和 NPPMODIS(图 7) 再将分辨率为 30 m 的 NPPTM 转换为分辨率为 1km 的 NPPTM(图 8) 经统计, 整个研究区 2006 年 7 月 NPPTM 的平均 值为 55 gcm 2moth 1(图 8) 其中, 稀疏灌丛 NPPTM 的平均值为 28 gcm 2moth 1; 低覆盖度草地的平均 NPPTM 为 40 gcm 2moth 1; 试验区域的主要植被 覆盖类型是中覆盖度草地, NPPTM 的平均值为 63 gcm 2 moth 1 ; 郁闭灌丛 NPP T M 的平均值为 67 gcm 2 moth 1 ; 高覆盖度草地的平均 NPP TM 为 69 gcm 2moth 1; 阔叶林 针阔混交林和针叶林的 平均 NPPTM 值最高, 分别为 78 75 和 72 gcm 2moth 1 同时, 对图 7 进行统计分析, 得出整个研究区 2006 年 7 月份 NPPMODIS 的平均值为 62 gcm 2 moth 1 其 中, 稀疏灌丛 NPPMODIS 的平均值为 34 gcm 2moth 1; 低覆盖度草地的平均 NPPMODIS 为 45 gcm 2moth 1; 研究区的主要植被覆盖类型中覆盖度草地的 NPPMODIS 平 均 值 为 65 gcm 2moth 1; 郁 闭 灌 丛 NPPMODIS 的平均值为 68 gcm 2moth 1; 高覆盖度草 图6 研究区 30m 分辨率 NPP 分布图(NPPTM ) 地的平均 NPPMODIS 为 71 gcm 2moth 1; 阔叶林 针 阔混交林和针叶林的平均 NPPMODIS 值最高, 分别为 79 77 和 74 gcm 2moth 1 从上面的计算结果看出, NPPMODIS 的值略高于 NPPTM, 而且发现前者的波动范围大于后者 这可能 是因为, 由于空间异质性和混合像元现象, 低分辨 率 MODIS 数据在估算 NPP 时忽略了像元中非主要 土地覆盖类型的信息, 而高分辨率 TM 数据在估算 NPP 时考虑了较多的像元中其他土地覆盖类型的信 息 研究区土地覆盖类型的分布特点会在很大程度 上影响 NPP 的估算结果, 根据研究区各种植被类型 图7 研究区 1km 分辨率 NPP 分布图(NPPMODIS) 分 布 特 点 的 不 同, 会 导 致 NPPMODIS 大 于 ( 或 小 于)NPPTM 在本文中, 研究区达日县的植被覆盖率 比较高, 达到 94%, 而且主要的植被覆盖类型为中 覆 盖 度 草 地 (60.96%), 其 次 为 低 覆 盖 度 草 地 (16.81%), 这种植被的分布特征, 使得整个试验区 域 NPPMODIS 的值略大于 NPPTM 5.3 Rj 和 Fij 的回归分析 在 NPP 空间尺度转换过程中, 需要确定标注为 j 土地覆盖类型像元的尺度效应校正因子 Rj 与该像 图8 研究区由 NPPTM 经 NPP 空间尺度转换算法得到 的 1km 分辨率 NPP 分布图(NPPTM) 元中非主要土地覆盖类型 i 所对应的 Fij 的回归关系, 采用支持向量机(SVM)算法建立了 Rj 和 Fij 之间的回 归模型
1088 Joural of Remote Sesig 遥感学报 2010, 14(6) (RMSE) SVM,, C RBF γ C γ,,,, C γ 9 SVM R j_corrected 1 F ( ) 9, SVM R j_corrected 1 F, R 2 0.81 5.4 NPP NPP TM NPP, SVM R j_corrected NPP MODIS, NPP MODIS_corrected 10 11 NPP NPP TM ( 9 R j F ij ) 9, NPP MODIS NPP TM, R 2 0.69, RMSE 3.47; NPP MODIS_corrected NPP TM, R 2 9 R j_corrected 1 F 10 NPP(NPP MODIS ) NPP TM 11 NPP(NPP MODIS_corrected ) NPP TM 0.84, RMSE 1.87, NPP 6 NPP, : (1) NPP,, NPP (2) (SVM) j R j i F ij, SVM,, SVM R j_corrected 1 F, R 2 0.81 (3) (TM) (MODIS), (MODIS) NPP NPP MODIS NPP TM, R 2 0.69, RMSE 3.47; NPP MODIS_corrected NPP TM, R 2 0.84, RMSE 1.87, NPP,, ( Ladsat), ( )
: 1089,, NPP, MODIS NPP 13% NPP NPP REFERENCES Aita S, Che J M ad Liu J. 2004. Spatial scalig of et primary productivity usig subpixel iformatio. Remote Sesig of Eviromet, 93(1 2): 246 258 Che J M, Liu J ad Cihlar J. 1999. Daily caopy photosythesis model through temporal ad spatial scalig for remote sesig applicatios. Ecological Modellig, 124(2 3): 99 119 Che L F, Gao Y H, Li L, Liu Q H ad Gu X F. 2007. Estimatio of daily et primary productivity for forest based o white sky data of MODIS. Sciece i Chia Ser. D, 37(11): 1515 1521 Ehleriger J R ad Field C B. 1993. Scalig Physiological Processes: Leaf to Globe. Bosto: Academic Li G Z, Wag M ad Zeg H J. 2006. A Itroductio to Support Vector Machies. Beijig: Publishig House of Electroics Idustry Li X W, Wag J F, Wag J D ad Liu Q H. 2001. Multi-Agle ad Thermal Iferred Remote Sesig. Beijig: Sciece Press Li Y N, Zhao X Q ad Cao G M. 2004. Aalyses o climates ad vegetatio productivity backgroud at Haibei Alpie Meadow ecosystem research statio. Plateau Meteorology, 23(4): 558 567 Li H D. 2008. A simple Itroductio to LIBSVM. Iteret: www.csie.tu.edu.tw/~cjli/libsvm Pierce L L ad Ruig S W. 1995. The effects of aggregatig sub-grid lad surface variatio o large-scale estimates of et primary productio. Ladscape Ecology, 10(4): 239 253 Nemai R R, Keelig C D ad Hashimoto H. 2003. Climate-drive icreases i global terrestrial et primary productio from 1982 to 1999. Sciece, 300(5625): 1560 1563 Pu J Y, Li Y N, Zhao L ad Yag S H. 2005. Seasoal chages of Kobresia Humilis Meadow biomass with climate factor. Acta Agrestia Siica, 13(3): 238 241 Raich J W, Rastetter E B ad Melillo J M. 1991. Potetial et primary productivity i South America: applicatio of a global model. Ecological Applicatios, 1(4): 399 429 Su L H, Li X W ad Huag Y X. 2001. A review o scale i remote sesig. Advace i Earth Scieces, 16(4): 544 548 Suleima A ad Crago R. 2004. Hourly ad daytime evapotraspiratio from grasslad usig radiometric surface temperatures. Agroomy Joural, 96: 384 390 Tog Q X, Zhag B ad Zheg L F. 2006. Hyperspectral Remote Sesig: Theory, Techology ad Applicatio. Beijig: Higher Educatio Press Turer D P, Dodso R D ad Marks D. 1996. Compariso of alterative spatial resolutios i the applicatio of a spatially distributed biogeochemical model over complex terrai. Ecological Modelig, 90(1): 53 67 Veturii V, Bisht G ad Islam S. 2004. Compariso of evaporative fractios estimated from AVHRR ad MODIS sesors over South Florida. Remote Sesig of Eviromet, 93(1 2): 77 86 Wag L W, Wei Y X ad Niu Z. 2008. Spatial ad temporal variatios of vegetatio i Qighai Provice based o satellite data. Joural of Geographical Scieces, 18(1): 73 84 Wag X Z, Tao B Z ad Qiu W N. 2006. Advaced Surveyig Adjustmet. Beijig: Surveyig ad Mappig Press We P Y, Shu G L ad Guag S Z. 2007. Derivig a light use efficiecy model from eddy covariace flux data for predictig daily gross primary productio across biomes. Agricultural ad Forest Meteorology, 143(3 4): 189 207 Yag F T, Wag Q J ad Shi S H. 1987. The allocatio of the biomass ad eergy i Kobresia Humilis Meadow, Haibei district, Qighai provice. Acta Phytoecologica Et Geobotaica Siica, 11(2): 106 112 Zhag Y Q, Liu C M ad Yu Q. 2004. Eergy fluxes ad the Priestley-Taylor parameter over witer wheat ad maize i the North Chia Plai. Hydrol. Process, 18: 2235 2246 Zhu W Q, Pa Y Z ad He H. 2006. Simulatio of maximum light use efficiecy for mai vegetatio i Chia. Chiese Sciece Bulleti, 51(6): 700 706 附中文参考文献,,,,. 2007. MODIS. D :, 37(11): 1515 1521,,. 2006.. :,,,. 2001.. :,,. 2004.., 23(4): 558 567. 2008. LIBSVM. www.csie.tu.edu.tw/ cjli/ libsvm,,,. 2005.., 13(3): 238 241,,. 2001.., 16(4): 544 548,,. 2006.. :,,. 2006.. :,,. 1987.., 11(2): 106 112,,. 2006.., 51(6): 700 706