Generating Daily Synthetic Landsat Imagery by Combining Landsat and MODIS Data

Transcription

1 Sensors 2015, 15, ; do: /s Artcle OPEN ACCESS sensors ISSN Generatng Daly Synthetc Landsat Imagery by Combnng Landsat and MODIS Data Mngquan Wu 1, *, Wenjang Huang 2, Zheng Nu 1 and Changyao Wang 1 1 The State Key Laboratory of Remote Sensng Scence, Insttute of Remote Sensng and Dgtal Earth, Chnese Academy of Scences, Bejng , Chna; E-Mals: nuzheng@rad.ac.cn (Z.N.); wangcy508@rad.ac.cn (C.W.) 2 Laboratory of Dgtal Earth Scences, Insttute of Remote Sensng and Dgtal Earth, Chnese Academy of Scences, Bejng , Chna; E-Mal: huangwj@rad.ac.cn * Author to whom correspondence should be addressed; E-Mal: wumq@rad.ac.cn; Tel./Fax: Academc Edtor: Assefa M. Melesse Receved: 14 July 2015 / Accepted: 14 September 2015 / Publshed: 18 September 2015 Abstract: Owng to low temporal resoluton and cloud nterference, there s a shortage of hgh spatal resoluton remote sensng data. To address ths problem, ths study ntroduces a modfed spatal and temporal data fuson approach (MSTDFA) to generate daly synthetc Landsat magery. Ths algorthm was desgned to avod the lmtatons of the condtonal spatal temporal data fuson approach (STDFA) ncludng the constant wndow for dsaggregaton and the sensor dfference. An adaptve wndow sze selecton method s proposed n ths study to select the best wndow sze and movng steps for the dsaggregaton of coarse pxels. The lnear regresson method s used to remove the nfluence of dfferences n sensor systems usng dsaggregated mean coarse reflectance by testng and valdaton n two study areas located n Xnjang Provnce, Chna. The results show that the MSTDFA algorthm can generate daly synthetc Landsat magery wth a hgh correlaton coeffcent (R) ranged from to between synthetc mages and the actual observatons. We further show that MSTDFA can be appled to 250 m 16-day MODIS MOD13Q1 products and the Landsat Normalzed Dfferent Vegetaton Index (NDVI) data by generatng a synthetc NDVI mage hghly smlar to actual Landsat NDVI observaton wth a hgh R of Keywords: spatal and temporal data fuson; remote sensng; MODIS; Landsat; FROM-GLC

2 Sensors 2015, Introducton Because Earth surface observatons can be obtaned perodcally by satellte remote sensng, ths technology has become a foremost technque for montorng land surface processes [1]. Snce 1978, satellte remote sensng of the land surface process has been domnated by polar-orbtng sensors ncludng the Advanced Very Hgh Resoluton Radometer (AVHRR) [2], Systeme Pour l Observaton de la Terre (SPOT) Vegetaton (VGT) [3], and the Moderate Resoluton Imagng Spectroradometer (MODIS) [4]. The temporal resoluton of these coarse resoluton sensors s one two days, whch means they can mage the entre Earth at one two-day ntervals. Owng to ther hgh temporal resolutons, the tme seres data of these sensors are wdely used n land surface processes dynamc montorng applcatons [5 8] such as land cover and land change [9,10], crop mappng and producton forecasts [11,12], dsasters such as fres [13], floods [14], and algal blooms, and forest [15] and grassland [16] ecosystems. However, these hgh temporal resoluton data have low spatal resoluton; therefore, the sgnals recorded by these sensors are not sutable for hghly spatally varant land surface processes montorng. Medum spatal resoluton sensors such as the Thematc Mapper (TM), Enhanced Thematc Mapper Plus (ETM+), and Operatonal Land Imager (OLI) sensors on Landsat satelltes are other types satellte data wdely used n land surface montorng for applcatons such as detaled land use and land cover mappng [17,18], envronmental montorng [19], and ecologcal system dynamc montorng [20,21]. However, these data have low temporal resoluton and cloud nterference; therefore, ther applcaton to land surface montorng leads to a shortage of vald data. Lecke [22] found that the probablty of acqurng cloud-free Landsat mages for a gven tme wth cloud cover <10% can be as low as 10%. Therefore, no sngle satellte can provde data to meet the challenges of hgh spatal and temporal land surface process montorng. To address ths problem, several spatal and temporal data fuson approaches have been proposed to generate hgh spatal and temporal data by fusng coarse and medum spatal resoluton data. Gao et al. [23] ntroduced the spatal and temporal adaptve reflectance fuson model (STARFM) for blendng MODIS and Landsat magery, and Roy et al. [24] used a sem-physcal fuson approach to fuse mult-temporal MODIS Landsat data. Several studes have appled the STARFM to urban envronmental varable extracton, vegetated dry-land ecosystem montorng, publc health studes, and daly land surface temperature generaton [25 29]. Hlker et al. [25] mproved the STARFM for the spatal temporal adaptve algorthm for mappng reflectance change (STAARCH) for producng synthetc magery and detectng changes. Zhu et al. [30] enhanced the STARFM for complex heterogeneous regons. Emelyanova et al. [31] assessed the accuracy of STARFM and ESTARFM n two landscapes wth contrastng spatal and temporal dynamcs. However, most of these models are based on the assumpton that the change n reflectance r of each land cover class s lnear, whch s not accurate for land cover types such as vegetaton [32]. In response, other scholars have proposed methods based on a lnear mxed model that assumes that the reflectance of each coarse spatal resoluton pxel s a lnear combnaton of the responses of each land cover class contrbutng to the mxture [33 35]. However, owng to dfferences n envronmental factors such as alttude, morphology, and sol type or management factors such as sowng date and fertlzaton, ths assumpton s not always vald. Zhukov et al. [36] and Masell [37] addressed ths problem by usng the neghborng pxel nformaton based on the assumpton that spectral propertes of

3 Sensors 2015, a land cover class do not show great varatons n the surroundng pxels. On the bass of ther work [36,37], Busetto et al. [32] developed a new method n whch the neghborng pxels are selected and weghted on the bass of ther Eucldean dstances from the target pxel; ths method also consders the spectral smlarty of the subcomponents wth those of the targets. However, the reflectance dsaggregated by these proposed methods s the mean reflectance of each land cover class n the dentfed subset s, whch s stll not equal to the real surface reflectance r of fne-resoluton pxels. To solve ths problem, Wu et al. [38] proposed a spatal and temporal data fuson approach (STDFA) based on the assumpton that the temporal varaton propertes of each fne pxel n the same class are constant. They appled ths method to the estmaton of the hgh spatal and temporal resoluton land surface temperature [39] and leaf area ndex [40]. They also valdated that ESTARFM and STDFA can be appled to combne Huanjng (HJ) charge coupled devce (CCD) and MODIS reflectance data together wth Gaofen satellte no. 1 (GF-1) wde feld of vew camera (WFV) and MODIS reflectance data [41]. Gevaert and García-Haro [42] ntroduced an unmxng-based algorthm and compared t wth STARFM. They recommended usng unmxng-based data fuson for stuatons n whch the spectral characterstcs of the medum-resoluton nput mages are downscaled. However, STDFA has several lmtatons. The dfferences n sensor systems are not consdered, and the wndow szes used to select coarse pxels nvolved n the soluton of the lnear mxed models s fxed. The best wndow szes for dfferent land cover classes may vary accordng to the dfferent spatal dstrbuton of each land cover class; a constant wndow sze may result n lower accuracy soluton of the lnear mxed models of some land cover classes. To address these lmtatons, the objectves of the present study are (1) to modfy STDFA by ntroducng sensor dfference correcton and adaptve wndow sze selecton methods; (2) to test and analyze the applcablty of the modfed spatal and temporal data fuson approach (MSTDFA) n other data such as MODIS MOD13Q1 products and Landsat Normalzed Dfferent Vegetaton Index (NDVI) data; and (3) to test and analyze the avalablty of Fner Resoluton Observaton and Montorng of Global Land Cover (FROM GLC) data n MSTDFA. 2. Methods 2.1. Method Inputs and Processng Steps To address the weaknesses of STDFA n the estmaton of daly synthetc Landsat magery, MSTDFA s proposed n ths paper. The nput of ths algorthm ncludes a Landsat mage, land cover and tme seres MODIS reflectance data that were acqured on the same day as was the Landsat mage, and at least one MODIS reflectance dataset acqured on the same day as was the Landsat mage selected for smulaton. The output of ths algorthm s a tme seres of synthetc Landsat mageres n whch the acquston date can be consdered the same as that of the tme seres MODIS data. The algorthm ncludes four steps: (1) best wndow sze selecton; (2) mean reflectance dsaggregaton; (3) sensor dfference adjustment; and (4) calculaton of each pxel s reflectance and outputtng of the daly synthetc Landsat magery. A flowchart of the algorthm s shown n Fgure 1. The algorthm s run by one band. To generate a multspectral synthetc Landsat magery, we need to apply ths method for each band, respectvely.

4 Sensors 2015, MODIS scene (T1) Land cover Map Landsat Image (T1) Next wndow sze Intalzaton Lnear regresson Inverson Dsaggregated mean surface reflectance Fractonal covers Mean reflectance of Landsat mage Synthetc Landsat mage (T2) Synthetc Landsat mage (Tn) Surface reflectance calculaton for every Landsat pxel 4) Calculaton of pxels reflectance correlaton coeffcent calculaton Lnear regresson Inverson Landsat mean reflectance Regresson Model of every wndow sze correlaton coeffcent of every wndow sze 3)Sensor dfference adjustment Best wndow sze selecton 1) Best wndow sze selecton MODIS scene (T1) Dsaggregated mean surface reflectance (T1) Mean reflectance of Landsat mage (T1) MODIS scene (T2) Lnear regresson Inverson Dsaggregated mean surface reflectance (T2) Mean reflectance of Landsat mage (T2) MODIS scene (Tn) Dsaggregated mean surface reflectance (Tn) Mean reflectance of Landsat mage (Tn) 2) Mean reflectance dsaggregaton Fgure 1. Flowchart of the modfed spatal and temporal data fuson approach (MSTDFA) algorthm. The processng steps of the three man blocks are explaned n Sectons Selectng the Best Wndow Sze Accordng to the unmxng theory, the reflectance of a coarse-resoluton spatal pxel s assumed to be a lnear combnaton of the responses of each land cover class contrbutng to the mxture [33]. The coarse spatal reflectance R (, t) of the landscape thus conssts of k dscrete land cover class c weghted by ther class fractonal cover as Constraned: k c 0 k R(, t ) f (, c) r ( c, t ) (, t ) (1) c f c 0 fc(, c) 1 and 1 f ( c, c ) 0 for all, where fc(, c ) s the fractonal cover of class c n coarse pxel, whch s usually assumed to not change over tme; r f ( c, t ) s the mean reflectance of fne-resoluton homogeneous pxels belongng to land cover class c at tme t; and ( t, ) s the resdual error term. If we know the coarse spatal reflectance R (, t) and the fractonal cover values, Equaton (1) can be solved wth the ordnary least squares technque and by generatng the dsaggregated mean surface reflectance value r f ( c, t ) for class c at tme t. Generally, the fractonal cover values were extracted from hgh-resoluton spatal land cover map.

5 Sensors 2015, Then, by nputtng the fractonal cover values and the tme seres coarse spatal reflectance from tme t1 to tme tn, the tme seres mean surface reflectance value r f ( c, t ) for each class was calculated by solvng Equaton (1) usng the ordnary least squares technque. Snce ths solved mean surface reflectance was dsaggregated by the ordnary least squares technque, we defned ths dsaggregated mean surface reflectance as r d ( c, t ) to dstngush wth the mean surface reflectance r f ( c, t ) calculated from real fne-resoluton pxels. The tme seres mean surface reflectance value r f ( c, t ) for each class ranges from 0 to 1. Dsaggregated mean surface reflectance out of ths constrant were not used to buld a lnear model between the dsaggregated mean surface reflectance and actual TM mean surface reflectance for the adjustment of sensor dfference. To reduce the nfluence of spatal varaton and geolocaton errors, the soluton of Equaton (1) was conducted n a rectangular wndow centered to the MODIS target pxel. Owng to the dfferent spatal dstrbuton of each land cover class, the best wndow sze s for each land cover class may be dfferent. To determne the best wndow sze s for each land cover class, rectangles of lengths of MODIS pxels for dfferent classes were tested. Frstly, for wndow length l, a subset s of MODIS pxels centered to the MODIS target pxel mk, fractonal cover data, and Landsat pxels was extracted. Then, the dsaggregated mean surface reflectance value r d ( c, t, m k) and real mean surface reflectance value of Landsat pxels r f ( c, t, m k) for the MODIS target pxel mk were calculated. Thrdly, along wth the MODIS target pxel beng moved to cover the entre r f c, t, m and a dsaggregated mean coarse MODIS mage, a real mean fne reflectance vector reflectance vector r d,, correlaton coeffcent between those two vectors for wndow length l. r f c, t, m and r d,, descrbed as follows: c t m par for land cover class c was calculated to allow calculaton of the c t m were r f c, t, m ( r f c, t, m1, r f c, t, m2,, r f c, t, mn ) (2) rd c, t, m ( rd c, t, m1, rd c, t, m2, rd c, t, mn ) (3) where m1, m2, and mn s the MODIS target pxel; n s the number of the MODIS pxel. Obvously, the best wndow length l for land cover class c wll have the hghest correlaton coeffcent. Thus, the wndow length l wth the hghest correlaton coeffcent R was set to the best wndow length for land cover class c Dsaggregatng Mean Reflectance After the best wndow sze for land cover class c s determned, the mean reflectance for land cover class c of target MODIS pxel mk can be calculated by solvng Equaton (1) usng the ordnary least squares technque by nputtng a subset s for wndow length l of the MODIS pxels centered to the MODIS target pxel, fractonal cover data, and Landsat pxels Adjustng Sensor Dfference Owng to sensor system dfferences n bandwdth, acquston tme, spectral response functons, geolocaton errors, and atmospherc correcton, there s a need to adjust the dsaggregated mean reflectance r d ( c, t ) for land cover class c to the real mean fne reflectance r f ( c, t ). In Secton 2.2, a

6 Sensors 2015, real mean fne reflectance vector r f,, r d,, c t m and dsaggregated mean coarse reflectance vector c t m par for each land cover class was generated. Ths allowed for constructon of a lnear model between the real mean fne reflectance vector and dsaggregated mean coarse reflectance vector by usng lnear regresson analyss, whch can be descrbed as: r f( c, t, m) a r d( c, t, m) b (4) where a and b are coeffcents of the lnear regresson model. Then, ths model was used to calculate real tme seres mean fne reflectance r f ( c, t ) from tme t2 to tme tn from the tme seres dsaggregated mean coarse reflectance r d ( c, t ) Calculatng Pxels Reflectance and Method Outputs Snce the tme seres mean fne reflectance r f ( c, t ) from tme t1 to tme tn was calculated, the tme seres reflectance of each fne resoluton pxel can be determned usng the SRCM model proposed by Wu et al. [38], whch s descrbed as r f( c, t) r f( c, t1) rf ( c, t, k) rf ( c, t1, k) (5) where r f ( c, t ) and r f ( c, t 1) s the mean fne reflectance at tme t and t1, rf( c, t, k ) and rf ( c, t1, k ) s the reflectance of pxel k of class c n target MODIS pxel mk at tme t and t1. Wth the r f ( c, t 1) and r ( c, t, k ) obtaned from the Landsat scene at the tme t1 and tme seres r f ( c, t ) calculated n Secton 2.4, f 1 MSTDFA allows the output of tme seres synthetc 30 m Landsat magery. 3. Method Tests and Results 3.1. Study Area Two study areas located n Xnjang Provnce, West Chna, were selected to test and valdate ths method (Fgure 2). The frst s Bole County, Xnjang Provnce, Chna, located n the valleys between Alatau and Gang Gger mountans, Boertala Rver. The area to the west of Bole s mountanous, whereas that to the east s plans. In the plans area, the man land use type s farmland, n whch the crop plots are usually large. Therefore, the land cover types n ths area are relatvely homogeneous. The second study area s Lunta County, Xnjang Provnce, Chna, located n southern Tanshan, northern Tarm Basn. The areas north of Lunta are hlls, whereas the mddle and the southern parts are plans. The crop plots n Lunta are very small; thus, the landscapes are heterogeneous.

7 Sensors 2015, Data and Pre-Processng Landsat Data and Pre-Processng Fgure 2. Locatons of the study areas. Three Landsat-5 TM datasets n Bole and Three Landsat-8 OLI datasets n Lunta were used n ths study (Table 1). All data were acqured n clear sky condtons and were provded by the Unted States Geologcal Survey (USGS). The Landsat data used n the Bole study area were surface reflectance products, whereas those used n Lunta were Level L1T products. The three Landsat L1T products were atmosphercally corrected by usng the Fast Lne-of-Sght Atmospherc Analyss of Spectral Hypercubes (FLAASH) Atmospherc Correcton Model n software ENVI 5.0. Then, the sx Landsat datasets were georeferenced by usng a second-order polynomal warpng approach based on the selecton of 43 ground control ponts (GCPs) usng a 1:10,000 topographc map by the nearest neghbor resamplng method wth the poston error wthn 0.74 Landsat pxels. Table 1. Satellte mages used n ths study. Study Area Bole Lunta Landsat-5 TM/ Landsat-8 OLI MODIS Acquston Date Path/Row Usage Acquston Date Usage 11 July /29 Reference Classfcaton 12 July 2011 Mean reflectance 27 July /29 Valdaton 28 July 2011 estmaton 13 September /29 Valdaton 14 September September /31 Valdaton 3 September 2013 Mean reflectance 6 October /31 Reference Classfcaton 7 October 2013 estmaton 22 October /31 Valdaton 21 October 2013 Landsat mages acqured on 11 July 2011 n Bole and on 6 October 2013 n Lunta were used as reference mages for buldng a lnear model between Landsat and MODIS mean reflectance and to calculate the reflectance of fne pxels from the mean reflectance. These mages were also used for land cover mappng, whch s explaned n Secton The subsequent Landsat mages were used to evaluate ths algorthm.

8 Sensors 2015, MODIS Data and Pre-Processng Sx daly MODIS surface reflectance products (MOD09GA, 500 m) obtaned n clear-sky condtons were used n ths study (Table 1). Ideally, the MODIS mage acquston date should be the same as the acquston date of Landsat data. However, the qualty of MODIS data acqured on the same date as the Landsat data was not good n Bole and Lunta; therefore, these MODIS data were replaced wth data of good qualty acqured one day earler or later than the Landsat data. These sx MODIS mages were reprojected from the natve Snusodal projecton to a UTM-WGS84 reference system and were reszed to the selected study area usng MODIS Reprojecton Tool (MRT) software. We also reszed the spatal resoluton from 500 m to 480 m wth a nearest neghbor resamplng method n MRT. All of these MODIS data were then georeferenced by a second-order polynomal warpng approach based on the selecton of 38 GCPs on 480 m Landsat mages wth a nearest neghbor resamplng method n whch the poston error was wthn 0.63 MODIS pxels. The 480 m Landsat mages were reszed from georeferenced Landsat mages by usng the pxel aggregate resamplng method Land Cover Data Two types of land cover data were used n ths study. The frst was mapped by usng the maxmum lkelhood classfcaton method from the reference Landsat mages wth 1196 feld survey data ncludng 334 plots n Bole and 862 plots n Lunta. The feld survey data n Bole and Lunta was obtaned n 2011 and 2013, respectvely. These land cover data were used n the generaton of hgh spatal and temporal synthetc Landsat mageres. Currently, abundant global and regonal nformaton of land cover and use are provded, for example, by Natonal Land Cover Database (NLCD) and FROM GLC data. Usually these sources wll be updated every fve years. To test the applcablty of these data n MSTDFA, the FROM GLC data n Bole mapped by usng Landsat-5 TM data acqured on 21 July 2009 was used n MSTDFA. The classfcaton accuracy of the FROM GLC data n Bole was evaluated by usng a confuson matrx wth regons of nterest (ROIs) selected by usng vsual nterpretaton methods and feld survey data. Table 2 shows the accuracy evaluaton results. The overall accuracy and Kappa coeffcent of the FROM GLC data n Bole s 78.06% and 0.65, respectvely. Table 2. Accuracy evaluatons of Fner Resoluton Observaton and Montorng of Global Land Cover (FROM GLC) data n Bole. Class Reference Data Water Forest Grass Shrub Impervous Cropland Bare Land Prod. Acc. (%) User Acc. (%) Water Forest Grass Shrub Impervous Cropland , Bare land

9 Sensors 2015, Results and Accuracy Assessment Results of Landsat Mean Reflectance Regressng A lnear model between Landsat mean reflectance and dsaggregated mean coarse reflectance at tme t1 (Bole: 11 July 2011; Lunta: 6 October 2013) was bult successfully by usng lnear regresson analyss. Table 3 shows the best wndow sze for each band and each land class. Table 4 shows that hgh-correlaton coeffcents R were acqured n these two study areas. Ths result demonstrates that the adaptve wndow sze and movng steps selecton method have the ablty to select the best wndow sze for the dsaggregaton of coarse pxels. Table 3. The best wndow sze for each band and each land class. Bole Best Wndow Sze (MODIS pxels, 500 m) Class Blue Green Red NIR SWIR1 SWIR2 Forest Corn Cotton Desert Bare land Water Buldng land Other crops Lunta Best Wndow Sze (MODIS pxels, 500 m) Class Blue Green Red NIR SWIR1 SWIR2 Cotton Water Buldng land Bare land Desert Corn

10 Sensors 2015, Table 4. Lnear model bult by lnear regresson analyss between Landsat mean reflectance and dsaggregated mean coarse reflectance. Bole: y = x b + a Blue Green Red NIR SWIR1 SWIR2 Class R 2 a b R 2 a b R 2 a b R 2 a b R 2 a b R 2 a b Forest Corn Cotton Desert Bare land Water Buldng land Other crops Lunta: y = x b + a Blue Green Red NIR SWIR1 SWIR2 Class R 2 a b R 2 a b R 2 a b R 2 a b R 2 a b R 2 a b Cotton Water Buldng land Bare land Desert Corn x: dsaggregated mean coarse reflectance; y: Landsat mean reflectance.

11 Sensors 2015, Results of Synthetc Landsat Image Generaton By usng MSTDFA, four synthetc Landsat mages were outputted that contaned sx bands ncludng blue, green, red, near nfrared (NIR), short-wave nfrared 1 (SWIR1), and SWIR2. The acquston date of these data can be consdered the same as that for MODIS. Fgure 3a shows the actual observaton of MODIS surface reflectance on the NIR band acqured on 28 July 2011 n Bole and on 21 October 2013 n Lunta, and Fgure 3b shows the synthetc surface reflectance magery on the NIR band generated by MSTDFA n the two study areas. Fgure 3c shows the actual observaton of Landsat NIR band surface reflectance acqured on 27 July 2011 n Bole and on 22 October 2013 n Lunta. Through vsual nterpretaton, we determned that the synthetc and actual Landsat data are hghly smlar and were unable to be dstngushed wth the unaded eye. Fgure 3. Comparson of Near-nfrared (NIR) band surface reflectance data of the Moderate Resoluton Imagng Spectroradometer (MODIS; left), synthetc Landsat mage (mddle); and actual Landsat mage (rght) acqured on 11 July 2011 n Bole (upper panels) and 6 October 2013 n Lunta (lower panels), respectvely Accuracy Assessment In ths study, four actual observatons of Landsat data were used to evaluate the accuracy. Closer smlarty of the synthetc Landsat mage to the actual mage relates to the hgher precson of the method. To quanttatvely evaluate the smlarty between the actual observatons and synthetc mages, correlaton analyss was used to calculate the correlaton between the synthetc Landsat magery and the actual observaton of Landsat data. Several ndcators such as the coeffcent (R), varance, mean absolute dfference (MAD), bas, and RMSE were calculated. Table 5 shows the results of ths analyss. As ndcated n Table 5, MSTDFA can generate synthetc Landsat mages wth hgh smlarty to the actual mages. Most synthetc Landsat mages had a hgh correlaton wth the actual Landsat mageres

12 Sensors 2015, wth a coeffcent (R) hgher than Fgure 4 shows the scatter plots between the actual and synthetc Landsat mages, whch were close to the 1:1 lne. These results ndcate a hgh smlarty between the actual and synthetc Landsat data and that MSTDFA has hgh accuracy n generatng synthetc Landsat mages. Table 5. Results of correlaton analyss between synthetc and actual Landsat mageres. Study Area Bole Lunta Date 27 July September 2013 Parameters R Var MAD RMSE Bas R Var MAD RMSE Bas Blue < Green < Red < NIR < SWIR < SWIR < Date 13 September October 2013 Parameters R Var MAD RMSE Bas R Var MAD RMSE Bas Blue < Green < < Red < < NIR < SWIR < < SWIR < < Fgure 4. Cont.

13 Sensors 2015, Fgure 4. Scatter plots between the actual and synthetc magers of Landsat at (a) Bole and (b) Lunta by usng the modfed spatal and temporal data fuson approach (MSTDFA). 4. Dscusson 4.1. Comparson to STDFA Table 6 and Fgure 5 show the results of the condtonal STDFA. It s evdent that n the Lunta area, MSTDFA had better accuracy than the condtonal STDFA n nearly all parameters. In the Bole area, MSTDFA had much better accuracy than the condtonal STDFA n SWIR1 and SWIR2 bands. However, the precson of MSTDFA was slghtly lower n the blue, green, and red bands. A comparson of Fgures 4 and 5 revealed that n the Lunta area, the scatter dagram of MSTDFA was closer to the 1:1 lne than that of STDFA. Two factors can explan these results. Frstly, MSTDFA has two mportant mprovements over STDFA. In partcular, the dfferences n sensor systems are consdered n MSTDFA. In Fgure 5, a hgh correlaton s shown between the synthetc and actual Landsat mages. However, a certan devaton appeared between the regresson lne and the 1:1 lne that was caused manly by dfferences n the sensor systems. Therefore, these devatons were elmnated n the scatter plots of MSTDFA. Secondly, many plots ndcate that land cover type changed n the Bole area from 11 to 27 July For example, the northwest corner of the study area s Ab Lake whch s a huge shallow lake. In addton, Boertala Rver flows east from the west of the study area. As the water level changed, a lot of bare land changed nto water. These changes n land cover type wll reduce the precson of the model, and they had a more severe nfluence for MSTDFA. In STDFA, two days of Landsat mages were used to detect the land cover change areas. The land cover change areas were classfed as other class. However, only one Landsat mage was used n MSTDFA and the change area cannot be detected. So, the land cover type change leads to more reflectance changes of blue, green, and red bands. Thus,

14 Sensors 2015, MSTDFA performed worse than STDFA n these bands. Detals of the nfluence of land cover change are descrbed n Secton 4.5. Table 6. Results of correlaton analyss between synthetc and actual Landsat mageres by usng the spatal and temporal data fuson approach (STDFA). Study Area Bole Lunta Date 27 July September 2013 Parameters R Var MAD RMSE Bas R Var MAD RMSE Bas Blue < Green < < Red < NIR < SWIR < SWIR < Date 27 July October 2013 Parameters R Var MAD RMSE Bas R Var MAD RMSE Bas Blue < Green < < Red < < NIR < SWIR < SWIR < Fgure 5. Cont.

15 Sensors 2015, Fgure 5. Scatter plots between actual and synthetc mages of Landsat n (a) Bole and (b) Lunta by usng the spatal and temporal data fuson approach (STDFA) Improvement Compared wth tradtonal STDFA, MSTDFA has two mprovements. Frstly, n the tradtonal STDFA, the best wndow sze for the soluton of Equaton (1) was set to a fxed value of 40 coarse pxels. However, owng to the dfferent spatal dstrbuton of each land cover class, the best wndow sze for each land cover class may be dfferent. Thus, a fxed wndow sze may not be approprate for all classes. In MSTDFA, an adaptve wndow sze and movng step length selecton method was used to avod ths problem. Ths method tested every wndow sze and moved the step length from mnmum to maxmum by usng the exhaustve method, and the correlaton coeffcent between the reference Landsat mean reflectance and the MODIS dsaggregated mean coarse reflectance of every test was calculated. The best wndow sze and movng step length were desgned to be those whch can lead to the maxmum correlaton coeffcent between the fne and coarse mean reflectance. Secondly, the dfferences n sensor systems are not consdered n tradtonal STDFA. In MSTDFA, the sensor dfferences were removed by usng lnear models between the Landsat mean reflectance and the dsaggregated mean reflectance. To show the mprovements by the above steps, we tested those methods step by step n Lunta at NIR band. Frstly, the basc STDFA model wth a fxed wndow sze of 40 MODIS pxels and wthout sensor dfference adjustment was used to generate a synthetc NIR mage. Then the adaptve wndow sze selecton method and sensor dfference adjustment were added step by step to generate a synthetc NIR mage. Fnally, the smlarty between those synthetc NIR mageres and the actual Landsat NIR mage were evaluated by correlaton analyss. Table 7 shows the results of correlaton analyss. From Table 7, we can see mprovements of correlaton coeffcent R and declnes of Varance, RMSE, MAD, and bas n each step. So, the tradtonal STDFA methods were enhanced by those mprovements.

16 Sensors 2015, Table 7. Accuracy mprovements of every step n MSTDFA method of NIR band n Lunta. Parameters STDFA Adaptve Wndow Sze Selecton Sensor Adjustment Total R Varance MAD RMSE Bas Landsat and MODIS Fuson Usng FROM GLC Data To determne the applcablty of the FROM GLC data n MSTDFA, the FROM GLC data n Bole mapped by usng Landsat-5 TM data acqured on 21 July 2009 was used n MSTDFA. By nputtng FROM GLC data rather than classfcaton data mapped usng the maxmum lkelhood classfcaton method, sx synthetc Landsat mages ware generated. Table 8 shows the results of accuracy assessment of these sx synthetc Landsat mages. As ndcated n Table 8, most synthetc Landsat mages had a hgh correlaton wth the actual Landsat mageres wth coeffcent R hgher than Fgure 6 shows the scatter plots between the actual and synthetc Landsat mages, whch were close to the 1:1 lne. These results ndcate a hgh smlarty between the actual and synthetc Landsat data and that MSTDFA has hgh accuracy n generatng synthetc Landsat mages. Therefore, the FROM GLC data can be used n MSTDFA. Table 8. Results of correlaton analyss between synthetc and actual Landsat mageres usng the modfed spatal and temporal data fuson approach (MSTDFA) wth nput of Fner Resoluton Observaton and Montorng of Global Land Cover (FROM GLC) data. Parameters R Varance MAD RMSE Bas Blue Green Red NIR SWIR SWIR Fgure 6. Cont.

17 Sensors 2015, Fgure 6. Scatter plots between the actual and synthetc mageres of Landsat usng modfed spatal and temporal data fuson approach (MSTDFA) wth nput of Fner Resoluton Observaton and Montorng of Global Land Cover (FROM GLC) Influence of the Image Extents The Ordnary Least Squares technque was used to dsaggregate the tme seres mean surface reflectance value r ( c, t ) for each class. However, the solutons wll affect the number of pxels an d mage contans from two aspects. It s easer to generate outlers n the soluton of Ordnary Least Squares n a small area. In addton, the mages wth dfferent extents wll produce dfferent dsaggregated mean reflectance. These effects were reduced by three steps n ths method. Frstly, abnormal dsaggregated mean surface reflectance value was not used. Then, all the normal dsaggregated mean reflectance for every target MODIS pxels was used to adjust the sensor dfference. Fnally, the sensor adjusted tme seres mean fne reflectance r f ( c, t, m k) of target MODIS pxels mk was only used for the fuson of fne pxels belongng to target MODIS pxels mk. To evaluate the nfluence of the mage extents, we appled MSTDFA n dfferent mage extents n NIR band acqured on 22 October 2013 n Lunta. Fgure 7 shows that the correlaton coeffcent R has a logarthmc relatonshp wth the sze of appled area. The bgger of the study area, the hgher the correlaton coeffcent R. So, we recommend usng the MSTARFM model n a large area and usng the MSTDFA model n areas greater than MODIS pxels. Fgure 7. Relatonshp between correlaton coeffcent R and study area sze.

18 Sensors 2015, Comparson of Actual NDVI and NDVI Calculated Usng Synthetc Data Ths method was appled to MODIS and Landsat NDVI data n Bole to test ts applcablty wth dfferent data. The MODIS NDVI data ncluded 250 m 16-day MOD13Q1 products acqured on 12 July 2011 and 28 July The Landsat NDVI data was calculated by usng the red and NIR bands of Landsat data acqured on 11 July By nputtng these NDVI data and the land cover data mapped by usng the maxmum lkelhood classfcaton method, a synthetc Landsat NDVI mage was generated by usng MSTDFA and condtonal STDFA. Table 9 shows the accuracy assessment results of the two synthetc Landsat NDVI mages. Both STDFA and MSTDFA can be used to fuse NDVI data, and the results of the latter were better than those of the former. Another method to generate synthetc NDVI data s to calculate NDVI usng the synthetc red and NIR data generated by STDFA and MSTDFA. Wu et al. [41] compared the two methods n the generaton of synthetc NDVI and leaf area ndex. Table 9. Accuracy assessment result comparson of Landsat Normalzed Dfferent Vegetaton Index (NDVI) fuson applcaton Lmtatons of the Method NDVI Generated by MSTDFA NDVI Generated by STDFA R Varance MAD RMSE Bas Although MSTDFA can generate daly synthetc Landsat mages wth hgh smlarty to actual Landsat mages, ths method has weaknesses. Frstly, all the spatal and temporal data fuson methods are based on the assumpton that the land cover classes do not change over tme. Weng et al. [28] found that the accuracy of those models wll be substantally reduced f ths assumpton s volated. As dscussed n Secton 4.2, many plots nclude land cover changes over tme n the Bole area. For example, the northwest corner of the study area s Ab Lake whch s a huge shallow lake. In addton, Boertala Rver flows east from the west of the study area. As the water level changes, a lot of bare land changes nto water. To evaluate the nfluence of land cover change, we extracted the land change plots by usng the dfferences n Landsat mages acqured on 11 July 2011 and 27 July The resdual mage between the synthetc and actual mages was also calculated. Then, we calculated the correlaton coeffcent between these two data sets to obtan a hgh correlaton coeffcent R of We also found that pxels wth large errors were located n the plots n whch the land cover types changed. Therefore, land cover class change has a very mportant nfluence n MSTDFA. To assess the nfluence, we masked these plots and calculated the accuracy assessment parameters agan, as shown n Table 10. A comparson of Tables 5 and 10 revealed that the maxmum fuson accuracy mprovement was 0.32 n correlaton coeffcent R f the plots are not consdered.

19 Sensors 2015, Table 10. Accuracy assessment result wth no consderaton of land cover change plots. R Varance MAD RMSE Bas Blue < Green < Red < NIR < SWIR < SWIR < Secondly, ths method consdered only the fuson of mult-sensor optcal mages. Therefore, t cannot provde effectve data under cloudy condtons because all optcal satelltes are affected by clouds. In ths stuaton, radar satellte must be consdered as a soluton [43 48]. Thrdly, a smple method n whch solutons out of the range of 0% 100% were not used was employed to satsfy the constrants of the solutons of the lnear mxed model. Optmzaton algorthms, such as the normalzaton algorthm whch uses all solutons, mnmzng the squared errors between the predcted response varable and observed data, can be used to mprove ths method [49]. Furthermore, nonlnear least squares regressons for spectral quanttatve analyss s also an mportant drecton to modfy ths method for future research [50]. 5. Conclusons In ths study, an MSTDFA was developed and valdated for two study areas located n Xnjang, Chna. By nputtng MODIS reflectance data, Landsat data, and land cover data, ths method s able to generate daly synthetc Landsat mages n whch the spatal resoluton s the same as that of the Landsat data and the temporal resoluton s the same as that of the MODIS data. A comparson wth the actual Landsat mage revealed the followng fndngs: (1) The adaptve wndow sze and movng step selecton method can select the best wndow sze for dsaggregaton of coarse pxels. The dsaggregated mean coarse reflectance had a strong lnear relatonshp wth the Landsat mean reflectance. (2) MSTDFA had hgher accuracy than STDFA but was more easly nfluenced by land cover change. Land cover data such as that of FROM-GLC can be used n MSTDFA. Synthetc Landsat mages wth hgh smlarty to actual Landsat mages wth a correlaton coeffcent R of 0.96 can be generated. (3) Land cover class change had a very mportant nfluence n MSTDFA, whch can lead to a reducton n the correlaton coeffcent R of 0.32 n the blue band. (4) MSTDFA can be appled n 250 m 16-day MODIS MOD13Q1 products and Landsat NDVI data. A synthetc NDVI mage wth very hgh smlarty to the actual NDVI observaton wth a hgh correlaton coeffcent R of 0.97 can be generated. Acknowledgments Ths work was supported by the Natonal Natural Scence Foundaton of Chna ( ), the Natonal Scence and Technology Major Project (2014AA06A511), the Major State Basc Research Development Program of Chna (2013CB733405, 2010CB950603), the Natonal Scence and

20 Sensors 2015, Technology Major Project of Chna, and the Yunnan Provncal Scence and Technology Program (2010AD004). The funders had no role n choosng the study desgn, n the collecton, analyss, and nterpretaton of the data, n the wrtng of the report, or n the decson to submt the artcle for publcaton. Author Contrbutons Mngquan Wu, Wenjang Huang, Zheng Nu and Changyao Wang conceved and desgned the experments. Mngquan Wu performed the experments and the manuscrpt draft. Wenjang Huang, Zheng Nu and Changyao Wang revsed the manuscrpt draft. All authors read and approved the fnal verson. Conflcts of Interest The authors declare no conflct of nterest. Abbrevatons MSTDFA Modfed Spatal and Temporal Data Fuson Approach MODIS Moderate Resoluton Imagng Spectroradometer STDFA Spatal and Temporal Data Fuson Approach NDVI Normalzed Dfferent Vegetaton Index AVHRR Advanced Very Hgh Resoluton Radometer SPOT Systeme Pour l Observaton de la Terre VGT Vegetaton TM Thematc Mapper ETM+ Enhanced Thematc Mapper Plus OLI Operatonal Land Imager STARFM Spatal and Temporal Adaptve Reflectance Fuson Model STAARCH Spatal Temporal Adaptve Algorthm for mappng Reflectance Change ESTARFM Enhanced Spatal and Temporal Adaptve Reflectance Fuson Model FROM-GLC Fner Resoluton Observaton and Montorng of Global Land Cover HJ Huanjng CCD Charge Coupled Devce GF-1 Gaofen satellte No. 1 WFV Wde Feld of Vew camera USGS Unted States Geologcal Survey FLAASH Fast Lne-of-Sght Atmospherc Analyss of Spectral Hypercubes GCPs Ground Control Ponts MRT MODIS Reprojecton Tool NLCD Natonal Land Cover Database NIR Near Infrared Reflecton SWIR shortwave nfrared MAD Mean Absolute Dfference RMSE Root Mean Square Error

21 Sensors 2015, References 1. Fensholt, R.; Anyamba, A.; Huber, S.; Proud, S.R.; Tucker, C.J.; Small, J.; Pak, E.; Rasmussen, M.O.; Sandholt, I.; Shsanya, C. Analysng the advantages of hgh temporal resoluton geostatonary MSG SEVIRI data compared to Polar Operatonal Envronmental Satellte data for land surface montorng n Afrca. Int. J. Appl. Earth Obs. Geonf. 2011, 13, Polt, E.; Cutler, M.E.J.; Rowan, J.S. Usng the NOAA advanced very hgh resoluton radometer to characterse temporal and spatal trends n water temperature of large European lakes. Remote Sens. Envron. 2012, 126, Masongrande, P.; Duchemn, B.; Dedeu, G. VEGETATION/SPOT: An operatonal msson for the Earth montorng; presentaton of new standard products. Int. J. Remote Sens. 2004, 25, Salomonson, V.V.; Barnes, W.L.; Maymon, P.W.; Montgomery, H.E.; Ostrow, H. MODIS: Advanced faclty nstrument for studes of the earth as a system. IEEE Trans. Geosc. Remote Sens. 1992, 27, Zhang, X.; Sun, R.; Zhang, B.; Tong, Q.X. Land cover classfcaton of the North Chna Plan usng MODIS EVI tme seres. ISPRS J. Photogramm. Remote Sens. 2008, 63, Carrão, H.; Gonçalves, P.; Caetano, M. Contrbuton of multspectral and multtemporal nformaton from MODIS mages to land cover classfcaton. Remote Sens. Envron. 2008, 112, Thenkabal, P.S.; Schull, M.; Turral, H. Ganges and Indus rver basn land use/land cover (LULC) and rrgated area mappng usng contnuous streams of MODIS data. Remote Sens. Envron. 2008, 95, Km, Y.; Kmball, J.S.; Ddan, K.; Henebry, G.M. Response of vegetaton growth and productvty to sprng clmate ndcators n the contermnous Unted States derved from satellte remote sensng data fuson. Agrc. Forest Meteorol. 2014, 194, Estel, S.; Kuemmerle, T.; Alcántara, C.; Levers, C.; Prshchepov, A.; Hostert, P. Mappng farmland abandonment and recultvaton across Europe usng MODIS NDVI tme seres. Remote Sens. Envron. 2015, 163, Eckert, S.; Hüsler, F.; Lnger, H.; Hodel, E. Trend analyss of MODIS NDVI tme seres for detectng land degradaton and regeneraton n Mongola. J. Ard Envron. 2015, 113, Chen, C.F.; Son, N.T.; Chang, L.Y. Montorng of rce croppng ntensty n the upper Mekong Delta, Vetnam usng tme-seres MODIS data. Adv. Space Res. 2012, 49, Zhang, J.; Feng, L.; Yao, F. Improved maze cultvated area estmaton over a large scale combnng MODIS EVI tme seres data and crop phenologcal nformaton. ISPRS J. Photogramm. Remote Sens. 2014, 94, Huesca, M.; Ltago, J.; Merno-de-Mguel, S.; Ccuendez-López-Ocaña, V.; Palacos-Orueta, A. Modelng and forecastng MODIS-based Fre Potental Index on a pxel bass usng tme seres models. Int. J. Appl. Earth Obs. Geonf. 2014, 26, Sakamoto, T.; Nguyen, N.V.; Kotera, A.; Ohno, H.; Ishtsuka, N.; Yokozawa, M. Detectng temporal changes n the extent of annual floodng wthn the Camboda and the Vetnamese Mekong Delta from MODIS tme-seres magery. Remote Sens. Envron. 2007, 109,

22 Sensors 2015, Sulla-Menashe, D.; Kennedy, R.E.; Yang, Z.; Braaten, J.; Krankna, O.N.; Fredl, M.A. Detectng forest dsturbance n the Pacfc Northwest from MODIS tme seres usng temporal segmentaton. Remote Sens. Envron. 2014, 151, Cao, R.; Chen, J.; Shen, M.; Tang, Y. An mproved logstc method for detectng sprng vegetaton phenology n grasslands from MODIS EVI tme-seres data. Agrc. Forest Meteorol. 2015, 200, Gong, P.; Wang, J.; Yu, L.; Zhao, Y.C.; Zhao, Y.Y.; Lang, L.; Nu, Z.; Huang, X.; Fu, H.; Lu, S.; et al. Fner resoluton observaton and montorng of global land cover: Frst mappng results wth Landsat TM and ETM+ data. Int. J. Remote Sens. 2013, 34, Ja, K.; Lang, S.L.; Zhang, N.; We, X.; Gu, X.; Zhao, X.; Yao, Y.; Xe, X. Land cover classfcaton of fner resoluton remote sensng data ntegratng temporal features from tme seres coarser resoluton data. ISPRS J. Photogramm. Remote Sens. 2014, 93, Fu, D.J.; Chen, B.Z.; Zhang, H.F.; Wang, J.; Black, T.A.; Amro, B.D.; Bohrer, G.; Bolstad, P.; Coulter, R.; Rahman, A.F.; et al. Estmatng landscape net ecosystem exchange at hgh spatal-temporal resoluton based on Landsat data, an mproved upscalng model framework, and eddy covarance flux measurements. Remote Sens. Envron. 2014, 141, Jenffer, K.; Su, Z.B.; Wolda, T.; Maathus, B. Estmaton of spatal-temporal ranfall dstrbuton usng remote sensng technques: A case study of Makanya catchment, Tanzana. Int. J. Appl. Earth Obs. Geonf. 2010, 12, S90 S Du, P.J.; Lu, S.C.; Xa, J.S.; Zhao, Y. Informaton fuson technques for change detecton from mult-temporal remote sensng mages. Inf. Fuson 2013, 14, Lecke, D. Advances n remote sensng technologes for forest surveys and management. Can. J. Forest Res. 1990, 21, Gao, F.; Masek, J.; Schwaller, M.; Hall, F. On the blendng of the Landsat and MODIS surface reflectance: Predctng daly Landsat surface reflectance. IEEE Trans. Geosc. Remote Sens. 2006, 44, Roy, D.P.; Ju, J.C.; Lews, P.; Schaaf, C.; Gao, F.; Hansen, M.; Lndqust, E. Mult-temporal MODIS Landsat data fuson for relatve radometrc normalzaton, gap fllng, and predcton of Landsat data. Remote Sens. Envron. 2008, 112, Hlker, T.; Wulder, M.A.; Coops, N.C.; Lnke, J.; McDermd, G.; Masek, J.G.; Gao, F.; Whte, J.C. A new data fuson model for hgh spatal- and temporal- resoluton mappng of forest dsturbance based on Landsat and MODIS. Remote Sens. Envron. 2009, 113, Lu, H.; Weng, Q.H. Enhancng temporal resoluton of satellte magery for publc health studes: A case study of West Nle Vrus outbreak n Los Angeles n Remote Sens. Envron. 2012, 117, Walker, J.J.; Beurs, K.M.; Wynne, R.H.; Gao, F. Evaluaton of Landsat and MODIS data fuson products for analyss of dry land forest phonology. Remote Sens. Envron. 2012,117, Weng, Q.H.; Fua, P.; Gao, F. Generatng daly land surface temperature at Landsat resoluton by fusng Landsat and MODIS data. Remote Sens. Envron. 2014, 145, Schmdt, M.; Lucas, R.; Buntng, P.; Verbesselt, J.; Armston, J. Mult-resoluton tme seres magery for forest dsturbance and regrowth montorng n Queensland, Australa. Remote Sens. Envron. 2015, 158,

23 Sensors 2015, Zhu, X.L.; Chen, J.; Gao, F.; Chen, X.; Masek, J.G. An enhanced spatal and temporal adaptve reflectance fuson model for complex heterogeneous regons. Remote Sens. Envron. 2010, 114, Emelyanova, I.V.; McVcar, T.R.; van Nel, T.G.; L, L.T.; van Djk, A.I. Assessng the accuracy of blendng Landsat MODIS surface reflectances n two landscapes wth contrastng spatal and temporal dynamcs: A framework for algorthm selecton. Remote Sens. Envron. 2013, 133, Busetto, L.; Meronb, M.; Colombo, R. Combnng medum and coarse spatal resoluton satellte data to mprove the estmaton of sub-pxel NDVI tme seres. Remote Sens. Envron. 2008, 112, Settle, J.J.; Drake, N.A. Lnear mxng and the estmaton of groundcover proporton. Int. J. Remote Sens. 1993, 14, Masell, F.; Glabert, M.A.; Conese, C. Integraton of hgh and low resoluton NDVI data for montorng vegetaton n Medterranean envronments. Remote Sens. Envron. 1998, 63, Duran, O.; Petrou, M. Subpxel temporal spectral magng. Pattern Recognt. Lett. 2014, 48, Zhukov, B.; Oertel, D.; Lanzl, F.; Renhackel, G. Unmxng-based multsensor multresoluton mage fuson. IEEE Trans. Geosc. Remote Sens. 1999, 37, Masell, F. Defnton of spatally varable spectral end members by locally calbrated multvarate regresson analyses. Remote Sens. Envron. 2001, 75, Wu, M.Q.; Nu, Z.; Wang, C.Y.; Wu, C.Y.; Wang, L. The use of MODIS and Landsat tme seres data to generate hgh resoluton temporal synthetc Landsat data usng a spatal and temporal reflectance fuson model. J. Appl. Remote Sens. 2012, do: /1.jrs Wu, M.Q.; L, H.; Huang, W.J.; Nu, Z.; Wang, C.Y. Generatng daly hgh spatal land surface temperatures by combnng ASTER and MODIS land surface temperature products for envronmental process montorng. Envron. Sc. Process. Impacts 2015, 17, Wu, M.Q.; Wu, C.Y.; Huang, W.J.; Nu, Z.; Wang, C.Y. Hgh-resoluton Leaf Area Index estmaton from synthetc Landsat data generated by a spatal and temporal data fuson model. Comput. Electron. Agrc. 2015, 115, Wu, M.Q.; Huang, W.J.; Nu, Z.; Wang, C.Y. Combnng HJ CCD, GF-1 WFV and MODIS Data to Generate Daly Hgh Spatal Resoluton Synthetc Data for Envronmental Process Montorng. Int. J. Envron. Res. Publc Health 2015, 12, Gevaert, C.M.; García-Haro, J. A comparson of STARFM and an unmxng-based algorthm for Landsat and MODIS data fuson. Remote Sens. Envron. 2015, 156, Reche, J.; Verbesselt, J.; Hoekman, D.; Herold, M. Fusng Landsat and SAR tme seres to detect deforestaton n the tropcs. Remote Sens. Envron. 2015, 156, Lu, M.L.; Lu, X.N.; L, J.; Dng, C.; Jang, J. Evaluatng total norganc ntrogen n coastal waters through fuson of mult-temporal RADARSAT-2 and optcal magery usng random forest algorthm. Int. J. Appl. Earth Obs. Geonf. 2014, 33, Zhang, Y.Z.; Zhang, H.S.; Ln, H. Improvng the mpervous surface estmaton wth combned use of optcal and SAR remote sensng mages. Remote Sens. Envron. 2014, 141, McAlpn, D.; Meyer, F.J. Mult-sensor data fuson for remote sensng of post-eruptve deformaton and depostonal features at Redoubt Volcano. J. Volcanol. Geotherm. Res. 2013, 259,