200

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 1, JANUARY 2008

A Neural Network Technique for Separating Land Surface Emissivity and Temperature From ASTER Imagery
Kebiao Mao, Jiancheng Shi, Senior Member, IEEE, Huajun Tang, Zhao-Liang Li, Xiufeng Wang, and Kun-Shan Chen, Fellow, IEEE

AbstractFour radiative transfer equations for Advanced Spaceborne Thermal Emission and Reection Radiometer (ASTER) bands 11, 12, 13, and 14 are built involving six unknowns (average atmospheric temperature, land surface temperature, and four band emissivities), which is a typical ill-posed problem. The extra equations can be built by using linear or nonlinear relationship between neighbor band emissivities because the emissivity of every land surface type is almost constant for bands 11, 12, 13, and 14. The neural network (NN) can make full use of potential information between band emissivities through training data because the NN simultaneously owns function approximation, classication, optimization computation, and self-study ability. The training database can be built through simulation by MODTRAN4 or can be obtained from the reliable measured data. The average accuracy of the land surface temperature is about 0.24 K, and the average accuracy of emissivity in bands 11, 12, 13, and 14 is under 0.005 for test data. The retrieval result by the NN is, on average, higher by about 0.7 K than the ASTER standard product (AST08), and the application and comparison indicated that the retrieval result is better than the ASTER standard data product. To further evaluate self-study of the NN, the ASTER standard products are assumed as measured data. After using
Manuscript received January 2, 2007; revised May 23, 2007. The work was supported in part by the National Science Foundation of China under Grants 90302008 and 40571101, by the Central Scientic Research Institution for Public Welfare through the Special Fund for Basic Research Work, by the project 863 of China under Grants 2006AA10Z241, and by the Ministry of Agriculture through the Open Fund of the Key Laboratory of Resource Remote Sensing and Digital Agriculture. K. Mao was with the State Key Laboratory of Remote Sensing Science, jointly sponsored by the Institute of Remote Sensing Applications of the Chinese Academy of Sciences and Beijing Normal University, Beijing 100101, China, and also with the Graduate School of the Chinese Academy of Sciences, Beijing 100049, China. He is now with the Key Laboratory of Resources Remote Sensing and Digital Agriculture, Ministry of Agriculture, Hulunber Grassland Ecosystem Observation and Research Station, Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China (e-mail: kebiaomao2004@hotmail.com). J. Shi is with the Institute for Computational Earth System Science, University of California at Santa Barbara, Santa Barbara, CA 93106 USA. H. Tang is with the Key Laboratory of Resources Remote Sensing and Digital Agriculture, Ministry of Agriculture, Hulunber Grassland Ecosystem Observation and Research Station, Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China. Z.-L. Li is with the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China, and also with the Laboratoire des Sciences de lImage, de lInformatique et de la Teledetection (UMR7005), 67412 Illkirch, France. X. Wang is with the Graduate School of Agriculture, Hokkaido University, Sapporo 060-8589, Japan. K.-S. Chen is with Center for Space and Remote Sensing Research, National Central University, Chungli 320, Taiwan, R.O.C. Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TGRS.2007.907333

AST09, AST08, and AST05 (ASTER Standard Data Product) as the compensating training data, the average relative error of the land surface temperature is under 0.1 K relative to the AST08 product, and the average relative error of the emissivity in bands 11, 12, 13, and 14 is under 0.001 relative to AST05, which indicates that the NN owns a powerful self-study ability and is capable of suiting more conditions if more reliable and high-accuracy ASTER standard products can be compensated. Index TermsAdvanced Spaceborne Thermal Emission and Reection Radiometer (ASTER) data, emissivity, land surface temperature (LST).

I. I NTRODUCTION HE Advanced Spaceborne Thermal Emission and Reection Radiometer (ASTER) is an imaging instrument aboard the Terra satellite, which was launched in December 1999 as part of the National Aeronautics and Space Administrations (NASAs) Earth Observing System (EOS). ASTER has 15 bands, which cover the visible, near-infrared, short-wave infrared, and thermal infrared regions, and the spatial resolution is from 15 to 90 m. It is mainly used to obtain detailed maps of land surface temperature (LST), emissivity, reectance, and elevation [1]. Many methods have been developed to retrieve the sea surface temperature and the LST from the National Oceanic and Atmospheric Administration (NOAA)/Advanced Very High Resolution Radiometer (AVHRR) and Moderate Resolution Imaging Spectroradiometer (MODIS) data [2][17]. The study of algorithms for retrieving the LST and emissivity from highresolution thermal images (like ASTER) is not too much [1], [18][21] because it is difcult to obtain the atmospheric parameters (like water vapor content). The retrieval of land surface emissivity and temperature is a typical ill-posed problem in geophysical parameter retrieval because the number of unknown parameters is always at least one more than the number of simultaneous equations that are available for solution. It is very difcult to exactly separate land surface emissivity and temperature from thermal radiance measurement if we do not utilize some prior knowledge. Many people [1], [10], [13], [16], [18], [20][29] made a lot of research for the separation of land surface emissivity and temperature. The detailed introduction for different LST/emissivity separation algorithms has been well discussed by Li and Becker [10] and Gillespie et al. [1]. Three algorithms of them [1], [10], [13] are widely used in application. Li and Becker [10] and

0196-2892/$25.00 2007 IEEE

MAO et al.: NN TECHNIQUE FOR SEPARATING LAND SURFACE EMISSIVITY AND TEMPERATURE

201

Wan and Li [13] proposed a multiband algorithm to retrieve land surface emissivity and LST from a low-resolution image of AVHRR and EOS/MODIS, respectively. These two algorithms assume that the emissivity is the same at daytime and nighttime. Gillespie et al. [1] hybridized three established algorithms, which rst estimated the normalized emissivities and then calculated emissivity ratios. An empirical relationship was used to predict the minimum emissivity from the spectral contrast of the ratio values, which permitted the recovery of the emissivity spectrum. On the other hand, this algorithm used an iterative method to eliminate the inuence of sky irradiance. Liang [28] made some analysis for retrieving the LST and water vapor content from AVHRR thermal imagery by using the neural network (NN). Guo and Zhou [29] used the MODIS LST to evaluate forest re risk of northeast China by using the NN. The retrieval result is not very good because the number of thermal band is too little. The standard algorithm for retrieving land surface emissivity and temperature from ASTER data is a combination of radiance-transfer-based atmospheric correction [30] and temperature/emissivity separation [1]. The accuracy of this algorithm depends on the empirical relationship between emissivity values and spectral contrast, compensation for reected sky irradiance, ASTER data calibration, and atmospheric compensation. In recent years, to eliminate the inuence of atmosphere and calibrate the thermal infrared bands of ASTER, many people did many works [31][34] for the ASTER data. In this study, we rst make an analysis for ill-posed problem and potential information between band emissivities for separating land surface emissivity and temperature from ASTER data in Section II, make some analysis for the NN to separate land surface emissivity and temperature in Section III, and then train and test NN in Section IV. Finally, we will make an evaluation for separating land surface emissivity and temperature in Section V. II. A NALYSIS OF I LL -P OSED P ROBLEM AND P OTENTIAL I NFORMATION The derivation of all algorithms for separating land surface emissivity and temperature is based on the thermal radiance of the ground and its transfer from the ground through the atmosphere to the remote sensor. Generally speaking, the ground is not a blackbody. Thus, ground emissivity has to be considered for computing the thermal radiance emitted by the ground. Atmosphere has important effects on the received radiance at the remote sensor level. Considering all these impacts, the general radiance transfer equation for remote sensing of LST can be formulated as follows:
Bi (Ti ) = i () i ()Bi (Ts ) + (1 i ()) Ri + Ri

Bi (Ts ) is the ground radiance, and Ri and Ri are the downwelling and upwelling atmospheric radiance in channel i, respectively. If we correct the measured radiance at sensor for the effects of the upwelling atmospheric radiance, then we get [10] Bi (Tgi ) = i ()Bi (Ts ) + (1 i ()) Ri

(2)

where Bi (Tgi ) is the radiance in channel i of a blackbody at the surface brightness temperature Tgi at ground level. For Ri , we can use expression (3) to represent it by assuming that the downwelling average atmospheric temperature is Ta , i.e.,
Ri = (1 i ( )) Bi Ta

(3)

where is the downwelling direction of the atmospheric radiance. Ri is slightly relative to the rst term in (2). The AST09 product (ASTER Standard Data Product) provides downwelling atmospheric radiance (Ri ). In fact, this term (Ri ) can be eliminated when we use an iterative method [1], [18] or the NN to solve retrieval equations. To make a simplication for expression (3), we use fi Ta instead of Ri . Thus, expression (2) can be rewritten as
Bi (Tgi ) = i ()Bi (Ts ) + (1 i ()) fi Ta .

(4)

The ASTER Standard Data Product AST09 (Level-2 RadianceTIR, Land_Leaving) has eliminated the upwelling radiance inuence for bands 1014, which lie within the TIR window of 812 m. Band 10 is inuenced by the atmosphere more than the other four thermal bands, and the emissivity in band 10 deviates much from the other four bands. Therefore, we just build four radiance transfer equations for AST09 bands 1114 as follows to retrieve land surface emissivity and temperature:
B11 (T11 ) = 11 ()B11 (Ts ) + (1 11 ()) f11 Ta B12 (T12 ) = 12 ()B12 (Ts ) + (1 12 ()) f12 Ta B13 (T13 ) = 13 ()B13 (Ts ) + (1 13 ()) f13 Ta B14 (T14 ) = 14 ()B14 (Ts ) + (1 14 ()) f14 Ta .

(5a) (5b) (5c) (5d)

(1)

where Ts is the LST, Ti is the brightness temperature in channel i at sensor, i () is the atmospheric transmittance in channel i at viewing direction (zenith angle from nadir), and i () is the ground emissivity in channel i at viewing direction . Bi (Ti ) is the radiance measured by sensor,

In (5), there are six unknowns [four band emissivities (i ), LST (Ts ), and Ta ]. It is impossible to solve (5) if we cannot nd two extra equations or conditions, which is a typical illposed problem [18]. As we all know, the emissivity of every land surface type is almost constant for thermal bands [2][17], which can be made as a classication symbol for distinguishing from other types [35][39]. Mao et al. [17], [20] proposed a method to overcome the ill-posed problem by building the local linear relationship between neighbor band emissivities for every type (soil, vegetation, water, etc.). In fact, we can build a linear or nonlinear relationship between neighbor band emissivities for every land surface type (every spectral curve) and obtain at least six equations for six unknowns [17], [20]. The brightness temperature is determined by emissivity under a given LST. This means that the brightness temperature in a different band owns the potential information between neighbor band emissivities.

202

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 1, JANUARY 2008

We did not make full use of this potential information because it is very difcult to be depicted by a few functions [17], [20]. III. A NALYSIS OF THE NN U SED TO S OLVE THE R ETRIEVAL P ROBLEM As shown from the analysis in Section II, we can get an accurate solution under ideal conditions for four equations with six unknowns if we make full use of the potential information between neighbor band emissivities for every land surface type [20]. In fact, the spectral cure of land surface is more complicated than the analysis in [17] and [20], which may make the solution unstable for general mathematics method because the approximated error can be transferred to another parameter when we solve retrieval equations. In previous retrieval algorithms, many people usually used an iterative or other optimized method to solve retrieval equations [1], [13], [18]. The analysis in [17] and [20] indicates that one or two empirical equations cannot accurately depict the relationship well for most land surface types. On the other hand, the simplication of nonlinear function (like Planck function) may produce some error. The NN is much different from the conventional retrieval algorithm. If the conventional algorithm cannot be exactly known, the costly and time-consuming process of developing the rules of the algorithm must be undertaken [17], [20]. For the cases of parameter retrieval from remote-sensing data, this may be quite difcult due to the many nonlinear and poorly understood factors involved. The NN is composed of a large number of highly interconnected processing elements (neurons) working in parallel to solve a specic problem [41]. The NN works through the interconnection of simple computational elements (or nodes) with activation functions that are usually nonlinear, monotonically increasing, and differentiable [40]. In contrast to conventional methods, the NN does not require that the relationship between the input parameters and the output parameters be known, which directly determines from the training data the relationship between the input to the network and the output from the networks [41]. Furthermore, a NN with only a single hidden layer of a sufcient number of nodes with nonlinear activation function can approximate any continuous scalar function under a given precision and a nite domain [40], [52]. Many studies [40][53] have shown that a multiple-layer network can be competent for classication [41][47] and inversion [41], [43], [48][53]. Due to the NN simultaneously owning function approximation, classication, optimization computation, and self-study ability, we can utilize the NN with simulation data by radiance-transfer model and reliable measured data to separate land surface emissivity and temperature. The trained NN can be taken as a set of nonlinear least mean squares interpolation formula for the discrete set of data points in the training set. Each neuron is adjusted such that the functional approximation created by the NN minimizes the following global error between the desired output and the summed output created by the network: Error = 1 2 [Tji Oji ]2
j i

where Tji is the ith desired output of the jth training pattern, and Oji is the corresponding activation level of the ith unit in the output layer. In the above expression, the values of i are summed only for output units.

IV. T RAINING AND T EST OF NN As shown from the analysis in Sections II and III, the NN is one of the best methods to separate land surface emissivity and temperature. In this section, we will train and test the NN through simulation data by MODTRAN4. The retrieval accuracy of the NN algorithm depends on the training data; therefore, the key point of the NN algorithm is how to build the reliable training database. There are three methods to get training and testing data. The rst method is that we use the radiance transfer model (LOWTRAN [54] and MODTRAN [55]) to simulate the data, which must accord to the characteristic of spectral emissivity (http://speclib.jpl. nasa.gov). The advantage of this method is that it can overcome the measurement error and accurately keep the relationship between geophysical parameters, and the disadvantage is that it does not consider the inuence of environment (like terrain). The second method is that we measure data (land surface emissivity and temperature) in eld when satellite overpasses, which is very good; however, it is very difcult to accurately measure the data. Generally speaking, LST and emissivity vary from point to point on the ground, and ground measurement is generally point measurement. The third method is that we can use the reliable and high-accuracy ASTER standard product in some study regions. In this study, we select the dynamic learning (DL) NN [43], [47] to separate land surface emissivity and temperature. This NN (DL) uses the Kalman ltering algorithm to increase the convergence rate in the learning stage and enhance separately the ability for highly nonlinear boundaries problem. The detailed introduction can be referred to [43]. The input of the NN is the brightness temperature (Ti , i = 11/12/13/14), and the output are land surface emissivity and temperature. To train and test the NN, the radiance transfer model (MODTRAN4) is applied to generate training and testing data sets. These simulation data sets can be viewed as reference data from a known ground truth. When necessary, additional data sets may be obtained from the reliable measured data. The detailed scheme is as follows. 1) According to spectral cure, the detailed information of the data used can be referred to the spectral database (http://speclib.jpl.nasa.gov). We use every band emissivity (about 160 kinds of land surface type) in bands 1114 as an input parameter of MODTRAN4. The range of the LST is from 270 to 320 K, and the correspondent air temperature near surface (at 2-m height) is arbitrarily assumed to be from 273 to 310 K. The range of atmospheric water vapor content change is from 0.2 to 4 g/cm2 for simulation, and six atmospheric proles are usedtropical atmosphere, midlatitude summer, midlatitude winter, subarctic summer, subarctic winter, and USA.1976.

(6)

MAO et al.: NN TECHNIQUE FOR SEPARATING LAND SURFACE EMISSIVITY AND TEMPERATURE

203

TABLE I SUMMARY OF RETRIEVAL ERROR

Fig. 2. ASTER1B image (bands 3, 2, and 1).

Fig. 3. AST08 product (land surface temperature).

Fig. 1. (a) Histogram of the difference between the land surface temperature retrieved (Tr) by the NN and the ground temperature (Tg). (b) Histogram of the difference between the emissivity retrieved (Er) by the NN and the ground emissivity (Eg).

between the simulated ground-truth data and the retrieval LST/emissivity, and the y-axis is the number of pixels. The average error [( n |Ts Tr |)/n] of the LST is under 0.24 K. i=1 The average error of emissivity is under 0.005, 0.005, 0.004, and 0.004 in bands 11, 12, 13, and 14, respectively. V. A PPLICATION AND E VALUATION To further evaluate the NN algorithm in the application, we compare the retrieval result by the NN with the ASTER standard product. We select Harbin as the study region, which is in Heilongjiang province, China. Fig. 2 is the image (2005/9/9) combined by ASTER bands 3, 2, and 1, which mainly includes city, river, village, cropland, and pond/swamp. Fig. 3 is the LST of the ASTER standard data product (AST08). The retrieval result is not very good in blocks A and C (Fig. 3), which is inuenced by a thin cloud (see Fig. 2). We use the DL NN trained above to retrieve the LST and emissivity from AST09 data. The retrieval result is like Fig. 4(a), and the

2) Randomly dividing the simulation data into two parts. The training data are 7387 sets, and the testing data are 1505 sets. 3) Training NN. According to the number of land types and trial and error, the size of two hidden layers with 800 nodes each is good in this study, which is mainly determined by the number of land surface types and the simplication of Planck function. The detailed test data set information can be referred to Table I. Fig. 1 shows the comparison between the retrieval results and the truth of LST/emissivity, given that the two hidden layers contain 800 nodes each. As shown in Table I and Fig. 1, the retrieval result is very good. The x-axis is the difference

204

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 1, JANUARY 2008

Fig. 4. (a) LST retrieved by the NN. (b) Difference between Fig. 3(a) and AST08. (c) Histogram of the difference between the land surface temperature retrieved by the NN and the AST08 product.

spatial distribution of the relative error between Figs. 3 and 4(a) (AST08) is like Fig. 4(b). The histogram of the relative error in Fig. 4(b) is like Fig. 4(c) (LST_SM-AST08), and the retrieval result by the NN is, on average, higher by about 0.7 K than the AST08. By comparing the spatial distribution of the LST shown in Fig. 3 with the one in Fig. 4 [particularly Fig. 4(b)], it could be found that the LST located over the left region in Fig. 4(a) is mostly higher than the one in Fig. 3, and it is in reverse in the right region. The most retrieval result in city and village in Fig. 4(a) is about 1 K higher than AST08 (Fig. 3), and the retrieval result in cropland in Fig. 4(a) is about 0.5 K lower than AST08. The distribution of the largest relative error is at the bank of the Songhuajiang River. As shown in Fig. 2, the main reason is the inuence of swamp at the bank of the river, the little island in the river, and the bridge over the river. Compared with blocks A and C in Figs. 3 and 4(a), which is inuenced by a thin cloud, the retrieval result by the NN is better, which can be shown in Fig. 2. It seems that the retrieval result by the NN is more reasonable than the ASTER standard data product (AST08).

To further evaluate the self-study ability of the NN, we select two little blocks (10 60, 5 80) in block C (shown in Fig. 3) from the same three images (AST09/AST08/AST05) as a compensation training data set of the training database. Fig. 5(a) is the retrieval result after training again by compensating some training data sets. The spatial distribution of the relative error is like Fig. 5(b) [Fig. 5(a), AST08], the average relative error of the LST is under 0.1 K [LST_SM_C-AST08, Fig. 4(c)], and the average relative error of emissivity in bands 11, 12, 13, and 14 is under 0.001 [like Fig. 5(c)]. As shown in Fig. 5(b), it is very interesting that the retrieval result in block A becomes higher, which is inuenced by a thin cloud and is underestimated in Figs. 3 and 4(a). The retrieval result indicates that the NN owns a powerful self-study and is capable of suiting for more conditions if we can obtain reliable measured data. It is very difcult to obtain the in situ ground-truth measurement of LST matching the pixel of the ASTER data at the satellite pass for validation of the algorithm. Generally speaking, LST varies from point to point on the ground, and ground measurement is generally point measurement. On the other hand, ASTER observes the ground at different angles, and

MAO et al.: NN TECHNIQUE FOR SEPARATING LAND SURFACE EMISSIVITY AND TEMPERATURE

205

Fig. 5. (a) LST retrieved by the NN after training again. (b) Difference between (a) and AST08. (c) Histogram of the difference between the emissivity (after compensating training data) retrieved by the NN and the AST05 product.

precisely locating the pixel of the measured ground in ASTER data is also a problem. In addition to these difculties, ground emissivity and the in situ atmospheric conditions also have to be known for validation. Since there are so many difculties in obtaining ground-truth data, validation with the use of groundtruth data is quite difcult. We just obtain four-point data from the meteorological branch in Harbin City, Suihua, Zhaodong, and Anda City. On the other hand, project 973 of China made ground measurement in Xiaotangshan and provided a data set of two sites (116.448 E and 40.182 N; 116.447 E and 40.177 N) from May to July 2004 [56]. We get an ASTER image in this region and extract two pixels from ASTER data through the longitude/latitude. The comparison between the retrieval result by the NN and the in situ LST is shown in Fig. 6, and the average accuracy is about 1.3 K. As we all know, the inuence factor for ground LST measurement is too much, particularly for a mixed pixel. We will make more eld measurements, and more comparison analyses will be reported in the future. On the other hand, another advantage of the NN is that the retrieval accuracy can be improved by compensating some training data.

Fig. 6. Validation results.

VI. C ONCLUSION In this paper, we make an analysis for an ill-posed problem that a single multispectral thermal measurement presents with N bands for N equations in at least N + 1 unknowns (N spectral emissivities and LST). The ill-posed problem can

206

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 1, JANUARY 2008

be overcome if we can make full use of the potential information between band emissivities for land surface types. Due to the NN simultaneously owning function approximation, classication, optimization computation ability, and also self-study ability, we use the NN to separate LST and emissivity. A new method is being explored in an attempt to synergize radiance transfer model and NN to retrieve geophysical parameters. This paper examines the combined use of the thermal emission model with NNs to perform parameter retrieval from remotely sensed data. The standard deviation errors are 0.35, 0.007, 0.006, 0.005, and 0.005 for LST and the emissivity of bands 11, 12, 13, and 14, respectively, for testing data simulated by MODTRAN4. Finally, we gave an evaluation in application, and the analysis indicates that the retrieval result by the NN is, on average, higher by about 0.7 K than the ASTER standard product (AST08). To further evaluate the retrieval ability (selfstudy) of the NN, we use AST09, AST08, and AST05 as the compensating training data; the average relative error of the LST is under 0.1 K relative to the AST08 product. The average relative error of the emissivity in bands 11, 12, 13, and 14 is under 0.001 relative to AST05. The comparison between the retrieval result by the NN and the in situ LST shows that the average accuracy is about 1.3 K. The analysis indicates that the NN is capable of suiting more conditions if we can obtain reliable measured data. The incorporation of the RM-NN to perform inversion is an important advancement in current retrieval techniques. The combination of radiance transfer and the NN makes it possible to perform inversion with higher accuracy and is more practical. ACKNOWLEDGMENT The authors would like to thank the following for their various help with this study: Z. Wan of University of California, Santa Barbara; S. Liang of the Department of Geography, University of Maryland, College Park; Y. Chang Tzeng of the Center for Space and Remote Sensing Research, National Central University, Taiwan, R.O.C.; and the ASTER Science Team for providing the ASTER Spectral Library data. They would also like to thank the anonymous reviewers for their valuable comments, which greatly improved the presentation of this paper. R EFERENCES
[1] A. R. Gillespie, S. Rokugawa, T. Matsunaga, J. S. Cothern, S. Hook, and A. Kahle, A temperature and emissivity separation algorithm for Advanced Spaceborne Thermal Emission and Reection Radiometer (ASTER) images, IEEE Trans. Geosci. Remote Sens., vol. 36, no. 4, pp. 11131126, Jul. 1998. [2] J. C. Price, Land surface temperature measurements from the splitwindow channels of the NOAA-7 AVHRR, J. Geophys. Res., vol. 89, no. D5, pp. 50395044, Aug. 1984. [3] F. Becker and Z.-L. Li, Towards a local split window method over land surface, Int. J. Remote Sens., vol. 11, no. 3, pp. 369393, 1990. [4] J. A. Sobrino and V. Caselles, A methodology for obtaining the crop temperature from NOAA-9 AVHRR data, Int. J. Remote Sens., vol. 12, no. 12, pp. 24612475, 1991. [5] J. A. Sobrino, Z.-L. Li, M. P. Stoll, and F. Becker, Improvements in the split window technique for land surface temperature determination, IEEE Trans. Geosci. Remote Sens., vol. 32, no. 2, pp. 243253, Mar. 1994. [6] C. Coll, V. Caselles, A. Sobrino, and E. Valor, On the atmospheric dependence of the split-window equation for land surface temperature, Int. J. Remote Sens. Environ., vol. 15, no. 1, pp. 105122, 1994.

[7] A. Vidal, Atmosphere and emissivity correction of land surface temperature measured from satellite data, Int. J. Remote Sens., vol. 12, no. 12, pp. 24492460, 1991. [8] Y. H. Kerr, J. P. Lagouarde, and J. Imbernon, Accurate land surface temperature retrieval from AVHRR data with use of an improved split window algorithm, Remote Sens. Environ., vol. 41, no. 2/3, pp. 197209, 1992. [9] C. Otlle and M. Stoll, Effect of atmospheric absorption and surface emissivity on the determination of land temperature from infrared satellite data, Int. J. Remote Sens., vol. 14, no. 10, pp. 20252037, 1993. [10] Z.-L. Li and F. Becker, Feasibility of land surface temperature and emissivity determination from AVHRR data, Remote Sens. Environ., vol. 43, no. 1, pp. 6785, 1993. [11] A. J. Prata, Land surface temperatures derived from the Advanced Very High Resolution Radiometer and the Along-Track Scanning Radiometer 2. Experimental results and validation of AVHRR algorithms, J. Geophys. Res., vol. 99, no. D6, pp. 13 02513 058, 1994. [12] Z. Wan and J. Dozier, A generalized split-window algorithm for retrieving land surface temperature measurement from space, IEEE Trans. Geosci. Remote Sens., vol. 34, no. 4, pp. 892905, Jul. 1996. [13] Z. M. Wan and Z.-L. Li, A physics-based algorithm for retrieving landsurface emissivity and temperature from EOS/MODIS data, IEEE Trans. Geosci. Remote Sens., vol. 35, no. 4, pp. 980996, Jul. 1997. [14] Z. H. Qin, G. D. Olmo, and A. Karnieli, Derivation of split window algorithm and its sensitivity analysis for retrieving land surface temperature from NOAA-Advanced Very High Resolution Radiometer data, J. Geophys. Res., vol. 106, no. D19, pp. 22 65522 670, 2001. [15] J. A. Sobrino, J. E. Kharraz, and Z.-L. Li, Surface temperature and water vapour retrieval from MODIS data, Int. J. Remote Sens., vol. 24, no. 24, pp. 51615182, Dec. 2003. [16] K. Mao, Z. Qin, J. Shi, and P. Gong, A practical split-window algorithm for retrieving land surface temperature from MODIS data, Int. J. Remote Sens., vol. 26, no. 15, pp. 31813204, Aug. 2005. [17] K. B. Mao, J. C. Shi, Z. H. Qin, P. Gong, W. Liu, and L. N. Xu, A multiple-band algorithm for retrieving land-surface temperature and emissivity from MODIS data, in Proc. IGARSS, Seoul, Korea, Jul. 2005, pp. 32693272. [18] S. Liang, An optimization algorithm for separating land surface temperature and emissivity from multispectral thermal infrared imagery, IEEE Trans. Geosci. Remote Sens., vol. 39, no. 2, pp. 264274, Feb. 2001. [19] K. B. Mao, J. C. Shi, Z. H. Qin, and P. Gong, An advanced and optimized split-window algorithm for retrieving land-surface temperature from ASTER data, in Proc. 9th ISPMSRS, Beijing, China, Oct. 2005. [20] K. B. Mao, J. C. Shi, Z. L. Li, Z. H. Qin, X. F. Wang, and L. M. Jiang, A multiple-band algorithm for separating land surface emissivity and temperature from ASTER imagery, in Proc. IGARSS, Denver, CO, Aug. 2006, pp. 13581361. [21] C. M. Juan and J. A. Sobrino, Feasibility of retrieving land surface temperature from ASTER TIR bands using two-channels algorithms: A case study of agriculture areas, IEEE Geosci. Remote Sens. Lett., vol. 4, no. 1, pp. 6064, Jan. 2007. [22] A. B. Kahle and R. E. Alley, Separation of temperature and emittance in remotely sensed radiance measurements, Remote Sens. Environ., vol. 42, no. 2, pp. 120, Nov. 1992. [23] S. J. Hook, A. R. Gabell, A. A. Green, and P. S. Kealy, A comparison of techniques for extracting emissivity information from thermal infrared data for geologic studies, Remote Sens. Environ., vol. 42, no. 2, pp. 123135, Nov. 1992. [24] K. Watson, Spectral ratio method for measuring emissivity, Remote Sens. Environ., vol. 42, no. 2, pp. 113116, Nov. 1992. [25] P. S. Kealy and S. Hook, Separating temperature and emissivity in thermal infrared multispectral scanner data: Implication for recovering land surface temperatures, IEEE Trans. Geosci. Remote Sens., vol. 31, no. 6, pp. 11551164, Nov. 1993. [26] T. Schmugge, S. J. Hook, and C. Coll, Recovering surface temperature and emissivity from thermal infrared multispectral data, Remote Sens. Environ., vol. 65, no. 2, pp. 121131, Aug. 1998. [27] T. Schmugge, A. French, J. C. Ritchie, A. Rango, and H. Pelgrum, Temperature and emissivity separation from multispectral thermal infrared observations, Remote Sens. Environ., vol. 79, no. 2/3, pp. 189198, Feb. 2002. [28] S. L. Liang, Retrieval of land surface temperature and water vapor content from AVHRR thermal imagery using articial neural network, in Proc. IGARSS, 1997, pp. 19591961. [29] G. M. Guo and M. Zhou, Using MODIS land surface temperature to evaluate forest re risk of northeast China, IEEE Geosci. Remote Sens. Lett., vol. 1, no. 2, pp. 98100, Apr. 2004.

MAO et al.: NN TECHNIQUE FOR SEPARATING LAND SURFACE EMISSIVITY AND TEMPERATURE

207

[30] F. D. Palluconi, G. Hoover, R. Alley, M. J. Nilsen, and T. Thompson, An atmospheric correction method for ASTER thermal radiometry over land, ASTER Standard Data Product AST09, in Level-2 RadianceTIR, Land_Leaving. Pasadena, CA: Jet Propulsion Lab., 1999. Revision 3. [Online]. Available: http://eospso.gsfc.nasa.gov/eos_homepage/for_ scientists/atbd/ docs/ASTER/atbd-ast-05.pdf [31] H. Tonooka and F. Palluconi, Verication of the ASTER/TIR atmospheric correction algorithm based on water surface emissivity retrieved, Proc. SPIE, pp. 5158, 2001. [32] H. Tonooka, F. Sakunma, M. Kudoh, and K. Iwafune, ASTER/TIR onboard calibration status and user-based recalibration, Proc. SPIE, pp. 191201, 2003. [33] H. Tonooka, F. Palluconi, S. Hook, and T. Matsunaga, Vicarious calibration of ASTER thermal infrared bands, IEEE Trans. Geosci. Remote Sens., vol. 43, no. 12, pp. 27332746, Dec. 2005. [34] H. Tonooka and F. Palluconi, Validation of ASTER/TIR standard atmospheric correction using water surface, IEEE Trans. Geosci. Remote Sens., vol. 43, no. 12, pp. 27692777, Dec. 2005. [35] Y. Yamaguchi and C. Naito, Spectral indices for lithologic discrimination and mapping by using the ASTER SWIR bands, Int. J. Remote Sens., vol. 24, no. 22, pp. 43114323, 2003. [36] L. C. Rowan and C. J. Mars, Lithologic mapping in the Mountain Pass, California area using Advanced Spaceborne Thermal Emission and Reection Radiometer (ASTER) data, Remote Sens. Environ., vol. 84, no. 3, pp. 350366, Mar. 2003. [37] L. C. Rowan, J. C. Mars, and C. J. Simpson, Lithologic mapping of the Mordor, NT, Australia ultramac complex by using the Advanced Spaceborne Thermal Emission and Reection Radiometer (ASTER), Remote Sens. Environ., vol. 99, no. 1/2, pp. 105126, Nov. 2005. [38] Y. Ninomiya, F. Bihong, and T. J. Cudahy, Detecting lithology with Advanced Spaceborne Thermal Emission and Reection Radiometer (ASTER) multispectral thermal infrared radiance-at-sensor data, Remote Sens. Environ., vol. 99, no. 1/2, pp. 127139, Nov. 2005. [39] R. G. Vaughan, S. J. Hook, W. M. Calvin, and J. V. Taranik, Surface mineral mapping at Steamboat Springs, Nevada, USA, with multiwavelength thermal infrared images, Remote Sens. Environ., vol. 99, no. 1/2, pp. 140158, Nov. 2005. [40] K. M. Hornik, M. Stinchcombe, and H. White, Multilayer feedforward networks are universal approximators, Neural Netw., vol. 2, no. 5, pp. 359366, 1989. [41] A. K. Fung, Microwave Scattering and Emission Models and Their Application. Boston, MA: Artech House, 1994, ch. 12, pp. 527555. [42] S. Y. Hsu, T. Masters, M. Olson, M. Tenorio, and T. Grogan, Comparative analysis of ve neural networks models, Remote Sens. Rev., vol. 6, pp. 319329, 1992. [43] Y. C. Tzeng, K. S. Chen, W. L. Kao, and A. K. Fung, A dynamic learning neural network for remote sensing applications, IEEE Trans. Geosci. Remote Sens., vol. 32, no. 5, pp. 10961102, Sep. 1994. [44] S. Haykin, W. Stehwien, C. Deng, P. Weber, and R. Mann, Classication of radar clutter in an air trafc control environment, Proc. IEEE, vol. 79, no. 6, pp. 742772, Jun. 1991. [45] P. D. Herrmann and N. Khazene, Classication of multiple spectral remote sensing data using a back-propagation neural network, IEEE Trans. Geosci. Remote Sens., vol. 30, no. 1, pp. 8188, Jan. 1992. [46] H. Bischof, W. Schneider, and A. J. Pinz, Multiple spectral classication of Landsat images using neural networks, IEEE Trans. Geosci. Remote Sens., vol. 28, no. 3, pp. 482489, May 1992. [47] K. S. Chen, Y. C. Tzeng, C. F. Chen, and W. L. Kao, Land-cover classication of multispectral imagery using a dynamic learning neural network, in Photogramm. Eng. Remote Sens., vol. 61, 1995, pp. 403408. [48] Y. Q. Jin and C. Liu, Biomass retrieval from high-dimensional active/passive remote sensing data by using articial neural network, Int. J. Remote Sens., vol. 18, no. 4, pp. 971979, Mar. 1997. [49] F. Aires, C. Prigent, W. B. Rossow, and M. Rossow, A new neural network approach including rst guess for retrieval of atmospheric water vapor, cloud liquid water path, surface temperature, and emissivities over land from satellite microwave observations, J. Geophys. Res., vol. 106, no. D14, pp. 14 88714 907, 2001. [50] T. Faure, H. Isaka, and B. Guillemet, Neural network retrieval of cloud parameters of inhomogeneous and fractional clouds feasibility study, Remote Sens. Environ., vol. 77, no. 2, pp. 123138, Aug. 2001. [51] M. Tedesco, J. Pulliainen, M. Takala, M. Hallikainen, and P. Pampaloni, Articial neural network-based techniques for the retrieval of SWE and snow depth from SSM/I data, Remote Sens. Environ., vol. 90, no. 1, pp. 7685, Mar. 2004. [52] W. J. Blackwell, A neural-network technique for the retrieval of atmospheric temperature and moisture proles from high spectral resolu-

[53]

[54]

[55] [56]

tion sounding data, IEEE Trans. Geosci. Remote Sens., vol. 43, no. 11, pp. 25352546, Nov. 2005. L. Tang, Z. Chen, S. Oh, R. J. Marks, II, and A. T. C. Chang, Inversion of snow parameters from passive microwave remote sensing measurements by a neural network trained with a multiple scattering model, IEEE Trans. Geosci. Remote Sens., vol. 30, no. 5, pp. 10151024, Sep. 1992. F. X. Kneizys, E. P. Shettle, W. O. Gallery, J. H. Chetwynd, L. W. Abreu, J. E. A. Sellby, S. A. Clough, and R. W. Fenn, Atmospheric transmittance/radiance: Computer code LOWTRAN 6, Air Force Geophys. Lab., Bedford, MA, Rep. AFGL-TR-83-0187(NTIS AD A137796), 1983. A. Berk, L. S. Bemstein, and D. C. Roberttson, MODTRAN: A moderate resolution model for LOWTRAN, Spectral Sci., Inc., Burlington, MA, Rep. AFGL-TR-87-0220, 1987. L. J. Han, Estimation of evapotranspiration by assimilating MODIS LST product into the CLM, Ph.D. dissertation, Beijing Normal Univ., Beijing, China, 2006. (In Chinese).

Kebiao Mao received the B.S. degree from Northeast Normal University, Changchun, China, in 2001, the M.S. degree from Nanjing University, Nanjing, China, in 2004, and the Ph.D. degree in geographic information systems from the Chinese Academy of Sciences, Beijing, China, in 2007. He is currently with the Key Laboratory of Resources Remote Sensing and Digital Agriculture, Ministry of Agriculture, Hulunber Grassland Ecosystem Observation and Research Station, Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing. He was with the State Key Laboratory of Remote Sensing Science, which was jointly sponsored by the Institute of Remote Sensing Applications, Chinese Academy of Sciences, and Beijing Normal University, and also with the Graduate School of the Chinese Academy of Sciences. His research interests include geophysical parameters retrieval (like land surface temperature and emissivity and soil moisture) and the assimilation of satellite data to the crop growth model.

Jiancheng Shi (SM02) received the B.A. degree from the University of Lanzhou, Lanzhou, China, in 1982 and the M.A. and Ph.D. degrees in geography from the University of California at Santa Barbara (UCSB) in 1987 and 1991, respectively. He is currently with the Institute for Computational Earth System Sciences, UCSB, as a Research Scientist. He is also an Adjunct Research Professor with the State Key Laboratory of Remote Sensing Science, which is jointly supported by the Institute for Remote Sensing Applications, Chinese Academy of Science, and Beijing Normal University. His research interests are microwave modeling of snow and soil signatures, image processing and analysis, and inversion models for retrieving physical parameters from remote sensing data.

208

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 1, JANUARY 2008

Huajun Tang received the B.S. degree in agronomy from Southwest Agricultural University, Chongqing, China, in 1982 and the M.S. and Ph.D. degrees (with greatest distinction) in physical land resources from the University of Ghent, Ghent, Belgium, in 1987 and 1993, respectively. After his graduation, he joined the Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing, China, where he became a Full Professor in 1997 and is currently a Research Professor and the Director General. He is also with the Key Laboratory of Resources Remote Sensing and Digital Agriculture, Ministry of Agriculture, Hulunber Grassland Ecosystem Observation and Research Station, Beijing. Over the years, he has served as a Consultant to a number of governmental agencies, nonprot organizations, and companies, including the World Bank, Food and Agriculture Organizations, and the United Nations Economic and Social Commission for Asia and Pacic. He has collaborated with colleagues and institutes in Belgium, the U.S., Australia, and Japan. He serves as an advisor to many organizations in China including the Ministry of Agriculture. He has published more than 80 papers in international and Chinese scientic journals. His research interests include land use and crop modeling and remote sensing and GIS applications in agriresources management. Dr. Tang is the recipient of a number of awards and honors, including being elected as a Corresponding Member of the Belgian Royal Academy of Science and President of the Chinese Society of Agricultural Resources and Regional Planning. He has organized and sponsored more than a dozen international conferences in the past ten years.

Xiufeng Wang received the B.S. degree from Northeast Agricultural University, Harbin, China, in 1982 and the Ph.D. degree from Hokkaido University, Sapporo, Japan, in 1994. She is a Lecturer with the Graduate School of Agriculture, Hokkaido University. Her main research interests are in the study of the analysis methods and its applications about satellite data for agricultural meteorology. Dr. Wang is a recipient of the Scientic Award from the Society of Agricultural Meteorology of Japan in 2005.

Zhao-Liang Li received the B.S. degree in photogrammetry from the WuHan Technical University of Surveying and Mapping, Wuhan, China, in 1985 and the M.S. degree in imaging processing and computer graphics and the Ph.D. degree in terrestrial environmental physics from the University of Louis Pasteur, Strasbourg, France, in 1987 and 1990, respectively. He is currently with the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, China, and also with the Laboratoire des Sciences de lImage, de lInformatique et de la Teledetection (UMR7005), Strasbourg. He has participated in many national and international projects such as MODIS (funded by the National Aeronautics and Space Administration) and the programs EAGLE (funded by the European Community) and SPECTRA (funded by the European Space Agency). He serves as a referee for several international journals. He has published more than 50 papers in international referred journals. His main expertise elds are in the retrieval of surface emissivity and temperature from satellite data and the assimilation of satellite data to land surface model. Dr. Li is a member of several scientic committees.

Kun-Shan Chen (S86M92SM98F07) received the B.S.E.E. degree from the National Taiwan Institute of Technology, Taipei, Taiwan, R.O.C., and the M.S. and Ph.D. degrees from the University of Texas, Arlington, in 1987 and 1990, respectively, all in electrical engineering. Since 1992, he has been with the faculty of the Center for Space and Remote Sensing Research, National Central University (NCU), Jung-Li, Taiwan, where he served as the Director from 2001 to 2004 and is currently an NCU Distinguished Professor. He has joint appointments at the Institute of Space Sciences and Institute of Communication Engineering at the same university. He serves as a Technical Consultant at several national research agencies in areas of satellite remote sensing, radar, and radio techniques. He has published over 60 referred journal and 120 conference papers in addition to three book chapters. He has been the Editor-in-Chief of the Journal of Photogrammetry and Remote Sensing since 2001. He serves on the Editorial Board of the Journal of Electromagnetic Waves and Applications and the Transactions of the Aeronautical and Astronautical Society of the Republic of China. His research activities include microwave remote sensing, image processing and analysis for satellite and aircraft remotesensing data, radio and microwave propagation, and scattering from terrain and ocean with applications to remote sensing and wireless communications. Dr. Chen is a member of the Electromagnetic Academy. He was a Technical Chairman of PIERS 1999, which was held in Taipei. In 2001, he was appointed as Chairman of the Commission F, Taipei, of Union of Radio Science International. He is currently an Associate Editor of the IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING.