33 1 2010 1 33 No. 1 Vol. AR ID LAND GEOGRAPHY Jan. 2010 1, 2, 1, 3 (1, 100190; 2, 100049; 3, 100101) : (RBF),,, 9 m 10 m 12 m, 4A 100,, : ; ; : TP732. 2 : A : 1000-6060 (2010) 01-0099 - 07 (99 105),,,, ( ), MetOp IASI,,, IASI, 1 1. 1 4A 4A /O P 4A (Automatized A t2 mospheric Absorp tion A tlas), 1-2,, 4A, IASI ( 645 cm - 1 2 760 cm - 1 ) IASI 0. 25 cm - 1, 4A IASI NOVELTIS Laboratoire de M t orologie Dynam ique (LMD ) 4A, 4A 4A /OP, 4A 1 4A /OP 3 : 2009-04 - 18; : 2009-08 - 04 : 863 (2006AA12Z121, 2006AA12Z148) (KZCX2 - YW - Q10-2) :, 2006,, : :,,. Email: xgjiang@aoe. ac. cn
1 00 33 1 4A /OP Fig. 1 D iagram Flow of 4A /OP 4A /OP / :, ; 4A /OP,, / ; 1. 2,,,, ;, 1. 3 Matlab, : ( 1) 4A /OP ; ( 2 ) TIGR ( Thermody2 nam ic Initial Guess Retrieval) 2000 11 ; (3) 1) 3 Matlab 4-8 Fig. 3 Structure of RBF Neural Network in Matlab ANN toolbox 3, 20 80 J. R, N 1, Moody C. Darken 9 N 2, P R, (RBF) W 1 W 2, : W 1 (N 1 R ), W 2 (N 2, N 1), b1 b2, : b1 (N 1 ), b2 (N 2 ), Y N 2 ζ dist ζ 10, radbas, b1 ζ dist(x, Y) ζ = (X - Y) 2, (1) : X, Y ζ radbas ( n) = e - n2, (2) d ist ζ,, n radbas a1, Y: 2 Fig. 2 Mapp ing of RBF Neural Network a1 j = exp ( - ( R (W 1 ji - P i ) 2 b1 j ) 2 ) i = 1 j = 1, N 1, (3)
1 : 101 Y K N 1 = (W 2 k j a1 j ) + b2 k k = 1, N 2, (4) j = 1,, ; 4A /OP,, : (1) N 1 W 1 b1, N 1,,,,,, b1, b1,, b1, (2) N 2, W 2, b2 N 2 BP,,,, Matlab,,, 1. 4, (1) ; (2), ; ( 3), ; (4) ; ( 5) 2 2. 1 TIGR2000 2311, 1 /2, 1 /2 MetOp, 15 ( 1) ASTER 12, JPL JHU USGS - Reston,, - 10 K + 15 K 1 Tab. 1 O bserva tion Angles of Sa tellite in Exper im en t No. Angle / No. Angle / No. Angle / 1 1 40 6 18 20 11 35 00 2 5 00 7 21 40 12 38 20 3 8 20 8 25 00 13 41 40 4 11 40 9 28 20 14 45 00 5 15 00 10 31 40 15 48 20 4 Fig. 4 Flow Chart of the Fast A lgorithm for simulating the measured radiance based on ANN 2. 2, 39, - -,, :,,
1 02 33 : R ( ) = R ( ) sec ( ) / sec ( R ), (5) : R 1, R 1 40 R ( ), (6) : j = i i ( ) = j = 1 e- sig j( ), (6) 39, 40, ( 7) : rad ( ) = 40 B i ( ) ( i - 1 ( ) - i ( ) ), (7) i = 1 rad, B 2. 3 4A /OP, 5 6 7 9 m ( 1 111. 0 cm - 1 ) 10 m ( 1 000. 0 cm - 1 ) 12 m ( 833. 25 cm - 1 ), 9 m: 0. 949 0; 10 m: 0. 971 9; 12 m: 0. 964 6, 1 40 2 2, 12 m, 10 m, 10 m 0. 1 K 45, 1156 3. 9% ; 0. 1 K 61, 1155 5. 3%, 0. 1 K 95%, 5 9 m ( a) 39 ; ( b) :, : ; ( c) : : Fig. 5 on 9 m ( a) No. of Neurons for H idden Layers; ( b) left: Absolute Errors of Total Op tical Dep th for Training Group; right: Absolute Errors of Total Optical Depth for Verification Group; ( c) left: Absolute Errors of B right Temper2 ature for Training Group; right: Absolute Errors of B rightness Temper2 ature for Verification Group.
1 : 103 6 10 m ( a) 39 ; ( b) :, : ; ( c) : : Fig. 6 on 10 m ( a) No. of Neurons for H idden Layers; ( b) left: Absolute Errors of Total Op tical Dep th for Training Group; right: Absolute Errors of Total Op tical Dep th for Verification Group; ( c) left: Absolute Errors of B right Temper2 ature for Training Group; right: Absolute Errors of B rightness Temperat2 ure for Verification Group. 7 12 m ( a) 39 ; ( b) :, : ; ( c) : : Fig. 7 on 12 m ( a) No. of Neurons for H idden Layers; ( b) left: Absolute Errors of Total Optical Depth for Training Group; right: Absolute Errors of Total Op tical Depth for Verification Group; ( c) left: Absolute Errors of B right Temper2 ature for Training Group; right: Absolute Errors of B rightness Temperat2 ure for Verification Group.
1 04 33 CPU: AMD A thlon 64 X2 Dual 4000 + 991MHz; : Fedo2 ra 7; 4A /OP : Fortran, pgf 90 ( ) ; :Matlab 7, 2 311, 11 s 4A /OP 60 60, 4A /OP 1 /3 60 N (N < 60), 4A /OP, 11 s N,, 100 10% 20%,, 2 Tab. 2 Rm se of Fa st Approach 9 m 10 m 12 m 0. 0012 0. 0012 0. 0041 0. 0045 0. 0010 0. 0011 0. 0098 0. 0103 0. 0393 0. 0458 0. 0054 0. 0061 3, ( ),,,,,,,,,, 9 m, 10 m 12 m,, ( References) 1 Scott N A, Chedin A. A fast line2by2line method for atmospheric absorp tion computations: the automatized atmospheric absorp tion atlas J. J App lmeteor, 1981, 20: 556-564. 2 Scott N A. A directmethod of computation of transm ission function of an inhomogeneous gaseous medium: description of the method and influence of various factors J. J Quant Spectrosc Radiat Transfer, 1974, 14: 691-707. 3 Chaumat L, DecosterN, Standfuss C, et al. 4A /OP Reference Doc2 umentation, NOV - 3049 - NT - 1178 - v3. 3 M. NOVELTIS, LMD /CNRS, CNES, 2006: 260. 4 Chevallier F, Cheruy F, Scott N A, et al. A neural network ap2 p roach for a fast and accurate computation of a longwave radiative budget J. Journal of Applied Meteorology, 1998, 37 ( 11) : 1385-1397. 5 Chevallier F, Morcrette J2J, Cheruy F, et al. U se of a neural2net2 wrok2based long2wave radiative2transfer scheme in the ECMW F at2 mospheric model J. Q J R Meteorol Soc, 2000, 126: 761-776. 6 Key J R, Schweiger A J. Tools for atmospheric radiative transfer: streamer and fluxnet J. Computers & Geosciences, 1998, 24 (5) : 443-451. 7 Krasnopolsky V M. New app roach to calculation of atmospheric model physics: accurate and fast neural network emulation of long2 wave radiation in a climate model J. Monthly W eather Review, 2004, 133 (5) : 1370-1383. 8 Krasnopolsky V M, Chevallier F. in environmental sciences. Some neural network app lications Part II: advancing computational effi2 ciency of environmental numerical models J. Neural Networks, 2003, 16: 335-348. 9 Moody J, Darken C. Learning with localized recep tive fields M / / Touretzky H inton, Sejnowski, eds. Proceedings of the 1988 Connec2 tionist Models Summer School. 1988. Morgan - Kaufmann Publishers, 10 Robert J, Schilling J, Carroll J. App roximation of nonlinear system with radial basis function neural networks J. J IEEE Transac2 tions on Neural Networks, 2001, 2 (1) : 21-28. 11 Chevallier F, Chedin A, Cheruy F, et al. TIGR2like atmospheric2 p rofile databases for accurate radiative2flux computation J. Q J R Meteorol Soc, 2000, 126: 777-785. 12 Korb A R, Dybwad P,W adsworthw, Salisbury J W. Portable FTIR spectrometer for field measurements of radiance and em issivity J. Applied Op tics, 1996, 35: 1679-1692.
1 : 105 Fa st ca lcula tion approach for the hyperspectra l infrared rad ia tive tran sfer m odel ba sed on artif ic ia l neura l network WU M in2j ie 1, 2, J IANG Xiao2Guang 1, TANG Bo2Hui 3 ( 1 Academ y of Opto2Electronics, CAS, B eijing 100190, China; 2 Graduate University of Chinese Academ y of Sciences, B eijing 100049, China; 3 Institute of Geographical Sciences and N atural Resources Research, CAS, B eijing 100101, China) Abstract: Surface temperature and surface emm isivity are two important parameters for earth environm ental resear2 ches. Theoretically they can be retrieved w ith radiance data received from satellite, but in p ractice, more constraint conditions are needed to solve this p roblem. sensing data p rovides a new way for this p roblem. Taking advantage of the large amount of channels, hyperspectral remote In order to develop a new model to separate surface temperature and surface emm isivity fast and accurately with the help of hyperspectral remote sensing data, a fast method to as2 sim ilate the radiance data received by hyperspectral sensors such as Infrared A tmospheric Sounding Interferom eter ( IASI) must be developed at first. Currently, som e kind of hyperspectral infrared atmospheric radiative transfer model ( RTM ) have been app lied in numerical weather p rediction (NW P) system s. Though these models have highly accuracy, they still can not meet the requirement for calculation speed. This paper developed a fast calcula2 tion app roach for the hyperspectral infrared radiative transfer model based on artificial neural network, which would significantly imp rove the calculation speed of RTM and at the sam e time be relatively accurate compared w ith other hyperspectral infrared atmospheric RTM s. In recent years, A rtificial Neural Network (ANN ) has been com bined in2 to some atmospheric RTM s, such as NeuroFlux in ECMW F s atmospheric model and NN emulation in NCAR CAM longwave atmospheric radiation parameterization. W ith the help of ANN technique, these models could enjoy a high2 ly calculation speed for radiance and other related parameters. However, existing methods could not be suitable for the fast calculation of hyperspectral models. Autom atized A tmospheric Absorp tion A tlas ( 4A ) is an accurate hyper2 spectral infrared radiative transfer model which is suitable for the simulation of hyperspectral infrared therm al sen2 sors such as IASI. But it sill takes too much tim e for model calculation. In the paper, radiance and atmospheric pa2 rameters calculated with 4A model were used as true value to judge the accuracy of fast calculation app roach, and the calculation speed of 4A would be compared w ith fast app roach too. In this paper, RB F ( Radial basis Function) neural network technique is introduced to design a fast calculation app roach which is used to accelerate the calcula2 tion speed of hyperspectral infrared thermal radiative transfer model. neural network have been found through a lot of numerical simulations and calculations. The p roper inputs and outputs for our p roposed In addition, a type of multi2 layer neural network structure has also been developed to fast calculate the top of atmosphere radiance for typ ical wavelengths in hyperspectral therm al infrared spectrum. spectively for three wavelengths: 9, 10 and 12 m. In this paper, three sets of neural networks were trained re2 Results of the experiment show that this fast calculation ap2 p roach could calculate the top of atmosphere radiance with the error less than 0. 1 K compared with 4 A, and its running speed is 100 tim es faster than that of 4 A for a single wavelength. for the choosing of spectral channels. This app roach also enjoys more flexibility Key W ords: atmospheric radiative transfer; RBF neural network; hyperspectral infrared thermal remote sensing