A Neural Network Technique for Separating Land Surface Emissivity and Temperature From ASTER Imagery

Four radiative transfer equations for Advanced Spaceborne Thermal Emission and Reflection 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, classification, 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 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.