Retrieval of land surface temperature and emissivity from ASTER1B data using a dynamic learning neural network

Abstract

The accuracy of a radiance transfer model neural network (RM-NN) for separating land surface temperature (LST) and emissivity from AST09 (the Advanced Spaceborne and Thermal Emission and Reflection Radiometer (ASTER) Standard Data Product, surface leaving radiance) is very high, but it is limited by the accuracy of the atmospheric correction. This article uses a neural network and radiance transfer model (MODTRAN4) to directly retrieve the LST and emissivity from ASTER1B data, which overcomes the difficulty of atmospheric correction in previous methods. The retrieval average accuracy of LST is about 1.1 K, and the average accuracy of emissivity in bands 11–14 is under 0.016 for simulated data when the input nodes are a combination of brightness temperature in bands 11–14. The average accuracy of LST is under 0.8 K when the input nodes are a combination of water vapour content and brightness temperature in bands 11–14. Finally, the comparison of retrieval results with ground measurement data indicates that the RM-NN can be used to accurately retrieve LST and emissivity from ASTER1B data.