An RM-NN algorithm for retrieving land surface temperature and emissivity from EOS/MODIS data

[1] Three radiative transfer equations are built for MODIS bands 29, 31, and 32, which involve six unknown parameters ( average atmospheric temperature, land surface temperature (LST), three band emissivities, and water vapor content). The relationships between geophysical parameters have been analyzed in detail, which indicates that neural network is one of the best methods to resolve these ill-posed problems ( LST and emissivity). Retrieval analysis indicates that the combined radiative transfer model (RM) with neural network (NN) algorithm can be used to simultaneously retrieve land surface temperature and emissivity from Moderate-Resolution Imaging Spectroradiometer ( MODIS) data. Simulation data analysis indicates that the average error of LST is under 0.4 K and the average error of emissivity is under 0.008, 0.006, and 0.006 for bands 29, 31, and 32, respectively. The comparison analysis between retrieval result by RM-NN and MODIS product algorithm indicates that the generalized split window LST overestimates the emissivity and underestimates land surface temperature. The retrieval results by RM-NN lie between the two products provided by NASA and closer to day/night LST algorithm after statistics analysis. The average error is 0.36 K relative to MODIS LST product (MOD11_L2) retrieved by generalized split window algorithm if we make a regression revision. The comparison of retrieval results with ground measurement data in Xiaotangshan also indicates that the RM-NN can be used to retrieve accurately land surface temperature and emissivity.