An NN-based atmospheric correction algorithm for Landsat/TM thermal infrared data

Land surface temperature (LST) is a key variable for studies of global or regional land surface processes, energy and water cycle, and thus, has important applications in various areas. Atmospheric correction is a major issue in LST retrieval using remote sensing data because the presence of the atmosphere always influences the radiation from the ground to the space sensor. Atmospheric correction of thermal infrared (TIR) data for land surface temperature retrieval is to estimate the three atmospheric parameters: transmittance, path radiance and the downward radiance. Typically the atmospheric parameters are obtained using atmospheric profiles combined with a radiative transfer model (RTM). But this approach is time-consuming and expensive, which is impractical for high-speed (near-realtime) operational atmospheric correction. An artificial neural network (NN) based atmospheric correction model for Landsat/TM thermal infrared data is proposed. The multi-layer feed-forward neural network (MFNN) is selected, in which the atmospheric profiles (temperature, humidity and pressure), elevation and scan angle are the input variables, and the atmospheric parameters are the output variables. The MFNN is combined with the radiative transfer simulation, using MODTRAN 4.0 and the latest global assimilated data. Finally, the transmittance and path radiance derived by the MFNN-based algorithm is compared with MODTRAN4.0 results. The RMSE for both parameters are 0.0031 and 0.035 W·m^-2·sr^-1·μm^-1, respectively. The results indicate that the proposed approach can be a practical method for Landsat/TM thermal data in both accuracy and efficiency.