Use of neural networks in radiometric studies of land surface parameters

Radiometric characteristics of the land surface nonlinearly depend on the surface state, so it is in general a great challenge to recover the surface state using mathematically-based schemes. Neural networks are known for their capability in dealing with nonlinear fittings. We investigate the use of a Dynamic Learning Neural Network (DLNN) in the retrieval of land surface parameters from radiometric signatures. Two case studies are considered. The first study is based on predictions from a 60-day summer dry-down simulation of the Land Surface Process/Radiobrightness (LSP/R) model, which manages land-air interactions and microwave radiative transfer in order to furnish temperature and moisture profiles of the vegetation and soil, and the corresponding brightness temperatures of the terrain. For the purpose of this investigation, the second study is based on LSP/R model predictions, which are used for model validation against a field campaign. Both cases utilize about 10% of the predictions from the LSP/R model to train the DLNN, and another 10% or so of the predictions as the ground truth to evaluate the DLNN retrievals. The training data include horizontally- and vertically-polarized brightnesses at 1.4, 19, and 37 GHz as the inputs of the DLNN, and the corresponding temperatures and moisture contents of the soil and canopy as the outputs. In the first study, we find that root mean square (rms) errors are less than 1% between DLNN retrievals and ground truth for all of the four surface parameters of interest. The rms errors are about 0.42 Kelvin for soil temperature (uppermost 5 mm), 0.11% for soil moisture (by volume), 0.034 Kelvin for canopy temperature, and 0.008 kg/m². In the second study, the rms errors are slightly greater but within a reasonable range of less than 2% for all of four parameters.