A neural-network approach to radiometric sensing of land-surface parameters

A biophysically-based land-surface process/radiobrightness (LSP/R) model is integrated with a dynamic learning neural network (DLNN) to retrieve the land-surface parameters from its radiometric signatures. Predictions from the LSP/R model are used to train the DLNN and serve as the reference for evaluation of the DLNN retrievals. Both horizontally polarized and vertically polarized brightnesses at 1.4 GHz, 19 GHz, and 37 GHz for an incidence angle of 53° make up the input nodes of the DLNN. The corresponding output nodes are composed of land-surface parameters, canopy temperature and water content, and soil temperature and moisture (uppermost 5 mm). Under no-noise conditions, the maximum of the root mean-square (RMS) errors between the retrieved parameters of interest and their corresponding reference from the LSP/R model is smaller than 28 for a four-channel case with 19 GHz and 37 GHz brightnesses as the inputs of the DLNN. The maximum RMS error is reduced to within 0.5% if additional 1.4 GHz brightnesses are used (a six-channel case). This indicates that the DLNN produces negligible errors onto its retrievals. For the realization of the problem, two different levels of noises are added to the input nodes. The noises are assumed to be Gaussian distributed with standard deviations of 1 K and 2 K. The maximum RMS errors are increased to 9.3% and 10.3% for the 1 K-noise and 2 K-noise cases, respectively, for the four-channel ease. They are reduced to 6.0% and 9.1% for the 1 K-noise and 2 K-noise cases, respectively, for the six-channel case. This is an implication that 1.4 GHz is a better frequency in probing soil parameters than 19 GHz and 37 GHz