A Two-Stream LSTM-Based Backscattering Model at L-Band and S- Band for Dry Soil Surfaces Under Large Roughness Conditions

Abstract

In this article, we report for the first time two radar measurements (Ji Mo 2008 and Min Qin 2009) on natural soil surfaces under large roughness, which were conducted by the China Research Institute of Radiowave Propagation. The desired HH and VV polarization backscatter data were measured by a truck-mounted scatterometer, which operates at L-band and S-band (i.e., 1.34 and 3.2 GHz, respectively). Simultaneously to radar acquisitions, the ground-truth data related to the rms height, the correlation length, and the dielectric constant were collected. Discrepancies between the simulations of the advanced integral equation model (AIEM) and the radar data have indicated the inadequacy of the AIEM model under large roughness conditions. To address this limitation, a new two-stream long short-term memory–based network was developed to receive the radar and surface parameters, termed radar-surface network (RSNet). The proposed network was trained on a hybrid dataset consisting of 1) a simulated dataset generated based on the AIEM under a wide range of conditions and 2) the radar data reported in Ji Mo 2008 and Min Qin 2009 combined with those simulated to make the dataset more relevant to natural conditions. After training, extensive experiments were performed to evaluate the performance of the proposed backscatter model. Comparisons demonstrate that the predictions of RSNet are generally in good agreement with both simulations and measured data, in terms of magnitude and trend, thus demonstrating that the proposed model can yield trustworthy and high-quality backscatter estimations at L-band and S-band for dry soil surfaces under large roughness conditions.

Keywords

• Advanced integral equation model (AIEM)
• backscattering coefficient
• long short-term memory (LSTM)
• soil surface