Two-Step Deep Learning Approach for Estimating Vegetation Backscatter: A Case Study of Soybean Fields

Abstract

Precisely predicting vegetation backscatter involves various challenges, such as complex vegetation structure, soil–vegetation interaction, and data availability. Deep learning (DL) works as a powerful tool to analyze complex data and approximate the nonlinear relationship between variables, thus exhibiting potential applications in microwave scattering problems. However, few DL-based approaches have been developed to reproduce vegetation backscatters owing to the lack of acquiring a large amount of training data. Motivated by a relatively accurate single-scattering radiative transfer model (SS-RTM) and radar measurements, we, for the first time to our knowledge, introduce a transfer learning (TL)-based approach to estimate the radar backscatter of vegetation canopy in the case of soybean fields. The proposed approach consists of two steps. In the first step, a simulated dataset was generated by the SS-RTM. Then, we pre-trained two baseline networks, namely, a deep neural network (DNN) and long short-term memory network (LSTM), using the simulated dataset. In the second step, limited measured data were utilized to fine-tune the previously pre-trained networks on the basis of TL strategy. Extensive experiments, conducted on both simulated data and in situ measurements, revealed that the proposed two-step TL-based approach yields a significantly better and more robust performance than SS-RTM and other DL schemes, indicating the feasibility of such an approach in estimating vegetation backscatters. All these outcomes provide a new path for addressing complex microwave scattering problems.