Roughness and vegetation change detection: A pre-processing for soil moisture retrieval from multi-temporal SAR imagery

Abstract

Abstract Multi-temporal analysis has been widely acknowledged as a promising method to derive soil moisture from radar backscatter observations. The method assumes that only soil moisture varies in the period of interest, while all other parameters such as vegetation water content and soil surface roughness are sufficiently time invariant. However, this assumption is not easy to satisfy in agricultural areas where cultivation practices such as ploughing and irrigation are irregularly conducted between radar acquisitions. The paper has proposed an unsupervised change detection method to serve as a pre-processing procedure for multi-temporal retrieval. Briefly, the temporal ratio of HV and the temporal difference of HV/VV and VV polarizations were selected as the optimal feature space, using a genetic algorithm based feature selection algorithm and an extensive synthetic data set. The change map is determined from a two-step procedure with the first step producing multiple over-detected change maps for the period of interest using the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) method. The second step merges the multiple change maps to remove the false alarms with a principle similar to the ensemble machine learning. Evaluation on a synthetic data set demonstrated that the proposed method can largely remove the error in multi-temporal soil moisture retrieval that is caused by abrupt roughness and vegetation changes. Evaluation on real radar data sets, including airborne L-band radar, RADARSAT-2 at C-band and COSMO SkyMed at X-band, demonstrated an accurate identification (>0.9) while yielding a low false-alarm rate (