Soil moisture retrieval over agricultural fields with machine learning: A comparison of quad-, compact-, and dual-polarimetric time-series SAR data.

• Quad-, Compact-, and Dual-Pol SAR modes are exploited for soil moisture retrieval. • Four machine learning regression methods are utilized and compared. • A forward feature selection algorithm is adopted to enhance performance. • Four years of C-band RADARSAT-2 data across three crop types are validated. • Using the Quad-Pol SAR mode and the RFR method produces the best results. Accurate measurement of soil moisture (SM) is crucial for understanding crop growing conditions, optimizing irrigation practices, and early detection of drought. Synthetic aperture radar (SAR) has proven effective in SM inversion for agricultural scenarios. Machine learning methods enable large-scale SM mapping by circumventing complex computations and exhibiting high nonlinear fitting capabilities. While previous studies have explored SM retrieval using SAR data and machine learning with dual-polarimetric (DP), compact-polarimetric (CP), or quad-polarimetric (FP) modes, a comprehensive comparative study of these polarization modes in crop scenario is lacking. In this study, we assessed SM inversion using three SAR polarimetric modes (DP, CP, and FP) in C-band across three crop types (wheat, corn, and soybean) using multi-year RADARSAT-2 data. Various polarimetric backscattering variables, polarimetric decomposition parameters, and vegetation indices under different polarimetric modes were extracted to construct the corresponding SAR feature sets. Four machine learning algorithms, including Bagged Decision Tree (BAGTREE), Random Forest Regression (RFR), Extreme Gradient Boosting (XGB), and Gaussian Process Regression (GPR), were used. Additionally, forward feature selection (FFS) procedure was employed to reduce redundant features and enhance accuracy. Results indicate that FP mode consistently demonstrated superior performance in SM retrieval, with CP mode slightly trailing behind, and DP mode yielding the least favourable outcomes. The FFS method consistently enhanced SM retrieval accuracy. Among the machine methods, RFR and GPR exhibited superior performance across all three crop scenarios and three polarimetric modes. Specifically, RFR achieved the best accuracy in the corn and soybean scenarios, with root mean square errors (RMSE) of 4.46 vol.% and 6.49 vol.%, respectively, while GPR excelled in the wheat scenario, with a RMSE of 4.29 vol.%. FFS outputs highlighted the notable contribution of polarimetric decomposition parameters derived from all three modes in SM inversion. This study provides valuable insights and serves as a technical guide for using C-band multi-mode SAR systems for SM retrieval applications. [ABSTRACT FROM AUTHOR]