Potential of Remote Sensing Images for Soil Moisture Retrieving Using Ensemble Learning Methods in Vegetation-Covered Area

Soil moisture (SM) plays a critical role in various fields such as agriculture, hydrology, and land-atmosphere interactions. This study aims to evaluate the performance of the categorical boosting algorithm (CatBoost) in comparison to other multiple-boosting algorithms for SM prediction. Appropriate feature selection is vital for achieving accurate predictions, and this study focuses on identifying relevant features and assessing CatBoost's suitability for the task. The study incorporates several boosting algorithms including Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and CatBoost to estimate SM. Results indicate that radar backscatter coefficient, soil roughness, and digital elevation model (DEM) are crucial features for SM retrieval. Comparatively, CatBoost outperforms GBDT, XGBoost, and LightGBM in various feature combinations. The most favorable results are obtained when utilizing all features as inputs for the algorithm. These optimal results yield a mean absolute error (MAE) of 2.40 vol.%, mean relative error (MRE) of 0.16 vol.%, root mean square error (RMSE) of 3.26 vol.%, and Pearson correlation coefficient of 0.73. Additionally, the study analyzes the inversion results for different ranges of SM and Normalized Difference Vegetation Index (NDVI). Within the range of SM from 0 to 25 vol.% and NDVI from 0 to 0.7, utilizing all features yields the most accurate results. Using CatBoost, this approach achieves an MAE of 1.52 vol.%, MRE of 0.12 vol.%, RMSE of 2.11 vol.%, and R of 0.81. The study suggests that applying boosting algorithms, especially CatBoost, holds promise in accurately estimating surface SM.