A data driven comparison of hybrid machine learning techniques for soil moisture modeling using remote sensing imagery

Soil moisture plays a very important role in agricultural production, water and ecosystem well-being particularly in rain-fed areas such as Tamil Nadu, India. This study evaluates and compares the performance of eleven machine learning models, Linear Regression (LR), Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting (GB), XGBoost (XGB), Artificial Neural Network (ANN), Long Short-Term Memory tuned with Ant Lion Optimizer (LSTM-ALO), LSTM optimized with the weighted mean of vectors optimizer (LSTM-INFO), Random Vector Functional Link optimized using Enhanced Reptile Optimization Algorithm (RVFL-EROA), Artificial Neural Network optimized via Elite Reptile Updating Network (ANN-ERUN), and Relevance Vector Machine tuned with Improved Manta-Ray Foraging Optimization (RVM-IMRFO) for predicting monsoon-season soil moisture using rainfall and topographic parameters (slope, aspect, and Digital Elevation Model (DEM)). The models were trained using rainfall data from the India Meteorological Department (IMD) and high-resolution soil moisture datasets. Model performance was assessed using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), and Combined Accuracy (CA). Among all models, XGBoost and Random Forest achieved the highest accuracy (RMSE = 0.018-0.019 m³/m³; NSE ≈ 0.983-0.984; KGE ≈ 0.988), followed closely by ANN and ANN-ERUN (RMSE ≈ 0.020 m³/m³; NSE ≈ 0.980). The hybrid models RVFL-EROA and RVM-IMRFO demonstrated moderate performance (RMSE = 0.045-0.052 m³/m³; NSE = 0.87-0.90), while LSTM-ALO and LSTM-INFO performed relatively lower due to optimizer sensitivity and data non-stationarity. Error distribution and scatter plots confirmed that ensemble and metaheuristic-enhanced models effectively captured the non-linear soil moisture variability in topographically diverse regions. This evidence shows that ANN-ERUN, RVFL-EROA and RVM-IMRFO as hybrid metaheuristic learning methods can be used to complement ensemble models like XGBoost and Random Forest to estimate soil moisture in data-sparse, heterogeneous landscapes. Higher-level hybrid tuning strategies and longer-term models should be investigated in future research in an effort to promote predictive robustness.