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

Area of Science:

• Agricultural Science
• Environmental Science
• Data Science

Background:

• Soil moisture is vital for agricultural productivity, water resources, and ecosystem health, especially in rain-fed regions like Tamil Nadu, India.
• Accurate soil moisture prediction is essential for effective water management and agricultural planning.

Purpose of the Study:

• To evaluate and compare the performance of eleven machine learning models for predicting monsoon-season soil moisture.
• To identify the most accurate models using rainfall and topographic parameters.

Main Methods:

• Eleven machine learning models were assessed, including ensemble methods (XGBoost, Random Forest) and hybrid models (LSTM-ALO, LSTM-INFO, RVFL-EROA, ANN-ERUN, RVM-IMRFO).
• Models were trained using India Meteorological Department rainfall data and high-resolution soil moisture datasets.
• Performance was evaluated using metrics such as Root Mean Square Error (RMSE), Nash-Sutcliffe Efficiency (NSE), and Kling-Gupta Efficiency (KGE).

Main Results:

• XGBoost and Random Forest demonstrated the highest accuracy in soil moisture prediction (RMSE ≈ 0.018-0.019 m³/m³, NSE ≈ 0.983-0.984).
• Artificial Neural Network (ANN) and ANN-ERUN models also showed strong performance (RMSE ≈ 0.020 m³/m³, NSE ≈ 0.980).
• Hybrid models like RVFL-EROA and RVM-IMRFO achieved moderate performance, while LSTM-based models showed lower accuracy due to optimizer sensitivity.

Conclusions:

• Ensemble and metaheuristic-enhanced machine learning models effectively capture non-linear soil moisture variability.
• XGBoost and Random Forest are highly effective for soil moisture prediction.
• Hybrid models like ANN-ERUN, RVFL-EROA, and RVM-IMRFO can complement ensemble methods for soil moisture estimation in data-sparse regions.