Machine Learning and Remote Sensing for Soil Moisture Prediction

Abstract

Artificial intelligence (AI) plays a crucial role in soil moisture prediction, essential for sustainable agriculture in arid regions where water scarcity and climate variability threaten crop yields. This study introduces an AI-driven framework to forecast soil moisture across five sites in Morocco's Draa Valley (Agdz, Tagounite, Tamegroute, Tansikht, Zagora). A dataset (2003–2024) was built by integrating historical climate records with remote sensing indicators using Google Earth Engine (GEE). Six models Random Forest (RF), XGBoost, CatBoost, k-Nearest Neighbors (KNN), Long Short-Term Memory (LSTM), and Temporal Convolutional Networks (TCN) were assessed with RMSE, NSE, MSE, and MAPE. Results showed that tree-based models clearly outperformed deep learning, with RF, XGBoost, and CatBoost achieving RMSE of 2.89%–9.11% and NSE > 0.965. Findings highlight the potential of AI-based soil moisture prediction to enhance irrigation scheduling, optimize water allocation, and support climate-resilient farming, offering a scalable solution for precision agriculture.