Explainable Hybrid CNN-LSTM Framework with IG-Based Attribution Analysis for Robust Multi-Layer Soil Moisture Prediction in Tropical Cocoa Plantations

Accurate root-zone soil moisture prediction is critical for precision irrigation and sustainable crop management, yet it remains challenging due to the complex spatiotemporal interactions between soil and climate. This study presents an optimized hybrid deep learning framework that combines Convolutional Neural Networks (CNN) for feature extraction with Long Short-Term Memory (LSTM) networks for temporal sequence modelling to predict soil moisture across five subsurface layers M1-M5 (5-105 cm) in Malaysian cocoa plantations. A systematic lag optimization revealed that a 7-day lag provides the best balance between complexity and predictive accuracy, enhancing temporal feature learning. To improve transparency, Integrated Gradients (IGs) based attribution analysis was applied to identify the most influential climatic and temporal drivers at varying depths. Robustness was further validated by injecting Gaussian noise (5-20%) into meteorological inputs, showing minimal accuracy degradation. Across all layers and zones, the optimized framework consistently achieved strong predictive performance (R² > 0.94) with low error values (average RMSE ≈ 0.72, MAE ≈ 0.41, MAPE < 3%). These findings demonstrate that hyper-tuned hybrid CNN-LSTM models, when combined with explainable AI and robustness analysis, can deliver reliable and interpretable soil moisture forecasts. The framework provides a practical foundation for data-driven irrigation scheduling and adaptive water management in tropical agro-ecosystems and can be extended to broader agricultural applications.