Climate-aware hybrid 1D-CNN-LSTM model for multi-layer soil moisture prediction in tropical Cocoa plantations

Accurate multi-layer soil moisture (SM) prediction is crucial for optimizing irrigation management and supporting sustainable Cocoa production, yet modelling SM dynamics in tropical environments remains challenging due to strong climatic variability, heterogeneous soils, and depth-specific interactions. To address these challenges, this study develops a climate-aware hybrid convolutional neural network, long short-term memory (CNN-LSTM) model capable of predicting SM across five root-zone layers (M1–M5) in Cocoa ( Theobroma cacao ) plantations in Malaysia. The model integrates multi-source environmental variables including temperature, humidity, rainfall, solar radiation, wind speed, wind gust, and dew point with in-situ multi-depth soil measurements to capture the complex spatial and temporal dynamics that characterize humid tropical agroecosystems. While the CNN component extracts localized spatial patterns, the LSTM component effectively learns long-term temporal dependencies, enabling accurate depth-specific SM forecasting. Model performance for moisture prediction was assessed using standard regression metrics (MSE, RMSE, MAE, MAPE, and R 2 ), with results showing consistently high accuracy across all soil layers and zones (R 2 > 0.94; average RMSE < 0.9). Time-series and scatter plot analyses further confirmed strong agreement between observed and predicted values. Importantly, unlike earlier research, this study not only predicts multi-layer root-zone soil moisture but also demonstrates reliable layer-wise forecasting up to 3 days ahead in a tropical, high-humidity Cocoa farm, providing a novel contribution to the existing literature. By offering robust, depth-resolved soil moisture predictions and short-term forecasts, this hybrid deep learning framework establishes a practical foundation for automated, climate-aware irrigation systems in perennial tropical crops. The results highlight how combining environmental feature engineering with advanced deep learning can strengthen data-driven decision support and enhance the resilience and sustainability of Cocoa production under Malaysia's humid, rainfall-variable climate. • Developed a climate-aware hybrid 1D-CNN-LSTM model for multi-layer root zone soil moisture prediction. • Integrated multi-source environmental and in-situ multi-depth soil data for tropical Cocoa farms. • Performed lag optimization analysis, and achieved high prediction accuracy across five soil layers (M1–M5) with R 2 above 0.94 in all spatial zones. • Demonstrated novel layer-wise accurate soil moisture forecasting up to 3 days ahead in tropical conditions. • Provided a practical foundation for automated, climate-smart irrigation in perennial Cocoa systems.