Reproducible Delineation of Agro-Ecological Units in a Tropical Indian District from Multi-Source Earth Observation Using Machine Learning and Deep Learning under Spatial Cross-Validation

Agro-ecological units provide a framework for agricultural planning, land-use management, and resource allocation. However, conventional AEU delineation relies heavily on expert-driven integration of climate, soil, terrain, and land-use information. This study presents a reproducible machine learning and deep learning framework for automated AEU delineation using multi-source Earth observation and environmental data. A harmonized 10 m spatial dataset was constructed by integrating Sentinel-2 optical reflectance and spectral indices, Sentinel-1 SAR backscatter, terrain derivatives, SoilGrids soil properties, bioclimatic variables, and multi-temporal land-cover information. The framework was evaluated for delineating three confirmed agro-ecological units (AEU2, AEU11, and AEU15) in Kasaragod district, Kerala, India. Five classification models, Random Forest (RF), XGBoost, Multilayer Perceptron (MLP), one-dimensional Convolutional Neural Network (1D-CNN), and U-Net, were trained and compared. To mitigate the effects of spatial autocorrelation and avoid overly optimistic accuracy estimates, a spatial block cross-validation strategy using 2 km blocks was employed. Results demonstrated that the U-Net model achieved the best performance, attaining an overall accuracy of 0.89, Cohen’s κ of 0.84, and a macro-averaged F1-score of 0.87. In comparison, Random Forest achieved an overall accuracy of 0.86 (κ = 0.79), followed by XGBoost (0.85), 1D-CNN (0.83), and MLP (0.81). Random train–test splits overestimated model performance by approximately 6–9 percentage points, highlighting the importance of spatially explicit validation. Feature importance analysis further revealed that Sentinel-2 spectral indices and terrain-derived variables were the most influential predictors of AEU boundaries. The findings demonstrate that integrating multi-source Earth observation data with advanced machine learning and deep learning techniques can effectively reproduce expert-defined agro-ecological boundaries at fine spatial resolution under rigorous spatial validation. The proposed framework offers a scalable, transferable, and automated alternative to traditional manual delineation approaches, supporting more efficient agro-ecological mapping and evidence-based agricultural planning in tropical regions.