Field-validated explainable AI-based UAV remote sensing for variable-rate fertilization in precision sugarcane nutrient management

Excessive and spatially inconsistent fertilizer application in sugarcane production reduces nutrient-use efficiency, increases production costs, and contributes to environmental degradation. Despite advances in UAV sensing, machine learning, and variable-rate fertilization, an explainable and field-deployable framework integrating these technologies remains underdeveloped for commercial sugarcane nutrient management. This study develops an explainable artificial intelligence (AI)-driven UAV remote sensing-based variable-rate fertilization. The framework integrates UAV multispectral sensing, XGBoost-based fertilizer recommendation, SHapley Additive exPlanations (SHAP), GIS-based prescription mapping, and embedded edge control. UAV imagery was transformed into georeferenced orthomosaics and management zones, from which vegetation indices, canopy temperature, and crop structural features were extracted to generate spatially explicit nitrogen recommendations. The resulting prescription map was executed by a Jetson Nano–ESP32 control system using GNSS positioning and a PWM-controlled fertilizer mechanism. Real-time operation was defined as online prescription interpretation and actuator control based on the pre-generated map, rather than continuous UAV image acquisition or real-time nutrient reconstruction. Field experiments compared conventional farmer practice, fixed-rate mechanized fertilization, and AI-driven variable-rate fertilization under commercial sugarcane conditions. The XGBoost model achieved R² > 0.91 with low prediction error, while SHAP analysis identified NDVI, canopy temperature, and chlorophyll-related indices as the dominant predictive factors. Compared with conventional practice, the proposed approach reduced nitrogen fertilizer use by 21.8%, increased nitrogen-use efficiency by 27.4%, and increased sugarcane yield by 11.6% (p < 0.05). Edge deployment achieved inference latency below 100 ms, demonstrating real-time field control. The proposed framework provides an interpretable and practical pathway toward data-driven, resource-efficient, and sustainable sugarcane fertilization.