Machine Learning-Driven Solutions for Enhancing Agricultural Sustainability: The Case of Mango Farms in Sicily

This thesis presents a novel interdisciplinary framework that transfers machine learning models originally developed for astroparticle physics to agricultural forecasting, with a specific application to mango farming in coastal Sicily. Techniques such as XGBoost, Long Short-Term Memory (LSTM) networks, and Residual Networks (ResNet), previously applied to reconstructing inclined muon events in water-Cherenkov detectors, were reconfigured to address agro-meteorological challenges including temperature prediction, wind component forecasting, and climate risk assessment.The models were trained and validated using a combination of ground sensor data, MODIS satellite-derived indices (e.g., Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), Aerosol Optical Depth (AOD)), and topographic features from Digital Elevation Models (DEM). A hybrid architecture combining ResNet and XGBoost was developed for high-resolution temperature forecasting, while a Bayesian Network was used to integrate probabilistic risk scenarios related to drought, wind, and heat stress. These tools were evaluated against unseen data from 2022–2024, achieving high predictive accuracy and robustness.The findings demonstrate that machine learning models optimized for spatial-temporal analysis in astrophysics can be successfully adapted for precision agriculture under climate stress. This interdisciplinary approach improves predictive decision-making, resource allocation, and crop resilience. The methodology offers a transferable blueprint for applying domain-agnostic ML to broader sustainability challenges, including environmental monitoring and food security.