DATA-DRIVEN MODELS BASED ON REMOTE SENSING AND MACHINE LEARNING FOR THE EFFICIENT MONITORING OF NATURAL RESOURCES

Climate change is driving irreversible transformations to the Earth’s climate, impacting ecosystems, water resources, and agriculture. Addressing these effects requires urgent action, but managing natural resources remains complex. Traditional Decision Support Systems (DSS) depend on costly, expert-driven ground data, limiting scalability and objectivity. Satellite Remote Sensing (RS) offers consistent geospatial data, though it faces challenges like low resolution and weather interference. Advances in Machine Learning (ML) and Deep Learning (DL) help overcome these issues, enabling large-scale data processing and predictive modeling. This thesis explores integrating satellite RS with ML/DL to develop data-driven models for sustainable natural resource management. Novel data pipelines were designed, from noisy raw data through processing and model validation, following Geospatial MLOps principles for scalable and automated geospatial workflows. The models were validated through case studies: Land Subsidence (LS) mapping in Murcia using Extra-Trees Classifier (0.96 precision), InSAR signal restoration in Carpi with Transformers (MAE 0.26 cm), river water detection with U-Net (MAE 0.072), and crop mapping in Apulia using PRISMA data with RF and 2D-CNN (95% accuracy). Results show significant improvements in prediction, classification, and resource monitoring. Future work includes model enhancement, integration of multi-source data, and climate scenario analysis for long-term planning.