Earth observation for agricultural extreme events: drought monitoring and flood damage assessment

Climate change is intensifying the frequency and severity of hydrometeorological extremes, increasingly exposing agriculture to drought and flood risks. This thesis investigates the integration of multi-source Earth Observation (EO) data, in situ observations, and machine learning (ML) techniques to improve the monitoring and assessment of agricultural impacts caused by these events. At the municipal scale, a novel indicator — the Hierarchical Robust Combined Drought Index (HRCDI) — was developed to enhance agricultural drought assessment using freely available EO datasets in Google Earth Engine. The HRCDI integrates meteorological and biophysical variables, including the Standardized Precipitation Evapotranspiration Index, soil moisture, land surface temperature, and vegetation indices, through a fuzzy logic-based decision framework. The index captures the hierarchical propagation of drought impacts, from climatic anomalies to vegetation stress, ensuring temporal coherence and dynamic monitoring. Applied to the Province of Foggia (Southern Italy) from 2017 to 2022, the HRCDI effectively reproduced the temporal evolution and spatial variability of drought conditions, demonstrating strong scalability and operational relevance for early warning systems. At the field scale, a workflow was developed to detect drought impacts using Sentinel-2 and MODIS data. A classification model trained on EO-derived indicators from both drought and non-drought years was validated with independent field inspections provided by the Institute of Service for the Agricultural and Food Market (ISMEA). The results confirm the feasibility of satellite-based detection of drought-affected fields and highlight the potential of EO–ML frameworks for objective and spatially consistent impact assessments. For flood events, an EO–ML workflow was implemented for agricultural damage assessment at field scale, focusing on the severe flood that hit Emilia-Romagna (Italy) in May 2023. A Random Forest classifier integrating Sentinel- 2 spectral indices, topographic information, flood extent maps, and in situ agricultural damage (ISMEA data) achieved robust performance, demonstrating the value of field-level observations for model training and transparent, data-driven compensation mechanisms. Finally, the thesis underscores the importance of harmonized and spatially consistent in situ datasets for model training, calibration, and validation, which are essential to ensure the reliability and transferability of EO–ML approaches. It also advocates extending this integrated framework to minor and compound extremes to better capture complex hazard interactions and strengthen agricultural resilience under increasing climate variability