Integration of Remote Sensing, Ground Data and Meteo-Climatic Variables for Agricultural Drought Monitoring: First Results of a Data-Driven Approach

Agricultural drought is one of the most critical effects of climate change. This work proposes a machine learning based approach for agricultural drought monitoring that integrates seven standard remote sensing indices computed from Sentinel-2 multispectral imagery, agricultural drought damage percentage assessed in situ and six meteo-climatic variables, including Standardized Precipitation Index and Standardized Precipitation Evapotranspiration Index. We applied the approach to 117 agricultural fields in Italy, using a multinomial logistic regression model to classify the fields into zero-risk, medium-risk, and high-risk drought damage classes. The overall performances of the proposed classification model, summarized by an F1 score equal to 0.61, are not particularly encouraging as the model struggles to distinguish between medium-risk damage and high-risk damage classes. Nonetheless, the model shows promising results in identifying fields with zero drought damage and could be applied to reduce the time and cost of in situ measurements by excluding fields with no damage from the ground data collection.