Integration of convolutional neural networks for flood risk mapping in Tuscany, Italy

Machine learning-based methodologies have depicted remarkable performance in digital processing of remote sensing imagery. In this work, we propose an integration of hazard susceptibility and vulnerability assessment in flood risk mapping using a CNN-based methodological framework. For this reason, we used nine predictor variables and a flood inventory from past flood events in a part of Tuscany region to train the model. Following a successful learning procedure, the performance of the proposed model was evaluated on a test dataset and depicted a promising prediction accuracy (95%). The analysis of the flood susceptibility map indicated that 4.7 and 2% of the entire study area depict very high and high susceptibility to future flood occurrences, respectively, corresponding to total areas of 44.06 and 19.33 km(2). Flood risk map depicts those land cover categories that will be severely affected in a future flood event. Among them, a large part of Livorno and a few industrial buildings were highlighted as areas of very high risk.