Supervised machine learning for flood extent detection with optical satellite data

Floods are the most impactful type of natural disaster with an ever increasing frequency and people at risk. Earth Observation data can help detect flood extents on a large scale in a timely manner. In this study we implement a Machine Learning algorithm consisting of a SENet and UNet to detect water and flood related damage in optical satellite data. The approach is applied to the devastating Pakistan floods from summer 2022 for which we trained three models and analysed the feasibility and transferability of the proposed approach. A locally trained model achieves excellent performance of IoU = 93.5% (Intersection over Union) while the best transferable model achieves IoU = 83.8%.