Spatiotemporal Flood Susceptibility Mapping using a Hybrid CNN-ConvLSTM Architectur

Improving the accuracy of flood forecasts is paramount to minimising flood damage. Machine learning (ML) models are increasingly being applied for flood forecasting. Such models are typically trained on large historic hydrometeorological datasets. In this work, we evaluate methods for selecting training datasets that maximise the spatio-temporal diversity of the represented hydrological processes. Empirical results showcase the importance of hydrological diversity in training ML models.

The advent of new satellite technologies improves our ability to study floods. While the depth of water at flooded buildings is generally the most important variable for flood researchers, extracting this accurately from satellite data is challenging. The software tool presented here accomplishes this, and tests show the tool is more accurate than competing tools. This achievement unlocks more detailed studies of past floods and improves our ability to plan for and mitigate disasters.