Machine learning classification of geomorphometric segments for floodplain detection and delineation

We propose a floodplain detection and delineation workflow based on the Copernicus DEM. It involves watershed segmentation of slope and a machine learning algorithm - Multilayer Perceptron (MLP) model, for the geomorphometrical variables of the segments. For the method’s accuracy, spatially separated training and testing areas were used to assess the generalization power. Seven classes of landforms were labeled to allow the MLP model to statistically identify the flat floodplain class in the feature space compared to the adjacent channels, hillslopes, and levees. The confusion matrix data shows good generalization power with 94% accuracy for the floodplain class detection. The results show promising perspectives to solve the problem of quaternary deposits mapping and flood risk assessment.