Automated hazard zone mapping with neuronal networks
Matthias Rauter, Mathematics
Gravitational mass flows, such as avalanches are common events in mountainous regions with high socio-economic impact. Within the last few decades, hazard zone maps have been deployed in many European countries to systematically approach this threat. These maps mark vulnerable zones in habitated areas and make this information publicly available. Hazard zone maps are created in a complex and expensive process, based on experience, empirical models and physical simulations. Hence, their coverage is limited to crucially important regions.
In this work we are investigating the potential of neuronal networks to predict natural hazards and thus generate hazard zone maps in an efficient and automatic way. Official hazard zone maps are utilised as training data, which allows us to transfer human-made hazard zone maps to regions that are not covered yet.
We, for the first time, use a convolutional neuronal network (CNN) to predict the vulnerability of a point in space, based on the topography surrounding it. By repeatedly extracting different sections of the terrain and predicting the respective vulnerability, the CNN can generate hazard zone maps.
The performance is evaluated by comparing predictions to hazard zone maps that the network has not seen before. First results are very promising. The network is able to predict most avalanches and widely matches the official hazard zone map. We achieve an accuracy of up to 85%, depending on network and terrain complexity. We are confident to archive further improvement by fine tuning the network architecture and increasing the training data.