Item

Abstract

Artificial Intelligence (AI) provides new potential to deliver more accurate and efficient tools for natural disaster and hazard management especially in terms of forecasting and prediction of natural hazards. Simultaneously, there is a great need for validation schemes that help users to understand and evaluate systems that use AI. Moreover, regulations are being developed to ensure a trustworthy use of AI in applications. To this regard, methods from explainable AI can provide valuable insights of an AI model to build trustworthiness and show technical validity. In this work, we study the rainfall-runoff prediction in Germany using a long short-term memory (LSTM) network and show how tools from explainable AI can be applied. More precisely, a single network is trained as a regional model on meteorological time-series data (e.g., precipitation, temperature, etc.) and geographical ancillary data (e.g., orography, soil type, land use, etc.). Further, input data and predictions of the model are spaced on a regular grid to yield spatially fine-grained predictions. Finally, layer-wise relevance propagation is used to show that the network processes the various types of input information in an interpretable and plausible way. The network also manages to use the geographical information to adapt its hydrological dynamics to a given location. This study is part of the project DAKI-FWS which is funded by the Federal Ministry of Economic Affairs and Climate Action in Germany to develop an early warning system to stabilize the German economy.