Authors:
Dimitri Bratzel
;
Stefan Wittek
and
Andreas Rausch
Affiliation:
Institute for Software and Systems Engineering, Clausthal University of Technology, Arnold-Sommerfeld-Str. 1, Clausthal-Zellerfeld, Germany
Keyword(s):
Machine Learning, Flood Prediction, Benchmark, Dataset, Unknown Events.
Abstract:
Global warming is causing an increase in extreme weather events, making flood events more likely. In order to prevent casualties and damages in urban areas, flood prediction has become an essential task. While machine learning methods have shown promising results in this task, they face challenges when predicting events that fall outside the range of their training data. Since climate change is also impacting the intensity of rare events (i.e. by heavy rainfall) this challenge gets more and more pressing. Thus, this paper presents a benchmark for the evaluation of machine learning-based flood prediction for such rare, extreme events that exceed known maxima. The benchmark includes a real-world dataset, the implementation of a reference model, and an evaluation framework that is especially suited analysing potential danger during an extreme event and measuring overall performance. The dataset, the code of the evaluation framework, and the reference models are publicated alongside this
paper.
(More)