Detecting flooded areas with machine learning techniques: case study of the Selška Sora river flash flood in September 2007

Abstract Floods seem to appear with increased frequency from one year to another. They create great damage to property and in some cases even result in lost lives. However, a quick and effective response by rescue services can greatly reduce the consequences. Machine learning techniques can reduce the time necessary for flood mapping. We test various machine learning methods to find the one with the highest classification accuracy. We also present the most important points for quick and effective machine learning procedures on remote sensing data. First, the data must be prepared correctly. We use satellite images, digital terrain models (DTMs), and the river network. The data in its primary form (e.g., bands of multispectral satellite images or DTMs) is insufficient. We also need certain derived attributes, such as the vegetation index or the slope derived from the DTM. Second, we must select suitable training samples and a suitable machine learning method. This approach to determining floods is presented in a case study of flash floods in the Selška Sora river valley. Machine learning techniques have proven successful in quickly determining flooded areas. The best results are produced by the J48 decision tree algorithm. The success of the ensemble machine learning methods is comparable to the J48 algorithm, while the JRip classification is not as good.