Deep learning (Mask R-CNN)- vs. Manual editing approach to detect flood damage in operational settings in Germany

Kurzfassung

Heavy rains occurred in the Western-part of Germany, where the relief is hilly and it has narrow valleys. The last flood in Ahr Valley happened in 1804 and caused 63 casualties. Nobody from this generation has seen or could image such catastrophe in his hometown. Rainfall was expected to reach between 180-220 l/m². The soil was so saturated before the rains that could not absorbe more water. The rivers Ahr, Erft, Kyll, and Sauer in the border with Luxemburg flooded. The water reached 6 m height. Germany has invested since many decades million of euros in dams and others to regulate the rivers flow and it has worked very nice all these years but due to climate change, rivers increased their flow extremely that got out of control. The German Aerospace Center (DLR) in cooperation with the German Federal Institutions coordinated actions to respond to this natural disaster. With our own light aircrafts and helicopters, we acquired optical images of <15 cm resolution in challenging geographic conditions. Parallel to this, the International Charter Space & Major Disasters, which is inside the Earth Observation Center by the DLR processed SAR images to detect floods. The results shown sparse floods (poor results) because the water has gone and the remained water shown rough surfaces due to the water movement and debris. In the Ahr Valley due to its hilly relief occurred flash floods and took away cars, houses and bridges. The destruction was visible but could not be detected automatically the whole affected area based on SAR images. In the emergency context it was required to detect floods in the Bad Neuahr-Ahrweiler in the Ahr Valley. Me and my colleague started to digitize the ground labels. This could be later used for training the deep learning network. The floods mask layer was sent to the federal institutions and organizations. Asap, we finished the emergency work in Ahr Valley, we saw the need to develop a new automatic workflow based on optical images of <15 cm resolution. Then, I started with this work and took the flood layer to train the model. My target was to detect floods in 88 km² aerial photography. The ground dataset has an area of 1.5 km² and the ground labels contains flood damage. The ground dataset was chipped 400 x 400 pixel and were trained with Mask R-CNN algorithm. The model was trained in Jupyter notebook with 30 epochs. The results needs refinement but detected very well the damage outside the rivers. The output flood damage's layer was post-processed before the creation of maps and delivery to the federal institutions, Bavarian Red Cross and THW. In an emergency context, an approximate of the affected area and the number of people, is enough to start planing the rescue operations. Detailed analysis takes more time, that's why such analysis are necessary for further operation's steps. Currently, I am writing a scientific article about how much time it is possible to reduce by means of the deep learning approach in the floods emergency context. This paper will give you more details about the model accuracy and the processing time. Thank you.