Flooded house detection for natural disaster managment

In this paper we propose a method for submerged houses roof detection on a region with a flood natural disaster by a drone on the fly. We try two different object detection architecture, YOLOv6 and DETR, to find out which one has better performance for a drone application. We propose three different image datasets, namely Floodnet dataset, Giannitsa dataset and RedRoofs dataset merged into one large dataset. Both architecture’s house roofs detectors (YOLOv6 and DETR) trained a) on mixed regions images (flooded and non-flooded regions) and tested on flooded region images and b) on non-flooded region images and tested on flooded region images. Our results depicts that both architecture’s house roof detectors performs very well in environments different from the ones they were trained for. Finally, we do not choose the house roofs detector with the best performance because of computational power requirements and electric power consumption limitations that there are on a drone. The selected house roofs detector requires significantly lower computational but is has slightly lower performance than best.