Deep learning based road optimization using uav for disaster regions

Disaster management is becoming increasingly important for disaster preparedness, disaster risk reduction and accelerating post-disaster recovery. Technological innovations and tools provide support to manage these processes effectively. Artificial Intelligence (AI) and Unmanned Aerial Vehicle (UAV) technologies, which have developed rapidly in recent years, are used in disaster situations. Rapid developments in artificial intelligence have enabled models to provide fast and accurate results, while the integration of UAVs with artificial intelligence has made these tools more competent. In this study, in the context of disaster management, the detection of open roads and the shortest path optimization for access to the target were carried out. For this purpose, by using computer vision technique, the condition of the roads was determined with the image data taken instantaneously from the UAV in the disaster area and the most suitable route for transportation to the targeted point was determined. In the studies in the literature, disaster management processes were carried out using satellite images and old deep learning algorithms were used. Unlike the literature, this study adopts a more effective and up-to-date approach by using drones and applying new deep learning methods. For model training, a unique dataset representing the disaster situation was created and labeled. A 5-fold data replication technique was applied on the labeled dataset. Using the dataset, 80% training, 10% validation and 10% testing rates were applied. Segmentation was performed using YOLOv8x, YOLOv8x-seg, YOLOv9e and YOLOv9e-seg models. After segmentation, the mask road images were converted into single pixel wide data by skeletonization image processing. In the shortest path optimization phase, Dijkstra, A*, BFS and DFS algorithms were applied. As a result of the models compared in the path detection process, it was determined that YOLOv9e-seg gave the fastest and most accurate results with an average processing speed of 624 ms and mAP value of 84.4%. Among the shortest path algorithms, Dijkstra and A* provided the fastest access to the target point in 385 ms and 387 ms on average. It is envisaged that the developed method will shorten the intervention times with the drone used for fast detection of the target point during the disaster. It will be possible to intervene faster and more effectively in disaster areas, thus minimizing loss of life and material damages.