YOLO-Based Terrain Classification for UAV Safe Landing Zone Detection

Autonomous mobile robots such as Unmanned Aerial Vehicles (UAVs) must be capable of navigating safely in unknown and dynamic environments. In emergency situations such as hardware failure or loss of communication links, UAVs must be able to land safely in an area that is flat and free of obstacles. Currently, most UAVs make use of global positioning system (GPS) receivers during mission and navigation which allows Return-To-Home features for advanced UAVs in emergency scenarios such as signal loss or low battery. However, problems arise if the UAV operates in a heterogeneous environment with no GPS signal accessible. In these GPS-denied areas, it is important to determine the terrain of the environment where the UAV is located to locate a safe space to land. This paper utilizes deep learning algorithms in YOLO architecture including YOLOv5, YOLOv6, YOLOv7, and YOLOv8 to determine the type of terrains obtained from aerial images. Based on the simulations done, the most recently developed YOLOv8 obtained the highest mean average precision (mAP@0.5:0.95) of 89.1, and F1 score of 90.8. Meanwhile, the YOLOv5, YOLOv6, and YOLOv7 obtained mean average precision (mAP@0.5:0.95) of 69.5, 78.1, and 68.8, respectively, and F1 scores of 77.8, 84.9, 85.7, and 81.7, respectively. With these results, it can be confirmed that YOLOv8 outweighs the performance of the other YOLO architecture models in terms of the mAP and F1 scores in determining the terrain.