Forest Vegetation Detection Using Deep Learning Object Detection Models

Abstract

Forest fires have become increasingly prevalent and devastating in many regions worldwide, posing significant threats to biodiversity, ecosystems, human settlements, and the economy. The United States (USA) and Portugal are two countries that have experienced recurrent forest fires, raising concerns about the role of forest fuel and vegetation accumulation as contributing factors. One preventive measure which can be adopted to minimize the impact of the forest fires is to cut the amount of forest fuel available to burn, using autonomous Unmanned Ground Vehicles (UGV) that make use of Artificial intelligence (AI) to detect and classify the forest vegetation to keep and the forest fire fuel to be cut. In this paper, an innovative study of forest vegetation detection and classification using ground vehicles’ RGB images is presented to support autonomous forest cleaning operations to prevent fires, using an Unmanned Ground Vehicle (UGV). The presented work compares two recent high-performance Deep Learning methodologies, YOLOv5 and YOLOR, to detect and classify forest vegetation in five classes: grass, live vegetation, cut vegetation, dead vegetation, and tree trunks. For the training of the two models, we used a dataset acquired in a nearby forest. A key challenge for autonomous forest vegetation cleaning is the reliable discrimination of obstacles (e.g., tree trunks or stones) that must be avoided, and objects that need to be identified (e.g., dead/dry vegetation) to enable the intended action of the robot. With the obtained results, it is concluded that YOLOv5 presents an overall better performance. Namely, the object detection architecture is faster to train, faster in inference speed (achieved in real time), has a small trained weight file, and attains higher precision, therefore making it highly suitable for forest vegetation detection.