Fast Ready-To-Harvest Cotton Detection and Classification with YOLOv8 in Greenhouse Crops

Autonomous robotic systems come as a new way of harvesting cotton, which is needed to preserve quality while reducing expenses, as opposed to traditional methods. New algorithmic solutions must be developed to detect cotton and discern between ready-to-harvest and not ready-to-harvest bolls. Using YOLOv8 as our object detection model, this paper presents a new cotton boll image dataset, RTH-CONDIS (Ready-To-Harvest Cotton Discerning Imageset), with a total of 409 raw images, enlarged with data augmentation for the training process. After testing different sizes of YOLOv8, YOLOv8s is the most promising version for this project, with a final detection performance of 0.902 for mAP50, a recall of 0.852, and a precision of 0.901. As a result, we get satisfactory prediction metrics, considering the dataset's size. This solution is suitable for real-time, resource-limited implementations, as is needed for tracking and counting applications on a mobile harvester robot.