Different Dataset Preprocessing Methods for Tree Detection Based on UAV Images and YOLOv8 Network

The problem that we are trying to solve is related to determining the best approach for preparing a dataset to train a neural network for the needs of tree detection on agricultural areas. Namely, in order to analyze the condition of perennial orchards, a dataset was created by photographing agricultural areas with a drone and it consists of unprocessed, large-format images showing plantations of different ages and different planting structures. After the process of annotating the trees on the captured images, different models of automatic image preprocessing and augmentation were applied to prepare the dataset for the further training process. In order to compare different approaches of input dataset preparation, the training process was performed in the same way using the YOLOv8 neural network. The evaluation process of the obtained results has been performed by comparing metrics such as precision, recall, mAP@0.5, mAP@0.5–0.95, and F1score, for different versions of input datasets.