Detecting and Mapping Individual Fruit Trees in Complex Natural Environments via UAV Remote Sensing and Optimized YOLOv5

The location and number of individual fruit trees (IFTs) are critical for investigations on planting areas, fruit yield predictions, and smart orchard planning and management. These data are conventionally obtained through manual and statistical investigations that require long, laborious, and costly efforts. Object detection models of deep learning could provide an opportunity to detect IFTs accurately, which is essential for rapidly obtaining these data and reducing human operation errors. This study proposed an approach for detecting IFTs and mapping their spatial distributions by integrating deep learning with unmanned aerial vehicle (UAV) remote sensing. UAV remote sensing was used to collect high-resolution images of fruit trees in pomelo orchards in Meizhou, South China. Based on these images, a new individual pomelo tree image sample dataset was created through manual interpretation and field investigation. The evaluation results revealed that YOLOv5s was the best model among the five YOLOv5 models (i.e., YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x, whose layers, parameters, and floating-point operations all increased with the depth and width of layers) of different scales considered for optimization. Moreover, the coordinate attention (CA) optimized YOLOv5 model (YOLOv5s-CA) is the best model (named FruitNet) with the best overall accuracy for detecting IPTs among all seven attention-optimized YOLOv5 models and other state-of-the-art object detection models, such as faster R-CNN and YOLOv8s. The IPTs in the study areas were detected using FruitNet, their number and planting area were counted, and their spatial distributions were mapped based on the predicted results of the IPTs. This study suggested that our proposed approach could provide key data and technical support for smart orchard management.