Deep learning improved YOLOv8 algorithm: Real-time precise instance segmentation of crown region orchard canopies in natural environment

Abstract

Autonomous detection and segmentation of tree canopies, trunks, and nontarget areas is crucial for optimizing pesticide use and minimizing off-target spray deposition in orchards for precision spraying. To improve the precision of pesticide application in orchards, a real-time orchard tree segmentation method using an improved YOLOv8 deep learning algorithm was proposed. First, to enhance precision, a dilated convolution approach was introduced into the backbone of the YOLO algorithm, and then, to better converge and regularize the model GELU activation function was added into the dilated convolution network. Finally, a total of 600 images of manually labeled orchard canopies were used, and the adaptive gradient algorithm (AdaGrad) was employed to adjust the weights and minimize the loss function. Additionally, the ReduceLROnPlateau scheduler was utilized to fine-tune the model during training. The performance of the proposed method was assessed through a comparative analysis. The results of the improved algorithm showed a mean average precision (mAP) of 93.3 % with an improvement of 1.3 %, precision (P) of 93.6 % with an improvement of 3.8 %, and inference time of 28 ms. To verify the effectiveness of the proposed method, the results were compared with those of YOLO family algorithms and Mask RCNN. An outdoor experiment was conducted to assess the algorithm’s performance to evaluate the deposition outcomes of the spray for variable-rate spraying, which proved to be reduced by 40 % compared to uncontrolled spraying. The deposit coverage was more than 20 % on trees of different sizes, which were positioned 1.7 m away from each side of the spraying robot. The obtained results confirm the satisfactory performance of the improved YOLOv8 algorithm and aim to offer valuable technical assistance for improving the efficiency and precision of an orchard spraying machine.