A method for multi-target segmentation of bud-stage apple trees based on improved YOLOv8

The bud stage is a crucial period in the growth and development of apple trees. Accurate detections of the physiological changes in the organs (branches, buds, leaves, and the connecting parts between buds and branches) are essential for the scientific management of orchards. In the field of intelligent orchard management, image segmentation is a fundamental method for obtaining the phenotypes of fruit tree organs, making it especially critical. To address this, we have created a dataset for apple tree organ segmentation during the bud stage and have incorporated several advanced convolutional network modules (ConvNeXt V2, Multi-scale Extended Attention Module (MSDA), Dynamic Snake Convolution (DSConv)) to enhance YOLOv8 and improve the accuracy of organ segmentation in complex natural environments. In the backbone network, we have integrated ConvNeXt V2 and MSDA modules to increase the extraction of contextual information and improve the network's ability to recognize multi-scale and multi-shaped targets. Additionally, we have embedded DSConv in the network's head and utilized deformable convolution to enable adaptive sampling of the feature map, capturing a wider range of context information and improving the local feature processing capability for objects of varying sizes, ultimately leading to improved segmentation accuracy. Our models significantly outperform existing models, achieving 82.58 % mean Precision (mP), 74.58 % mean Recall (mR), 77.94 % mean Dice(mDice), 64.91 % mean IoU(mIoU), and 79.75 % mean Average Precision (mAP). Ablation studies confirm the contributions of each module to intelligent orchard management and suggest potential benefits for precise agricultural decision-making and operations.