Evaluation of lychee winter shoot length using UAV remote sensing technology

Lychee is an important cash crop in southern China. The excessive growth of winter shoots in the early winter season will lead to an increase in nutrient consumption, which in turn affects flower bud differentiation and fruit yield. To address the issue of low efficiency in traditional manual measurement methods, this study proposes an automated detection method using UAV remote sensing technology and an improved YOLOv8n_OBB_SEB algorithm. Through multi-dimensional optimization, this method successfully solves the issue of the small size of winter shoots, similar color to branches, and leaf occlusion in the orchard environment. The specific improvements include: using the SAHI algorithm for image slicing to assist inference to improve the recognition ability of small targets; embedding the Starblock in the StarNet model into the C2f module and replacing the original C2f module in the Backbone, which reduces the number of parameters and strengthens the feature extraction ability; replacing the Concat module in the Neck part with the BiFPN structure to optimize multi-scale feature fusion; introducing the EMA attention mechanism and embedding it into the C2f module in the Neck part to achieve pixel-level attention allocation and enhance the distinguishability between the target and the background. The experimental results show that on the lychee winter shoot test set, the detection accuracy of the improved YOLOv8_OBB_SEB algorithm reaches 89.2%, which is 20.7% higher than that of the original YOLOv8_OBB algorithm. Compared with other mainstream algorithms, YOLOv8_OBB_SEB shows stronger competitiveness and robustness. Through inference detection, the four coordinates of the target rotation box can be obtained, and the actual size can be calculated by converting the pixel height to estimate the real length of the lychee winter shoots. According to the estimation results, this paper divides the winter shoots into two groups: those requiring drug intervention and those not requiring drug intervention. The specific judgment standard is that when the length of the winter shoot exceeds 3 centimeters, it is classified into the group requiring drug intervention, and when the length of the winter shoot is less than 3 centimeters, it is classified into the group not requiring drug intervention. Remote sensing data of 24 lychee trees were collected on December 3, 2024. The spraying requirements were determined through manual field surveys, which were then compared and verified with the model inference results. Finally, it was concluded that the accuracy of the model reached 83.3%. This classification method provides reliable decision support and a clear decision-making basis for the precise management of winter shoots.