DDB-YOLO: A lightweight deep learning approach for individual tree crown segmentation from UAV imagery towards high-fidelity vegetation index derivation of rubber tree

Accurate monitoring of rubber tree growth and phenology is essential for understanding their effects on latex yield and quality. In densely planted plantations, high canopy density, overlapping crowns, and complex crown architectures present substantial challenges to conventional methods. To overcome these limitations, we built an integrated framework that combines a lightweight deep learning model for precise crown segmentation with a vegetation index computation strategy for spectral purification. This study proposes an instance segmentation model—DDB-YOLO, whose core architecture design centers on synergistic integration. It incorporates the DDCNv4 module for geometric adaptation and multi-scale feature extraction. Subsequently, the high-quality features are fed into the BRAttention module to address complex boundary and global context issues. Simultaneously, the DSConv module is adopted to systematically balance the computational load introduced by BRAttention, ensuring computational efficiency. This deeply optimized design achieved significant results in performance: Compared to the baseline YOLOv8n-Seg, DDB-YOLO improved Box mAP50 by 3.2 % and Mask mAP50 by 3.7 %, ultimately reaching accuracies of 79.4 % and 79.3 %, respectively. This design fully embodies outstanding resource optimization while maintaining high precision: it features a parameter count of 2.46 M and FLOPs of 9.66 G, effectively resolving the segmentation challenges in high-density plantations. Using these segmented crowns, we established the Individual Tree Crown Vegetation Index (ITC-VI), which computes vegetation indices exclusively within the crown, thereby minimizing background interference. ITC-VI outperforms conventional Area-Averaged Vegetation Indices (AAVI) in capturing phenological dynamics and distinguishing subtle physiological differences among six rubber cultivars. This framework has built a bridge to link crown-level growth and phenology with latex yield and quality, which provides precise individual tree-level measurements to support precision management and sustainable production in rubber plantations. The approach is computationally efficient, scalable, and readily applicable to field-level monitoring in high-density plantations.