YOLO-RSTS: A Precise Segmentation Model for Detecting Preservative and Stimulant Spraying Regions on Rubber Trees

Jincan Zhu¹Yu Feng¹Fengming Liu¹Lee Seng Hua²Haocen Zhao¹Bangqian Chen^3,4Kou Weili¹Jian Rong^1*Guiliang Chen⁵Dingfei Xu⁶

• ¹Southwest Forestry University, Kunming, China
• ²Universiti Teknologi MARA Cawangan Pahang, Bandar Tun Razak, Malaysia
• ³Chinese Academy of Tropical Agricultural Sciences Haikou Experimental Station, Haikou, China
• ⁴Rubber Research Institute, Chinese Academy of Tropical Agricultural Sciences, Haikou, China
• ⁵Yunnan Institute of Tropical Crops, Jinghong, China
• ⁶Yunxiang Investment Co., Ltd.,, Luang Namtha Province, Laos

The application of preservatives and ethylene stimulants on rubber trees is crucial for enhancing yield and prolonging their lifespan, with preservatives and ethylene stimulants acting on different application areas. Traditional manual spraying methods are inefficient and unsuitable for the effective management of large-scale rubber plantations. Therefore, the development of high-precision automatic spraying systems has become an urgent necessity. However, conventional segmentation models face challenges such as complex bark textures and variations in lighting conditions, which result in ambiguous boundaries of the sprayed areas, thereby affecting both recognition accuracy and robustness. To address this challenge, this study proposes an improved segmentation model based on the YOLOv12n-Seg framework, named YOLO-RSTS (YOLO for Rubber Spraying Target Segmentation), tailored for accurately distinguishing preservative and stimulant spraying regions on rubber trees. The model incorporates three novel modules: CrossScaleDSC, CG-Attention, and C2f-DSC. CrossScaleDSC improves long-range dependency modeling by combining depthwise separable convolution, bottleneck structures, and skip connections, efficiently enhancing global context perception. CG-Attention applies dual attention to refine spatial and channel features, suppress background noise, and enhance class separation. C2f-DSC embeds depthwise separable convolution into the C2f structure, enabling fine-grained multi-scale feature extraction with low complexity. Additionally, RFCAConv and DWConv are integrated into the backbone and head to strengthen spatial diversity and contextual representation. On a self-constructed dataset for segmenting preservative and stimulant spraying areas on rubber trees, the proposed YOLO-RSTS model achieves notable improvements over the baseline YOLOv12n. Precision increased by 6.3% (from 0.819 to 0.882), mAP0.50 improved by 6.3% (from 0.788 to 0.851), and Recall rose by 8.1% (from 0.740 to 0.821), indicating enhanced segmentation accuracy and region completeness. Moreover, these gains were achieved with a 14.5% reduction in parameter count (from 2.72M to 2.33M). Notably, compared to the latest YOLOv13n, YOLO-RSTS also outperforms it in both mAP0.50 by 7.5% (from 0.791 to 0.851) and F1 score by 9.2% (from 0.820 to 0.896), further reinforcing its superior performance. The proposed approach offers a promising solution for vision-based autonomous spraying, and holds great potential for advancing intelligent rubber plantation management.