Real-Time and Accurate State Recognition of Latex Collection Bowls From Unmanned Aerial Vehicle Imagery Using an Improved You Only Look Once Detector With Multi-Scale Attention

• Proposes YOLO-MCCR for real-time detection of latex bowl status using UAV imagery. • Integrates four novel modules to enhance feature extraction and multi-scale perception. • Achieves superior [email protected] (0.787) and recall (0.723) on a self-built field dataset. Natural rubber plantations are mostly located in hilly and mountainous areas with complex terrain and dense understory vegetation, making manual inspection inefficient and risky. Natural rubber latex is a viscous milky white to grayish-white liquid colloid. Although unmanned aerial vehicles enable large-scale non-contact visual acquisition, traditional methods struggle with stable recognition under small targets, occlusion, and complex backgrounds. To address these constraints, a detector termed You Only Look Once(YOLO)-Modified for Latex Collection Condition Recognition (YOLO-MCCR) is developed for latex collection bowl state recognition from unmanned aerial vehicle imagery. Built upon YOLO11, the proposed approach improves the network in four aspects. High-Capacity Reparameterized Four-Branch Cross Stage Partial Efficient Layer Aggregation Network(RepNCSPELAN4-high) strengthens fine-grained features, Spatial Pyramid Pooling–Fast with Squeeze-and-Excitation Network Version 2 (SPPF-SENetv2) enhances multi-scale semantic representation, Cross Stage Partial with Parallel Spatial Attention–Context Anchor Attention (C2PSA-CAA) performs context-anchor attention modeling, and multi-scale deep attention (MSDA) is embedded in the neck to enable adaptive cross-scale feature reweighting.Experiments on a self-built dataset indicate that YOLO-MCCR achieves a mean average precision of 0.787 at an intersection-over-union threshold of 0.5, representing a 3.8-percentage-point improvement over YOLO11n, and attains a recall of 0.723 with a 1.4-percentage-point gain, indicating reduced missed detections. Real-time inference is maintained at 36.6 frames per second with a compact weight size of 7.6 megabytes, supporting practical inspection. Comparisons with mainstream object detection models and visualization results validate the robustness and fine-grained perception capability of this method in rubber plantation scenes, supporting unmanned natural rubber latex harvesting.