Accurate and robust weed detection is a critical prerequisite for precision agriculture and site-specific weed management. However, real-world agricultural environments pose significant challenges to existing object detectors due to severe illumination variability, high inter-class similarity between crops and weeds, and the prevalence of small and occluded targets at early growth stages. To address these challenges, this paper proposes DSD-YOLOv11, a domain-adaptive and structurally refined detection framework tailored for complex field scenarios. Specifically, a physics-based data augmentation strategy is first introduced to simulate realistic illumination conditions and soil background variations, effectively broadening the training distribution without increasing model complexity. In addition, a lightweight Feature Enhancement Module (FEM) is selectively injected at the P3 detection layer, where high-resolution features are preserved. The FEM integrates a SpatialAttentionLite mechanism with a projection-based feature alignment strategy, enabling precise enhancement of fine-grained spatial cues while maintaining compatibility with pre-trained backbones. An epoch-aware alpha controller is further designed to ensure stable optimization by gradually activating the enhancement pathway during training. Extensive experiments on a real-world agricultural weed dataset demonstrate that the proposed method consistently outperforms baseline YOLOv11 models across multiple evaluation metrics. Notably, DSD-YOLOv11 achieves an absolute mAP@50 improvement of +12.73 percentage points over the native baseline without data augmentation (reaching 87.14%, where the physics-based augmentation contributes +7.94 percentage points and the FEM module contributes an additional +4.79 percentage points over the augmented YOLO11n baseline), while operating at 84.2 FPS on a desktop GPU (NVIDIA RTX 4090; NVIDIA Corporation, Santa Clara, CA, USA) and 7.2 FPS on an edge computing platform (NVIDIA Jetson Nano; NVIDIA Corporation, Santa Clara, CA, USA) with only marginal parameter increases. These results indicate that the proposed framework provides an effective and efficient solution for weed detection in unstructured agricultural environments.
Xu et al. (Wed,) studied this question.