HMTS-YOLO: Detection of Weed Apex and Root/Stem-Base Treatment Regions for Precision Laser Weeding

Abstract Under complex field backgrounds, weed apex and root treatment regions are readily affected by soil texture, surface cracks, target-scale variation, and morphological differences, resulting in missed detections, false detections, and bounding-box regression errors, particularly for small-scale and elongated targets. To improve dual-class treatment-region detection for field weeds, a dataset comprising 4,582 images and 13,421 bounding boxes was curated from existing field images and treatment-region annotations, and HMTS-YOLO (Hierarchical Morphology-aware Texture-Suppression YOLO) was developed using YOLOv8s as the baseline. The proposed model introduces STS-C2f, LDM-C2f, and DMA-SPPF at the P2, P3, and P5 feature levels, respectively. These modules improve feature representation across target scales and morphologies through shallow-layer soil-texture suppression, mid-level dual-morphology detail enhancement, and deep multi-scale context modeling. On the validation set, HMTS-YOLO achieved a Precision of 72.736%, Recall of 68.988%, mAP50 of 73.129%, and mAP50-95 of 35.177%, representing improvements of 3.619, 2.665, 4.341, and 4.514 percentage points, respectively, over the baseline. The model contains 11.572 M parameters and requires 69.157 GFLOPs. Class-wise evaluation showed that the mAP50-95 values for apex and root reached 42.212% and 28.142%, respectively, corresponding to gains of 4.969 and 4.059 percentage points over the baseline. These results demonstrate that feature-level mechanisms for texture suppression, morphology modeling, and contextual fusion can effectively improve the overall detection performance of apex and root treatment regions in complex field scenes.