Objective To propose an improved algorithm to enhance the accurate detection and identification ability of weed target, and solve the issues of insufficient real-time performance, small target scale, as well as severe overlapping occlusion in weed detection at the Glycine max (soybean) seedling stage.

Method The YOLOv11n was study selected as the baseline model and an improved target detection method, YOLOv11-FSi, was proposed. This method reconstructed the traditional C3k2 module by introducing FSConv, and enhanced the model’s small target feature representation ability through the collaborative interaction and feature fusion of frequency and spatial domain branches. A SEAM channel-spatial attention module was embedded at the output end of the feature pyramid network to strengthen feature representation and target discrimination under complex backgrounds and occlusion conditions. Simultaneously, the Inner-MPDIoU loss function was introduced to optimize bounding box regression, improving small target localization accuracy and detection robustness.

Result After the improvement, the precision, recall, and mAP_0.5 reached 78.3%, 72.2%, and 81.7%, respectively, representing improvements of 4.2, 4.5, and 4.0 percentage points compared to the baseline model YOLOv11n. Furthermore, the model had only 2.8 M parameters and a computational cost of 7.8 G, maintaining excellent lightweight characteristics. Edge device deployment tests showed that after TensorRT optimization and acceleration, the improved model achieved an inference speed of up to 126.6 frames per second, meeting the requirements for real-time detection.

Conclusion The proposed YOLOv11-FSi model achieves high-precision identification of weeds in soybean field at the seedling stage, while maintaining lightweight characteristics, providing a hnical support for intelligent and precise weeding operations in soybean fields.