Weed management is crucial to ensure crop yield and quality. However, the morphological and textural similarities between weeds and crops, multi-scale features, differences in growth stages, and uneven illumination in real-life scenarios pose significant challenges for detection. To address these challenges, this paper constructed actual scene weed dataset collected from agricultural fields with six common weed species. Then, this study presents an improved deep learning model, YOLOv8-LSI, designed for weed detection in real-world agricultural settings. In this paper, the C2f_LSK module is constructed by introducing a large convolutional kernel as a way to expand the effective sensory field of the model. Additionally, an improved SE_up channel attention module is integrated into the Neck network, improving inter-channel information interaction. Furthermore, we employ the Inner-CIoU loss function, which regulates the scaling size of auxiliary bounding boxes through a scaling factor ratio, accelerating model convergence without introducing additional penalty terms. Based on the self-collected test set, the results demonstrate that the YOLOv8l-LSI model achieves precision (P) and recall (R) rates exceeding 89%, with a mean average precision (mAP50) of 93.9% and mAP(50–95) of 81.3%. Compared to other YOLO series algorithms, our method exhibits superior performance in detecting widespread weeds with flat or entangled growth postures. Overall, our proposed method demonstrates excellent robustness and generalization, enabling high-precision weed detection in diverse environmental conditions.