Weed29: A Weed Dataset for Machine Learning

Weeds are a major threat to crop growth, necessitating accurate identification for intelligent weeding methods such as mechanical removal, targeted spraying, and laser-based approaches. However, existing weed datasets are often limited in species diversity and geographic coverage. To address this, we propose Weed29, a large-scale weed image dataset containing 15,204 annotated images across 29 common weed species from cornfields, rapeseed fields, and lawns in southwestern China, captured under diverse environmental conditions and growth stages. Building upon this dataset, we designed an improved detection model, YOLOv7-CA, by integrating Coordinate Attention into the YOLOv7 backbone. The proposed model achieved a mean Average Precision (mAP) of 93.4 percent, outperforming existing benchmarks in both accuracy and inference speed. These results demonstrate the practical potential of Weed29 and YOLOv7CA in real-time, field-level weed recognition. Limitations remain in detecting small or visually similar weeds under extreme conditions, which we aim to address in future work. Overall, this study offers a valuable dataset and model for advancing precision agriculture and intelligent weeding systems.