Weeds, which are undesirable plants growing alongside crops, pose serious risks to the global natural environment. They compete with native species, lead to land degradation, and diminish productivity in both agricultural and forested areas. This requires developing an efficient and optimal AI-based weed management and control approach. This study introduces the natural, extensive, and publicly accessible image dataset IndianPotatoWeeds, encompassing weed species in the Indian rangelands, and applies deep learning techniques to detect weeds. The dataset, comprising over 1500 RGB images, facilitates the creation of reliable detection and classification techniques for effective robotic weed management. The study evaluates the dataset’s segmentation, detection, and classification performance using state-of-the-art deep learning models YOLO-v3 and Mask RCNN. Both models performed well and achieved mean average precision (mAP@50) of 0.70 for Yolov3 and 0.75 for Mask RCNN. These results emphasize the dataset’s utility for building effective weed detection systems. This research enhances weed management strategies, resulting in better crop yields and reduced herbicide usage. It also provides a foundation for future studies by making the dataset available for further development and exploration in the field.
