TomatoweedDet: a real-field multi-class weed detection dataset and YOLO benchmark for tomato production systems

This study presents an approach for the object detection of multiple weeds in tomato production systems based on deep learning. A comprehensive dataset has been collected in three provinces of Türkiye (Balıkesir, Ankara, and Aksaray) under real-world field conditions. The data set has 32,607 images and 44,165 bounding boxes annotations. The two weed species included in the dataset are, to our knowledge, underrepresented in the current deep learning-based agricultural object detection literature. Drone and smartphone cameras took pictures at different times of the day (morning, noon, and afternoon) of different soil textures, light levels, and weather conditions, such as rain, mud, and shadows. The dataset reflects agricultural diversity as it exists in the real world, unlike previous studies that relied on controlled experimental environments. The model was trained using YOLO-based deep learning algorithms within the PyTorch framework. The metrics Precision, Recall, mAP@0.5, and mAP@[0.5:0.95] were used to evaluate the performance of the models. In this study, seven different YOLO architectures were comparatively evaluated on the TomatoWeedDet dataset created under real field conditions. The results show that the YOLOv8l model demonstrates high performance in the multi-class weed detection task and has significant potential for precision weed management applications. The model that was created could be used in mobile or embedded systems to monitor weeds in real time with drones. The proposed system enables targeted herbicide application and less use of chemicals. This study advances research on weed detection using deep learning. It also helps to make precision and sustainable farming systems a reality.