Weed Detection and Segmentation in Cotton Fields Using Object Detection and SAM

Weeds are unwanted plants that grow on farms and compete with crops for nutrition and other growth factors. Weeds are everything other than the desired crop. This significantly reduced the yield of the desired crops. Therefore, the elimination of weeds is an essential practice in farming. Traditionally, there are various ways farmers do this, such as weeding with manual labor, mechanical tools, and spraying herbicides that are hazardous to humans. However, as technology progresses, new methods are being developed for precision agriculture. Cotton is a soft, fluffy staple fiber. It is essential to effectively control weeds in cotton fields to enhance cotton production. In this study, we propose a lightweight solution for weed detection and segmentation using object detection models and SAM. For object detection, there are various models based on CNN and Transformer architectures. After training and testing these models on CottonWed12 dataset, YOLOv10-s achieved overall 0.97135 mAP50 and 0.94655 mAP50-95, YOLOv10-n achieved overall 0.95869 mAP50 and 0.92467 mAP50-95 , YOLOv5-s achieved overall 0.96295 mAP50 and 0.84951 mAP50-95, and YOLOv8-n achieved overall 0.96671 mAP50 and 0.93159 mAP50-95. Then providing output bounding box coordinates of these model to SAM we can obtain quality segmentation of detected weed. By combining these deep learning models into mini robots, farmers can eliminate weeds with a precise dose of herbicide applied at the sight of weeds.