A Novel Approach to Weed Detection Using Segmentation and Image Processing Techniques

Agronomic techniques ensure crop safety and maximize yield. One common goal of these techniques is to control weeds, which can compete with crops for nutrients, water, and light and can thus reduce crop yield. For training our model, we used an open-source Sugarbeets dataset that was available online. We chose 1,300 images featuring vegetation and weeds to est our system. The dataset consisted of multiple stages of crops right from sowing till it matures. For testing, we selected 1300 images from this huge dataset with crop sizes ranging between two to-five centimeters. Further, we applied augmentation techniques to increase the dataset size from 1300 images to 3900 images. Our paper proposes a hybrid approach that combines semantic segmentation and the YOLO (You Only Look Once) algorithm to distinguish between plants and unwanted vegetation, for which we achieved an accuracy of 93% in our model between crops and weeds.