WEED IDENTIFICATION USING VEGETATION INDICES and MULTISPECTRAL UAV IMAGING

Weeds are undesirable plants in a field that compete with desirable plants. This competition results in a significant reduction in the expected yield. Depending on the nature of weeds, different chemical herbicides are used to avoid this reduction in the yield. However, excessive use of chemical herbicides can be hazardous for the environment. Therefore, effective site-specific use of the herbicides is desirable to minimize the negative impacts on the environment. Multiple existing studies have proposed methods to identify weed concentration regions from the field images. These studies are primarily focused on the identification of weeds at early phenological stages. However, as far as we know, there is no study to identify the weeds at later phenological stages for the maize field as high similarity in the weeds and the crop makes it difficult to distinguish them. Therefore, this study proposes a novel pipeline to identify and mask weed concentration regions in the images, collected through UAV at the later phenological stage. The image dataset is collected on three different days and at three different altitudes. The proposed pipeline uses U-Net for precise and fast semantic segmentation in the image. Moreover, instead of generating ground truth images manually or from software, we used two vegetation indices GNDVI and NDRE images as ground truth images. GNDVI-based pipeline successfully identified weeds with a 0.81 IOU score whereas NDRE could achieve only a 0.75 IOU score.