Enhancing Weed Detection and Segmentation with Advanced Deep Learning Algorithms leveraging RGB-NIR Imaging

This study explores the potential of improving weed identification accuracy and reducing herbicide usage through the use of computer vision analysis, specifically incorporating imaging datasets that include both RGB and near-infrared (NIR) channels. The thesis investigates the application of deep learning models, namely YOLO and U-Net, to improve the detection of weeds and their semantic segmentation in precision agricultural contexts. It utilizes two meticulously annotated datasets: ACRE for object detection and Sunflower for semantic segmentation. Additionally, the study considers the effect of varying NDVI thresholds on segmentation performance, finding that fine-tuning these thresholds can be key to enhancing outcomes. The customized U-Net model deployed for segmentation offers promising results, outpacing several leading models in effectiveness, particularly with integrating NIR data. Combining RGB images with NIR channels has shown a notable boost in the deep learning models’ ability to differentiate between weeds and crops—an essential step forward for precision agriculture. The findings reveal marked progress in segmentation accuracy, with the UNet-ResNet50 model incorporating RGB+NIR data achieving significant benchmarks: a mean Intersection over Union (IoU) score of 0.88, a Crop IoU of 0.93, and a Weed IoU of 0.71, respectively with an improvement of 2.7%, 4%, and 3%. Conversely, YOLO models have demonstrated solid capabilities in object detection tasks utilizing only RGB data. YOLOv8 achieves a mean average precision (mAP) of 0.48, which is competitive with the state-of-the-art models in object detection. While surpassing some traditional methods, these encouraging outcomes highlight the promising role that multispectral imaging could play in transforming precision agriculture. This research suggests a path toward an era where the merging of sophisticated imaging technologies fundamentally changes precision agriculture. In summary, the study presents evidence that combining NIR data with RGB channels can significantly boost the precision and effectiveness of weed segmentation systems, contributing to the development of more eco-friendly agricultural practices.