Journal of Agronomy and Field Innovations

Background: Practices employed in conservation agriculture to support soil health like minimum soil disturbance, permanent residue cover, and diversified farming practices promote weed growth thus timely control of weeds is critical for maintaining crop yield without reverting to extensive tillage or general application of herbicides.
Objective: The study presents a summary of automatic weed mapping system based on computer vision that makes use of images obtained via different platforms and deep learning technology to facilitate herbicide application in conservation agriculture.
Methods: A modular frame that combines UAV-based image taking, ground images, progress made before taking images, extraction of features, and qualification of images (CNN, ResNet, EfficientNet, DenseNet, MobileNet, Vision Transformer, YOLO, Mask R-CNN, U-Net) is constructed and tested on publicly available data.
Result: The results showed that object detection and segmentation frameworks like YOLO models exhibited performance over 90% in terms of detection metrics while also having low latencies suitable for real-life applications, whereas the transformer-based models displayed a superior ability to cope with occlusion and variability in illumination, but at the expense of higher computational complexity. The combination of the GIS technology and the weed mapping system made it possible to create weed maps which provided variable-rate prescriptions of herbicides that led to a significant reduction in anticipated herbicide application.
Conclusion: The weed mapping system based on computer vision technology can be regarded as economical and practical in creating the site-specific weed management strategies that are in accordance with the sustainability principles of conservation agriculture. However, the diversity of datasets, super small object detection as well as real-time deployment still pose challenges to further research in this area.