AI-Enabled Vision Transformer for Automated Weed Detection: Advancing Innovation in Agriculture

Abstract

Precision agriculture is focusing on automated weed detection in order to improve the use of inputs and minimize the application of herbicides. The presented paper outlines a Vision Transformer (ViT) model for weed detection in crop fields, that tackle difficulties stemming from the resemblance of crops and weeds, especially in complex, diversified settings. The model was trained via pixel-level annotation of the images obtained using high-resolution UAV imagery shot over an organic carrot field with crop, weed, and background. Due to the nature of the mechanism in ViTs that includes self-attention, which allows it to capture long-range spatial dependencies, this approach can very well distinguish crop rows from inter-row weed clusters. To solve the problem of class imbalance and improve the generality of the patches, techniques of data preprocessing such as patch extraction and augmentation were used. The effectiveness of the proposed approach has been confirmed by an accuracy of 89.4% in classification, exceeding the efficiency of basic models such as U-Net and FCN in practical application conditions. This proposed ViT-based approach is a marked improvement in crop management; and provides the prospect for selective weed control, in support of more sustainable agriculture. This model can also be integrated into AI-based tractors for real-time weed management in the field.