A U-Net Improved Version for Crop and Weed Segmentation from Aerial Images

Area of Science:

• Agricultural Science
• Computer Vision
• Environmental Science

Background:

• Precision Agriculture aims to optimize herbicide application for economic and ecological benefits.
• Excessive herbicide use poses risks including soil degradation, water contamination, biodiversity loss, and health concerns.
• Automated crop-weed segmentation is crucial for sustainable agricultural practices.

Purpose of the Study:

• To propose a novel deep learning architecture for accurate crop and weed segmentation in tobacco plantations.
• To develop a single-stage framework that enforces logical consistency between crop and weed predictions.
• To evaluate the proposed architecture's performance using various modern encoder backbones.

Main Methods:

• A U-Net variant architecture incorporating deep supervision, a Vegetation Global Context block, and a dual-headed output was designed.
• The model separately predicts vegetation and crop masks, with weed regions derived as the difference.
• The architecture was tested with ConvNeXtV2, FastViT, RepViT, and MambaVision encoder backbones.

Main Results:

• The proposed architecture achieved high segmentation accuracy, with Dice scores of 94.24% for crops and 93.72% for weeds.
• It demonstrated superior performance compared to previous segmentation approaches.
• Inference time was significantly reduced due to the single-stage pipeline, enabling real-time deployment.

Conclusions:

• The novel U-Net variant architecture effectively improves crop and weed segmentation accuracy in tobacco fields.
• The single-stage, dual-headed approach ensures logical consistency and enhances efficiency.
• The model's real-time capabilities make it suitable for practical precision agriculture applications, supporting sustainable farming.