# Weed Segmentation in Soybean Fields and Variable-Rate Herbicide Prescription Map Generation Based on UAV Imagery and Improved YOLOv11-seg Model

## TL;DR

This paper introduces an improved YOLOv11-seg model for precise weed detection in soybean fields using UAV imagery, enabling automated herbicide application maps for precision agriculture.

## Contribution

A complete end-to-end workflow for UAV-based weed segmentation and prescription map generation is established, with a novel model architecture for improved accuracy and efficiency.

## Key findings

- The improved model achieves mAP@0.5(Box) = 0.89 and mAP@0.5(Mask) = 0.84 for weed segmentation.
- The model outperforms YOLOv8s-seg and YOLOv12s-seg with lower computational cost (25.3 GFLOPs, 8.3 M parameters).
- Integration with ArcGIS Pro enables automated decision-making for precision herbicide application.

## Abstract

Weeds pose a major threat to soybean yield during the early seedling stage, where accurate identification of their spatial locations and contours is essential for precise field management. This study proposes an improved UAV-based YOLOv11-seg framework for high-precision weed segmentation in soybean fields.

A real-field weed dataset was established under complex agricultural environments. A UAV-inspection-oriented, task-driven improved YOLOv11-seg weed segmentation method is proposed. The core of this method lies in the targeted integration and adaptation of existing modules to optimize small-target perception. To enhance detection accuracy, the backbone and neck C3K2 modules were replaced with RCSOSA (reparameterized convolution based on channel shuffle and one-shot aggregation). A Spatially Enhanced Attention Module (SEAM) was integrated into the C2PSA block to better distinguish small weeds from soybean seedlings, while the inverted Residual Mobile Block (iRMB) and adaptive down-sampling module (ADown) improved feature representation and reduced detail loss in low-contrast scenes.

Experimental results show that the proposed model achieves mAP@0.5(Box) = 0.89 and mAP@0.5(Mask) = 0.84, surpassing mainstream models such as YOLOv8s-seg and YOLOv12s-seg, with lower computational cost (25.3 GFLOPs, 8.3 M parameters).

The main contribution of this study lies in establishing a complete and practical end-to-end engineering workflow, spanning from accurate UAV image recognition to the generation of variable-rate application prescription maps. By integrating with the ArcGIS Pro platform, this solution achieves a fully automated pipeline from perception to decision-making, offering reliable technical support for intelligent weed control during the seedling stage in precision agriculture.