Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lightweight Multispectral Crop-Weed Segmentation for Precision Agriculture

Zeynep Galymzhankyzy, Eric Martinson

TL;DR

This work tackles robust crop–weed segmentation under diverse field conditions by introducing a lightweight transformer–CNN hybrid that processes RGB, NIR, and Red-Edge data through modality-specific encoders and a dynamic gated fusion mechanism. The approach achieves strong multispectral segmentation with 8.7 million parameters, yielding a mean IoU of $mIoU$ = 78.88% on the WeedsGalore dataset, substantially outperforming RGB-only baselines while maintaining real-time feasibility for UAV edge deployment. Compared with heavier baselines like DeepLabv3+ and MaskFormer, the model offers a favorable accuracy–efficiency trade-off suitable for site-specific weed management in precision agriculture. The work highlights the importance of adaptive modality weighting and multi-scale context, with future directions including domain adaptation, self-supervised pretraining, and hardware-aware optimization for field-scale deployment.

Abstract

Efficient crop-weed segmentation is critical for site-specific weed control in precision agriculture. Conventional CNN-based methods struggle to generalize and rely on RGB imagery, limiting performance under complex field conditions. To address these challenges, we propose a lightweight transformer-CNN hybrid. It processes RGB, Near-Infrared (NIR), and Red-Edge (RE) bands using specialized encoders and dynamic modality integration. Evaluated on the WeedsGalore dataset, the model achieves a segmentation accuracy (mean IoU) of 78.88%, outperforming RGB-only models by 15.8 percentage points. With only 8.7 million parameters, the model offers high accuracy, computational efficiency, and potential for real-time deployment on Unmanned Aerial Vehicles (UAVs) and edge devices, advancing precision weed management.

Lightweight Multispectral Crop-Weed Segmentation for Precision Agriculture

TL;DR

This work tackles robust crop–weed segmentation under diverse field conditions by introducing a lightweight transformer–CNN hybrid that processes RGB, NIR, and Red-Edge data through modality-specific encoders and a dynamic gated fusion mechanism. The approach achieves strong multispectral segmentation with 8.7 million parameters, yielding a mean IoU of = 78.88% on the WeedsGalore dataset, substantially outperforming RGB-only baselines while maintaining real-time feasibility for UAV edge deployment. Compared with heavier baselines like DeepLabv3+ and MaskFormer, the model offers a favorable accuracy–efficiency trade-off suitable for site-specific weed management in precision agriculture. The work highlights the importance of adaptive modality weighting and multi-scale context, with future directions including domain adaptation, self-supervised pretraining, and hardware-aware optimization for field-scale deployment.

Abstract

Efficient crop-weed segmentation is critical for site-specific weed control in precision agriculture. Conventional CNN-based methods struggle to generalize and rely on RGB imagery, limiting performance under complex field conditions. To address these challenges, we propose a lightweight transformer-CNN hybrid. It processes RGB, Near-Infrared (NIR), and Red-Edge (RE) bands using specialized encoders and dynamic modality integration. Evaluated on the WeedsGalore dataset, the model achieves a segmentation accuracy (mean IoU) of 78.88%, outperforming RGB-only models by 15.8 percentage points. With only 8.7 million parameters, the model offers high accuracy, computational efficiency, and potential for real-time deployment on Unmanned Aerial Vehicles (UAVs) and edge devices, advancing precision weed management.
Paper Structure (11 sections, 5 figures, 1 table)

This paper contains 11 sections, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deep Learning Methods for Weed Segmentation
  4. Multispectral Fusion Strategies
  5. UAVs and Real-Time Constraints
  6. Multispectral and Multitemporal Data
  7. Proposed Methodology
  8. Model Architecture
  9. Training and Evaluation
  10. Comparison with Baselines
  11. Conclusion

Figures (5)

  • Figure 1: WeedsGalore dataset: acquisition and annotation workflow. Reprinted from celikkan2025weedsgalore.
  • Figure 2: WeedsGalore dataset: example tile with RGB, NIR, and Red-Edge channels, alongside the remapped 3-class segmentation mask.
  • Figure 3: Architecture of the lightweight transformer--CNN hybrid, showing modality-specific encoding, Transformer refinement, gated fusion, and pyramid-based decoding.
  • Figure 4: Segmentation by the MSI model: RGB input (left), ground truth (center), predicted mask (right). Note the accurate delineation of crop boundaries.
  • Figure 5: Segmentation by the RGB-only model: limited differentiation of dense weeds from crops.