ASVLB-Net: A lightweight network for multispectral weed segmentation with NDVI-guided adaptive fusion.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Remote Sensing

Background:

• Accurate crop-weed segmentation is crucial for precision spraying in agriculture.
• Visual similarity and occlusion between crops and weeds pose significant challenges.
• Existing methods often lack the required accuracy or efficiency for practical applications.

Purpose of the Study:

• To develop a lightweight yet highly accurate multispectral segmentation network for crop-weed identification.
• To enhance the discriminability of vegetation regions and capture fine-grained features.
• To improve segmentation performance in challenging scenarios like overlapping areas.

Main Methods:

• Proposed ASVLB-Net, a novel lightweight network utilizing normalized difference vegetation index (NDVI) priors.
• Introduced the Adaptive Spectral-Vegetation Fusion (ASVF) module for adaptive feature allocation.
• Employed a U-shaped architecture with a Layer-wise Concatenated Multi-Scale Feature (LCMF) encoder and a Bottleneck-SCSA module for attention.

Main Results:

• ASVLB-Net achieved a mean intersection over union (mIOU) of 86.5% and a mean precision of 91.89%.
• The network significantly outperformed several state-of-the-art (SOTA) models in crop-weed segmentation.
• ASVLB-Net demonstrated high efficiency with only 0.47 million parameters.

Conclusions:

• ASVLB-Net offers a robust solution for precision weeding applications using UAV multispectral imagery.
• The proposed network significantly improves segmentation accuracy and robustness.
• This advancement supports more effective precision agriculture strategies.