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

Accurate crop-weed segmentation is critical for precision spraying but is hindered by visual similarity and occlusion. This study aims to develop a lightweight yet high-accuracy multispectral segmentation network to address these challenges.

We propose ASVLB-Net, a novel lightweight multiscale feature extraction network guided by normalized difference vegetation index (NDVI) priors for crop-weed segmentation in multispectral imagery. First, we design the Adaptive Spectral-Vegetation Fusion (ASVF) module at the input stage. It achieves adaptive allocation of spectral and spatial features, enhancing the discriminability of vegetation regions. Additionally, we adopt a U-shaped architecture based on the Layer-wise Concatenated Multi-Scale Feature (LCMF) encoder. It efficiently extracts fine-grained features and boundary information with low computational cost. Finally, we use the Bottleneck -SCSA (Spatial and Channel Synergistic Attention) module to capture long-range dependencies, improving segmentation in overlapping areas. Experimental results demonstrate that ASVLB-Net achieves a mean intersection over union (mIOU) of 86.5% and a mean precision of 91.89%, outperforming several state-of-the-art (SOTA) models, while maintaining high efficiency with only 0.47 million parameters.

The ASVLB-Net significantly improves segmentation accuracy and robustness in unmanned aerial vehicle (UAV) multispectral imagery, offering a robust solution for precision weeding applications. © 2026 Society of Chemical Industry.