Hierarchical vision mamba U-net for farmland semantic segmentation from remote sensing imagery
TL;DR
A hierarchical Vision Mamba U-Net (HVM-UNet) integrating three key components: a cross-scanning Visual State Space (CSVSS) block to improve scanning performance, a lightweight global feature fusion (GFF) module to replace traditional skip connections for enhanced detail preservation, and a multiscale spatial attention module (MSSA) for refined feature aggregation.
Abstract
Farmland semantic segmentation (FSS) from remote sensing imagery (RSIs) is a critical yet challenging task in precision agriculture. CNN-based methods suffer from limited receptive fields and poor long-range dependency modeling, while Transformers are limited by quadratic computational complexity. To address these issues, this paper proposes a hierarchical Vision Mamba U-Net (HVM-UNet) integrating three key components: a cross-scanning Visual State Space (CSVSS) block to improve scanning performance, a lightweight global feature fusion (GFF) module to replace traditional skip connections for enhanced detail preservation, and a multiscale spatial attention module (MSSA) for refined feature aggregation. Extensive experiments on benchmark datasets demonstrate that HVM-UNet achieves superior segmentation accuracy with linear complexity, outperforming both CNN-based and Transformer-based approaches and offering a robust solution for precision agriculture.