A multi-scale remote sensing semantic segmentation model with boundary enhancement based on UNetFormer.

Area of Science:

• Geoscience
• Computer Vision
• Remote Sensing

Background:

• Accurate semantic segmentation of remote sensing data is crucial for geoscience applications.
• Challenges include blurred target edges and scale variability in high-resolution imagery, limiting segmentation accuracy.

Purpose of the Study:

• To propose a novel network, the Boundary-Enhanced Multi-Scale Semantic Segmentation Network (BEMS-UNetFormer), to address segmentation accuracy issues in remote sensing data.
• To improve the recognition of target edges and the integration of multi-scale features.

Main Methods:

• Developed an improved Boundary Awareness Module (BAM) for extracting edge information from low-level features.
• Integrated edge information into decoding using an improved Boundary-Guided Fusion Module (BFM).
• Incorporated a Multi-Scale Cascaded Atrous Spatial Pyramid Pooling (MSC-ASPP) module for multi-scale feature mining.

Main Results:

• Tested on Potsdam and Vaihingen datasets, achieving 86.12% and 83.10% Mean IoU (MIoU), respectively.
• Demonstrated significant improvements over the baseline model (1.38% and 1.79% increase in MIoU).
• Achieved high IoU and F1 Scores for small targets like 'Car' and specific categories ('Building', 'LowVeg') in challenging datasets.

Conclusions:

• The proposed BEMS-UNetFormer effectively enhances semantic segmentation accuracy for remote sensing data.
• The method shows superior performance in segmenting small-scale targets and precise boundaries compared to existing approaches.