MSS-MambaNet: A Mamba Framework for Building Extraction from Multi-Phase Disaster Imagery

Area of Science:

• Remote Sensing
• Computer Vision
• Disaster Management

Background:

• Building extraction from disaster imagery is crucial for emergency response.
• Multi-phase disaster imagery presents challenges due to cross-phase heterogeneity in building characteristics.
• Existing methods struggle with stable semantic segmentation under complex damage conditions.

Purpose of the Study:

• To develop a robust method for building extraction from multi-phase disaster imagery.
• To address the challenges posed by phase-dependent variations and complex damage conditions.
• To improve the accuracy and stability of semantic segmentation in disaster scenarios.

Main Methods:

• Proposed MSS-MambaNet, a novel deep learning architecture for building extraction.
• Incorporated a multi-scale architecture to enhance perception of diverse building morphologies.
• Introduced Dual-Domain Cross-Gated Fusion (DDCGF) and Pixel-Aware Dynamic Weighting (PADW) strategies for improved feature discrimination and segmentation consistency.

Main Results:

• MSS-MambaNet achieved state-of-the-art performance in building extraction from multi-phase disaster imagery.
• The model obtained an average mIoU of 92.78% and mF1 of 96.25%.
• Demonstrated effectiveness with only 12.37 million parameters, indicating efficiency.

Conclusions:

• MSS-MambaNet effectively handles the heterogeneity of multi-phase disaster data.
• The proposed method provides a stable and efficient solution for building extraction in disaster scenarios.
• Results highlight the potential for improved emergency response and post-disaster assessment.