MSTCA-Net: A Novel Impervious Surface Extraction Method Based on a Multistage Transformer

Abstract

The area of impervious surfaces serves as a critical metric to gauge urbanization levels and evaluates ecological health in a given region. However, in some areas with low-density impervious surfaces, these impervious surfaces are often small in size and distributed scatteredly, making them easily mistaken for the background. Therefore, accurately extracting these small and scattered impervious surfaces remains a significant challenge. In our previous work, TCA-Net pioneered the application of transformers for extracting impervious surfaces. However, owing to the limitations of transformers, TCA-Net struggled to precisely identify small, scattered impervious surfaces. In many cases, some minor roads extracted are discontinuous and incomplete. To enhance the transformer's performance in extracting impervious surfaces, we introduce MSTCA-Net, a dual-branch network. MSTCA-Net incorporates a multiscale transformer that preserves local details while capturing global context, and a Unet branch augmented with a coordinate attention mechanism, which captures intricate details while minimizing information redundancy. Experimental results show that MSTCA-Net proposed in this work can greatly outperform the traditional CNN models and our previous work, especially in extracting small and scattered impervious surfaces.

Keywords

• Coordinate attention
• impervious surface extraction
• semantic segmentation
• transformer