Built-Up Area Extraction from GF-3 SAR Data Based on a Dual-Attention Transformer Model

Abstract

Built-up area (BA) extraction using synthetic aperture radar (SAR) data has emerged as a potential method in urban research. Currently, typical deep-learning-based BA extractors show high false-alarm rates in the layover areas and subsurface bedrock, which ignore the surrounding information and cannot be directly applied to large-scale BA mapping. To solve the above problems, a novel transformer-based BA extraction framework for SAR images is proposed. Inspired by SegFormer, we designed a BA extractor with multi-level dual-attention transformer encoders. First, the hybrid dilated convolution (HDC) patch-embedding module keeps the surrounding information of the input patches. Second, the channel self-attention module is designed for dual-attention transformer encoders and global modeling. The multi-level structure is employed to produce the coarse-to-fine semantic feature map of BAs. About 1100 scenes of Gaofen-3 (GF-3) data and 200 scenes of Sentinel-1 data were used in the experiment. Compared to UNet, PSPNet, and SegFormer, our model achieved an 85.35% mean intersection over union (mIoU) and 94.75% mean average precision (mAP) on the test set. The proposed framework achieved the best results in both mountainous and plain terrains. The experiments using Sentinel-1 shows that the proposed method has a good generalization ability with different SAR data sources. Finally, the BA map of China for 2020 was obtained with an overall accuracy of about 86%, which shows high consistency with the global urban footprint. The above experiments proved the effectiveness and robustness of the proposed framework in large-scale BA mapping.