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

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