Mapping of Urban Areas from SAR Images via Semantic Segmentation

This letter proposes a new semantic segmentation network called IGFUnet to map the urban areas from large scale SAR images. The network is based on U-net structure which is especially suitable for semantic segmentation of small data set. Inception module is adopted as the encoder to extract rich and dense features. We develop a global feature attention module (GFA) as a channel attention module to perform feature correction on each decoding layer. Combing the inception module and GFA module can take full use of the feature map and efficiently integrate features from different scales. Qualitative and quantitative experiments based on Sentinel-1A have proved the effectiveness of our IGFUnet model.