Object-Level Semantic Segmentation on the High-Resolution Gaofen-3 FUSAR-Map Dataset

Abstract

Land cover classification with SAR images mainly focuses on the utilization of fully polarimetric SAR (PolSAR) images. The conventional task of PolSAR classification is single-pixel-based region-level classification using polarimetric target decomposition. In recent years, a large number of high-resolution SAR images have become available, most of which are single-polarization. This article explores the potential of object-level semantic segmentation of high-resolution single-pol SAR images, in particular tailored for the Gaofen-3 (GF-3) sensor. First, a well-annotated GF-3 segmentation dataset “FUSAR-Map” is presented for SAR semantic segmentation. It is based on four data sources: GF-3 single-pol SAR images, Google Earth optical remote sensing images, Google Earth digital maps, and building footprint vector data. It consists of 610 high-resolution GF-3 single-pol SAR images with the size of 1024 × 1024. Second, an encoder–decoder network based on transfer learning is employed to implement semantic segmentation of GF-3 SAR images. For the FUSAR-Map dataset, an optical image pretrained deep convolution neural network (DCNN) is fine-tuned with the SAR training dataset. Experiments on the FUSAR-Map dataset demonstrate the feasibility of object-level semantic segmentation with high-resolution GF-3 single-pol SAR images. Also, our algorithm obtains fourth place about the PolSAR image semantic segmentation on the “2020 Gaofen Challenge on Automated High-Resolution Earth Observation Image Interpretation.” The new dataset and the encoder–decoder network are intended as the benchmark data and baseline algorithm for further development of semantic segmentation with high-resolution SAR images. The FUSAR-Map and our algorithm are available at github.com/fudanxu/FUSAR-Map/.