A Multi-Channel Attention Network for SAR Interferograms Filtering Applied to TomoSAR

Abstract

Tomographic synthetic aperture radar (TomoSAR) is an advanced synthetic aperture radar (SAR) interferometric technique that can retrieve 3-D spatial information. However, the performances of 3-D reconstruction could be degraded due to the noise in interferograms, which makes the filtering crucial before the tomographic reconstruction. As known, filters for single-channel interferograms are common, but those for multi-channel interferograms are still rare. In this paper, we propose a multi-channel attention network to denoise the multi-channel interferograms applied for TomoSAR, which is built on the basis of multi-channel attention blocks. An important feature of the block is the local context mixing before the computation of attention maps across channels, which explores the intra-channel local information and the inter-channel relationship of the multi-channel interferograms. Based on this architecture, the proposed method can effectively filter the noise while preserving the structures in interferograms, thus improving the performance of tomographic reconstruction. The network is trained by simulated data and the promising results of both simulated and real data validate the effectiveness of our proposed method.