Spatial-channel dual autoregressive model for remote sensing image compression

Abstract

Remote sensing images are typically large in size and contain abundant ground object information as well as complex spatial texture structures. These characteristics result in high storage and transmission costs, which necessitates highquality image compression methods. When compressing remote sensing images, compression models must preserve critical object boundaries and texture details under limited bandwidth while reducing the bit rate as much as possible, which poses significant challenges to the contextual modeling capability of entropy models. Although deep learningbased image compression methods have developed rapidly in recent years, they still do not fully exploit spatial-domain information. To address this issue, this paper proposes a spatial-channel dual autoregressive context modeling strategy for remote sensing images, termed Hierarchy Spatial Image Compression (HSIC). By leveraging the advantages of the window-based attention mechanism in Swin Transformer, the proposed method maintains channel-wise autoregressive modeling while progressively utilizing contextual information in the spatial domain through hierarchical modeling. This enables the model to capture potential spatial redundancy from coarse to fine granularity, thus significantly reducing bit rates without sacrificing reconstruction quality. Experimental results demonstrate that, compared with traditional image coding methods and existing learning-based image compression approaches, HSIC achieves superior rate-distortion performance across multiple remote sensing image datasets. Compared with the VTM encoder, HSIC reduces BD-rate by approximately 13.90% and 13.5% on the AID and NWPU VHR-10 datasets, respectively, and by 13.2% and 14.56% on the WorldView-3 Multispectral and Panchromatic datasets, respectively.