PanSAM-HAT: A Hybrid Attention Transformer With Panchromatic Spectral Attention Module for Compression Artifacts Removal in Satellite Imagery

Hyun-Ho Kim, Jaemyung Kim, Doo-Chun Seo, Jae-Heon Jeong, Yongwoo Kim

Abstract: Multispectral (MS) satellite images are often subject to spatial–spectral variable compression (SSVC) because of constraints in transmission bandwidth and sensor scanning capabilities. This compression process, which varies across spatial locations and spectral bands, can introduce severe distortions and block artifacts, particularly under unknown compression ratios. To address these challenges, we propose a hybrid attention transformer (PanSAM-HAT) that incorporates a novel panchromatic spectral attention module (PanSAM) for effective artifact removal in SSVC-compressed MS imagery. PanSAM adaptively integrates spectral information from MS and spatial structure from high-resolution panchromatic (PAN) images through a pixel-wise attention mechanism, enabling robust restoration of spatial and spectral fidelity. Additionally, we present JPEG LAM, a novel attribution mapping technique designed to analyze and interpret the performance of compression artifact removal (CAR) networks. JPEG LAM introduces a compression-aware interpolation path and gradient-based block boundary analysis to highlight the network’s attention to JPEG artifacts. Extensive experiments on both synthetic and real-world remote sensing datasets demonstrate that our proposed PanSAM-HAT consistently outperforms existing convolutional neural network-based and transformer-based methods in reducing compression artifacts and enhancing image quality in SSVC scenarios.

External IDs:doi:10.1109/access.2025.3602150