Spectral and Spatial Graph Learning for Multispectral Solar Image Compression

TL;DR

This work tackles the challenge of compressing multispectral solar imagery with high spectral fidelity under bandwidth limits. It introduces iSWGE to explicitly model inter-spectral relationships and WSGA-C to sparsify spatial attention while leveraging CBAM, integrated into a transform-based learned compression framework with a channel-wise entropy model. The combined approach yields notable gains: up to $20.15\%$ reduction in MSID, and PSNR and MS-SSIM improvements at various bitrates, outperforming strong baselines on the SDOML dataset across six EUV channels. The method demonstrates sharper, spectrally faithful reconstructions at comparable bitrate, offering a practical path for efficient data transmission in solar missions and similar scientific imaging contexts.

Abstract

High-fidelity compression of multispectral solar imagery remains challenging for space missions, where limited bandwidth must be balanced against preserving fine spectral and spatial details. We present a learned image compression framework tailored to solar observations, leveraging two complementary modules: (1) the Inter-Spectral Windowed Graph Embedding (iSWGE), which explicitly models inter-band relationships by representing spectral channels as graph nodes with learned edge features; and (2) the Windowed Spatial Graph Attention and Convolutional Block Attention (WSGA-C), which combines sparse graph attention with convolutional attention to reduce spatial redundancy and emphasize fine-scale structures. Evaluations on the SDOML dataset across six extreme ultraviolet (EUV) channels show that our approach achieves a 20.15%reduction in Mean Spectral Information Divergence (MSID), up to 1.09% PSNR improvement, and a 1.62% log transformed MS-SSIM gain over strong learned baselines, delivering sharper and spectrally faithful reconstructions at comparable bits-per-pixel rates. The code is publicly available at https://github.com/agyat4/sgraph .

Figures (6)

• Figure 1: Multispectral solar images are partitioned into spatial windows. Graphs are constructed both within each window and across spectral channels to jointly capture spatial and spectral dependencies.
• Figure 2: Overview of the proposed multispectral compression framework. The encoder $g_a$ fuses spectral features from iSWGE with spatial features via concatenation (C), followed by WSGA-C blocks to form compact latent representations. The latent $y$ is quantized (Q) and entropy-coded using a channel-wise autoregressive model. The decoder $g_s$ reconstructs spectral and spatial features through parallel paths: convolutional layers recover spatial layout, while iSWGE refines spectral embeddings. The two streams are averaged to produce the final output $\hat{x}$, ensuring fidelity in both spectral and spatial domains.
• Figure 3: Detail diagram of (a) the iSWGE module and (b) the WSGA-C module
• Figure 4: Spectral graph construction in iSWGE: grouped convolutions extract per-band features, pooled nodes form graph vertices, and edges encode differences for spectral graph processing.
• Figure 5: Rate-distortion comparison of the proposed model against baselines: (a) PSNR vs. BPP and (b) log-transformed MS-SSIM vs. BPP. The (iSWGE + WSGA-C) model consistently outperforms BL1 and BL2 across bitrates, delivering higher spatial and perceptual quality.