Astro-L3C: boosting lossless solar image compression with Kolmogorov–Arnold-guided learning

Abstract Rapid advances in astronomical observation technology are generating data volumes that increasingly strain existing storage and transmission infrastructure. Solar observation data poses a particular challenge due to its complex spatiotemporal correlations and multi-scale structures, requiring strict lossless compression to preserve scientific integrity. Conventional compression methods fail to capture the highly nonlinear intensity distributions arising from diverse solar phenomena, and varying observation conditions further increase the risk of overfitting to specific data patterns. To address these challenges, we present Astro-L3C, a framework that advances learned lossless compression (L3C) through a fast Kolmogorov–Arnold network, a dynamic Tanh-based ResBlock, and Fisher information regularization, which jointly enable expressive nonlinear probability modeling, stable feature extraction, and robust generalization across diverse solar phenomena and observing conditions. Experimental evaluation on new vacuum solar telescope data demonstrates that Astro-L3C reduces bits-per-subpixel (bpsp) by 4.6% compared with standard L3C. Benchmarked against established techniques, our method achieves bpsp reductions of 43.84% and 74.30% relative to super-Resolution based Compression and integer discrete flow respectively, confirming consistent improvements for solar observation data compression. This framework provides a new pathway for lossless compression of high-volume solar observation data.