Robust Information Bottleneck for Satellite Edge Inference Over MIMO Channel

The exponential growth in data generation from low-Earth-orbit (LEO) satellites has created a critical downlink bottleneck, rendering traditional download-then-process paradigms infeasible for time-sensitive applications. To address this, we propose the source-channel robust information bottleneck (SC-RIB) framework for task-oriented satellite communication over fading multiple-input multiple-output (MIMO) channels. Unlike conventional joint source-channel coding schemes that prioritize reconstruction fidelity, SC-RIB leverages the information bottleneck principle to extract minimal yet sufficient semantic features for downstream inference. Crucially, we introduce a dual-robustness mechanism: (1) Source-Robustness, enforced via Fisher information regularization, ensures encoder insensitivity to input perturbations (e.g., cloud residues); and (2) Channel-Robustness, achieved through a receiver-side invariance penalty, encourages feature stability against stochastic channel fading without requiring instantaneous channel state information. Extensive experiments on the LEVIR-CD and RaVÆn datasets demonstrate that SC-RIB outperforms the baselines, exhibiting superior generalization across varying channel signal-to-noise ratio and Rician κ -factors.