ELMZip Compresses Satellite Images Onboard for Efficient Downlink.

Small satellites, such as CubeSats, face significant challenges in downlinking the large volumes of multispectral imagery they acquire, primarily due to high data volumes and limited communication windows. Traditional image compression methods often struggle to adapt to the complex, nonlinear statistics of multi-band, multi-resolution satellite data. To address this bottleneck, researchers propose ELMZip, a novel framework for efficient, resolution-free onboard neural representation and compression. ELMZip is built upon Extreme Learning Machines (ELM) and domain decomposition strategies. A key innovation is its formulation of the fitting process as a convex least-squares problem, which eliminates the need for computationally intensive backpropagation, making it suitable for resource-constrained onboard processing. ELMZip employs an asymmetric transmission protocol, sending only compact output weights rather than full network parameters, which drastically reduces the downlink payload. Unlike previous neural representation approaches that rely on iterative optimization, ELMZip achieves substantial compression efficiency while maintaining high reconstruction fidelity. This capability allows for immediate image reconstruction and analysis, maximizing data return for resource-limited platforms and advancing real-time AI-powered Earth observation.