Metasurface‐Encoded Single‐Pixel Hyperspectral Imaging

Hyperspectral imaging in the visible spectrum offers significant potential for diverse applications, but is often constrained by bulky hardware and limited robustness in low‐light conditions. To overcome these challenges, a simulation‐based proof‐of‐concept for a metasurface‐encoded single‐pixel hyperspectral imaging system (MESH) is presented, in which structured spatial modulation is combined with a compact set of 50 broadband metasurface filters designed using a binary pattern generation strategy to ensure low interfilter correlation. Hyperspectral datacubes comprising 301 channels from 400 to 700 nm are reconstructed via a sparsity‐constrained optimization algorithm, while a physics‐enhanced deep learning model is further introduced to enable fast and accurate recovery. Simulation results demonstrate that MESH achieves a spectral resolution of 1.17 nm. Even at a total compression ratio of 2.1%, the deep learning model maintains high reconstruction quality, with a peak signal‐to‐noise ratio of 30.96 dB, structural similarity of 0.8526, and spectral angle mapping of 0.0742 rad, indicating accurate intensity recovery, structural preservation, and spectral integrity. The present study provides a simulation‐based verification of feasibility and design guidelines, laying the groundwork for future experimental validation of the MESH system, which is expected to further demonstrate its practical applicability and performance for deployment in low‐light and resource‐constrained environments.