ABSTRACT Metasurface‐based hyperspectral imaging offers a compact solution for spectral sensing, but its performance in spatial multiplexing is often limited by position‐dependent spectral encoding and reconstruction instability. Here, we present an adaptive hyperspectral sensing framework that integrates a metasurface encoder with a position‐aware reconstruction strategy to address spatial misalignment and input–weight mismatch. Introducing a position calibration layer into the reconstruction network enables robust spectral recovery across super‐pixels using a single trained model. The system reconstructs hyperspectral data across the visible range from 400 to 780 nm with a wavelength sampling interval of 4 nm. Experimental results show consistently high accuracy, with fidelity values over 99% and RMSE below 0.05 across multiple spatial regions. Statistical evaluation of 300 samples confirms the robustness and spatial consistency of the framework under practical conditions. In addition to learning‐based reconstruction, we establish a physics‐guided baseline by restricting reconstruction to an effective spectral subspace and calibrating the transmission matrix with a limited number of independent spectra, enabling deterministic recovery via SVD inversion. These results establish a compact, physically grounded platform for metasurface‐enabled hyperspectral sensing, providing a foundation for future extensions toward advanced spectral and polarization‐resolved imaging.
Xia et al. (Tue,) studied this question.