Objective　Traditional　spectral　imaging　systems,　which　rely　on　spatial　or　temporal　scanning,　face　significant　limitations　in　practical　applications,　including　narrow　spectral　coverage,　low　photon　throughput　due　to　sequential　acquisition,　and　the　unresolved　tradeoff　between　spatial　resolution　and　spectral　fidelity.　To　address　these　challenges,　we　propose　a　metasurface-enabled　snapshot　compressive　spectral　imaging　system,　which　fundamentally　redefines　the　system　architecture　through　two　key　innovations:　1)　a　spectrally-encoded　metasurface　array　that　performs　parallel　light　field　manipulation　across　28　spectral　channels　within　the　visible　range　(450-650　nm);　2)　a　physics-informed　window　channel　attention-deep　unfolding　network　(WCA-DUN)　that　synergizes　computational　optics　with　deep　learning　for　real-time　hyperspectral　cube　reconstruction.　Compared　to　existing　snapshot　spectral　imaging　systems　based　on　coded　apertures　or　dispersive　optical　elements,　our　approach　leverages　the　unique　capability　of　metasurfaces　to　engineer　spectral-spatial　responses　at　subwavelength　scales,　enabling　a　compact,　lightweight,　and　integrated　spectral　imaging　system.　Methods　In　this　paper,　we　design　a　snapshot　compressive　spectral　imaging　system　based　on　a　metasurface　array.　By　leveraging　this　metasurface　array　for　efficient　spectral　encoding　and　integrating　the　proposed　WCA-DUN　algorithm,　real-time　hyperspectral　image　reconstruction　with　high　spectral　resolution　can　be　achieved.　For　the　metasurface　design,　a　randomly　generated　binary　pattern　is　used　to　construct　a　meta-atom　library.　This　approach　enriches　the　meta-atom　library　while　ensuring　the　minimum　feature　size,　making　fabrication　easier.　Meta-atoms　with　low　correlation　are　selected　using　the　Pearson　correlation　coefficient　as　criteria.　For　the　reconstruction　algorithm,　we　propose　the　WCA-DUN　algorithm,　which　integrates　window　segmentation　into　channel　attention,　achieving　a　larger　receptive　field　to　capture　more　global　information.　In　addition,　it　combines　amplitude　and　phase　feature　learning　of　the　optical　field　to　suppress　artifacts　and　minimize　cross-talk　between　spectral　channels.　Results　and　Discussions　Our　system　demonstrates　excellent　robustness　to　noise.　The　reconstruction　results　of　our　proposed　system　decrease　by　only　0.61　dB　in　PSNR　under　Gaussian　noise,　while　the　results　of　other　existing　systems　(GAP-TV,　ADMM-net,　and　PnP)　decrease　by　4.58,　0.56,　and　9.1　dB,　respectively.　Moreover,　our　system　achieves　a　significantly　faster　reconstruction　speed.　Compared　to　systems　using　traditional　iterative　algorithms,　our　proposed　system　improves　the　reconstruction　speed　by　three　orders　of　magnitude.　In　addition,　our　algorithm　ensures　both　high　reconstruction　speed　and　spectral　reconstruction　accuracy,　outperforming　depth-unfolding-network-based　algorithms.　Specifically,　a　reconstruction　rate　of　30　Hz　can　be　achieved　with　a　spectral　resolution　of　1　nm.　Conclusions　In　summary,　we　propose　a　novel　snapshot-based　compressive　spectral　imaging　system　based　on　a　metasurface　array,　enabling　real-time　hyperspectral　image　reconstruction　with　high　spectral　resolution.　By　leveraging　the　unique　capabilities　of　metasurfaces　to　manipulate　light　at　subwavelength　scales,　the　designed　metasurface　array　facilitates　the　development　of　a　compact,　lightweight,　and　high-performance　spectral　camera　with　superior　spatiotemporal　resolution.　To　achieve　real-time　and　high-fidelity　hyperspectral　reconstruction,　we　introduce　WCA-DUN,　a　deep　unfolding　network　framework　that　synergistically　integrates　computational　optics　with　deep　learning.　This　advanced　approach　not　only　enhances　reconstruction　accuracy　but　also　significantly　improves　processing　efficiency,　making　real-time　applications　feasible.　The　proposed　system　provides　a　groundbreaking　solution　to　the　long-standing　tradeoff　between　spatial,　temporal,　and　spectral　resolutions　in　spectral　imaging.　By　overcoming　the　fundamental　limitations　of　traditional　systems,　it　enables　broader　spectral　coverage,　higher　photon　throughput,　and　superior　resolution　without　the　need　for　bulky　optical　components　or　mechanical　scanning.　While　our　current　implementation　focuses　on　the　visible　range,　the　system　architecture　and　design　methodology　can　be　easily　extended　to　other　spectral　regions,　including　the　near-infrared　and　terahertz　ranges.　This　adaptability　makes　it　a　promising　candidate　for　a　wide　range　of　applications,　such　as　aerospace　exploration,　remote　sensing,　biomedical　imaging,　and　machine　vision,　where　compact　and　high-resolution　spectral　imaging　is　crucial.　We　believe　that　this　innovation　will　pave　the　way　for　the　next　generation　of　spectral　imaging　technologies,　driving　advancements　across　multiple　scientific　and　industrial　fields.　©　2025　Chinese　Optical　Society.　All　rights　reserved.