Large-scale　crowd　flow　prediction　is　a　critical　task　in　urban　management　and　public　safety.　However,　achieving　accurate　and　efficient　prediction　remains　challenging.　Most　existing　models　overlook　spatial　heterogeneity,　employing　unified　parameters　to　fit　diverse　crowd　flow　patterns　across　different　spatial　units,　which　limits　their　accuracy.　Meanwhile,　the　massive　spatial　units　significantly　increase　the　computational　cost,　limiting　model　efficiency.　To　address　these　limitations,　we　propose　a　novel　model　for　large-scale　crowd　flow　prediction,　namely　the　Stratified　Compressive　Sensing　Network　(SCS-Net).　First,　we　develop　a　spatially　stratified　module　that　posterior　adaptively　extracts　the　underlying　spatially　stratified　structure,　effectively　modeling　spatial　heterogeneity.　Then,　we　develop　compressive　sensing　modules　to　compress　redundant　information　from　massive　spatial　units　and　learn　shared　crowd　flow　patterns,　enabling　efficient　prediction.　Finally,　we　conduct　experiments　on　a　large-scale　real-world　dataset.　The　results　demonstrate　that　SCS-Net　outperforms　deep　learning　baseline　models　by　35.25-139.2%　in　MAE　and　26.3-112.4%　in　RMSE　while　reducing　GFLOPs　by　53-1067　times　and　shortening　training　time　by　3.1-83.2　times　compared　to　prevalent　spatio-temporal　prediction　models.　Moreover,　the　spatially　stratified　structure　extracted　by　SCS-Net　offers　valuable　interpretability　for　spatial　heterogeneity　in　crowd　flow　patterns,　providing　deeper　insights　into　urban　functional　layouts.