Backfill　of　cemented　waste　rock　into　underground　mined-out　areas　is　an　effective　way　to　eliminate　solid　wastes　and　potential　hazards　in　mines.　To　understand　the　backfill　strength　distribution　law　throughout　the　stope,　similarity　simulation　experiments　were　conducted　for　direct-irrigating　cemented　waste　rock　backfill,　and　OpenCV　and　neural　network　were　employed　to　analyze　particle　segregation　and　the　spatial　distribution　of　backfill　strength.　Results　show　that　distinct　gravitational　segregation　leads　to　an　uneven　and　heterogeneous　distribution　of　natural　graded　waste　rocks　in　a　similar　model.　Backfill　strength　near　sidewalls　and　bottom　of　the　model　surpasses　that　of　other　areas.　In　the　vertical　direction,　the　average　backfill　strength　increases　with　depth,　ranging　from　1.15　MPa　at　the　topmost　layer　to　1.91　MPa　at　the　bottommost　layer.　Horizontally,　the　average　backfill　strength　near　model　boundaries　is　consistently　higher　than　that　at　the　model　center,　irrespective　of　the　layer　depth　and　orientation.　Neural　network　prediction　on　spatial　backfill　strength　proves　reliable,　exhibiting　an　average　relative　error　of　4.12%,　compared　to　the　traditional　surface　fitting　with　a　10.20%　error.　Verification　tests　affirm　the　capability　of　the　neural　network　model　to　accurately　predict　the　anisotropic　and　nonlinear　distribution　of　backfill　strength　in　a　large　stope.