Loess　landform　attaches　importance　to　surface　processes　and　soil　erosion.　Taking　watersheds　as　basic　landform　units,　gullies　(negative　terrain　skeletons)　were　generally　quantified　for　loess　landform　recognition.　However,　ridges　(positive　terrain　skeletons)　were　rarely　considered,　neglecting　their　discriminations　in　telling　different　loess　landform　types　apart.　Considering　ridges　and　gullies　in　combination,　this　study　proposed　Watershed　Dual　Skeleton　Networks　(WDSN)　for　loess　landform　recognition.　Specifically,　hydrologic　analysis　and　network　theory　were　applied　to　extract　WDSN.　Then,　10　network　indices　were　used　to　quantify　ridges　and　gullies　respectively.　With　the　Light　Gradient　Boosting　Machine　(LightGBM),　recognition　results　showed　that　the　WDSN-based　approach　had　a　superior　performance,　achieving　an　overall　accuracy　of　93.33%　and　a　Kappa　coefficient　of　0.90.　Compared　to　single　ridge　or　gully　networks　and　traditional　terrain　indices,　WDSN　can　express　topographic　differences　more　comprehensively　among　loess　tableland,　loess　ridge　and　loess　hill.　©　2024　Copyright　held　by　the　owner/author(s).