Landform　recognition　is　of　great　significance　to　human　construction,　geological　structure　research,　environmental　governance　and　other　related　fields.　Traditional　recognition　methodology　is　mainly　based　on　pixel　unit　or　object-oriented　recgnition,　which　existed　limitations.　Landform　recognition　based　on　the　watershed　unit　has　become　a　new　hotspot　in　this　field　because　of　its　surface　morphology　integrity　and　clear　geographical　significance.　However,　the　traditional　methods　of　landform　recognition　based　on　terrain　factors　are　often　simple　or　repeatable　in　the　geological　description,　which　cannot　be　used　to　describe　the　spatial　structure　and　quantify　the　topological　relationship　characteristics　of　the　watershed　unit.　The　slope　spectrum　method　was　used　to　solve　the　problem　that　it　was　difficult　to　determine　the　stable　area　of　watershed　unit,　and　181　small　watersheds　were　extracted　through　hydrological　analysis.　Based　on　the　theory　of　complex　network　and　geomorphology,　the　concept　of　watershed　weighted　complex　network　and　8　quantitative　indexes　were　put　forward　to　simulate　and　quantify　the　spatial　structure　of　the　watershed.　Finally,　XGBoost　machine　learning　algorithm　is　adopted　for　landform　recognition.　XGBoost　machine　learning　algorithm　based　on　decision　tree　is　used　for　landform　recognition.　The　experiment　shows　a　well　performance　on　the　landform　recognition　of　the　main　landform　types　on　the　Loess　Plateau,　with　the　Kappa　coefficient　of　86.00%　and　the　overall　accuracy　of　88.33%.　Compared　with　the　landforms　having　obvious　morphological　features,　the　complex　network　method　considers　the　characteristics　of　spatial　structure　and　topological　features,　resulting　in　higher　recognition　accuracy　and　kappa　coefficient　of　90%~100%.　Compared　with　previous　studies,　the　recognition　results　show　high　accuracy,　which　verifies　that　the　method　based　on　watershed　weighted　complex　network　is　an　effective　method　with　high　accuracy　for　landform　recognition　based　on　watershed.　©2022,　Science　Press.　All　right　reserved.