Urban　land　use　information　can　be　effectively　extracted　from　high-resolution　satellite　images　for　many　urban　applications.　A　significant　challenge　remains　the　accurate　partition　of　fine-grained　land-use　units　from　these　images.　This　paper　presents　a　novel　method　for　deriving　these　units　based　on　unsupervised　graph　learning　techniques　using　high-resolution　satellite　images　and　open　street　boundaries.　Our　method　constructs　a　graph　to　represent　spatial　relations　between　land　cover　objects　as　graph　nodes　within　a　street　block.　These　nodes　are　characterized　by　spatial　composition　and　structure　features　of　their　surrounding　neighborhood.　We　then　apply　unsupervised　graph　learning　to　partition　the　graph　into　subgraphs,　which　represent　communities　spatially　bounded　by　street　boundaries　and　correspond　to　land　use　units.　Next,　a　graph　neural　network　is　used　to　extract　deep　structural　features　for　land　use　classification.　Experiments　were　conducted　using　high-resolution　satellite　images　from　the　cities　of　Fuzhou　and　Quanzhou,　China.　Results　showed　that　our　method　surpassed　traditional　grid　and　street　block　techniques,　improving　land　use　classification　accuracy　by　24%　and　9%,　respectively.　Furthermore,　it　achieved　classification　results　comparable　to　those　using　reference　land　use　units,　with　an　overall　accuracy　of　0.87　versus　0.89.