Predicting　urban　crowd　flow　spatial　distributions　plays　a　critical　role　in　optimizing　urban　public　safety　and　traffic　congestion　management.　The　spatial　dependency　between　regions　and　the　temporal　dynamics　of　the　local　crowd　flow　are　two　important　features　in　urban　crowd　flow　prediction.　However,　few　studies　considered　geographic　characteristic　in　terms　of　spatial　features.　To　fill　this　gap,　we　propose　an　urban　crowd　flow　prediction　model　integrating　geographic　characteristics　(FPM-geo).　First,　three　geographic　characteristics,　proximity,　functional　similarity,　and　road　network　connectivity,　are　fused　by　a　residual　multigraph　convolution　network　to　model　the　spatial　dependency　relationship.　Then,　a　long　short-term　memory　network　is　applied　as　a　framework　to　integrate　both　the　temporal　dynamic　patterns　of　local　crowd　flow　and　the　spatial　dependency　between　regions.　A　4-day　mobile　phone　dataset　validates　the　effectiveness　of　the　proposed　method　by　comparing　it　with　several　widely　used　approaches.　The　result　shows　that　the　root　mean　square　error　decreases　by　15.37%　compared　with　those　of　the　typical　models　with　the　prediction　interval　at　the　15-min　level.　The　prediction　error　increases　with　the　crowd　flow　size　in　a　local　area.　Moreover,　the　error　reaches　the　top　of　the　morning　peak　and　the　evening　peak　and　slopes　down　to　the　bottom　at　night.