Short-imminent　heavy　rainfall　is　a　weather　phenomenon　with　large　rainfall　intensity　in　a　short　period　of　time,　and　its　effective　prediction　can　minimize　the　damage　caused　by　short-imminent　heavy　rainfall.　Fujian　province　is　a　region　with　short-imminent　heavy　rainfall.　The　accurate　identification　of　short-imminent　heavy　rainfall　has　been　the　key　research　goal　of　Fujian　provincial　meteorological　researchers.　In　this　paper,　based　on　ECMWF　numerical　forecast　products　and　ground　observed　rainfall　data,　RBF　interpolation　is　adopted　to　achieve　data　assimilation,　and　data　are　divided　according　to　the　ground　observed　rainfall　data.　Due　to　the　unbalanced　distribution　of　samples,　unlike　the　traditional　oversampling　method,　this　paper　uses　the　Synthetic　Minority　Oversampling　Technique　(SMOTE)　to　generate　short-imminent　heavy　rainfall　data,　combined　with　Focal　loss　and　DBNs　to　construct　the　shortimminent　heavy　rainfall　classification　model　FL-DBNs.　The　experimental　results　show　that　the　TS　value　of　FL-DBNs　is　increased　by　47.8%　in　the　cumulative　rainfall　of　1-30mm,　and　the　PO　and　ETS　are　greatly　improved.　Accumulated　rainfall　of　more　than　30mm　has　reached　a　certain　classification　requirements,　and　solved　the　problem　that　it　is　difficult　for　ECMWF　to　predict　rainfall　above　30mm,　which　has　certain　guiding　significance　for　short-imminent　heavy　rainfall　classification.　©　2019　IEEE.