Studies　of　adverse　human　health　effects　due　to　exposures　to　particles　with　an　aerodynamic　diameter　smaller　than　2.5　μm　(PM2.5)are　often　limited　by　sparse　ground　in　situ　measurements.　Satellite　remote　sensing　technique　provides　an　effective　tool　for　PM2.5　assessment　in　areas　where　surface　PM2.5　monitoring　network　is　not　available.　The　MODIS　aerosol　optical　depth　(AOD)with　spatial　resolution　of　10　km　has　been　widely　used　in　retrieving　PM2.5　concentrations　in　a　large　scale,　but　it　is　insufficient　for　city-scale　PM2.5　studies.　In　this　study,　we　used　the　newly　released　AOD　product　with　higher　resolution　of　3　km　incorporating　meteorological　fields　from　Goddard　Earth　Observing　System-Forward　Processing　(GEOS-FP)and　road　density　for　ground-level　PM2.5　estimation　in　Fuzhou,　i.e.　a　coastal　city　of　China.　A　two-stage　statistical　model　combing　linear　mixed　effects　model　(LME)and　support　vector　regression　model　(SVR)was　proposed　in　this　study,　and　a　10-fold　cross　validation　approach　was　employed　for　model　validation.　We　obtained　an　overall　R2　of　0.81,　root　mean　square　error　(RMSE)of　8.83　μg/m3　in　model　fitting,　and　R2　of　0.77,　RMSE　of　9.51　μg/m3　in　model　validation.　The　retrieved　PM2.5　presented　a　spatial　pattern　with　high　concentrations　in　urban　areas　and　low　values　in　suburban　or　mountainous　areas.　We　also　found　that　the　spatial　distribution　of　PM2.5　concentrations　showed　a　very　strong　correlation　with　the　terrain　features,　forest　cover,　road　density,　and　industrial　pollution　sources.　The　results　revealed　that　the　combined　LME-SVR　model　using　3　km　AOD　along　with　GEOS-FP　field　and　road　density　could　achieve　high　accuracy　in　PM2.5　estimation,　and　would　be　helpful　for　air　quality　monitoring　in　Fuzhou.　©　2019　Elsevier　Ltd