The　fine　particulate　matter　(PM　2.5)　problem　has　become　one　of　the　biggest　topics　of　concern　in　many　cities,　such　as　Beijing　in　China.　This　study　attempts　to　employ　a　self-organizing　fuzzy　neural　network　(SOFNN)　to　predict　the　hourly　concentration　of　PM　2.5　using　multi-source　data.　The　multi-source　data　including　meteorological　data,　pollutant　concentration　data　and　image　data　are　first　obtained　through　different　methods.　Then,　a　novel　SOFNN,　which　is　optimized　by　an　improved　gradient　descent　algorithm,　is　proposed　based　on　the　mutual　information　(MI)　and　sensitivity　analysis　(SA).　More　importantly,　MI　and　SA　are　introduced　to　effectively　adjust　the　structure　of　fuzzy　neural　network　without　changing　the　performance　of　the　original　network.　Finally,　the　MI-　and　SA-based　SOFNN　is　deployed　to　infer　PM　2.5　concentration　at　the　study　area.　The　experimental　results　show　that　our　proposed　model　can　achieve　a　more　compact　network　structure,　higher　prediction　performance　and　lower　computation　time　as　compared　with　state-of-the-art　or　popular　methods.　©　2020,　Springer　Nature　Switzerland　AG.