PM2.5,　which　severely　affects　human　health,　is　one　of　the　most　important　indices　for　air　quality　estimation.　There　have　been　limited　studies　on　the　simple,　fast　and　cheap　PM2.5　concentration　prediction　and　thus　this　paper　presents　a　model　of　PM2.5　prediction　based　on　image　contrast-sensitive　features.　Two　types　of　features　were　extracted　from　the　images　and　utilized　to　estimate　the　PM2.5　concentration.　Then,　we　establish　a　recurrent　fuzzy　neural　network　model,　the　parameters　of　which　are　trained　by　using　the　gradient　descent　algorithm　with　an　adaptive　learning　rate.　Experiment　results　indicate　that　the　recurrent　neural　network　has　better　prediction　performance　than　traditional　radial　basis　function　and　fuzzy　neural　network.　©　2018　Technical　Committee　on　Control　Theory,　Chinese　Association　of　Automation.