The　conventional　method　of　PM2.5　prediction　requires　high-precision　instruments　to　obtain　data　on　the　concentration　of　pollutants,　resulting　in　a　high　prediction　costs.　In　this　work,　we　attempt　to　use　image　data　to　estimate　PM2.5　concentration.　The　concentration　of　atmospheric　PM2.5　is　closely　linked　to　the　image＇s　dark　channel　intensity,　contrast,　and　color　difference　of　HSI.　The　increase　in　atmospheric　PM2.5　concentration　leads　to　a　decrease　in　the　non-sky　area　dark　channel　intensity,　image　contrast,　and　HSI　spatial　color　difference.　In　this　paper,　a　Column-Generation　PM2.5　prediction　model　based　on　image　mixture　kernel　was　proposed　by　analyzing　the　relationship　between　PM2.5　and　image　features.　First,　the　sampling　period　was　taken　as　1　h,　and　8:00-17:00　was　taken　as　the　sampling　range　daily.　The　scene　images　were　recorded　in　different　weather　conditions,　and　five　image　features　were　extracted,　including　contrast,　dark　channel　intensity,　and　HSI　color　difference.　Secondly,　the　image　data　has　the　characteristics　of　large　sample　size　and　uneven　distribution,　and　the　prediction　model　consists　of　a　single　kernel　function,　which　makes　it　difficult　to　meet　the　prediction　accuracy　requirement.　Therefore,　the　linear　kernel　function,　polynomial　kernel　function,　and　Gauss　kernel　function　were　chosen　to　construct　a　composite　model　according　to　the　concept　of　kernel　structure　from　simple　to　complex.　Then　each　kernel＇s　Gram　matrix　was　calculated　based　on　training　samples,　and　all　gram　matrices　were　placed　into　a　mixture　kernel　matrix.　Using　the　column　generation　algorithm　and　mixture　kernel　matrix,　the　prediction　model　was　developed　and　the　parameters　of　the　model　were　solved.　Finally,　simulation　experiments　were　performed;　the　results　show　that　the　prediction　model　based　on　the　image　mixture　kernel　of　Column-Generation　PM2.5　can　meet　the　prediction　accuracy　requirements.　The　model　has　higher　prediction　accuracy　and　better　model　stability　in　comparison　with　the　single-kernel　prediction　model.　A　computational　complexity　analysis　shows　that　the　prediction　model　based　on　the　image　mixture　kernel　of　column-generation　PM2.5　has　no　significant　increase　in　computational　complexity　in　comparison　with　the　one-kernel　prediction　model.　©　2020,　Science　Press.　All　right　reserved.