Sensors　and　Internet　of　things　(IoT)　have　been　widely　used　in　the　digitalized　orchards.　Traditional　disease-pest　recognition　and　early　warning　systems,　which　are　based　on　knowledge　rule,　expose　many　defects,　discommodities,　and　it　is　difficult　to　meet　current　production　management　requirements　of　the　fresh　planting　environment.　On　purpose　to　realize　an　intelligent　unattended　alerting　for　disease-pest　of　fruit-melon,　this　paper　presents　the　convolutional　neural　network　(CNN)　for　recognition　of　fruit-melon　skin　lesion　image　which　is　real-timely　acquired　by　an　infrared　video　sensor,　which　network　is　grounded　upon　so-called　momentum　deep　learning　rule.　More　specifically,　(1)　a　suite　of　transformation　methods　of　apple　skin　lesion　image　is　devised　to　simulate　orientation　and　light　disturbance　which　always　occurs　in　orchards,　then　to　output　a　self-contained　set　of　almost　all　lesion　images　which　might　appear　in　various　dynamic　sensing　environment;　and　(2)　the　rule　of　variable　momentum　learning　is　formulated　to　update　the　free　parameters　of　CNN.　Experimental　results　demonstrate　that　the　proposed　presents　a　satisfying　accuracy　and　recall　rate　which　are　up　to　97.5　%,　98.5　%　respectively.　As　compared　with　some　shallow　learning　algorithms　and　generally　accepted　deep　learning　ones,　it　also　offers　a　gratifying　smoothness,　stableness　after　convergence　and　a　quick　converging　speed.　In　addition,　the　statistics　from　experiments　of　different　benchmark　data-sets　suggests　it　is　very　effective　to　recognize　image　pattern.