Forest　is　one　of　the　most　important　natural　resources.　Therefore,　people　are　eager　to　acquire　the　information　of　forest　resources　quickly,　accurately,　and　efficiently,　and　monitor　the　changes　in　time,　to　realize　the　scientific　management　and　effective　utilization　of　forest　resources.　The　combination　of　remote　sensing　technology　and　image　segmentation　technology　can　effectively　solve　this　problem.　This　paper　proposed　a　convolutional　neural　network　RC-Unet,　which　is　based　on　Unet　architecture.　Based　on　the　region　correlation　of　forest,　we　firstly　put　forward　an　innovative　theory,　which　use　8　pieces　of　surrounding　images　in　Sudoku　style　to　predict　the　center　image.　More　information　can　be　provided　for　forest　detection　in　the　middle　area　by　simultaneously　input　this　image　and　8　surrounding　images.　Our　experiments　were　performed　on　the　Ziyuan-3　image　data,　in　northeast　of　Qinghai　province,　which　is　belongs　to　the　Sanjiangyuan　region　of　China　and　is　an　important　area　for　the　concentrated　protection　of　natural　resources　in　China.　The　remote　sensing　database　of　plateau　forest　was　established　creatively.　Comparing　to　decision　tree　and　Unet,　RC-Unet　showed　the　best　results　in　evaluation　metrics　included　accuracy,　recall　and　F1-score.　Our　finding　has　taken　the　advantage　of　specificity　of　forest,　which　promote　the　detection　of　forest　areas　in　remote　sensing.　©　2020　ACM.