The　number　and　categories　of　the　existing　cloud　image　segmentation　datasets　are　limited,　and　the　segmentation　model　is　not　strong　targeted　enough　and　occupies　large　memory,　resulting　in　which　has　a　not　high　precision　and　efficiency.　Considering　these　problems,　a　dataset　GBCD-GT　with　large　amount　of　data　and　multiple　cloud　images　is　constructed　in　this　paper.　On　this　basis,　a　ground-based　cloud　image　segmentation　model　BFSegNet　based　on　bilateral　feature　fusion　is　proposed.　The　model　extracts　detail　features　and　semantic　features　respectively　through　detail　branch　and　semantic　branch,　and　then　fuses　the　two　features　together　through　feature　fusion　module,　finally　realizes　cloud　image　segmentation　through　up-sampling.　After　multiple　groups　of　comparative　experiments,　it　is　shown　that　the　model　BFSegNet　can　achieve　accurate　cloud　image　segmentation　under　the　premise　of　a　lower　number　of　parameters,　making　the　pixel　accuracy　up　to　94.39%　and　the　mean　intersection　over　union　up　to　73.26%.　Moreover,　the　prediction　time　of　single　image　of　the　model　is　only　1.216s,　which　improves　the　segmentation　efficiency　of　the　model.　It　lays　a　foundation　for　the　practical　application　of　the　model.　©　2022　IEEE.