Tobacco　is　a　significant　revenue-generating　crop　in　China.　Early　mapping　of　tobacco　can　be　highly　valuable　in　estimating　yields　and　assessing　real-time　losses　due　to　disasters.　We　use　medium　resolution　Sentinel-1　SAR　time　series　remote　sensing　images　from　Ninghua　County　in　Fujian　Province,　China,　covering　the　entire　phenological　period　of　tobacco　from　February　to　July　in　2020,　as　input　data　for　the　model　Attention　Long　Short-Term　Memory　Fully　Convolutional　Network　(ALSTM-FCN).　To　determine　the　earliest　identifiable　timing　(EIT)　of　tobacco,　we　conduct　three　experiments　respectively　using　VH,　VV　and　VV+VH　data　and　gradually　increase　the　length　of　time　series　in　training　until　the　tobacco　was　harvested,　generating　corresponding　forty-two　models.　Results　show　that　the　overall　accuracy　(OA)　of　VV+VH　bipolarized　data　volume　is　stable　above　0.85　when　data　increases　to　April.　Using　bipolarized　data,　the　EIT　of　tobacco　can　be　determined　at　the　beginning　of　April,　during　mid-growing　period.　The　McNemar　test　results　shows　an　increasing　trend　with　the　increase　of　time-series　length.　Compared　using　single　tunnel　data,　the　bipolarized　data　performs　better　with　90.76%　of　OA.　Overall,　our　study　demonstrates　the　potential　of　ALSTM-FCN　for　early　identification　of　tobacco　growth　and　highlights　the　importance　of　using　bipolarized　data　for　such　applications.　　©　2023　IEEE.