By　analyzing　a　large　number　of　radar　images　collected　by　Radar　Depth　Sounder　instruments,　a　detailed　study　of　under-ice　layer　structures　can　be　obtained.　However,　how　to　extract　the　ice　layer　structures　from　the　ice　sheet　radar　topology　sequences　quickly　and　accurately　is　still　a　challenge.　This　paper　proposes　a　method,　INNet,　which　is　based　on　the　improved　I3D　network　and　combined　with　the　attention　module,　to　automatically　extract　the　ice　layer　structures　at　the　pixel　level.　The　inception　units　and　non-local　attention　modules　are　used　in　the　task　of　automatically　extracting　the　under-ice　structures　for　the　first　time.　And　the　experimental　results　on　the　dataset　of　Center　of　Remote　Sensing　of　Ice　Sheets　have　shown　the　effectiveness　of　INNet,　whose　results　show　a　12.9%　reduction　on　the　measurement　of　average　mean　absolute　column　error　compared　with　the　state-of-the-art　in　deep　learning.　©2021　IEEE.