Complex　single-phase　ground　fault　(SPGF)　is　a　challenging　problem　for　early　detection　and　type　recognition　in　resonant　distribution　networks.　This　paper　proposes　a　novel　semantic-segmentation-based　approach　that　leverages　the　morphological　information　of　zero-sequence　voltage　signals　to　extract　diverse　semantic　features　representing　fault　inception　(FI),　fault　disappearance　(FD),　and　short-term　transient　fault　(STF).　A　1D-UNet　model　is　employed　to　classify　each　sample　point　into　one　of　these　categories,　which　enables　the　determination　of　the　moment　and　duration　of　SPGF.　Based　on　these　features,　three　types　of　SPGF　are　recognized:　permanent　fault　(PF),　long-term　transient　fault(LTF),　and　short-term　transient　fault　(STF).　Due　to　its　low　power　consumption　and　costeffectiveness,　an　industrial　prototype　integrated　with　the　proposed　approach　has　been　developed　using　a　Raspberry　Pi　board.　The　proposed　approach　achieves　an　overall　accuracy　of　over　94　%　in　classifying　sample　points　across　diverse　categories.　Specifically,　the　individual　accuracies　for　detecting　sample　points　belonging　to　FI,　FD,　and　STF　were　0.978,　0.968,　and　0.971,　respectively.　From　an　engineering　application　perspective,　the　proposed　approach　effectively　identifies　the　moment　of　fault　occurrence,　whether　it　is　PF,　LTF,　or　STF.　The　maximum,　minimum,　and　median　triggering　deviations　were　10.8　ms,-6.4　ms,　and-0.4　ms,　respectively,　significantly　outperforming　existing　methods　in　terms　of　fault　moment　triggering　deviation.　The　experimental　results　demonstrate　that　the　proposed　approach　works　effectively　for　early　detection　and　type　recognition　of　SPGF,　showcasing　significant　potential　for　further　expansion　and　broader　application.