In　the　field　of　remote　sensing　imagery,　road　extraction　is　one　of　the　key　technologies　supporting　for　Landuse　Landcover　classification.　In　this　paper,　a　new　semantic　segmentation　neural　network　named　SAT　U-Net　is　proposed　for　road　extraction　from　remote　sensing　imagery.　The　new　improved　network　replaces　the　sigmoid　layer　in　the　U-Net　with　a　self-adaptive　threshold　method　proposed　to　self-adaptively　adjust　the　road　thresholds　for　segmentation　results　of　U-Net.　The　proposed　method　is　combined　with　the　strength　of　U-Net　architecture　to　retain　the　complete　road　spatial　features,　thus　overcomes　the　problem　of　unconnected　and　blurry　roads　in　the　segmentation　results.　To　prove　the　effectiveness　and　utility　of　the　proposed　network,　it　was　experimented　on　the　test　set　of　a　public　road　dataset　and　compared　with　U-Net　in　five　different　environments.　Experimental　results　demonstrate　that　the　proposed　method　is　superior　to　U-Net　and　presents　clearer　and　more　complete　road　structures.　©　2019　IEEE.