Spatial　information　of　agricultural　fields　is　essential　to　many　agricultural　applications,　particularly　for　precision　agriculture.　This　paper　investigates　the　use　of　multi-task　learning　methods　to　delineate　agricultural　fields　from　very　high-resolution　satellite　images.　In　this　study,　we　introduce　a　new　boundary-aware　multi-task　neural　network,　i.e.,　BDA-Net,　and　combined　with　an　edge　detection　model　Richer　Convolutional　Features　(RCF)　for　agricultural　fields　delineation.　More　specifically,　BDA-Net　learns　three　tasks,　a　core　task　for　mask　prediction,　two　auxiliary　tasks　are　boundary　prediction　and　distance　maps　estimation.　The　auxiliary　tasks　are　contributing　to　capturing　the　boundary　and　shape　of　the　fields　to　produce　a　smooth　mask.　We　conducted　experiments　on　two　different　areas　in　the　Yongxing　town　and　Shuibei　Jie　town　of　Fujian　province,　China.　We　acquired　two　GF-2　PMS　satellite　images　with　0.8m　spatial　resolution　for　the　Shuibei　Jie　and　Yongxing　areas.　We　compared　BDA-Net　with　the　existing　UNet　and　evaluated　the　obtained　results　in　terms　of　object-based　evaluation　measures.　Our　results　show　that　BDA-Net　has　the　lowest　global　total　error　(0.302)　compared　with　UNet　for　the　Yongxing　datasets　and　it　can　also　apply　to　the　Shuibei　Jie　area.　Moreover,　BDA-Net　obtained　closed　fields　boundary　in　one　segmentation　and　thus　avoiding　post-processing.　©　2022　IEEE.