In order to ensure smooth navigation of unmanned surface vessels ( USVs) for water missions , accurate extraction of river information is required , so a semantic segmentation network model of river is investigated. To ad- dress the problem of inter-class inconsistency and intra-class inconsistency in river image segmentation , a segmenta- tion network DBDL- Net is proposed in the paper , in which a double -branch decoding structure and a double loss function are designed to capture semantic and spatial information respectively; a lightweight module with multi-scale residuals is also designed in the coding part to reduce parameters on the one hand and capture feature information at different scales on the other. Finally , the model is ablated and compared with experiments on the USVInland dataset. The experimental results show that the accuracy and the mIoU of DBDL-Net finally reach 93. 619% and 87. 682% , and DBDL-Net also has better overall performance compared with other advanced segmentation networks.