A speedy and reliable system detecting safe waters is essential to enable autonomous navigation of unmanned surface vessels (USVs). However, existing detection methods are prone to misidentifying trees and buildings on river banks as water and blurring water boundaries due to reflections from the water surface. This study proposes a dual-task network architecture for boundary-aware and waters segmentation(BAWS network) to address the above issues. The BAWS network consists of an encoder module, a decoder module and a boundary-aware module. The attention refinement module (ARM) focuses on water information to improve water detection. A feature fusion module (FFM) is introduced better to integrate visual features with boundary features in the decoder. In addition, boundary-aware modules and boundary loss functions are proposed to force the network to focus on detailed information about the boundary of the waters. To validate the effectiveness of the algorithm, this study was tested using three different types of typical inland river datasets. The results show that the method achieves a maximum of 97.19% mean intersection ratio (MIoU) and 0.776 root mean square error (RMSE) of the boundary on the USVInland dataset. This indicates that the method can produce clearer predictions at the watershed boundary and improve the performance of watershed segmentation with good generalisation performance.