<p>Unmanned boats will encounter many static and dynamic obstacles during navigation, and only real-time obstacle sensing can ensure safe navigation and long endurance of unmanned boats. In this paper, LGCGNet is proposed to perform real-time water surface semantic segmentation on the images captured by the on-board camera. In order to ensure that the model adapted to obstacles with extremely variable scales, a local-global module is proposed in this paper. The local-global module consisted of residual dense dilated module and context-enhanced separable self-attention. Residual dense dilated module enabled the enhancement of local detail information and context-enhanced separable self-attention enabled model receptive field expansion. In addition, the sub-pixel downsampling module is used to avoid the loss of feature information to improve segmentation accuracy. Experiments on the MaSTr1325 dataset showed that LGCGNet apprpached the segmentation accuracy of state-of-the-art semantic segmentation models with only 689,000 parameters and 9.068G floating point operations per second, with an mIoU of 84.14%. In addition, the processing speed of LGCGNet is 34.86FPS, which meets the frame rate conditions of commercially available photovoltaic equipment. The experiments demonstrated that the LGCGNet proposed in this paper strike a good balance between achieving high accuracy, reducing model size and improving real-time performance.</p>