Waterway navigation scene segmentation is crucial for safe and efficient autonomous vessel operation. To address the issues of large parameter sizes in current high-resolution semantic segmentation models, the difficulty in deployment on embedded devices, and inaccurate segmentation in navigation scenarios, this study proposes an improved Fast-SCNN-based real-time semantic segmentation network for waterway scenes. The proposed network is enhanced by incorporating a dual attention mechanism in the context branch, which allows feature extraction along spatial and channel dimensions to be guided more effectively, thereby improving accuracy. A pyramid pooling module based on decomposed dilated convolution is designed to enhance the model's multi-scale feature representation capabilities while ensuring real-time performance is maintained. Additionally, a loss function combining Focal Loss and Dice Loss is introduced to enhance focus on hard-to-classify samples and improve prediction accuracy for target regions, thereby significantly enhancing the model's ability to recognize minority-class targets. Experimental results have demonstrated that the improved model achieves a segmentation accuracy (mIoU) of 80.6% on the constructed On_Water_Dataset, with an inference speed of 90.1 frames per second, Segmentation accuracy increased by 16.9% relative to the original model. Deployment and experimental verification on embedded devices have confirmed that real-time performance is maintained while effectively segmenting complex backgrounds and multi-scale objects in waterway navigation scenarios.