Abstract: To address the issues of inaccuracy, insufficient robustness, and low automation in traditional seafloor line detection algorithms, this paper for the first time applies semantic segmentation technology from the field of image processing to the task of seafloor line extraction from multibeam water-column images, and proposes a high-precision, fully automatic seafloor line extraction algorithm for such images based on the U-Net architecture. First, the nearest-neighbor interpolation algorithm is used to correct outliers and convert raw sonar data into a structured, gridded format suitable for training, thereby constructing a high-quality dataset. Furthermore, by incorporating phase-based detection cues, an edge-beam replacement and fusion strategy is proposed to enhance detection robustness under low signal-to-noise ratios and diverse seafloor topographies. Building upon the U-Net architecture, we design a dual-loss collaborative training mechanism to mitigate the class imbalance problem arising from the extreme sparsity of seafloor line pixels; this significantly improves recognition sensitivity for sparse target pixels while effectively suppressing edge blurring and missed detections. Experimental results demonstrate that the proposed method achieves high-precision, efficient, and fully automatic seafloor line extraction across varied underwater environments, exhibiting strong generalizability. Under task-specific evaluation metrics designed for seafloor line segmentation, the method achieves over a 4-percentage-point improvement in detection accuracy compared to conventional approaches and maintains segmentation accuracy consistently above 90% on all primary test sets.