Dam hazard identification plays a critical role in water conservancy safety monitoring. However, existing methods encounter significant challenges in multi-source geophysical data fusion and precise identification due to diverse data sources, substantial feature variations, and ambiguous target boundaries. To address these issues, an improved U-Net neural network (U-Net)-based multi-source collaborative identification framework was proposed. Independent feature extraction modules coupled with noise-assisted mechanisms were designed to enhance feature consistency and complementarity across transient electromagnetic, ground-penetrating radar, and underwater sonar data. The model’s capability in identifying minority classes, high-noise backgrounds, and hazard boundaries was strengthened through enhanced skip connections, weighted cross-entropy loss function, and cosine annealing learning rate scheduling. Field application at the Lixinxu embankment in Jiujiang City, Jiangxi Province demonstrates that the proposed method achieves over 87% recognition accuracy in typical hazard zones, with significant improvements in mean intersection over union (mIoU) and other metrics compared to conventional approaches. The results validate the method’s stability and adaptability in processing multi-source geophysical data under complex geological conditions, providing reliable technical support for intelligent dam hazard identification systems.