Frequency–spatial feature learning with spectrally guided data augmentation for submarine cable fault detection under data scarcity

Submarine cable fault detection is challenged by severe fault-sample scarcity, class imbalance, and degraded underwater imaging conditions, including low contrast, turbidity, and background clutter. To address these challenges, this paper proposes a unified frequency–spatial framework for submarine cable fault detection. A frequency-guided generative adversarial network is introduced to generate spectrally consistent minority-fault samples, mitigating class imbalance and enhancing fault-sample diversity. Building upon this, a frequency–spatial multi-scale cable fault detection network is developed to improve the detection of weak fault patterns through joint modeling of frequency-domain responses and multi-scale spatial features. The resulting framework enables more reliable detection of subtle and small-scale fault patterns under low-contrast and visually degraded underwater conditions. Experimental results show that the proposed method consistently outperforms representative baseline detectors on the primary imbalanced dataset, while additional validation on expanded simulated and near-shore datasets further supports its robustness and promising transferability.