Robust and Accurate Magnetic Anomaly Detection Under Complex Noise Using Synchrosqueezed Short-Time Fourier Transform and Attention-Based Residual Network

Magnetic anomaly detection (MAD) is a vital passive sensing technique widely applied in underwater target detection, geological exploration, and military defense. A major challenge lies in accurately identifying weak signals in complex scenarios, particularly under low signal-to-noise ratios (SNR) conditions. Traditional and machine learning-based methods suffer from limited robustness and reliance on hand-crafted features. Although deep learning approaches enhance feature extraction, they often struggle with insufficient time–frequency resolution and noise sensitivity. In this study, a novel MAD method integrating the synchrosqueezed short-time Fourier transform (SSFT) with a residual neural network (ResNet), enhanced by a convolutional block attention module (CBAM) simultaneously, is proposed. SSFT is used to convert 1D time series into the 2D time-frequency spectrum with a high energy concentration. While CBAM is used to adaptively emphasizes salient features. Numerical experiments demonstrate that the proposed method achieves a detection accuracy of 77.4% and 6.1 false alarms per kilometer at an SNR of –3 dB, outperforming conventional methods (OBF: 63.5%, ME: 50%) and learning-based models (SVM: 70.6%, 1D-CNN: 67.5%, WVD-CNN: 69.25%, BiLSTM: 59.85%). The effectiveness and robust superior of the proposed method, achieving a detection accuracy of 94% with the lowest false alarms per kilometer, is further validated by field survey scenarios.