Ship wakes generated in relatively shallow estuaries are subjected to complex dynamic processes such as nonlinear wave evolution and wave–wave interaction. Identifying ship-wakes in coastal waters is important for maritime and coastal management, including waterway planning, shoreline protection, and navigation-related problems in busy waters. In this study, a machine learning framework was developed to identify ship-wakes using supervised training algorithms. Wave data were measured at two stations in the Pearl River Estuary of South China. The two stations are 1,500 m apart, one is at near-field of the navigation channel, and the other is at far-field. Two machine learning techniques, namely, the multilayer perceptron (MLP) model and support vector machine (SVM) model, are employed. We used two data formats in the training and testing processes; one is the digital data of wave time series, the other is the spectrogram derived from the time series. Tests suggested that spectrogram is a more appropriate format for both models versus the time-series format. The SVM model has higher cross-validation scores and higher computational costs compared to the MLP model. The trained models using the near-field data can be used for predictions at the far-field location with high accuracy. Sensitivity tests revealed that long primary waves and primary wake chirps are the most critical components in the spectrogram among overall ship-wake characteristics for ship-wake recognition. The time and frequency resolutions in the spectrogram only have minor effects on model performance.
Read full abstract