Few-shot source range estimation based on acoustic arrival structures in deep sea

Theoretical analysis indicates that variations in source range cause significant differences in acoustic arrival structures in the depth-time domain. Therefore, this paper proposes for the first time to use the arrival structure of the acoustic field as input features and to employ a two-dimensional convolutional neural network (2D-CNN) for underwater acoustic source localization from an image recognition perspective. To address the source range estimation with limited observed data, a transfer learning model is constructed. Based on pre-training with simulated data generated by the acoustic propagation model, the model is further trained on small-sample observed acoustic data from the detection area. Through sensitivity analysis, the impact of mismatched environmental parameters on source localization is investigated. Simulation results demonstrate that the proposed method exhibits superior performance and stronger robustness compared to traditional matched-field processing. Experimental results from the South China Sea show that the localization performance is much better compared to the conventional matched field processing (MFP) and the 2D-CNN method. For range predictions, the mean absolute percentage error (MAPE) is less than 5 % for the source up to 30 km.