Dual-decoder neural network based for end-to-end prediction of acoustic transmission loss in deep-sea environments

Efficient transmission loss (TL) computation in deep-sea environments is essential for underwater target detection and ocean monitoring. Traditional computational methods rely on complex numerical solvers, which are associated with high computational complexity and limited efficiency. Neural network-based end-to-end prediction methods improve computational efficiency, but accurately modeling deep-sea acoustic TL remains challenging. Existing methods often fail to represent both the global structure and local details of the acoustic field, which limits overall prediction accuracy. In this study, a dual-decoder parallel network model named DualDecoder-PUNet is proposed. Based on the U-Net architecture, it integrates a dual-decoder and a gated decoupled attention module for coordinated modeling of global and local features. In addition, deep-sea sound speed profiles are included as model input to enhance generalization across diverse deep-sea environments. Experimental results show that DualDecoder-PUNet achieves a root mean square error of approximately 1.2 dB on representative deep-sea terrain datasets, while reducing computation time by more than 91% compared with traditional numerical models. Gradient-based visualization helps identify key features influencing prediction and improves interpretability. The proposed DualDecoder-PUNet enables efficient and accurate prediction of deep-sea acoustic fields with reduced computational cost, and applies to deep-sea target detection.