Low-Complexity Multi-Step Signal Prediction for Underwater Acoustic Communications: A Joint Reservoir Computing and Transfer Learning Approach

Underwater acoustic communication faces significant challenges due to multipath propagation, signal attenuation, and environmental variability. Accurate signal prediction supports adaptive transmission by forecasting upcoming samples and enabling efficient resource allocation. This work presents a complexity-reduced echo state network (ESN) framework for underwater acoustic signal prediction using real-world shallow-water at-sea experimental data. By reducing the reservoir to 200 neurons and optimizing hyperparameters, the model achieves strong performance with low computational complexity. Single-step prediction on the first hydrophone (H1) yields a test R2 of 0.9978. Furthermore, multi-step forecasting on H1 is performed using a recursive prediction strategy and evaluated up to 6 steps ahead, achieving R2 = 0.8232 at step 6. Using transfer learning, we extend the H1-trained model across 14 additional hydrophones in the vertical array by retraining only the readout layer, which reduces retraining time and computational complexity. Cross-dataset validation between different shallow-water environments further demonstrates robustness to changes in propagation conditions. Across the vertical array, single-step transfer learning achieves R2 ≥ 0.997 for most hydrophones, while cross-dataset transfer yields R2 values typically in the range 0.93-0.97.