A CNN-Transformer-TCN-MHAP Hybrid Framework for Predicting Marine Environmental Noise Based on Meteorology

Marine ambient-noise spectrum levels vary nonlinearly across meteorological and oceanic conditions, making short-term prediction difficult under non-stationary and multi-scale temporal dynamics. This study develops a meteorology-driven causal temporal framework, termed CTTM, for predicting ambient-noise spectrum levels from ERA5 reanalysis variables and NOAA Ocean Noise Reference Station observations.The model combines convolutional encoding,Transformer-based dependency modeling, temporal convolutional networks, and multi-head attention pooling to represent local variability, long-range dependence, causal multi-scale dynamics, and temporally weighted aggregation.Hourly wind speed, sea-surface temperature,sea-level pressure, and precipitation were aligned with calibrated acoustic observations to construct the prediction dataset.The task was evaluated at 250,500,1000,and 1500Hz,with NRS02 used as the primary station and NRS03 used for direct cross-station testing. Rolling-origin validation was introduced to assess sensitivity to temporal partitioning. At 1000Hz on NRS02, CTTM achieved an RMSE of 1.10 dB, an MAE of 0.79 dB, and an R² of 0.959,outperforming recurrent and Transformer-based baselines, including LSTM, BiLSTM, BiGRU, PatchTST, and iTransformer, under the same evaluation protocol.Multi-frequency experiments showed frequency-dependent predictability, with 1000Hz providing the most consistent observed-predicted agreement. In direct NRS02-to-NRS03 transfer, the model retained predictive skill without retraining, although accuracy decreased under the contrasting station conditions. These results support a reproducible workflow for evaluating meteorology-driven ambient-noise prediction across selected frequencies and NOAA-NRS observation settings, while indicating that broader spatial generalization and physical interpretability require additional stations,vessel-activity information, ocean-dynamic variables,and quantitative attribution analyses.