Global wave parameter prediction via Hybrid CNN-LSTM networks and Multiparameter Ocean-atmosphere data

Accurate predictions of significant wave height (SWH) and principal wave direction (PWD) are essential for maritime safety and operational oceanography, particularly given the limitations of computationally expensive numerical wave models under extreme conditions. This study develops a hybrid convolutional neural network-long short-term memory (CNN-LSTM) deep learning framework for global wave parameter prediction using multiparameter ocean–atmosphere data. The model processes 10-meter wind components, sea surface temperature, mean sea level pressure, and bathymetry from ERA5 reanalysis data (2020–2024) to capture spatial and temporal dependencies in ocean‒atmosphere interactions. The hybrid architecture combines CNN spatial feature extraction with LSTM temporal modelling. Model performance is exceptional for SWH prediction, with correlation coefficients ranging from 0.909–0.938 and RMSE values ranging from 0.549–0.552 meters. PWD prediction shows moderate-to-good performance ( r = 0.805–0.889), with characteristic challenges inherent to directional variables. Spatial analysis reveals optimal performance in mid- to high-latitude storm-dominated regions (correlations > 0.92) and reduced skill in tropical convective zones and coastal environments. Independent validation against 54 NDBC buoy stations confirmed the short-term operational viability, with an SWH correlation of 0.932 (RMSE = 0.83 m) and a PWD correlation of 0.889 (RMSE = 12.24°). Temporal validation has demonstrated successful capture of seasonal variability and storm events, although systematic underestimation of extreme wave heights in coastal regions represents a key limitation. Compared with complex nearshore conditions, the model exhibits superior performance in open ocean swell-dominated environments. This framework provides computationally efficient global wave forecasting with performance comparable to that of traditional numerical models while requiring significantly reduced computational resources.