A neural network framework for extrapolating sea surface spectra from wave pressure signals

This study investigates the prediction of sea surface elevation spectra from pressure measurements using frequency-specific artificial neural networks (ANNs). Unlike traditional time-domain approaches, the proposed methodology employs spectral analysis to decompose pressure signals into discrete frequency components and applies compressive sampling for outlier handling at the preprocessing stage. A separate ANN is trained for each frequency, utilising the water depth and other scalar features—including the statistical properties of the pressure head (mean, variance, peak period, and the narrow-bandedness parameter) - as inputs. The models are trained on 3005 samples, validated on 1252 samples, and tested on 751 samples. After model development, the ANNs were applied to a separate set of 100 previously unseen samples to evaluate the extrapolation framework. Model performance was assessed using a comprehensive set of error metrics. Results indicate that incorporating spectral features of pressure signals into ANN architectures provides a robust and efficient framework for sea surface spectra prediction, with potential applications in ocean and coastal engineering. • A frequency-domain ANN framework extrapolates sea surface spectra from wave pressure signals. • Each spectral component is predicted by a dedicated, frequency-specific neural network. • A modular architecture of 100 models supports frequency-wise validation and uncertainty control. • Field-measured data and spectral preprocessing enhance robustness and practical relevance. • Full-spectrum extrapolation achieves NRMSE below 25 % across most frequencies in wind-driven seas.