Abstract: The energy level in the high-frequency tail of the ocean wave spectrum is physically coupled with sea surface wind stress, which provides a theoretical foundation for wind field parameter inversion based on ocean wave observations. Nevertheless, the accuracy of traditional empirical methods is limited under mixed sea conditions. Furthermore, existing deep neural network (DNN) models only extract features from individual frequency points and fail to exploit the correlation information among different frequencies of the wave spectrum. Accordingly, convolutional operations are introduced in this paper to globally extract the implicit inter-frequency features of the wave spectrum.Based on the in-situ observational data from the National Data Buoy Center (NDBC), a lightweight end-to-end regression framework named MobileNetV4-1D is constructed for temporal feature mining of ocean wave spectra, to realize synchronous inversion of wind speed and wind direction. Subsequently, comparative experiments are conducted among physical empirical methods, conventional DNN models and the proposed MobileNetV4-1D model to systematically evaluate their inversion accuracy and robustness in wind field retrieval. In particular, the prediction scenario of models on unknown future monthly data in practical operational applications is emphatically simulated. The experimental results demonstrate that the temporal generalization performance of MobileNetV4-1D is significantly superior to that of the conventional DNN: the root mean square error (RMSE) of wind speed inversion is reduced to 1.43 m⋅s⁻¹ (a decrease of 9.0%), and the RMSE of wind direction inversion is lowered to 31.06°(a decrease of 32.7%). This study verifies that the adapted MobileNetV4-1D model achieves favorable inversion accuracy with enhanced temporal generalization ability, which provides reliable technical support and theoretical references for the operational application of ocean wind field inversion.

Key words: wave spectrum, inverse wind, MobileNetV4, one-dimensional convolution