Retrieval of wave period from altimetry: Deep learning accounting for random wave field dynamics

Abstract As an extension of wave period inversion models based on altimetry data by an artificial neural network approach, a Deep Neural Network approach (DNN) mean wave period model is developed by introducing new parameters as DNN inputs. In addition to conventional altimeter-derived parameters such as significant wave height (SWH) and the normalized radar cross-section (NRCS) sigma0, the spatial (along-track) SWH gradient and SWH standard deviation (STD) for standard 1-s altimetry data are assumed to be responsible for random wave field dynamics and, thus, for the observed characteristic mean wave period. A comparison with in situ measurements of wave buoys shows higher accuracy of the novel DNN models by using these new parameters. The wave period estimation from DNN model is also consistent with the latest wave reanalysis and indicates less bias when compared to the buoys. The global mean wave steepness distribution from the DNN model shows good agreement with those provided by the wave reanalysis. The sensitivity of input variables sigma0, SWH, SWH gradient, and SWH STD on the results of the DNN model are also investigated. Perspectives on the DNN method for developing universal mission-independent wave period models are discussed. Objectives of the work The mean wave period is an important parameter for characterizing wave properties and can be retrieved from altimeter observations. Considering random wave field dynamics, we present a deep neural network (DNN) model for mean wave period retrieval by introducing the SWH gradient and standard deviation as the inputs. The mean wave period estimation of the DNN model shows good agreement with the latest wave reanalysis and less bias compared to buoys. In an attempt to make the DNN mean wave period retrieval model universal, a method of applying DNN model cross-altimetry missions is presented.