A deep learning model for joint prediction of three-dimensional ocean temperature, salinity and flow fields

Abstract

Temporal prediction of three-dimensional spatial fields of ocean temperature, salinity and flow is important for a number of civilian and military applications. For accurate prediction of regional ocean environments, combing an accurate dynamic ocean model with optimized in-situ observations of ocean states provided by a network of marine vehicles (such as unmanned surface vehicles, autonomous underwater vehicles, and underwater gliders) is an important approach. To realize fast and accurate three-dimensional spatial prediction of regional ocean environments via ocean models and fast optimization of ocean observation strategies of marine vehicles guided by ocean models, this paper proposes a data driven dynamic model for temporal prediction of three-dimensional regional ocean environments. The proposed model is based on the deep learning method, and is developed based on the Convolutional-LSTM (long short-term memory) network. The feature of the proposed model is that the correlation of temperature, salinity and flow is considered and implemented in the neural network, with which the joint prediction of three-dimensional ocean temperature, salinity and flow fields is achieved. Data set form a numerical ocean model is used to conduct the training of the model and the results demonstrate that the proposed model could provide more accurate prediction than implementing prediction of single ocean temperature, salinity or flow field. The proposed model could be integrated with marine vehicles to form an accurate, high-resolution three-dimensional regional ocean prediction and fast-response adaptive sensing system.