Research On Deep Learning-based Marine Spatiotemporal Prediction

Spatiotemporal prediction is a very important part of marine scientific research,and it is of great significance for the marine environmental monitoring,early warning of marine disasters and marine emergency rescue.The temporal and spatial continuity and heterogeneity of marine processes lead to significant uncertainties and complex nonlinearities.Traditional marine spatiotemporal prediction methods are driven by empirical models,which are difficult to accurately extract the spatiotemporal characteristics of big data and effectively fit nonlinear features,which limits the accuracy of nonlinear marine environment spatiotemporal prediction in the era of big data.With the rapid development of marine monitoring technologies,the massive marine data with high precision,high frequency and large coverage has exploded.The marine big data has brought unprecedented opportunities and challenges to marine spatiotemporal prediction.However,the existing observation data are often incomplete in the representation of complex spatiotemporal marine processes,which is embodied in the incomplete data representation,incomplete feature representation and incomplete process representation.Aiming at the incomplete representation problem,a deep learning-based spatiotemporal marine prediction theory and method system has been established in this thesis based on the nonlinear leaning and fitting capabilities of deep learning and the characteristics of marine big data.The methods are applied in the buoy monitoring of Zhejiang costal area and Pacific Decadal Oscillation forecasting.The experimental results are analyzed and effectiveness of the models are verified.The main contents of this thesis are as follows:(1)Missing values will lead to inaccurate modelling,untimely analysis and prediction of marine prediction.In order to solve the problem of missing values in incomplete data representation,the multi-dimensional information of marine monitoring data is introduced to construct multiple hybrid views.A Matrix Completion-based Multiview Learning method(MC-MVL)is proposed for missing values imputation of marine spatiotemporal data.The complex correlation between different single views is fully considered in MC-MVL,and the problem of block missing problem is solved through the cooperation of different views.(2)In order to solve the error accumulation problem of continuous prediction caused by incomplete feature representation of complex monitoring series and difficulty of direct modelling,a Seasonal Gated Recurrent Unit(SGRU)network is proposed to realize the continuous forecasting of complex marine observation time series.The deep features can be sufficiently extracted through time series decomposition strategy.Then,a multi-pathway GRU network is constructed to continuously predict different components.The cumulative error can be effectively reduced and improve the accuracy of multi-step continuously prediction of time series.(3)A Gradient Convolutional Gated Recurrent Unit(GConv GRU)network is proposed for continuously spatiotemporal forecasting of marine dynamic systems.This model takes the advantages of powerful series feature extraction capabilities of the recurrent neural network and the spatial feature extraction advantages of the convolutional neural network.The marine spatiotemporal coupling,geographic spatial correlations and gradient information are considered in this model.Thereby it is able to extract the spatiotemporal evolution characteristics of the sea surface temperature anomaly and achieve end-to-end forecast of PDO accurately.(4)In view of the problems of incomplete process representation of complex marine spatiotemporal process,the weak explanation and prior information of deep learning based prediction methods,a Geo Spatial-Temporal Embedding Network(GSTEN)is proposed for marine environment spatiotemporal prediction.The Takens' embedding theorem and generalized embedding theorem are introduced in this method to design a Geospatial temporal information transformation equations.This thesis fully demonstrates the effectiveness and practicability of fusing dynamic system theory and the deep learning method in marine spatiotemporal prediction.In summary,this thesis expects to make theoretical innovations and method breakthroughs in the marine spatiotemporal prediction,effectively improve the accuracy and application capabilities of marine environment prediction based on deep learning,and promote the development of cross-research and integration of geographical information science,artificial intelligence and marine science.