Research Of Analysis And Prediction For Marine Environment Data Based On Machine Learning

As an important part of building a smart ocean,big data processing of marine environment plays an important role in rationally developing marine resources,accurately predicting marine disasters,and effectively protecting marine environment.Therefore,the environmental change of a certain sea area in China is considered as the research object.Based on machine learning theory,the marine environmental data collected from buoys are analyzed and predicted.The obtained results have been applied in the public service platform of marine environment observation and monitoring technology based on big data by the marine geological resources survey center of Hebei province.The main research contents are as follows:Firstly,the correlation analysis methods of marine environment data are studied.A global analysis mechanism based on gray correlation is established for marine red tide disaster,considering the buoy data of water quality,hydrology,meteorology and nutrient salt that affect marine environment change.It can calculate the correlation degree among influencing factors,aiming at evaluating influence on the occurrence of marine disaster.Furthermore,a local analysis mechanism based on the Apriori algorithm is proposed,where the maximum frequent item set is built by mining the association rules between divers influencing factors.Combining these two mechanisms,16 kinds of main factors are obtained for marine red tide disaster.Secondly,the prediction methods of marine single factor are studied,and support vector regression with smoothness priority is proposed.In structure,the model is composed of a nonlinear smoother and a least square support vector machine.The nonlinear smoothing device can effectively deal with the outliers and noises of the marine environment data from buoy collectors.The support vector machine serves as a nonlinear approximator to predict marine environment data series.The experimental results show that the model has good nonlinear approximation performance,where the normalized root mean squared error of is less than 10%,and the Pearson correlation coefficient is more than 90%.Finally,the multi-factor prediction methods are studied under the framework of statistical learning.A warning scheme of marine red tide disaster is developed based on an adaptive vector machine regression model.Essentially,this is a multi-input singleoutput prediction model,where the input is a variety of red tide influencing factors,and the output is the single determining factor of red tide.Besides,this model employs an adaptive kernel mapping mechanism,which can effectively select the optimal kernel function for specific tasks.The experimental results show that this proposal can accurately predict the occurrence of red tide,where the normalized root mean squared error is less than 7%,and the Pearson correlation coefficient is more than 95%.Figure 37;Table 18;Reference 65...