Research On Density Clustering Algorithms And Its Application In Oceanic Mesoscale Phenomenon

In the era of big data,the explosive growth of data volume can not be coped with traditional manual processing methods.However data mining technology can quickly and efficiently discover potential patterns and knowledge within data.Among them,clustering algorithm has been widely used in data processing as a common data processing method.Its goal is to extract knowledge from data without category tags,discover potential associations between data and classify them.Mesoscale eddy,as a ubiquitous phenomenon in the ocean,generates a large amount of data every day.For either detection or tracking of mesoscale eddy,the earlier mesoscale eddy methods mostly use the method of processing data in the course of manual detection and tracking.However,the re-recurring algorithm for manually data processing usually has various shortcomings.Using data mining algorithms to process ocean data can effectively improve processing efficiency.Therefore,the research of mesoscale eddy detection and mesoscale eddy tracking algorithm based on data mining algorithm has important theoretical and practical significance.Foucs on the density clustering,to solve the problems mentioned above,combine the needs of mesoscale eddy research.We propose the algorithms to overcome the shortcomings of existing density clustering algorithms,and applied these algorithms to the processing of marine mesoscale eddy data.The main contribution can be summarized as follows:(1)A density clustering algorithm based on minimum spanning tree:The traditional density clustering algorithm usually uses fixed global parameters,so they are not competent for sample processing in sparse density regions.In this paper,by combining with the minimum spanning tree method of graph theory,the data set is divided by the density as the evaluation index,found the potential connection between the clusters through pruning to achieve the purpose of processing of similar density clusters in the data set as neighbour.(2)A density clustering algorithm based on k-nearest neighbor graph:Existing density clustering algorithms usually require at least two parameters to ensure the successful completion of the clustering process,and the parameters are mostly set as non-integer,so the parameter testing process is complicated.By constructing the k-nearest neighbor graph,the effective similarity information is filtered,and the test complexity in parameter setting stage is reduced by parameter adaptive method,meanwhile ensuring the accuracy of the algorithm.(3)A mesoscale eddy detection algorithm based on density clustering:In the process of mesoscale eddy detection,based on the robustness of density clustering to irregular clusters,density clustering is applied to cluster the data set which has been removed the non-eddy region,and the potential mesoscale eddy data set is screened out.And the stability of the results is guaranteed by adding stability criteria.Then the closed contours that meet the requirements are found.As a result,the above method eliminates the problem of sensitivity of threshold setting and sensitivity test required in parameter setting in existing algorithms,and solves the problem of instability of screening results.(4)A mesoscale eddy tracking algorithm based on time-scale density clustering:Density clustering can be used to separate the characteristics of dense and sparse regions of the samples.So,based on the characteristic of density clustering,the mesoscale eddy trajectory is separated from the environment region,through the distance metric matrix is changed by time scale and amplitude constraints,to find the potential trajectory.Aiming at the path with time conflict,combining Kalman filter method to eliminate noise points,improve the accuracy of tracking process.Through the above methods,the local optimal problem caused by the serial problem of the existing algorithm is eliminated,and the batch processing capability and the operation speed of the algorithm are improved.