A Research On The Preprocessing Methods Of Spatio-temporal Serial Data

With the continuous development of 3S technology,the volume of spatio-temporal serial data increases rapidly.Meanwhile data acquisition technology and storage capacities are constantly improved and more different kinds of spatio-temporal data become the research objects of geography.As the typical representation of Big Earth Data from space,spatio-temporal serial data are widely used in multiple research fields,which continue to receive increased attention.These multivariate spatio-temporal serial data have a great significance for scientific researches,such as spatio-temporal analysis,resources and environmental monitoring and earth spatial cognition.Data low-quality problems become more obvious when spatio-temporal serial data volume drastically increased.Caused by various restrictive factors related to data acquisition methods,data transmission and storage,incomplete data and high levels of noise problems in time series frequently appear,and the scale differences problems of multivariate spatial data become more significant.Low-quality problems have become bottlenecks for spatio-temporal data analysis,modelling and knowledge discovery,and provide new challenges.Traditional processing methods of the low-quality problems mostly assume that spatio-temporal data are subsection smooth and directly process these data using classical statistical analysis methods.However,spatio-temporal serial data have both local correlations and gradients,which also exist in spatio-temporal mutations,and nonlinear and non-stationary features make traditional methods have some limitation to process low-quality problems.Aimed at missing value,high noise and scale differences in spatio-temporal data,this paper deeply discusses processing methods of these low-quality problems: 1)Missing value imputation study: using nonlinear regression analysis method fillmultiple missing values in time series;using improved non-local means(NLM)algorithm achieve single missing value imputation and fully consider nonlocalsimilarity of the data laws in time series;2)Denoising study: using an NLM algorithm deals with Gaussian noise in time seriesthrough the analysis of noise statistical model and fully excavate similarity of thedata laws to remove noise;3)Scale transformation study: using the Gaussian pyramid method solves scaledifferences geographical spatial data and achieves spatial resolution reducing byGaussian smoothing fuzzy features.Based on the above theory methods,we utilize several time series and spatial data in Jing-Jin-Ji Region to make experimental analysis,and validate the proposed algorithms effectiveness.From the experimental results analysis,when missing value imputation,nonlinear regression analysis method and improved NLM algorithm can solve the problems such as traditional statistical method processing data on a smaller scale and complicated model parameters,and also overcomes traditional methods limitations in only considering the effects of local adjacent data and adapt for long time series well;NLM algorithm removes Gaussian noise effectively and keeps non-linear and non-stationary trend in geographical time series excellently;Gaussian pyramid method solves the problems of traditional methods in easily ignoring data extreme change and spatial similarity mutation of research objects,and has a good advantage in keeping with the local detail characteristics of spatial data.