Multiple-Point Geostatistics For Spatial Information Analysis And Applications

Statistical analysis is one of the basic methods in geography area. The theory of geostatistics was proposed by a French researcher Matheron in1963. It accounts for the characteristics of geospatial data (i.e., spatial distribution and correlations of variables) and has been applied widely in various geography research areas. Geostatistics includes a set of regionalized variable theory and can infer statistics of spatial variables using prediction or simulation, in which variogram or covariance function is used as a measurement tool to establish spatial dependence or correlation model of spatial variables. Early applications of geostatistics were in geology exploration, and now are widely applied in ecological resource, agriculture and forest, etc. Multiple-point geostatistics (MPG) proposed by Guardiano and Srivastava in1993is a development of geostatistics. The main advantage of MPG is introducing the concept of training image, from which the multiple-point statistical information can be searched and stored. The information can be regarded as the prior knowledge for reconstruction and spatial prediction so that spatial continuity and variability can be expressed. Training image substitutes the function of traditional variogram (or covariance), and overcomes the limitation of variogram that can only describe the spatial relationships between pairs of points. MPG is used for oil and facies detection in most cases. Recently, some researchers applied MPG to process remotely sensed images, but not common in methods or applications.The purpose of this thesis is to introduce the MPG theory to remote sensing, and to use multiple-point geostatistical simulation to explore the methods for image classification, image data refinement, and super-resolution reconstruction. It is expected to address the advantage of MPG over traditional geostatistics through testing the proposed methods by experiments, and to widen the application of MPG as well. The details of the research include:(1) The thesis first presented the research status of remote sensing and GIS. Then the problems of those key areas were summarized; the existing methods were compared and analyzed. Thus, the objective of this research, the main contents, the key methods and significance were expanded and discussed.(2) The basic theory and some concepts of traditional geostatistics and MPG were introduced. The classical kriging methods in traditional geostatistics and multiple-point simulation were elaborated. (3) Post-processing of classification from remotely sensed imagery was performed using multiple-point geostatistical simulation. The spatial context was extracted from training image to establish spatial dependence and to improve the classification accuracy. High-order statistical analysis was performed on the representative classes. Spatial filtering post-processing and context-based Markov random field classifier were compared to the MPG post-processing, the characteristics of those classes with complicated distribution were analyzed.(4) A spatial data integration method in the condition of sparse sampling was studied. The pattern-based multiple-point geostatistical simulation was applied to predict DEM (digital elevation model) data with a high spatial resolution, that is, data refinement. Traditional geostatistical methods using various kriging predictions were compared to MPG refinement, the abilities of extracting and utilizing spatial structures of terrain data were analyzed.(5) Super-resolution image reconstruction was studied. The change-of-support (COS) model based on the regularization was introduced in pattern-based MPG, so the key technique of downscaling was developed. The method was compared to other methods without considering COS model in several combined images with different scale factors. The continuity, the visual effect and error statistics of the constructed image were studied.The innovations of the thesis are:(1) A post-processing method based on multiple-point spatial structural statistics for land cover classification of remotely sensed imagery was proposed. It is suggested that MPG has the advantage in achieving a higher classification accuracy as well as processing those classes (land cover categories) with curvilinear spatial distribution over the traditional spectral-and contexture-based classification.(2) A DEM data refinement method was proposed based on multiple-point structural feature. The method utilizes the self-similar characteristics of training image represented as terrain structure. The ability of capturing spatial structure was improved based on local pattern matching. The experiment shows that this method can result in more accurate integration results than traditional kriging predictions.(3) A super-resolution image reconstruction method was proposed based on ATPCK and MPG. It realized statistical downscaling based on a data-driven data support transformation model, which is, reconstructing of multispectral images at the fine scale. The method compensates the drawback of lacking of data support transformation model in MPG simulation, and expands the application for super-resolution in continuous image data.There is much improvement left for MPG in remotely sensed image processing, multi-source data integration and refinement, and super-resolution image resconstruction. According to the results of this thesis, there are several aspects that can be further addressed:(1) Sensitivity analysis of the parameters of training, including image accuracy, extent and non-stationarity of training image, etc. The selection of unconditional training image should be further considered.(2) Quantitative accuracy analysis of the constraint for MPG simulation. It is suggested that the secondary variable, which represented as the probability field for categorical data or varying locally mean for continuous data, has great impact on the accuracy of simulation. The accuracy analysis of conditioning for simulation is still weak.(3) Self-adjustment of the non-model based simulation algorithm should be tested. It is expected to select the most appropriate parameters for simulation based on some quantitative indices such as land cover map continuity of categorical data, and the information entropy of the continuous image data.(4) The algorithm efficiency is expected to be further improved. The proposed methods are expected to apply to wide extent and multi-temporal spatial data.(5) More tests can be applied to improve the applicability of these methods. If a set of stable and mature techniques is developed, more application can be exploited. For example, the real multi-scale data can be collected to solve problem in the land surface system such as scale transform using MPG.