Efficient Visualization Of Multivariate Spatial Data

Multivariate spatial data visualization plays an important role in the fields of computing fluid dynamics, combustion simulation, medical imaging and meteorological simulation. With the rise of applications in large-scale scientific and engineering areas, the power of high-performance computers is enhanced rapidly, and the devices and methods of data acquisition become increasingly intelligent and wide-ranging. A dataset which meets the needs of specialists is usually multivariate, high-resolution and time-varying. TB or PB level multivariate spatial data has been common. However, the gap between the mentioned situation and the capability of multivariate spatial data visualization is growing. Most recent studies focused on the traditional univariate domain visualization and they are difficult to be applied to visualize and analyze the mentioned multivariate spatial data.Multivariate spatial data visualization is designed to express and analyze the variables and their relationships efficiently. The evolution rules of complex scientific phenomenons are displayed and explored by visualizing multivariate spatial data, and this is expected to assist scientists in discovering new physical phenomenons and rules. The main challenges are derived from the features of multivariate spatial data itself, including the huge amount of data size, the complexity of internal structures of data, the overlap and occlusion among variables in 3D data space.In this thesis, we consider 3D multivariate scalar field data as the main object of study. Com-bined with visualization pipeline, the key issues of efficient multivariate spatial data visualization have been discussed. The main contributions of this thesis are listed as follows:● We propose a compression domain volume rendering (CDVR) technique based on hierarchi-cal vector quantization and perfect spatial hashing for multivariate volume data. First, the original volume is split to blocks, and a Laplace decomposition process is applied for each block to obtain a hierarchical decomposition. The hierarchical decompositions of the data are compressed using vector quantization level by level. Then the obtained index volumes and vector quantization codebooks are utilized to get a recovery volume. The difference between the original volume and the recovery volume is considered as sparse residual volume; And then the perfect spatial hashing is employed to compress the sparse residual volume; Finally, the index volumes, the vector quantization codebooks and the results of perfect spatial hashing are transferred to the GPU for real time CDVR.● We propose a novel multivariate volume visualization scheme based on multi-class blue noise sampling with three visualization modes, DVR mode, iso-surface mode and cutting plane mode. First, the volumes of variables are converted to 2D view-dependent density fields, and all the separate density fields form a 2D mixed density field. In the screen-space density field, multi-class blue noise sampling is utilized to obtain a multi-class sampling distribution with no overlapping, which makes each pixel associate with one variable. Finally, based on the multi-class sampling distribution and one visualization mode, we execute a ray casting procedure. This method can effectively avoid the overlapping among variables and the variable-related color blending which exists in traditional multivariate visualization methods, and help users to perceive the features of the data effectively and accurately.● We propose an interactive environment for exploring multivariate spatial data based on semantic lens. First, we make high level abstractions of features and structures, and define the abstractions by well known semantics. Second, we develop a tool named semantic lens. By setting different semantics inside and outside the lens, we can observe different features inside and outside the lens separately. Meanwhile, we provide two methods to fuse different features in transition regions under users’ control. One is traditional color blending and the other is structure-based method. With the tools, the spatial occlusions among variables are alleviated, as well as the misleading caused by color blending.