Acceleration, Parameters Optimization And Application Of Kriging Interpolation

Kriging interpolation is an effective interpolation method. As an important part of geostatistics, it has been widely applied in geology, meteorology, geography and cartography etc. Compared with the traditional interpolation methods(IDW interpolation, Natural Neighbor interpolation, nearest neighbor interpolation, local polynomial method, moving average, linear interpolation triangulation method and RBF), Kriging considered characteristics of regionalized variable, i.e. variation features and structural characters. Based on the full consideration of the positional relationship between the spatial sample points and the spatial relationships among points to be estimated, the variogram theory and structural analysis is introduced. Kriging is an optimal unbiased estimation with very high precision. In recent years, kriging interpolation method is continuous improved, researchers devoted more attention on it and a large number of practical applications was developed. Now, Kriging has become an important application tools and research focus in related disciplines.To a large interpolation area or a large number of samples, kriging interpolation method will consume a great amount of computation and memory resources, which severely restricts the application scope of kriging. Therefore, it is necessary to reduce kriging time based on parallel platforms. With the development of parallel computing technology, rapid interpolation to a large interpolation area and large samples is possible. GPGPU is a recently developed new method of parallel computation. Compared with cluster platform, it has its unique advantages: low-cost, low power consumption and other advantages; compared with the multi-core platform, it can run more concurrent threads. In this research, kriging interpolation method is realized based GPU platform, which considered GPU architectural features and characteristics of the kriging algorithm; establishing a kriging interpolation technology architecture on CPU-GPU heterogeneous platforms; variogram parameters were optimized by a hybrid of genetic algorithm and particle swarm algorithm; Finally, kriging interpolation method was applied in the prediction of leaf area index(LAI). The research results of this dissertation can be used to not only accelerate the Kriging interpolation, but also provide some references for other GPU algorithms. The main work is as follows:First, parallel granularity selection. Different granularity is analyzed by using Amdahl’s Law, and optimal granularity was concluded. It also provides a reference for the other parallel algorithms on granularity problems.Second, Reduction of redundant storage and computation. Based on the fact that adjacent unknown points usually have the same neighbors, a optimizing strategy called RMR(Redundant matrix reduction) was propose. In this method, only a coefficient matrix was constructed and calculated the inverse matrix to that points having same neighbors. In this process, RMR method can simultaneously reduce GPU computing algorithm storage and consumption.Third, nearest neighbor searching. To the large interpolation area, the solution is divided the area into many small parts. For the above reason, a fast scalable window search method(quick extensible window searching, QEWS) is proposed to search nearest neighbors. Compared with the KD-tree, QEWS is more suitable for the small sample size.Fourth, acceleration of kriging on CPU-GPU heterogeneous platforms. When GPU is busy, CPU is usually in a wait state. This study jointed CPU and GPU to accelerate kriging algorithm, the results show that this method can effectively improve speed of kriging interpolation.Fifth, estimation of variogram parameters. Variogram parameters has an important impact on the final result of the kriging interpolation. In this study, variogram parameters were estimated by a bybrid of genetic algorithm and particle swarm algorithm.Sixth, prediction of LAI. The time series of MODIS LAI include linear and nonlinear components of a single point, which cannot be accurately modeled and predicted by either linear method or nonlinear method. In this study, a hybrid of SARIMA(a time series analysis method) and BP neural network is used to model the linear component and the nonlinear component of MODIS LAI time series respectively. The sum of above two method is as the final result. Experimental results show that the combination of SARIMA and BP neural network can improve the prediction accuracy. To regional LAI, the LAI of each point is predicted by SARIMA and the residue of predicted result is fitted by kriging. Experimental results show that the proposed hybrid method of Kriging and SARIMA can improve prediction accuracy.