一种基于并行SNESIM的空间数据重建方法

A Reconstruction Method for Spatial Data Using Parallel SNESIM

空间数据的应用领域正在不断扩大.数据插值可以有效重建空间未知数据.数据插值就是一个数据再生的过程,即由原始数据再生出具有更高分辨率的数据.插值方法分为"确定"性插值和"不确定"性插值方法.不确定性插值方法的不确定性一方面表现在选用的插值方式具有随机性,另一方面表现在插值参数的选取和确定需要依赖于概率统计原则.多点随机模拟法(multiple-point simulation,MPS)是实现空间数据不确定插值重建的重要手段.单一标准方程模拟(single normal equation simulation,SNESIM)作为一种常用的MPS方法,目前已经用于多个领域的离散型空间数据三维重建.但是由于SNESIM给CPU和内存带来的负荷较大,大大限制了其实际应用.为了克服这种局限性,基于统一计算设备架构(compute unified device architecture,CUDA)实现SNESIM的并行化,并在计算训练图像(training image,TI)熵的基础上选择合适的数据模板尺寸;同时,通过整合软硬数据提高重建质量.与以往基于CPU的重建方法相比,基于CUDA的SNESIM并行算法显示出更好的空间数据重建效率.

The application of spatial data is becoming increasingly large. Interpolation can effectively reconstruct the unknown data in space, which is actually a process of data reproduction, and also a process of reproducing data with higher resolution from original data. Interpolation methods are divided into two branches： definite interpolation and indefinite interpolation. On one hand, the uncertainty of indefinite interpolation shows in selecting certain stochastic interpolation ways; on the other hand, the uncertainty is reflected by selecting the interpolation parameters using probability principles. Multiple-point simulation （MPS） is an important indefinite interpolation method in reconstructing spatial data, and single normal equation simulation（SNESIM）, as a frequently used MPS method, has been used in three-dimensional reconstruction of categorical spatial data in many fields currently. However, due to the large burdens on CPU and memory brought in by SNESIM, its practical application has been limited greatly. To overcome this limitation, SNESIM is parallelized using compute unified device architecture（CUDA）. A proper size of data template is chosen using the entropy theory of training image （TI） and the reconstruction quality is improved by the integration of soft data and hard data. Compared with the CPU-based SNESIM method, the CUDA-based one shows the better reconstruction efficiency of spatial data.