Stochastic reconstruction of spatial data using LLE and MPS
详细信息 查看全文
- 作者：Ting Zhang ; Yi Du ; Bo Li ; Anqin Zhang
- 关键词：Locally linear embedding ; Stochastic reconstruction ; Dimensionality reduction ; Nonlinear ; Spatial data
- 刊名：Stochastic Environmental Research and Risk Assessment
- 出版年：2017
- 出版时间：January 2017
- 年：2017
- 卷：31
- 期：1
- 页码：243-256
- 全文大小：
- 刊物类别：Earth and Environmental Science
- 刊物主题：Math. Appl. in Environmental Science; Earth Sciences, general; Probability Theory and Stochastic Processes; Statistics for Engineering, Physics, Computer Science, Chemistry and Earth Sciences; Computa
- 出版者：Springer Berlin Heidelberg
- ISSN：1436-3259
- 卷排序：31
文摘
Spatial data are widely used in many scientific and engineering fields, such as remote sensing, environment monitoring, weather forecast and mineral exploitation. However, direct measurements of such spatial data sometimes are difficult to achieve due to the expensive cost of equipment or current limited technology, so stochastic reconstruction or simulation of spatial data are necessary based on the principles of statistics. As a typical statistical modeling method, multiple-point statistics (MPS) has been successfully used for stochastic reconstruction by reproducing the features from training images (TIs) to the reconstructed regions. However, because these features mostly have intrinsic nonlinear relations, the traditional MPS methods using linear dimensionality reduction are not suitable to deal with the nonlinear situation. In this paper a new method using locally linear embedding (LLE) and MPS is proposed to resolve this issue. As a classical nonlinear method of dimensionality reduction in manifold learning, LLE is combined with MPS to reduce redundant data of TIs so that the subsequent reconstruction can be faster and more accurate. The tests are performed in both 2D and 3D reconstructions, showing that the reconstructions can reproduce the structural features of TIs and the proposed method has its advantages in reconstruction speed and quality over typical methods using linear dimensionality reduction.