S3-GEP: Scalable Spatiotemporal Statistics for Global Environmental Phenomena

利用今天大量的开放地球观测数据，可以在全球范围内对环境现象进行全面的地质统计分析。然而，随机场操作的计算复杂性、全局不现实的模型假设（如平稳性、可分离性和各向同性）以及当数据大小超过本地存储容量时的数据管理问题，目前限制了数据在应用程序中的实际使用，并导致不可共享、不可复制的研究。例如，使用几tb量级的全球高程数据作为协变量信息来模拟降水的空间变化，需要快速的推理统计方法和大量的数据管理工作。在这个项目中，我们的目标是开发对全球环境现象进行地质统计推断的有效方法，这些方法：(i)在共享无架构上进行计算缩放，（ii）考虑非平稳、不可分和各向异性的时空依赖关系，以及（iii）能够整合多个数据源，包括遥感图像和原位观测。因此，我们将建立将空间马尔可夫随机场与时间平流扩散过程相结合的地统计模型，并将新算法应用于高度可扩展的数据管理和分析系统SciDB。开发的方法将在两个用例中进行演示，包括创建全球高分辨率降水数据集和使用外部自变量模拟土地利用变化。除了全球建模之外，第二个用例将强调项目结果如何促进与最新的高分辨率卫星衍生数据集的合作。为此，来自Sentinel-2和TanDEM-X任务的数据将用于全国范围的土地利用变化分析。预期的主要贡献包括分布式计算环境中有效的地质统计推断算法，全球尺度上非平稳性和各向异性建模方法，以及可重复的大规模环境数据管理和分析的开源软件工具。

With today's amount of open Earth observation (EO) data, comprehensive geostatistical analyses of environmental phenomena can be brought to global scale. However, computational complexity of random field operations, globally unrealistic model assumptions like stationarity, separability, and isotropy, and data management issues when data size exceeds local storage capacity currently limit the practical use of the data in applications and result in unshareable, irreproducible research. As an example, using global elevation data in the order of a few terabytes as covariate information in modeling the spatial variation of precipitation requires fast methods to inferential statistics and extensive effort in data management. In this project, we aim at developing efficient methods for geostatistical inference on global environmental phenomena that (i) computationally scale well on shared nothing architectures, (ii) consider non-stationary, non-separable, and anisotropic spatiotemporal dependencies, and (iii) are capable of integrating multiple data sources including remote sensing imagery and in-situ observations. Therefore, we will build geostatistical models that combine spatial Markov random fields with temporal advection-diffusion processes and adapt novel algorithms to the highly scalable data management and analytics system SciDB. Developed methods will be demonstrated in two use cases including the creation of a global high resolution precipitation dataset and modeling land use change with external independent variables. In addition to global modeling, the second use case will emphasize how project results facilitate working with latest, high-resolution satellite-derived datasets. For this, data from the Sentinel-2 and TanDEM-X missions will be used in a national scale land use change analysis. Expected key contributions include algorithms for efficient geostatistical inference in distributed computing environments, approaches for modeling non-stationarity and anisotropy on global scale, and open source software tools for reproducible large-scale environmental data management and analyses.