Prospective Interest of Deep Learning for Hydrological Inference
Prospective Interest of Deep Learning for Hydrological Inference
复制标题
DOI：
10.1111/gwat.12557
复制
发表时间：
2017-07
期刊：
影响因子：
2.6
通讯作者：
J. Marçais;J. de Dreuzy
中科院分区：
文献类型：
--
作者：
J. Marçais;J. de Dreuzy
与地下水资源有关的引入决策需要对复杂的、多自由度的动力系统进行表征、建模和预测。尽管如此，这些系统具有大规模的结构，这些结构是基于应用于基本物理量(例如，质量、动量和能量)的守恒原理的分层性质而产生的。由于水文系统本质上是异质的，有时甚至是混乱的，因此出现了困难。这并不是水文学所特有的，但它是天然或人工复杂系统的共性。已经提出了两种方法来处理这些复杂的系统。无论它们被称为模型驱动还是数据驱动，是显式的还是隐式的，它们在方法上都有很大的不同。在显式的模型驱动方法中，流程以每个特征规模进行物理建模，并逐步向上扩展。模式、等价属性和有效的法律通过升级逐步显现。这在随机水文学中得到了广泛的应用，用来推导等效渗透率，并在渗流理论中用来确定普遍的标度定律(Stauffer和Aharony，1992)。通过经典但通常不适定的反问题调整模型参数来整合大规模观测(Zimmerman等人。1998年)。在隐式数据驱动的方法中，对所开发的模型的结构进行了最少的假设(Hastie等人。2003年；蒙哥马利2006年)。相反，这种方法依赖于基于统计学和人工智能的通用数据驱动分析。在众多的
Introduction Decision making relative to groundwater resources requires the characterization, modeling, and prediction of complex and dynamical systems with many degrees of freedom. Nonetheless, these systems have largescale structure that emerges from hierarchical properties based on conservation principles applied to fundamental physical quantities (e.g., mass, momentum and energy). Difficulties arise as hydrologic systems are inherently heterogeneous and sometimes chaotic. This is not specific to hydrology, but it is generic to natural or manmade complex systems. Two approaches have been proposed to handle these complex systems. Whether they are called model-driven or data-driven, explicit or implicit, they differ widely by methodology. In the explicit model-driven approach, processes are physically modeled at each characteristic scale and progressively scaled up. Patterns, equivalent properties, and effective laws emerge progressively through upscaling. This has been performed extensively in stochastic hydrology to derive equivalent permeability and in percolation theory to identify universal scaling laws (Stauffer and Aharony 1992). Large-scale observations are integrated by adapting the model parameters through the classic, but generally ill-posed, inverse problem (Zimmerman et al. 1998). In the implicit data-driven approach, minimal assumptions are made on the structure of the models developed (Hastie et al. 2003; Montgomery 2006). Rather, this approach relies on generic data-driven analysis based on statistics and artificial intelligence. Among numerous