Examining groundwater-flood and soil moisture-flood relationships across scales using national-scale data mining, deep learning and knowledge distillation

在美国的许多地方，已经证明地下水位和土壤湿度是通过洪水产生的机制联系在一起的，土壤湿度量化了土壤的湿度。当地下水接近地表时，水不能渗入地下，因此被迫迅速流向河流，造成更高的洪水风险。然而，地下水和洪水之间的关系被发现是高度多样化的，很难预测。根据地形、地下水深度和许多其他因素的不同，洪水有时会导致地下水增加，而在其他情况下，地下水可能会导致洪水。以前从选定的试验性流域进行的研究还没有导致对控制过程的全面和可转移的理解。这个项目将采用大数据、机器学习的方法来加强我们对这种关系的理解，使我们能够启发式地利用以前未得到充分利用的地下水数据来进行洪水预测和减少损失。机器学习算法使用来自全国范围的地下水和径流数据的学习模式，将创建可信的地下水-洪水关系。该项目将利用现有卫星任务的大水文学数据，创建共享的本科课程模块，以增强学生使用大数据的能力，并提高他们对全球水问题的认识。这项研究通过回答以下首要问题来推进水文科学：在集水区尺度上，集水区的地下水位是否为洪水阈值函数和基流提供了预测能力？我们将分多个小步骤来解决这个问题。我们将确定在美国大陆可以找到的地下水-降雨-径流(GW-P-Q)关系的类型。这些关系通过不同滞后和不同时间尺度上的水位深度与洪水阈值(和基流量)之间的相关性来量化。我们将寻求决定GW-P-Q关系类型的因素，以及这些关系是否在季节和年份中保持稳定。我们将采用两种方法：人工指导的分类分析和基于深度学习(DL)的知识蒸馏方案，这是一组快速发展的技术，支持最近人工智能的激增。在第一种方法中，我们将使用分类和回归树来识别可以解释GW-P-Q关系的因素。在基于DL的方法中，我们将使用所有可用的数据来训练大陆尺度的时间序列DL模型来预测流量。这种方法解决了分类树中没有足够的数据可用于分支节点的问题。通过一种新颖的知识提炼过程，我们将在深度网络中获得的知识转换为更可解释的格式，包括显式的数学公式。这项研究的结果将提供对GW-P-Q关系的全面理解，其中出现了区域模式和物理控制。除了获得新的知识外，一个重要的副产品是经过训练的DL模型。它们可以作为洪水预报工具，将最近的土壤水分和地下水观测整合在一起，这些观测到现在还没有得到开发。教育活动将通过让本科生处理、可视化和解释大型水文数据来配合研究活动。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

In many parts of the United States, it has been shown that groundwater levels and soil moisture, which quantifies the wetness of the soil, are connected via the mechanism of flood production. Water cannot infiltrate into the ground when groundwater is close to the surface and is thus forced to quickly run off to rivers, creating higher flooding risks. However, the relationship between groundwater and floods has been found to be highly diverse and difficult to predict. Depending on terrain, groundwater depth, and many other factors, floods lead groundwater increase in some cases while groundwater can lead floods in others. Previous research from selected experimental watersheds have not resulted in a comprehensive and transferable understanding of the controlling processes. This project will take a big-data, machine learning approach to enhance our understanding of this relationship, allowing us to heuristically exploit previously under-utilized groundwater data for flood predictions and reducing damages. Using learning patterns from national-scale groundwater and streamflow data, the machine learning algorithms will create plausible groundwater-flood relationships. Taking advantage of the big hydrologic data from available satellite missions, this project will create shared undergraduate course modules to enhance student's ability to work with big data and increase their awareness of global water issues.This research advances hydrologic science by answering the following overarching question: at catchment scales, do groundwater levels in the catchment provide predictive power for flood threshold functions and baseflow? We will address this question in multiple small steps. We will identify the kinds of groundwater-rainfall-runoff (GW-P-Q) relations that can be found over the Continental United States. These relations are quantified by the correlations between water table depths and flood thresholds (and baseflow) at different lags and time scales. We will seek the factors dictate the type of GW-P-Q relations and whether these relations are stable across seasons and years. We will employ two approaches: a human-directed classification analysis, and a knowledge distillation scheme based on deep learning (DL), a rapidly advancing group of techniques supporting the recent surge in artificial intelligence. In the first approach, we will use classification and regression tree to identify factors that could explain the GW-P-Q relations. In the DL-based approach, we will train continental-scale time series DL models using all available data to forecast discharge. This approach addresses the issue with classification trees in which not enough data are available for branch nodes. Through a novel knowledge distillation procedure, we transfer the knowledge gained in the deep network to more interpretable formats, including explicit mathematical formula. Results from the study will provide a comprehensive understanding of GW-P-Q relations where regional patterns and physical controls emerge. Besides gaining new knowledge, a significant by-product is the trained DL models. They can be used as a flood forecasting tool to integrate recent soil moisture and groundwater observations, which have not been exploited until now. The educational activity will mesh with the research activity by engaging undergraduate students in handling, visualizing and interpreting big hydrologic data.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.