水资源危机是社会不稳定和石油争端的主要推手。最近的中东纠纷和非洲萨赫勒问题表明水资源匮乏，尤其是对于底层人民来说，会导致人口迁徙，道德纠纷和文明冲突。因为缺乏足够的关于水资源的数据，尤其是地下水的数据，世界各国领导和水资源管理者无法对水资源分配做出有效的决策。尤其是非洲国家常年因为不知道地下有多少水且有多深而受苦。两亿非洲人民依靠压水井来维持每天对水的需求。据估算，在非洲国家有超过100万的压水井。我们的工作展示了如何利用压水井一类的抽水工具来测量地下水深度，无需其他昂贵的设备。通过在压水井上安装加速度计来测量用户使用压水井时的振动模式，地下水的深度可以通过机器学习技术被有效地估计出来。为了实施这次实验，我们在实验室环境下搭建了一个5米高的模拟器，所以井里水的深度可以被我们严格控制。在井的上方我们安装了一个压水器。当我们使用压水井的时候，水管振动的数据被加速度计记录下来。这个振动的数据之后被用来训练机器学习模型，用来预测水井里的水到压水器下面的深度。我们的模型可以精确地预测深度，误差在17.9cm的范围，也就是水井深度的3.8%。我们的研究证明了这种已有的设备，也就是压水井，可以用来在发展中国家预测地下水深度。这种系统很有意义，因为它能为那些被气候变化负面影响巨大的地区提供地下水的数据。
Water crises are playing a more prominent role in destabilizing societies and fueling conflict. Recent violence in the Middle East and the Sahel region of Africa exemplifies how inadequate access to water – particularly among society’s most vulnerable members – can lead to migration, ethnic conflict, and civil strife. And without enough data about water resources – specifically groundwater resources – world leaders and water resource managers are not well equipped to make informed decisions about how to responsibly manage the resource. African nations in particular suffer from a lack of information about how much water lies below the Earth’s surface as well as how deep it is. Yet 200 million Africans rely on hand-operated water pumps to extract groundwater for their daily domestic needs. It is estimated that there are over one million of these handpumps on the African continent. This thesis demonstrates that existing groundwater-extracting infrastructure like handpumps can be employed as useful sites to measure groundwater levels without building new, costly infrastructure. By attaching an accelerometer to the handle of a hand-operated water pump and collecting the vibration patterns when a user pumps the handpump, the depth-to-the-water-level below the handpump can be reasonable estimated using machine learning techniques. To perform this study, a five-meter-tall simulation well was constructed in a laboratory setting such that the water level in the well could be strictly controlled. A handpump was attached to the top of the well. When water was pumped from the well, vibration data was collected using an accelerometer. This vibration data was then used to train a machine learning model to predict the distance from the handpump down to the top of the water level in the well. This model could accurately predict the depth to the water level within a median absolute error of 17.9 cm, or 3.8% of the entire well depth. This research suggests that existing rural infrastructure, namely handpumps, can be used to estimate groundwater depth in much of the developing world. Such a system is particularly beneficial because it addresses the lack of groundwater data in regions of the world that are most vulnerable to the adverse effects related to climate change.
