A contribution to drought resilience in East Africa through groundwater pump monitoring informed by in-situ instrumentation, remote sensing and ensemble machine learning
A contribution to drought resilience in East Africa through groundwater pump monitoring informed by in-situ instrumentation, remote sensing and ensemble machine learning
复制标题
DOI：
10.1016/j.scitotenv.2021.146486
复制
发表时间：
2021-03-24
影响因子：
9.8
通讯作者：
Coyle, Jeremy
中科院分区：
文献类型：
--
作者：
Thomas, Evan;Wilson, Daniel;Coyle, Jeremy
非洲之角普遍存在的干旱继续威胁着数百万人获得安全和负担得起的水。为了改进对水泵功能的监测，已在肯尼亚和埃塞俄比亚干旱地区的480个电动地下水水泵上安装了遥测连接传感器，目的是改进对这些供水的监测和支助操作和维护。在本文中，我们描述了两个分类系统的开发和验证，旨在识别这些电动泵的功能和非功能，一个是专家通知的条件分类器，另一个是利用机器学习。鉴于地表水可用性与地表水泵使用之间已知的关系，分类器将原位传感器数据与降雨和地表水的遥感指标结合起来。我们的验证表明，专家分类器的整体泵状态灵敏度（真阳性率）为82%，机器学习器为84%。当泵被使用时，两个分类器都有100%的真阳性率性能。当不使用泵时，专家分类器的特异性（真阴性率）约为50%，机器学习器的特异性超过65%。如果将这些检测功能集成到维修服务中，如果提供预算资源和制度激励措施，该地区干旱期间泵的正常运行时间可能会从60%提高到近85%，从而将泵停机的相对风险降低40%。(c) 2021提交人。这是一篇基于CC by-nc-nd许可（http://creativecommons.org/licenses/by-nc-nd/4.0/）的开放获取文章。
The prevalence of drought in the Horn of Africa has continued to threaten access to safe and affordable water for millions of people. In order to improve monitoring of water pump functionality, telemetry-connected sensors have been installed on 480 electrical groundwater pumps in arid regions of Kenya and Ethiopia, designed to im -prove monitoring and support operation and maintenance of these water supplies. In this paper, we describe the development and validation of two classification systems designed to identify the functionality and non-functionality of these electrical pumps, one an expert-informed conditional classifier and the other leveraging machine learning. Given a known relationship between surface water availability and groundwater pump use, the classifiers combine in-situ sensor data with remote sensing indicators for rainfall and surface water. Our val-idation indicates a overall pump status sensitivity (true positive rate) of 82% for the expert classifier and 84% for the machine learner. When the pump is being used, both classifiers have a 100% true positive rate performance. When a pump is not being used, the specificity (true negative rate) is about 50% for the expert classifier and over 65% for the machine learner. If these detection capabilities were integrated into a repair service, the typical up -time of pumps during drought periods in this region could potentially, if budget resources and institutional incen-tives for pump repairs were provided, result ina drought-period uptime improvement from 60% to nearly of 85% -a 40% reduction in the relative risk of pump downtime. (c) 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).