FloodNet: Low-Cost Ultrasonic Sensors for Real-Time Measurement of Hyperlocal, Street-Level Floods in New York City
FloodNet：用于实时测量纽约市街道级超本地洪水的低成本超声波传感器
首次出版：2024 年 5 月 7 日 https://doi.org/10.1029/2023WR036806
Abstract 摘要
Flooding is one of the most dangerous and costly natural hazards, and has a large impact on infrastructure, mobility, public health, and safety. Despite the disruptive impacts of flooding and predictions of increased flooding due to climate change, municipalities have little quantitative data available on the occurrence, frequency, or extent of urban floods. To address this, we have been designing, building, and deploying low-cost, ultrasonic sensors to systematically collect data on the presence, depth, and duration of street-level floods in New York City (NYC), through a project called FloodNet. FloodNet is a partnership between academic researchers and NYC municipal agencies, working in consultation with residents and community organizations. FloodNet sensors are designed to be compact, rugged, low-cost, and deployed in a manner that is independent of existing power and network infrastructure. These requirements were implemented to allow deployment of a hyperlocal, city-wide sensor network, given that urban floods often occur in a distributed manner due to local variations in land development, population density, sewer design, and topology. Thus far, 87 FloodNet sensors have been installed across the five boroughs of NYC. These sensors have recorded flood events caused by high tides, stormwater runoff, storm surge, and extreme precipitation events, illustrating the feasibility of collecting data that can be used by multiple stakeholders for flood resiliency planning and emergency response.
洪水是最危险、最昂贵的自然灾害之一，对基础设施、流动性、公共卫生和安全都有很大影响。尽管洪水具有破坏性影响，而且预测气候变化会导致洪水加剧，但市政当局几乎没有关于城市洪水发生率、频率或范围的定量数据。为了解决这个问题，我们一直在设计、建造和部署低成本的超声波传感器，通过一个名为 FloodNet 的项目，系统地收集纽约市（NYC）街道洪水的存在、深度和持续时间的数据。FloodNet 由学术研究人员和纽约市市政机构合作开展，并与居民和社区组织进行协商。FloodNet 传感器结构紧凑、坚固耐用、成本低廉，其部署方式不受现有电力和网络基础设施的影响。考虑到城市洪水通常因当地土地开发、人口密度、下水道设计和拓扑结构的不同而以分布式的方式发生，这些要求的实施允许部署一个超本地、全市范围的传感器网络。到目前为止，纽约市的五个行政区已经安装了 87 个 FloodNet 传感器。这些传感器记录了由涨潮、雨水径流、风暴潮和极端降水事件引起的洪水事件，说明了收集数据的可行性，这些数据可供多个利益相关方用于洪水恢复规划和应急响应。
Key Points 要点
-
Low-cost, ultrasonic sensors were designed and built to monitor the profiles of hyperlocal, street-level floods
设计并制造了低成本超声波传感器，用于监测超本地街道级洪水的概况 -
Sensor hardware, network architecture, and data ingestion, processing, and visualization tools were designed to maximize data usability
传感器硬件、网络架构以及数据摄取、处理和可视化工具的设计旨在最大限度地提高数据的可用性 -
The FloodNet project is installing flood sensors across New York City to collect data for community, city agency, and research stakeholders
FloodNet 项目正在纽约市各地安装洪水传感器，为社区、城市机构和研究利益相关者收集数据。
1 Introduction 1 引言
Of the many hazards that are expected to increase with climate change, flooding is one of the most dangerous and costly (National Academies of Sciences, Engineering, and Medicine, 2019), and can have a large influence on public health, infrastructure, and mobility in urban areas (Galloway et al., 2018). For example, climate change is projected to increase the frequency and magnitude of precipitation events (Lenderink & Van Meijgaard, 2010; Sillmann et al., 2013), which can overwhelm urban rivers, streams, and drainage systems, resulting in pluvial or fluvial flooding (National Academies of Sciences, Engineering, and Medicine, 2019). Additionally, sea level rise associated with climate change has been observed to increase coastal flooding during high tides. In New York City (NYC), for example, there is an increasing trend in the number of days with cumulative rainfall over 1.75 inches since 1950 (Depietri & McPhearson, 2018), and in community-reported evidence of coastal flooding due to high tides (Science and Resilience Institute at Jamaica Bay, 2023). While flooding from extreme events - such as hurricanes, typhoons, and cyclones - can cause catastrophic damage, it has been estimated that floods from smaller, yet more frequent, storms and high tide events have a large cumulative impact as well (Moftakhari et al., 2017; National Academies of Sciences, Engineering, and Medicine, 2019).
在许多预计会随着气候变化而增加的灾害中，洪水是最危险和代价最高的灾害之一（美国国家科学、工程和医学研究院，2019 年），而且会对城市地区的公共卫生、基础设施和流动性产生巨大影响（Galloway 等人，2018 年）。例如，预计气候变化将增加降水事件的频率和规模（Lenderink 和 Van Meijgaard，2010 年；Sillmann 等人，2013 年），这可能会淹没城市河流、溪流和排水系统，导致冲积或河道洪水（美国国家科学、工程和医学院，2019 年）。此外，据观察，与气候变化相关的海平面上升会加剧涨潮时的沿海洪水。例如，在纽约市（NYC），自 1950 年以来，累计降雨量超过 1.75 英寸的天数呈上升趋势（Depietri & McPhearson，2018 年），社区报告的涨潮导致沿海洪水的证据也呈上升趋势（牙买加湾科学与恢复力研究所，2023 年）。虽然极端事件（如飓风、台风和气旋）造成的洪水会造成灾难性的破坏，但据估计，规模较小但更为频繁的风暴和涨潮事件造成的洪水也会产生巨大的累积影响（Moftakhari 等人，2017 年；美国国家科学、工程和医学院，2019 年）。
Urban flooding is affected by a number of local factors including upstream land use and development; the size of vegetated and impervious areas; population density and human contributions to the sewer system; drainage and sewer design; and physical characteristics, including elevation and watershed topography. The combination of these factors make flood events highly specific to a city block, street corner, or other subsection of a neighborhood (New York City Stormwater Resiliency Plan, 2021). The highly localized and ephemeral nature of urban flooding make it challenging to measure, as monitoring needs to happen at a hyperlocal resolution and in real-time. As such, very little data exist on the frequency and extent of urban surface flooding (Galloway et al., 2018), and there is an unmet need for hyperlocal information on the presence, depth, and duration of street-level floods.
城市洪水受到许多当地因素的影响，包括上游土地利用和开发、植被区和不透水区的面积、人口密度和人类对下水道系统的贡献、排水和下水道设计以及物理特征，包括海拔高度和流域地形。这些因素的综合作用使得洪水事件对一个街区、街角或社区的其他分区具有高度的特殊性（纽约市雨水恢复计划，2021 年）。城市洪水的高度局部性和短暂性使其测量具有挑战性，因为需要以超局部分辨率进行实时监测。因此，有关城市地表洪水的频率和范围的数据非常少（Galloway 等人，2018 年），而有关街道级洪水的存在、深度和持续时间的超本地信息需求尚未得到满足。
There is a range of stakeholders that require real-time, accurate, and reliable data on flood events, including flood frequency, depth, and profile in flood-prone locations (Silverman et al., 2022). In discussions with city agencies, community members, and researchers, we have identified needs that include: data for the development and validation of hydrological models that predict flooding (city agencies and researchers; e.g., pluvial flood models produced as part of the New York City Stormwater Resiliency Plan (2021)); flood data for resilience and transportation planning, emergency response, issuing flash flood warnings, and tracking flood frequency over time (city agencies); data that can be used for day-to-day decision making and longer-term advocacy when faced with living with flood water (community members); and data that can signal when to collect samples when evaluating flood water quality (researchers) (Silverman et al., 2022). While some data sources exist, such as crowd-sourced flood data and US Geological Survey (USGS) stream gauges, each has limitations that prevent them from meeting all these needs (Helmrich et al., 2021).
许多利益相关者都需要有关洪水事件的实时、准确和可靠的数据，包括洪水频率、深度和洪水易发地点的剖面（Silverman 等人，2022 年）。在与城市机构、社区成员和研究人员的讨论中，我们发现他们的需求包括：用于开发和验证预测洪水的水文模型的数据（城市机构和研究人员；例如，作为水文模型的一部分而制作的冲积洪水模型）、作为纽约市暴雨恢复计划（2021 年）的一部分而制作的冲积洪水模型）；用于恢复和交通规划、应急响应、发布山洪暴发预警以及跟踪洪水频率的洪水数据（城市机构）；可用于日常决策以及在面临洪水时进行长期宣传的数据（社区成员）；以及可在评估洪水水质时提示何时收集样本的数据（研究人员）（Silverman 等人，2022 年）。虽然存在一些数据源，如众包洪水数据和美国地质调查局 (USGS) 的溪流测量仪，但每种数据源都有其局限性，无法满足所有这些需求（Helmrich 等人，2021 年）。
Crowd-sourced data, including social media posts and reports from community members or community scientists–such as those made to the Community Flood Watch Project in NYC (Science and Resilience Institute at Jamaica Bay, 2023) and NYC’s 311 service request line (NYC 311, 2023)–are limited in that they require that someone observes the event and that the observer creates a report (Helmrich et al., 2021; Smith & Rodriguez, 2017). Additional challenges with crowd-sourced data are that the data are not always correctly or precisely geotagged, reports rarely include flood depth readings over the duration of an event, and observers are discouraged from collecting data during hazardous conditions that might be present during high tides or extreme rain events. Existing water level sensors, such as USGS stream gauges, overcome some of the challenges above by collecting continuous, real-time data on water depths in the locations in which they are deployed. However, many of these sensors are deployed to monitor water bodies as opposed to flooding at the street-level, and are too expensive and bulky for large scale implementation in urban areas, which precludes their ability to be deployed in a network to capture hyperlocal flooding across a city scale.
众包数据，包括社区成员或社区科学家在社交媒体上发布的帖子和报告，例如向纽约市社区洪水观察项目（牙买加湾科学与恢复力研究所，2023 年）和纽约市 311 服务请求热线（纽约市 311，2023 年）提供的数据，其局限性在于需要有人观察事件并由观察者创建报告（Helmrich 等人，2021 年；Smith & Rodriguez，2017 年）。众包数据面临的其他挑战还包括：数据并不总是正确或精确地标注在地理标签上；报告很少包括事件持续时间内的洪水深度读数；不鼓励观察者在涨潮或极端降雨事件中可能出现的危险情况下收集数据。现有的水位传感器，如美国地质调查局（USGS）的溪流测量仪，通过收集所部署地点的连续、实时水深数据，克服了上述一些挑战。但是，这些传感器中的许多都是用于监测水体，而不是街道层面的洪水，而且过于昂贵和笨重，不适合在城市地区大规模使用，这就排除了将其部署到网络中以捕捉城市范围内超本地洪水的能力。
Therefore, our aim was to develop real-time flood sensors that overcome the limitations described above. In particular, the goals of this work were to design a sensor that is low-cost (allowing deployment of a large sensor network and increasing accessibility for community science), accurate (enabling detection of flood depths as low as 25 mm), flexible for multiple use cases and installation scenarios (not requiring power or connectivity infrastructure), robust to withstand long-term deployment in the urban environment, and easy to construct with an open-source design. Community engagement and collaboration with NYC municipal agencies are also core components of the FloodNet project and have been instrumental in assisting project implementation, including selection of sensor installation locations, development of tools for meaningful data access and visualization, and collation of growing knowledge about the experience and impacts of flooding within communities most at risk.
因此，我们的目标是开发能够克服上述限制的实时洪水传感器。特别是，这项工作的目标是设计一种低成本（允许部署大型传感器网络并提高社区科学的可及性）、精确（可检测低至 25 毫米的洪水深度）、灵活（不需要电源或连接基础设施）、坚固耐用（可在城市环境中长期部署）以及易于构建的开源设计的传感器。社区参与以及与纽约市市政机构的合作也是 FloodNet 项目的核心组成部分，在协助项目实施方面发挥了重要作用，包括传感器安装位置的选择、有意义的数据访问和可视化工具的开发，以及对高危社区内洪水体验和影响的不断增长的知识进行整理。
In this paper, we describe technical aspects of our project, FloodNet, including the design of FloodNet’s ultrasonic-based sensor hardware, network architecture, data ingestion and analysis pipeline, and data visualization tools.
在本文中，我们将介绍我们的项目 FloodNet 的技术方面，包括 FloodNet 基于超声波的传感器硬件设计、网络架构、数据摄取和分析管道以及数据可视化工具。
2 Prior Research on Flood Sensing
2 有关洪水传感的先前研究
Previous research has been conducted to design water level sensors using various sensing modalities, each of which has opportunities and limitations for measuring flood depths. Moreover, many previous flood sensing efforts have focused on monitoring water bodies, and have not deployed sensors in an urban street-level setting where the street is dry most of the time. Here we discuss the strengths and weaknesses of prior flood monitoring strategies, which have provided motivation and justification for the FloodNet sensor design and hardware choices.
以往的研究使用各种传感模式设计水位传感器，每种模式都有测量洪水深度的机会和局限性。此外，以前的许多洪水传感工作都侧重于监测水体，而没有在大部分时间街道都是干燥的城市街道环境中部署传感器。在此，我们将讨论先前洪水监测策略的优缺点，这些优缺点为 FloodNet 传感器的设计和硬件选择提供了动力和理由。
Generally, water level sensors can be divided into contact and non-contact designs (Kang et al., 2021). Contact sensors include bubble or float gauges (Kang et al., 2021), pressure sensors (Garcia et al., 2015), and capacitive or conductivity-based sensors (Chetpattananondh et al., 2014). Pressure sensors placed within drainage infrastructure, for example, have been used to detect street-level flooding through monitoring the surcharging of sewers (Gold et al., 2023). Contact sensors, however, have a number of limitations that make them ill-suited for wide-spread and long-term sensing of street-level floods. For one, contact sensor readings are potentially influenced by the composition, temperature, and turbidity of the water being monitored (Chetpattananondh et al., 2014). Moreover, given that these sensors must contact floodwaters that often have poor water quality, they are subject to fouling and require frequent maintenance, which are challenges for long-term installation. Additionally, the logistics of deploying contact sensors on the sidewalk or street in an urban environment in a manner that allows water contact are challenging and can make sensors susceptible to damage or vandalism.
一般来说，水位传感器可分为接触式和非接触式设计（Kang 等人，2021 年）。接触式传感器包括气泡或浮子测量仪（Kang 等人，2021 年）、压力传感器（Garcia 等人，2015 年）以及基于电容或电导率的传感器（Chetpattananondh 等人，2014 年）。例如，放置在排水基础设施内的压力传感器已被用于通过监测下水道的淤积情况来探测街道洪水（Gold 等人，2023 年）。然而，接触式传感器有许多局限性，不适合大范围、长期地感知街道洪水。首先，接触式传感器的读数可能会受到被监测水体的成分、温度和浊度的影响（Chetpattananondh 等人，2014 年）。此外，由于这些传感器必须接触通常水质较差的洪水，因此容易结垢，需要经常维护，这对长期安装是个挑战。此外，在城市环境中的人行道或街道上以允许与水接触的方式部署接触式传感器的后勤工作也具有挑战性，可能会使传感器容易受到损坏或破坏。
Much recent development of water level sensor technologies has focused on non-contact sensing modalities such as camera (Filonenko et al., 2015; Hiroi & Kawaguchi, 2016; Lo et al., 2015), ultrasonic (Loftis et al., 2018; Mousa et al., 2016), and LiDAR-based sensors (Loftis et al., 2018; Paul et al., 2020), given their relative low cost, ease of installation, compatibility with wireless connectivity, and lower maintenance requirements when compared with sensors that contact flood waters.
水位传感器技术的最新发展主要集中在非接触式传感方式上，如摄像头（Filonenko 等人，2015 年；Hiroi & Kawaguchi，2016 年；Lo 等人，2015 年）、超声波（Loftis 等人，2018 年；Mousa 等人，2016 年）和基于激光雷达的传感器（Loftis 等人，2018 年；Paul 等人，2020 年），因为与接触洪水的传感器相比，这些传感器成本相对较低、易于安装、与无线连接兼容且维护要求较低。
Visual sensing techniques that utilize camera-based technologies (such as preexisting video surveillance (CCTVs) or traffic cameras) paired with image analysis algorithms and computer vision have been used to generate real-time flood data and predictive flood alerts (Arshad et al., 2019; Filonenko et al., 2015; Hiroi & Kawaguchi, 2016; Jafari et al., 2021; Jan et al., 2022; Lo et al., 2015; Moy de Vitry et al., 2019; Sabbatini et al., 2021). Camera-based sensors can be mounted to existing infrastructure and provide quantitative complexity unmatched by one-dimensional data capture, but they come with a host of challenges. For one, image analysis and computer vision approaches can struggle with low-quality images, and cameras affected by heavy rain, fog, or glare may not be able to provide useful data. These sensing systems also have limited success in processing imagery in low light, limiting their nighttime effectiveness (Lo et al., 2015). Data from camera-based sensors may not include flood depth measurements, and existing cameras are not always positioned in locations that are most at risk for flooding (Helmrich et al., 2021). Moreover, these technologies use a large amount of power and are highly reliant on availability of existing power and communication infrastructure, and are therefore less independent and flexible than low-powered counterparts described below (Lo et al., 2015). This also renders them vulnerable to extreme weather events, when they may be needed the most. Finally, the use of cameras raises ethical concerns relating to privacy, surveillance and consent. Collected images would need to be anonymized, risks associated with unintended uses of imagery would need careful consideration, and stringent data security protocols would be required.
利用基于摄像头的技术（如已有的视频监控（CCTV）或交通摄像头）与图像分析算法和计算机视觉相搭配的视觉传感技术已被用于生成实时洪水数据和预测性洪水警报（Arshad 等人，2019 年；Filonenko 等人，2015 年；Hiroi & Kawaguchi，2016 年；Jafari 等人，2021 年；Jan 等人，2022 年；Lo 等人，2015 年；M、2019；Filonenko 等人，2015；Hiroi & Kawaguchi，2016；Jafari 等人，2021；Jan 等人，2022；Lo 等人，2015；Moy de Vitry 等人，2019；Sabbatini 等人，2021）。基于摄像头的传感器可以安装在现有的基础设施上，并提供单维数据采集所无法比拟的定量复杂性，但它们也带来了一系列挑战。首先，图像分析和计算机视觉方法在处理低质量图像时可能会遇到困难，受大雨、大雾或强光影响的相机可能无法提供有用的数据。这些传感系统在弱光下处理图像的成功率也很有限，从而限制了它们在夜间的有效性（Lo 等人，2015 年）。基于相机的传感器提供的数据可能不包括洪水深度测量值，而且现有的相机并不总是安装在洪水风险最大的位置（Helmrich 等人，2021 年）。此外，这些技术使用大量电力，高度依赖现有的电力和通信基础设施，因此其独立性和灵活性不如下文所述的低功耗技术（Lo 等人，2015 年）。这也使它们在最需要它们的时候容易受到极端天气事件的影响。最后，使用摄像机会引发与隐私、监视和同意有关的伦理问题。收集到的图像需要进行匿名处理，与图像的意外使用相关的风险需要仔细考虑，并且需要严格的数据安全协议。
Another camera-based approach to flood monitoring is the use of satellite imagery to track flood water inundation. The unparalleled field of view and global coverage offered make satellite-based sensing particularly suitable for large-geographical-scale flood event tracking (Li et al., 2018). However, its shortcomings - including the high cost of long term image acquisition, low temporal and spatial resolution, difficulty in gathering water depth data, and, importantly for pluvial flood events, its inability to obtain imagery through cloud cover (Olthof & Svacina, 2020) - make it impractical for urban flood monitoring.
另一种基于相机的洪水监测方法是利用卫星图像跟踪洪水淹没情况。无与伦比的视场和全球覆盖率使卫星传感特别适合大地理范围的洪水事件跟踪（Li 等人，2018 年）。然而，卫星图像的缺点--包括长期图像采集成本高、时间和空间分辨率低、难以收集水深数据，以及对于冲积洪水事件而言，重要的是无法通过云层获取图像（Olthof & Svacina，2020 年）--使其无法用于城市洪水监测。
The increasing affordability of ultrasonic and LiDAR time-of-flight sensors has yielded other novel flood monitoring approaches. Each of these sensor modalities detects water levels by emitting a pulsed sound or light wave and measuring the round trip travel time of the pulse after it reflects from a surface and returns to the sensor; this travel time can be converted to distance based on the speed of the ultrasonic or LiDAR pulse.
由于超声波和激光雷达飞行时间传感器的价格越来越低廉，因此出现了其他新型洪水监测方法。每种传感器模式都通过发射脉冲声波或光波来探测水位，并测量脉冲从表面反射后返回传感器的往返时间；根据超声波或激光雷达脉冲的速度，可将该往返时间转换为距离。
LiDAR has become a popular tool to measure the physical world and is used for oceanography, archeology, topographical modeling, and urban planning. Paul et al. recently used a LiDAR-based flood sensor to measure water levels with relative error around 0.1% and at angles as small as 40° to the planar surface (Paul et al., 2020). However, while LiDAR is becoming more affordable, it is still too expensive to deploy a large number of sensors for spatial coverage at the city-scale, and it can be affected by both the relative calmness and opacity of the water being measured (Paul et al., 2020). Additionally, LiDAR is best used to measure complex surfaces and is difficult to scale in a wireless environment, given the amount of raw data it produces.
激光雷达已成为测量物理世界的常用工具，被用于海洋学、考古学、地形建模和城市规划。Paul 等人最近使用基于激光雷达的洪水传感器测量了水位，相对误差约为 0.1%，与平面的角度小至 40°（Paul 等人，2020 年）。不过，虽然激光雷达的价格越来越实惠，但要在城市范围内部署大量传感器进行空间覆盖，成本仍然太高，而且激光雷达还会受到被测水域相对平静和不透明度的影响（Paul 等人，2020 年）。此外，激光雷达最适合用于测量复杂的表面，而且由于其产生的原始数据量大，很难在无线环境中进行扩展。
Alternatively, several flood monitoring projects have had success in using ultrasonic distance sensors given their low cost, low power consumption, and opportunities for flexible deployment due to their ability to use solar power and transmit data over wireless networks (Bartos et al., 2018; Kang et al., 2021; Loftis et al., 2018; Mousa et al., 2016). Examples of flood monitoring projects that utilize ultrasonic sensors include StormSense, an IoT water-level monitoring system in Hampton Roads, Virginia that employs wireless ultrasonic sensors that are deployed over water bodies and connected to a Long Range Wide Area Network for data transmission (Loftis et al., 2018), Open Storm with installations in Ann Arbor, Michigan and Dallas-Fort Worth, Texas (Bartos et al., 2018), and an ultrasonic sensor project carried out by a research team from King Abdullah University of Science and Technology in Saudi Arabia that utilized ultrasonic sensors installed over street traffic, to monitor and model flood events and traffic flow (Mousa et al., 2016).
另外，一些洪水监测项目也成功使用了超声波距离传感器，因为它们成本低、功耗低，而且能够使用太阳能并通过无线网络传输数据，因此可以灵活部署（Bartos 等人，2018 年；Kang 等人，2021 年；Loftis 等人，2018 年；Mousa 等人，2016 年）。利用超声波传感器的洪水监测项目包括 StormSense，这是弗吉尼亚州汉普顿路的一个物联网水位监测系统，该系统采用无线超声波传感器，部署在水体上方，并连接到远程广域网进行数据传输（Loftis et al、2018）、在密歇根州安阿伯市和德克萨斯州达拉斯-沃斯堡市安装的 Open Storm（Bartos 等人，2018），以及沙特阿拉伯阿卜杜拉国王科技大学研究团队开展的超声波传感器项目，该项目利用安装在街道交通上空的超声波传感器，对洪水事件和交通流量进行监测和建模（Mousa 等人，2016）。
Due to the opportunities presented by ultrasonic-based sensors, we employed ultrasonic range finders as the basis for the FloodNet sensor design described below. While prior studies have used ultrasonic sensor networks for riverine (Bartos et al., 2018) and urban flood monitoring (Hiroi & Kawaguchi, 2016; Kang et al., 2021; Lo et al., 2015; Loftis et al., 2018; Mousa et al., 2016), there are limited examples of such systems being deployed at street-level on a city-wide scale, let alone in a metropolitan area as dense and granularly complex as New York City. FloodNet, as a result, offers a novel approach to developing a scalable, reliable, and informative flood-sensing network to aid in urban resilience efforts.
基于超声波传感器带来的机遇，我们采用超声波测距仪作为下文所述 FloodNet 传感器设计的基础。虽然之前的研究已将超声波传感器网络用于河道（Bartos 等人，2018 年）和城市洪水监测（Hiroi & Kawaguchi，2016 年；Kang 等人，2021 年；Lo 等人，2015 年；Loftis 等人，2018 年；Mousa 等人，2016 年），但在全市范围内街道级部署此类系统的例子非常有限，更不用说像纽约市这样密集且粒度复杂的大都市地区了。因此，FloodNet 为开发可扩展、可靠且信息丰富的洪水传感网络提供了一种新方法，以协助城市抗灾工作。
3 FloodNet Sensor Network
3 FloodNet 传感器网络
The FloodNet project is a collaboration between academic researchers at New York University and the City University of New York with NYC municipal agencies: NYC Mayor’s Office of Climate and Environmental Justice, NYC Office of Technology and Innovation, and NYC Department of Environmental Protection. FloodNet has been focused on (a) the design, construction, and deployment of sensors to record street-level floods in NYC; (b) the development of data analysis tools to accurately process flood data; and (c) community engagement, data sharing, and the communication of flood data to various stakeholders in meaningful ways. The FloodNet sensors (Figure 1) are novel in that they were designed to be compact, rugged, low-cost, and independent of existing urban power and network infrastructure, allowing the project to go beyond the limitations of some previously developed ultrasonic flood sensing systems. These specifications make the deployment of a hyperlocal, city-wide sensor network possible.
FloodNet 项目是纽约大学和纽约市立大学的学术研究人员与纽约市市政机构的合作项目：纽约市市长气候与环境正义办公室、纽约市技术与创新办公室和纽约市环境保护部。FloodNet 的工作重点是：(a) 设计、建造和部署传感器，以记录纽约市的街道洪水；(b) 开发数据分析工具，以准确处理洪水数据；(c) 社区参与、数据共享，以及以有意义的方式向各利益相关者传播洪水数据。FloodNet 传感器（图 1）设计新颖，结构紧凑、坚固耐用、成本低廉，并且独立于现有的城市电力和网络基础设施，使该项目能够超越之前开发的一些超声波洪水传感系统的局限性。这些技术指标使得部署超本地、全城范围的传感器网络成为可能。
(a) Closeup of the FloodNet sensor, showing the (i) ultrasonic sensor cone, (ii) sensor housing, (iii) solar panel for battery charging, and (iv) antenna for data transmission. (b) Signage that is installed with each sensor. (c) FloodNet engineers installing a sensor in the field. (d) FloodNet sensor and sign installed on a U-channel pole in the Bronx, NY.
(a) FloodNet 传感器特写，显示 (i) 超声波传感器锥体、(ii) 传感器外壳、(iii) 电池充电太阳能电池板和 (iv) 数据传输天线。(b) 与每个传感器一起安装的标识牌。(c) FloodNet 工程师在现场安装传感器。(d) 安装在纽约布朗克斯区 U 型槽电线杆上的 FloodNet 传感器和标志。
更具体地说，FloodNet 传感器网络的设计符合以下标准：
-
Sense water depth within an accuracy of <±25 mm
感知水深的精度在 <±25 毫米范围内 -
Collect and transmit data to a central server every ≈1 min
每 ≈1 分钟收集并向中央服务器发送一次数据 -
Operate autonomously in the environment for long periods of time
在环境中长时间自主操作 -
Operate independent of existing power and networking infrastructure
独立于现有电力和网络基础设施运行 -
Comprise low-cost components for sensor network scalability
由低成本组件组成，实现传感器网络的可扩展性
In addition to sensor hardware, the FloodNet system includes a data ingestion and analysis pipeline, user-facing dashboards that display sensor readings in real-time, and a tool that can send automated flood alerts to subscribing government agencies, community members, or other stakeholders when a flood is detected. All information on the sensor design, including build instructions, quality control practices, and data analysis pipelines, is open-source and provided in a GitHub repository (FloodNet, 2024b) (https://github.com/floodnet-nyc/flood-sensor). While FloodNet is currently focused on NYC, FloodNet sensor designs could be utilized in other locations.
除传感器硬件外，FloodNet 系统还包括数据摄取和分析管道、可实时显示传感器读数的面向用户的仪表板，以及可在检测到洪水时向订阅的政府机构、社区成员或其他利益相关者自动发送洪水警报的工具。传感器设计的所有信息，包括构建说明、质量控制实践和数据分析管道，都是开源的，并在 GitHub 存储库中提供（FloodNet，2024b）（https://github.com/floodnet-nyc/flood-sensor）。虽然 FloodNet 目前主要集中在纽约市，但 FloodNet 传感器设计也可用于其他地区。
The following sections explain how FloodNet sensors meet the design criteria listed above, with technical details on the sensor design, operation, deployment scenarios, and networking infrastructure.
下文将介绍 FloodNet 传感器如何满足上述设计标准，并提供有关传感器设计、运行、部署方案和网络基础设施的技术细节。
3.1 Sensor Core Components
3.1 传感器核心部件
FloodNet sensors (Figure 1) use the Maxbotix MB7389 ultrasonic range-finder to measure the distance to the horizontal surface below the sensor, with a detection range between 30 and 500 cm, and 1 mm resolution. The sensor transmits 42 kHz ultrasound pulses, which it detects on return to the sensor after reflection from any large surface directly below. Sensor accuracy is a key design criterion for FloodNet, and the MB7389 sensor’s datasheet specifies the accuracy as being within 1% of the measured distance, which equates to a range of 30 mm (i.e., ±15 mm) with a typical mounting height of 3 m, satisfying design criterion 1; data illustrating this accuracy are described in Section 14.
FloodNet 传感器（图 1）使用 Maxbotix MB7389 超声波测距仪测量传感器下方水平表面的距离，探测范围在 30 厘米到 500 厘米之间，分辨率为 1 毫米。传感器发射 42 kHz 超声波脉冲，从正下方的任何大表面反射后返回传感器进行检测。传感器精度是 FloodNet 的一个关键设计标准，MB7389 传感器的数据表规定其精度在测量距离的 1% 以内，相当于 30 毫米的范围（即 ±15 毫米），典型安装高度为 3 米，满足设计标准 1；第 14 节中描述了说明该精度的数据。
Data are collected by the sensor and transmitted every 60 s, satisfying design criterion 2. Data transmission occurs through the use of Long Range Wide Area Network (LoRaWAN), which is a low power, wide area (LPWA) networking protocol designed to wirelessly connect battery operated devices to the internet. The microcontroller unit (MCU) platform at the core of the sensor is the Heltec HTCC-AB02 development board that uses the Heltec HTCC-AM02 as its MCU/LoRaWAN radio module, which integrates a Semtech SX1262 LoRaWAN radio and PSoC 4000 series ARM Cortex M0 MCU. The MCU is programmed to record range measurements and transmit them to a nearby internet connected gateway using LoRaWAN networking technology, discussed below.
传感器每 60 秒收集并传输一次数据，满足设计标准 2。数据传输采用长距离广域网（LoRaWAN），这是一种低功耗、广域（LPWA）网络协议，旨在将电池供电设备无线连接到互联网。传感器核心的微控制器（MCU）平台是 Heltec HTCC-AB02 开发板，它使用 Heltec HTCC-AM02 作为 MCU/LoRaWAN 无线电模块，集成了 Semtech SX1262 LoRaWAN 无线电和 PSoC 4000 系列 ARM Cortex M0 MCU。MCU 经编程可记录测距结果，并利用 LoRaWAN 网络技术将其传输到附近的互联网连接网关。
The power consumption of the sensor averages 2 mA at 3.3 V, peaking at ≈50 mA when collecting measurements, making it suitable for long-term battery-powered field operation. Between measurements, the MCU and radio are set to a low power sleep mode with the power to the ultrasonic sensor switched off using the in-built transistor switch of the Heltec development board; current consumption is ≈20 μA when in sleep mode. This allows for long-term operation with a modest battery capacity. The sensor uses a lithium ion polymer (LiPo) 400 mAh single cell battery. To satisfy design criterion 3, a UV resistant 0.6 W solar panel is integrated to charge the battery, allowing for indefinite operation. A custom solar panel mount was designed and 3D printed in resilient ABS plastic (Figure 1). The sensor’s power consumption is discussed in more detail in Section 7.
传感器在 3.3 V 电压下的平均功耗为 2 mA，收集测量数据时的峰值为 ≈50 mA，因此适合长期使用电池供电的现场操作。在测量间隙，微控制器和无线电设置为低功耗睡眠模式，并通过 Heltec 开发板的内置晶体管开关关闭超声波传感器的电源；睡眠模式下的电流消耗为 ≈20 μA。因此，只需适度的电池容量即可实现长期运行。传感器使用锂离子聚合物 (LiPo) 400 mAh 单芯电池。为满足设计标准 3，传感器集成了一个 0.6 W 抗紫外线太阳能电池板，用于为电池充电，从而实现无限期运行。我们设计了一个定制的太阳能电池板支架，并用弹性 ABS 塑料进行了 3D 打印（图 1）。第 7 节将详细讨论传感器的功耗。
For ease of assembly and to securely accommodate the development board and battery, a custom breakout printed circuit board (PCB) was designed and fabricated with headers for the development board to be mounted to, and screw terminals for the ultrasonic sensor and solar panel wiring.
为了便于组装，并安全地安装开发板和电池，我们设计并制作了一个定制的分离式印刷电路板（PCB），上面有用于安装开发板的接线头，以及用于超声波传感器和太阳能电池板接线的螺丝端子。
A common constraint in the deployment of sensor networks is a reliance on existing infrastructure such as power and networking. By designing a sensor network that makes use of ultralow power sensing, solar energy harvesting and wireless networking technologies, we have alleviated these requirements, thereby greatly expanding the potential deployment locations for these sensors and meeting design criterion 4. As street flooding occurs in myriad and distributed urban locations, this flexibility is critical for urban flood detection.
部署传感器网络的一个常见限制因素是对现有基础设施（如电源和网络）的依赖。通过设计一种利用超低功耗传感、太阳能收集和无线网络技术的传感器网络，我们降低了这些要求，从而大大扩展了这些传感器的潜在部署地点，并满足了设计标准 4。由于街道洪水发生在无数分散的城市地点，这种灵活性对于城市洪水探测至关重要。
The bill of materials for the sensor, excluding mounting hardware costs, is listed in Table 1. At the time of writing (September 2023), the total cost of the sensor parts was <$200 USD, allowing the sensors to meet design criterion 5. The build process for the sensor is included on the project's public Github page (FloodNet, 2024a) (https://floodnet-nyc.github.io).
表 1 列出了传感器的材料清单（不包括安装硬件成本）。在撰写本文时（2023 年 9 月），传感器部件的总成本小于 200 美元，因此传感器符合设计标准 5。传感器的构建过程包含在该项目的 Github 公共页面上（FloodNet, 2024a）( https://floodnet-nyc.github.io)。
表 1.一个 FloodNet 传感器单元的材料清单 (BOM)，包括编写时的成本
| Item | Cost (USD) 成本（美元） |
|---|---|
| Ultrasonic sensor 超声波传感器 | $99.95 |
| Mounting hardware 安装硬件 | $22.96 |
| Heltec AB02 development board Heltec AB02 开发板 |
$14.40 |
| Antenna | $12.85 |
| Breakout components 分线组件 | $11.44 |
| Solar panel (0.6 W) 太阳能电池板（0.6 瓦） | $9.00 |
| LiPo battery (400 mAh) 锂电池（400 毫安时） | $6.95 |
| Solar mount 太阳能支架 | $6.00 |
| Custom breakout board 定制突破板 | $0.50 |
| Cable glands 电缆接头 | $0.45 |
| Total | $184.50 |
3.2 Sensor Deployment 3.2 传感器部署
The FloodNet sensor network was designed to monitor and collect data from various urban flood scenarios, and is indifferent to the cause of the flooding or water level change. For example, the sensors can measure the profiles of floods caused by stormwater runoff, high tides, storm surge, water infrastructure failures, or compound events, and are able to measure water depths over ground-level or water bodies, as long as the sensor is placed above and at a 90° angle to the surface to be monitored. Nonetheless, sensor deployment locations have a few key requirements: (a) mounting infrastructure (e.g., sign post, pole, wall, overhang, etc.) located directly above the surface to be monitored; (b) a LoRaWAN gateway within range for ingestion of the transmitted data; and (c) sunlight exposure to recharge the sensor’s battery. This section describes how these requirements inform the sensor installation process. Note that we plan for new versions of the FloodNet sensor to transmit data over a cellular network (described in Section 15), which would eliminate the need for a nearby LoRaWAN gateway.
FloodNet 传感器网络的设计目的是监测和收集各种城市洪水情况下的数据，而与洪水或水位变化的原因无关。例如，传感器可以测量由雨水径流、涨潮、风暴潮、水利基础设施故障或复合事件引起的洪水概况，并且能够测量地面或水体的水深，只要传感器放置在要监测的地表上方并与地表成 90° 角即可。不过，传感器的部署位置有几个关键要求：(a) 位于要监测的水面正上方的安装基础设施（如标志杆、柱子、墙壁、悬挂物等）；(b) 范围内的 LoRaWAN 网关，以便接收传输的数据；以及 (c) 阳光照射，以便为传感器的电池充电。本节将介绍这些要求如何影响传感器的安装过程。请注意，我们计划让新版本的 FloodNet 传感器通过蜂窝网络传输数据（详见第 15 节），这样就不需要在附近安装 LoRaWAN 网关了。
U-channel posts, used to mount street signs, are some of the most abundant pieces of street hardware in NYC (Figure 1). The NYC Department of Transportation (DOT) granted FloodNet permission to mount sensors on these sign posts. The posts are typically >3 m tall, providing a mounting height out of reach of passersby, and include 5 cm spaced mounting holes running vertically up the U-channel. FloodNet sensors must be mounted such that the ultrasonic beam is transmitted perpendicular to the ground surface. This configuration is important because ultrasonic pulses transmitted at an angle greater than 5° from vertical may not be returned to the sensor following reflection.
用于安装街道标志的 U 型槽柱是纽约市最常见的街道硬件之一（图 1）。纽约市交通局 (DOT) 允许 FloodNet 在这些标志柱上安装传感器。这些标志柱通常>3 米高，提供了一个路人够不到的安装高度，并在 U 型槽上垂直开有 5 厘米间距的安装孔。FloodNet 传感器的安装必须使超声波束垂直于地表传播。这种配置非常重要，因为超声波脉冲的传输角度与垂直面的夹角大于 5°，在反射后可能无法返回传感器。
While plentiful, U-channel sign posts present some sensor installation limitations. First, sign posts are typically located on the sidewalk (i.e., not the roadway), and therefore may not be located directly above the lowest elevation in an area of interest (i.e., the location most likely to flood first). As such, in many locations, sensors are unable to measure flooding before the water depth exceeds the street curb height (usually ≈ 15 cm in NYC); this offset is noted in the metadata for each sensor, to aid the usability and interpretation of measured flood data. U-channel posts with vegetation beneath pose another challenge, as the shifting vegetation can produce variance in sensor measurements during non-flooded conditions. U-channel posts can also be susceptible to damage (i.e., being pushed off-vertical) or removal; these changes can be detected in real-time sensor readings and must be monitored as part of ongoing operation and maintenance. Despite the limitations of U-channel sign posts, their plenitude in NYC make them great candidates for sensor mounting. In situations where U-channel posts are unusable or unavailable, sensors can also be mounted on other types of street infrastructure, such as street lights or utility poles, using modified mounting hardware.
U 型槽标志桩虽然数量众多，但在传感器安装方面存在一些限制。首先，标志桩通常位于人行道上（即不是道路上），因此可能无法直接位于相关区域的最低标高之上（即最有可能首先淹没的位置）。因此，在许多地方，传感器无法在水深超过街道路缘石高度（纽约市通常≈ 15 厘米）之前测量洪水；每个传感器的元数据中都注明了这一偏移量，以帮助测量洪水数据的可用性和解释。下面有植被的 U 型槽支柱会带来另一个挑战，因为在非洪水条件下，移动的植被会导致传感器测量值出现偏差。U 型槽标柱还容易受到损坏（例如，被推离垂直面）或拆除；这些变化可在传感器的实时读数中检测到，必须作为持续运行和维护的一部分进行监控。尽管 U 型槽标志桩有其局限性，但其在纽约市的广泛应用使其成为安装传感器的理想选择。在 U 型槽标志柱无法使用或无法安装的情况下，也可以使用改进的安装硬件将传感器安装在其他类型的街道基础设施上，如路灯或电线杆。
All installed sensors are accompanied by signage that provides information about the FloodNet project, the type of data collected by the sensor and where to access it, and an explanation that no identifying information is collected by the sensor (Figure 1). Sensor maintenance occurs on a regular and ad hoc basis. Regular visits to visually inspect functioning sensors currently occur yearly; the frequency may change as we expand the network and learn more about specific maintenance needs for the hardware. The health of each sensor is monitored remotely using its battery charge, data packet upload success ratio, frequency of non-returned distance measurements, and ratio of noise to valid distance measurements as metrics. Poorly functioning sensors are visited on an ad hoc basis and replaced with new functioning sensors.
所有已安装的传感器都配有标牌，提供有关 FloodNet 项目的信息、传感器收集的数据类型和访问位置，并说明传感器不会收集任何身份识别信息（图 1）。传感器的维护既有定期的，也有临时的。目前，我们每年都会定期访问传感器，目测其运行情况；随着网络的扩展和对硬件具体维护需求的进一步了解，访问频率可能会有所改变。我们使用电池电量、数据包上传成功率、未返回距离测量值的频率以及噪声与有效距离测量值的比率等指标对每个传感器的健康状况进行远程监控。对功能不佳的传感器进行临时访问，并用新的功能正常的传感器进行替换。
3.3 Data Collection 3.3 数据收集
For each data point collected by the sensor, seven distance measurements are recorded at 150 ms intervals, based on the time recorded between sending and receiving the ultrasound pulse. A median of these seven measurements is recorded as the measured distance, as a means of internal filtering to exclude any erroneously small or large range measurements that could be reflections from smaller surfaces, such as the base of a light pole, and ensure that ranging is to the largest surface below the sensor, such as the street, sidewalk, or floodwater surface.
对于传感器收集的每个数据点，根据超声波脉冲发送和接收之间的时间记录，以 150 毫秒的间隔记录七个距离测量值。这七个测量值的中位数被记录为测量距离，作为内部过滤的一种手段，以排除可能来自较小表面（如灯杆底座）反射的任何错误的小范围或大范围测量值，并确保测距到传感器下方的最大表面，如街道、人行道或洪水表面。
The sensor also includes an internal temperature sensor that is used to correct distance measurements for the temperature-dependent speed of the ultrasound pulse through air (i.e., speed of sound increases with an increase in air temperature). Of note is that this temperature compensation can over-correct if, for example, the sensor is exposed to direct sunlight, causing it to absorb heat and measure warmer temperatures than ambient air. In this case, the sensor assumes that the ultrasonic pulse travels faster and for a longer distance than it actually does and will record an exaggerated distance to the ground (this phenomenon is further described in Section 13).
传感器还包括一个内部温度传感器，用于根据超声波脉冲穿过空气的速度（即声速随空气温度升高而增加）对距离测量进行修正。值得注意的是，如果传感器直接暴露在阳光下，导致其吸收热量并测量出比环境空气温度更高的温度，则温度补偿可能会过度校正。在这种情况下，传感器会假定超声波脉冲的传播速度和传播距离比实际传播速度和传播距离要长，从而记录下夸大的到地面的距离（第 13 节将进一步描述这种现象）。
After determining the median distance value, this measurement is transmitted to a nearby internet-connected gateway using LoRaWAN networking technology, described in the next section.
在确定距离中值后，该测量值将通过 LoRaWAN 网络技术传输到附近的互联网连接网关，详情将在下一节中介绍。
3.4 Networking and Data Transmission
3.4 网络和数据传输
Every sensor using LoRaWAN for networking must be within range of an internet-enabled gateway for data transmission and ingestion. FloodNet team members install these LoRaWAN gateways in flood prone areas to receive sensor data payloads; gateways can receive data from any sensors installed within a ≈2 km radius. The internet-connected gateways forward these small (≈28 byte) data payloads to our LoRaWAN network provider, The Things Network, which packages the data and delivers them via the MQ Telemetry Transport (MQTT) lightweight publish/subscribe messaging protocol to the project servers. The gateways themselves can be mounted inside or outside, but their 1 m antennas are mounted outside as high as possible for optimum coverage. Typically, gateways are mounted on the roofs of buildings on mounting points such as railings or existing vertical poles to ensure that the gateway and antenna are securely fastened.
每个使用 LoRaWAN 进行联网的传感器都必须在支持互联网的网关范围内，以便传输和接收数据。FloodNet 团队成员在洪水易发地区安装这些 LoRaWAN 网关，以接收传感器的数据有效载荷；网关可以接收安装在半径≈2 千米范围内的任何传感器的数据。与互联网连接的网关将这些小型（≈28 字节）数据有效载荷转发给我们的 LoRaWAN 网络提供商 The Things Network，后者将数据打包，并通过 MQ Telemetry Transport（MQTT）轻量级发布/订阅消息协议将数据传送到项目服务器。网关本身可以安装在室内或室外，但其 1 米长的天线尽可能安装在室外，以获得最佳覆盖范围。通常情况下，网关安装在建筑物屋顶的安装点上，如栏杆或现有的垂直杆上，以确保网关和天线固定牢靠。
Gateways require a continuous power source, such as an AC power outlet. The gateway typically used (i.e., MikroTik LtAP) measures 17 × 17 × 4 cm, weighs ≈ 500 g, and consumes around 5 W of power (similar to a small phone charger). The gateway also requires an internet connection to upload sensor payloads. A wired ethernet connection is preferred, but a gateway that incorporates a cellular modem and SIM card with low data rate plan can be used where ethernet connectivity is difficult to obtain. The data throughput of the device is minimal, at around 5 MB/day, which would not exceed a data plan bandwidth of 64 kbps. FloodNet has mounted gateways on the rooftops of local businesses, community organizations, academic institutions, public schools, and apartment buildings.
网关需要持续的电源，如交流电源插座。通常使用的网关（即 MikroTik LtAP）尺寸为 17 × 17 × 4 厘米，重量≈ 500 克，耗电量约为 5 瓦（类似于小型手机充电器）。网关还需要互联网连接才能上传传感器有效载荷。最好使用有线以太网连接，但在难以获得以太网连接的地方，也可以使用包含蜂窝调制解调器和低数据速率计划 SIM 卡的网关。该设备的数据吞吐量很小，约为 5 MB/天，不会超过 64 kbps 的数据计划带宽。FloodNet 已经在当地企业、社区组织、学术机构、公立学校和公寓楼的屋顶上安装了网关。
Each sensor transmits its distance measurement payload (i.e., the median of seven independent distance measurements, as described in Section 5) via LoRaWAN to the nearest gateway, which forwards it to The Things Network that then routes the data to our project servers. To allow for an extensible platform for sensor data ingestion, storage, analysis, visualization, and sharing, FloodNet uses a server setup that runs a set of open source tools. More specifically, the FloodNet data pipeline is composed of a combination of Docker containers including: (a) a load-balanced HTTP reverse proxy for the efficient routing of secure web traffic from a set of subdomains to each service running on the server; for example, https://dataviz.floodnet.nyc (Traefik); (b) a layer for data ingestion, routing, and processing (NodeRed); (c) a time-series database for data storage (InfluxDB); (d) a dashboard platform for data visualization (Grafana); and (e) a public facing dashboard built on the FieldKit platform at https://dataviz.floodnet.nyc. The data pipeline is illustrated in Figure 2.
每个传感器通过 LoRaWAN 将其距离测量有效载荷（即第 5 节所述的七个独立距离测量值的中位数）传输到最近的网关，再由网关转发到物联网，然后由物联网将数据路由到我们的项目服务器。为了使传感器数据摄取、存储、分析、可视化和共享平台具有可扩展性，FloodNet 使用了一套运行开源工具的服务器设置。更具体地说，FloodNet 数据管道由 Docker 容器组合而成，其中包括：(a) 负载平衡 HTTP 反向代理，用于将安全网络流量从一组子域高效路由到服务器上运行的每个服务；例如，https://dataviz.floodnet.nyc (Traefik)；(b) 数据摄取、路由和处理层 (NodeRed)；(c) 用于数据存储的时间序列数据库 (InfluxDB)；(d) 用于数据可视化的仪表盘平台 (Grafana)；(e) 基于 https://dataviz.floodnet.nyc 的 FieldKit 平台构建的面向公众的仪表盘。数据管道如图 2 所示。
FloodNet data pipeline including deployed sensors, gateway, backend services, and data dashboards.
FloodNet 数据管道，包括部署的传感器、网关、后端服务和数据仪表板。
Of note, and described further in Section 15, is that the logistics of operating a system requiring internet-connected, continuously powered gateways for data transmission are challenging, and has led our team to decide to transition to the use of a cellular network for data transmission in subsequent versions of the FloodNet sensor.
值得注意的是，我们的团队决定在后续版本的 FloodNet 传感器中过渡到使用蜂窝网络进行数据传输。
3.5 Sensor Power Consumption
3.5 传感器功耗
Figure 3 shows the sensor’s current consumption during all possible states of the sensor operating in LoRaWAN Class A, the default LoRaWAN class in which the end device (i.e., the sensor) always initiates the communication. The sensor’s data collection and transmission periodicity is ≈1 min; the total duration of each cycle varies by a few milliseconds due to the varying receive windows after a transmission, allowing the opportunity for bi-directional communication. In Figure 3, regions (a), (b), (d), and (f) are wake periods, and (c), (e), and (g) are sleep periods. The current consumption spike at the beginning of period (a) is when the sensor awakens from a deep sleep state. During every sensing period, seven measurements are taken from the ultrasonic range sensor, with seven corresponding current consumption peaks observed during period (a) in Figure 3. After the measurements are collected, a median is applied to these seven measurements and the result is transmitted in period (b). A short receive window (Rx1; period (d)), follows the transmission period (Tx) and a sleep period (c). Next, a second receive window Rx2, period (f), is typically opened one second after Rx1. The sensor enters a low power state between the receive windows to save power, period (e). The sleep period (g) lasts until the next sensing state (a) occurs.
图 3 显示了传感器在 LoRaWAN A 类（默认 LoRaWAN 类，其中终端设备（即传感器）始终启动通信）中运行的所有可能状态下的电流消耗。传感器的数据收集和传输周期为 ≈1 分钟；由于传输后的接收窗口不同，每个周期的总持续时间会有几毫秒的变化，从而为双向通信提供了机会。图 3 中，(a)、(b)、(d) 和 (f) 区域为唤醒周期，(c)、(e) 和 (g) 区域为睡眠周期。(a)段开始时的电流消耗峰值是传感器从深度睡眠状态中唤醒时出现的。在每个传感周期内，超声波测距传感器会进行七次测量，在图 3 中的(a)周期内会观察到七个相应的电流消耗峰值。收集测量值后，对这七个测量值进行中值处理，并在周期（b）中传输结果。在传输期 (Tx) 和休眠期 (c) 之后是一个较短的接收窗口 (Rx1；周期 (d))。接下来，第二个接收窗口 Rx2（周期 (f)）通常在 Rx1 之后一秒打开。在接收窗口之间，传感器进入低功耗状态，以节省电能，即周期 (e)。休眠期 (g) 一直持续到下一个传感状态 (a) 出现。
(a) Sensor current consumption in milliamperes (mA) during different operational states. The total cycle is approximately 60 s; only the first 10 s is presented, the remainder of the time the sensor is in sleep mode. (b) Battery charge-drain cycles for select sensors, over the month of January 2021.
(a) 传感器在不同工作状态下的电流消耗（毫安）。总周期约为 60 秒；仅显示前 10 秒，其余时间传感器处于睡眠模式。(b) 2021 年 1 月期间部分传感器的电池充放电周期。
Table 2 details the average current consumption of different sensor states during a single cycle illustrated in Figure 3a. For one such sleep-wake cycle, the wake-time was ≈2.7 s (i.e., the sum of periods (a), (b), (d), and (f)) and sleep-time was ≈60.5 s (i.e., the sum of periods (c), (e), and (g)), making the duty cycle approximately 4.5%. The average consumption of one sleep-wake cycle is ≈708 μA. This average varies from cycle to cycle but is within the range of 700–750 μA. Based on these observations, with a fully charged 400 mAh battery with no additional power or means of charging, the estimated lifetime of the sensor can be up to 22 days and 17 hr. This headroom in power consumption can accommodate periods of reduced sunlight, prolonged cover on the solar panel (e.g., snow), or a panel failure.
表 2 详细列出了图 3a 所示单个周期内不同传感器状态的平均电流消耗。在这样一个睡眠-唤醒周期中，唤醒时间≈2.7 秒（即周期（a）、（b）、（d）和（f）之和），睡眠时间≈60.5 秒（即周期（c）、（e）和（g）之和），占空比约为 4.5%。一个睡眠-唤醒周期的平均耗电量为 708 μA。这一平均值随周期而变化，但在 700-750 μA 范围内。根据上述观察结果，如果使用充满电的 400 毫安时电池，且没有额外的电源或充电方法，传感器的使用寿命估计可达 22 天 17 小时。这样的耗电余量可以应对日照时间减少、太阳能电池板被长时间覆盖（如雪）或电池板故障等情况。
表 2.图 3a 中详述的不同传感器工作状态的典型持续时间和平均电流消耗量
| Period | Sensor state 传感器状态 | Duration (ms) 持续时间（毫秒） | Avg. current consumption 平均电流消耗 |
|---|---|---|---|
| (a) | Sensing | 2,500 | 14.1 mA 14.1 毫安 |
| (b) | Up-link transmission (Tx) 上行链路传输 (Tx) |
67 | 115 mA 115 毫安 |
| (c) | Sleep after Tx until Rx1 发送后休眠，直到 Rx1 |
5,000 | 22.4 µA 22.4 µA |
| (d) | Receive window 1 (Rx1) 接收窗口 1 (Rx1) |
52 | 18.3 mA 18.3 毫安 |
| (e) | Sleep between Rx1 and Rx2 在 Rx1 和 Rx2 之间休眠 |
1,000 | 18.5 µA 18.5 µA |
| (f) | Receive window 2 (Rx2) 接收窗口 2 (Rx2) |
59 | 19.3 mA 19.3 毫安 |
| (g) | Sleep until next sensing cycle 休眠至下一个感应周期 |
54,500 | 21.3 µA 21.3 µA |
-
Note. These values vary from cycle to cycle but on average are below 1 mA average current consumption.
注。这些值因周期而异，但平均消耗电流低于 1 mA。
Battery operation and solar charging of the deployed sensors have been successful thus far. The sensors utilize a 0.6 W solar panel, mounted at an angle of 45°. Figure 3b illustrates the battery levels of three sensors deployed in locations with different cloud coverage, shade, and mounting conditions, which contribute to the differences in their battery trends. The three sensors also displayed differences in average voltage levels due to expected variations in lithium polymer battery condition. The sensor deployed at 5th Avenue and Hoyt Street had the best power harvesting of the three sensors illustrated in Figure 3b due to less shade from buildings and trees. All three sensors had minimal battery charging from January 1 to 3, 2021, due to consistent heavy cloud cover. Nonetheless, the sensor batteries were able to recharge in the following days as the cloud cover reduced. In conditions of full sun exposure, a single day of sun is sufficient to charge the 400 mAh battery completely. However, the sensor deployed at 6th Avenue and Waverly Place had less favorable power harvesting conditions, as observed by the slower upward charging trend after the heavy cloud cover period. Nonetheless, given the headroom provided through the battery size and low power consumption design, the sensors were able to operate through adverse weather conditions and maintain a healthy battery voltage level.
迄今为止，已部署传感器的电池操作和太阳能充电都很成功。传感器使用的是 0.6 瓦太阳能电池板，安装角度为 45°。图 3b 显示了部署在不同云层覆盖、遮阳和安装条件下的三个传感器的电池电量，这也是造成其电池趋势不同的原因。由于锂聚合物电池状态的预期变化，三个传感器的平均电压水平也存在差异。在图 3b 所示的三个传感器中，部署在第五大道和霍伊特街的传感器由于受到建筑物和树木的遮挡较少，因此电量采集效果最好。在 2021 年 1 月 1 日至 3 日期间，由于云层一直很厚，所有三个传感器的电池充电量都很小。尽管如此，随着云量的减少，传感器电池在接下来的几天里还是能够重新充电。在阳光充足的条件下，一天的日照足以为 400 毫安电池完全充电。然而，部署在第六大道和韦弗利广场的传感器的电量收集条件并不理想，这一点可以从浓云覆盖期过后较慢的充电上升趋势观察到。尽管如此，考虑到电池容量和低功耗设计所提供的余量，传感器能够在恶劣天气条件下正常工作，并保持健康的电池电压水平。
4 Data Processing and Visualization
4 数据处理和可视化
4.1 Flood Sensor Data Ingestion and Calibration
4.1 洪水传感器数据输入和校准
为了将传感器收集到的距离测量值转换成清晰的洪水深度剖面图，需要执行多个数据处理阶段；数据处理流水线如图 4 所示。首先，超声波脉冲未返回传感器的任何数据点（传感器将其记录为最大量程值（5,000 毫米））都会被标记为 "未定义"，这是缺失数据点的名称。然后将 t 时刻收集到的剩余距离测量值（z）转换为洪水深度（D）；该计算在 NodeRed 中进行。要计算 D，必须知道在稳定和无洪水条件下传感器与下方地面之间的距离。该值通过每天凌晨 5 点的动态校准程序确定，在该程序中，对前三个晚上（晚上 10 点到凌晨 5 点之间；n ≈ 1,260 ）收集的 z 值进行分析，以确定中值（z night−median ）。白天的测量值不包括在校准范围内，因为它们可能与温度有关，详见第 3 节。如果 z 的标准偏差超过 5 毫米（表示洪水或与噪声或不稳定条件有关的较大偏差），则使用前一天计算出的 z night−median 。为了计算 D，传感器当时的距离测量值（z）要减去 z night−median （等式 1）。
Sequential stages used to collect and filter flood sensor data. Data processing during data collection occurs on the sensor, whereas filter stages 0–3 occur on the FloodNet servers in NodeRed.
用于收集和过滤洪水传感器数据的顺序阶段。数据收集期间的数据处理在传感器上进行，而过滤阶段 0-3 则在 NodeRed 的 FloodNet 服务器上进行。
This dynamic calibration approach allows the project to adapt to seasonal variation in baseline znight−median readings or changes in sensor height, caused, for example, by a shift in U-channel post position.
这种动态校准方法可使项目适应基线 z night−median 读数的季节性变化或传感器高度的变化，例如，U 型通道柱位置的移动。
4.2 Flood Depth Data Processing
4.2 洪水深度数据处理
After converting distance measurements to depths, the sensor data are processed through a series of filter stages to filter anomalous data related to: (a) low-level measurement noise within our acceptable range of error, (b) the impact of direct sunlight on temperature compensation, and (c) objects located below the sensor, which is a type of noise more likely to be experienced by water level sensors monitoring flooding on streets and sidewalks than those installed over water bodies, making data filtering critically important for the FloodNet project. In the following, a measured “event” is defined as a series of one or more depth readings greater than 10 mm.
在将距离测量值转换为深度后，传感器数据将通过一系列过滤阶段进行处理，以过滤与以下方面有关的异常数据：(a) 在我们可接受的误差范围内的低水平测量噪声，(b) 阳光直射对温度补偿的影响，以及 (c) 位于传感器下方的物体，与安装在水体上方的水位传感器相比，监测街道和人行道洪水的水位传感器更容易受到这种噪声的影响，因此数据过滤对 FloodNet 项目至关重要。在下文中，测量到的 "事件 "被定义为一系列大于 10 毫米的一个或多个深度读数。
The first filter stage is designed to target the first two anomalies listed above. To correct for low-level measurement noise, all reported depths less than 10 mm are assigned a value of zero. This filter stage also works to filter measurements impacted by solar irradiance incident on the sensor housing, causing the ultrasonic sensor’s internal temperature sensor to read greater than ambient temperatures. When this occurs, the ultrasonic sensor’s calculation of distance is slightly increased as the speed of sound in air used in the calculation is exaggerated. As such, measured zt > znight−median, resulting in negative calculated depth values (i.e., measurements that dip below ground level). This phenomenon has also been observed by others utilizing ultrasonic sensors for water level monitoring (Mousa et al., 2016). Assigning all depths less than 10 mm a value of zero corrects for this anomaly. Regardless, the resulting measurement error caused by this effect is typically within the range of our target accuracy (i.e., ±25 mm; see Section 13).
第一滤波阶段是针对上述前两种异常情况设计的。为了校正低水平的测量噪声，所有报告的深度小于 10 毫米的值均为零。该滤波级还可过滤因太阳辐照入射到传感器外壳上而导致超声波传感器内部温度传感器读数高于环境温度的测量结果。出现这种情况时，超声波传感器的距离计算会略有增加，因为计算中使用的空气声速被夸大了。因此，测量到的 z > z night−median ，导致计算出的深度值为负值（即测量值低于地面）。利用超声波传感器进行水位监测的其他人员也观察到了这种现象（Mousa 等人，2016 年）。将所有小于 10 毫米的深度赋值为零可以纠正这种异常现象。无论如何，由这种效应引起的测量误差通常都在我们的目标精度范围内（即 ±25 毫米；见第 13 节）。
The second filter stage applies a gradient-based filter to target non-flood depth measurements caused by objects located under the sensor. When people, animals, or large objects—such as trash, bicycles, vehicles, etc.—pass beneath or are placed under a sensor, there is an immediate increase in depth measured from ground surface to the height of the object. These non-flood measurements can be distinguished from floods by their profile, given that floods have a gradual onset (Figure 5); this difference in profile allows the use of a filter that assesses the change in depth values between two time-adjacent depth measurements (Δdepth/Δtime) on a rolling basis. The upper threshold for allowable change in depth over time was set at 254 mm per minute, which is a rate six times greater than the fastest rate of flood onset we have measured thus far in NYC (i.e., during Hurricane Ida). Any data point that exceeds the allowable Δdepth/Δtime threshold is labeled as “undefined” and is not included in the data visualization or alerting platforms. Of note, however, is that this filter stage can miss non-flood events when a sensor is experiencing weak network connectivity and there are longer periods of time between adjacent depth measurements, reducing the chance that the gradient threshold is exceeded.
第二滤波阶段采用基于梯度的滤波器，针对传感器下方物体引起的非洪水深度测量。当人、动物或大型物体（如垃圾、自行车、车辆等）经过传感器下方或被置于传感器下方时，从地表到物体高度的测量深度会立即增加。这些非洪水测量值可以通过其轮廓与洪水区分开来，因为洪水是逐渐开始的（图 5）；这种轮廓上的差异允许使用滤波器，以滚动方式评估相邻两次深度测量值之间的深度值变化（Δ深度/Δ时间）。允许深度随时间变化的上限值设定为每分钟 254 毫米，这个速度比我们迄今为止在纽约市测得的最快洪水爆发速度（即伊达飓风期间）高出六倍。任何超过允许的 Δ深度/Δ时间阈值的数据点都会被标记为 "未定义"，不会被纳入数据可视化或警报平台。但值得注意的是，当传感器网络连接能力较弱，相邻深度测量之间的间隔时间较长时，该过滤阶段可能会错过非洪水事件，从而降低超过梯度阈值的几率。
Examples of flood and non-flood event data recorded by FloodNet sensors. An “event” is defined as a series of one or more depth readings greater than 10 mm.
FloodNet 传感器记录的洪水和非洪水事件数据示例。事件 "的定义是一系列大于 10 毫米的一个或多个深度读数。
To address non-flood data points that are not captured by the gradient filter, we characterized three primary types of noise events that appear in the sensor data - (a) blips, (b) boxes, and (c) pulse chains (described below) - and designed a series of real-time filters to target them. This third filter stage applies three filters in the following order: a blip filter, followed by a box filter, followed by another blip filter to clean non-flood data points that the box filter can leave behind. Each filter is described below.
为了处理梯度滤波器无法捕捉到的非洪水数据点，我们对传感器数据中出现的三种主要噪声事件进行了特征描述--(a) 突波、(b) 盒子和 (c) 脉冲链（如下所述）--并设计了一系列针对它们的实时滤波器。第三滤波阶段按以下顺序应用三个滤波器：一个 "突波 "滤波器，接着是一个 "方波 "滤波器，然后是另一个 "突波 "滤波器，以清除 "方波 "滤波器可能留下的非洪水数据点。下文将对每个过滤器进行说明。
Blips occur when sensor readings increase for a single measurement then return to the original (or similar) value at the next time point. These single, anomalous measurements are likely caused by a person, vehicle, or object temporarily located beneath the sensor in the moment a measurement is collected. For sensors that are consistently uploading data (i.e., collecting and transmitting data every ≈ 1 min), these readings can be confidently filtered from the dataset because they are clearly transient. Blip events are characterized by three consecutive readings with values of D1 = D, D2 = D + ΔD, and D3 = D ± 0.1ΔD, where D is any depth measurement ≥0, and ΔD is a change in depth greater than 2 mm (i.e., D2-D1 > 2 mm).
当传感器读数在单次测量中上升，然后在下一个时间点恢复到原始值（或类似值）时，就会出现突波。这些单次异常测量值很可能是由于在采集测量值的瞬间，传感器下方临时有人、车或物体造成的。对于持续上传数据的传感器（即每隔 ≈ 1 分钟收集和传输一次数据），这些读数可以从数据集中有把握地过滤掉，因为它们显然是瞬时的。昙花一现事件的特征是三个连续读数的值分别为：D 1 = D、D 2 = D + ΔD、D 3 = D ± 0.1ΔD，其中 D 为任何深度测量值≥0，ΔD 为深度变化大于 2 毫米（即 D 2 -D 1 > 2 毫米）。
To recognize this condition, a blip metric for depth measurement D2 is calculated following Equation 2. To apply the filter in real-time, any depth measurement that is at least 2 mm greater than the previous measurement is held in memory and not released to the database or visualization platform until the next measurement (D3) arrives, to assess whether the depth measurement is a blip. If the blip metric is less than 0.1, point D2 is characterized as a blip and is set to “undefined”, otherwise, the measurement remains unmodified.
为了识别这种情况，可根据公式 2 计算出深度测量值 D 2 的闪烁度量值。为了实时应用滤波器，任何比上一次测量值大至少 2 mm 的深度测量值都会被保存在内存中，直到下一次测量值（D 3 ）到达时才会被释放到数据库或可视化平台，以评估深度测量值是否为突波。如果模糊度量值小于 0.1，则 D 2 点被定性为模糊点，并设置为 "未定义"，否则，测量值保持不变。
在某些情况下，由于网络连接不畅导致可用数据点的频率较低，可能导致洪水事件只能在单个数据点中测量到，从而模仿突发事件的特征。为了降低过滤掉数据点较少的洪水事件的风险，如果在 D 1 和 D 3 时间跨度超过 6 分钟的情况下检测到突波，过滤器将查询我们的天气数据库，查找纽约市任何雨量计的降水报告，只有在过去一小时内没有降水报告的情况下才会进行过滤。用于检查沿海洪水可能性的潮汐数据仍在开发中。
当传感器读数立即增加，然后在一段持续的时间内保持恒定（或接近恒定）值时，就会出现方框。当传感器下方放置静态物体（如停放的车辆、垃圾等）时，就会出现这种情况。盒式滤波器的设计目的是检测深度测量值的急剧增加，随后的一个或多个测量值保持在深度初始增加值的 10%以内。具体来说，在长度为 N 的深度值序列中，盒状事件的特征是 D 1 = 0、D 2 = ΔD、D n = D 2 ± 0.1ΔD，对于 [3, N] 中的每个序列测量值 n，直到某个测量值 D N 偏离最初增加的高度 (ΔD)超过 10%。为识别这种情况，可根据公式 3 计算出测量值 D n 的箱形度量，其值小于 0.1 表示 D n 是箱形的一部分。在实时应用中，如果满足 D 1 和 D 2 的条件，则 D 2 将保留在内存中，直到接收到下一个点 D 3 。如果 D 3 和随后的深度测量值（直到 D N ）符合所述的箱体条件，则 D 2 至 D N 将被设置为 "未定义"。为了防止错误地将洪水定性为方框，例如，鉴于洪水剖面中的积水可能呈现出高原的外观，从而错误地过滤掉洪水的顶部，我们限制方框过滤器只过滤起始深度为零（即 D 1 = 0）的非洪水事件。当网络连接不畅导致数据点稀少时，箱式过滤器将遵循闪烁过滤器的逻辑，仅在测量数据点过去一小时内没有降水的情况下继续过滤数据点。
Pulse chains are characterized as a noisy series of blips and boxes (and blips on top of boxes) that cannot be easily categorized into the other two categories. These anomalous measurements are more challenging to filter and are not addressed by the filters explicitly, but can be captured partially by a combination of both the blip and box filters. Thus, some pulse chains may remain after the application of the three filter stages. Ongoing research is being conducted to optimize data analysis strategies to recognize and filter out this measurement noise.
脉冲链的特点是由一连串的 "突波 "和 "方波"（以及 "方波 "上的 "突波"）组成，这些 "突波 "和 "方波 "不能简单地归入其他两类。这些异常测量值的滤波难度较大，滤波器无法明确处理，但可以通过组合使用 "猝发 "和 "方框 "滤波器来部分捕捉。因此，在应用了三个阶段的滤波器后，可能会保留一些脉冲链。目前正在进行研究，以优化数据分析策略，识别并过滤掉这些测量噪声。
4.3 Data Visualization and Alerting
4.3 数据可视化和警报
Raw data and filtered flood depth measurements are stored in a database that is optimized for large-scale, time-series data ingestion and retrieval. Currently, the FloodNet project utilizes two real-time data visualization platforms that retrieve data from the database: a Grafana-based platform that the FloodNet team uses internally, and a public-facing dashboard openly available on the web (https://dataviz.floodnet.nyc). The design of the public-facing dashboard incorporated feedback collected from city agency personnel and community residents to ensure data visualization is useful and meaningful to both stakeholder groups.
原始数据和过滤后的洪水深度测量值存储在一个数据库中，该数据库针对大规模、时间序列数据的摄取和检索进行了优化。目前，FloodNet 项目使用两个实时数据可视化平台从数据库中检索数据：一个是 FloodNet 团队内部使用的基于 Grafana 的平台，另一个是在网上公开的面向公众的仪表盘 ( https://dataviz.floodnet.nyc)。面向公众的仪表板的设计纳入了从市政机构人员和社区居民那里收集到的反馈意见，以确保数据可视化对这两个利益相关群体都有用且有意义。
Following feedback from stakeholders, the data visualization platforms enable users to view sensor data alongside other sources of information that contextualize how flood data were collected and provide important touchpoints for understanding the impacts of flooding. For example, users can learn about the infrastructure on which a sensor is mounted, when it was installed, and whether it collects measurements over a sidewalk, road, or waterway. Additionally, tide gauge and precipitation data are provided to contextualize how environmental factors impact street-level flooding. We are conducting ongoing research to learn how to best incorporate qualitative documentation of flood events (such as photos, videos, and written accounts) into our visualization platforms. A core tenet of the FloodNet project is open access to collected data. To facilitate this, a public facing API is under development, with an accompanying data download portal.
根据利益相关者的反馈，数据可视化平台使用户能够在查看传感器数据的同时查看其他信息来源，从而了解洪水数据的收集方式，并为了解洪水的影响提供重要的接触点。例如，用户可以了解安装传感器的基础设施、安装时间，以及是否在人行道、道路或水道上收集测量数据。此外，我们还提供了验潮仪和降水量数据，帮助用户了解环境因素对街道洪水的影响。我们正在进行持续研究，以了解如何将洪水事件的定性记录（如照片、视频和文字说明）最好地融入我们的可视化平台。FloodNet 项目的核心宗旨是开放所收集的数据。为此，我们正在开发一个面向公众的应用程序接口（API）和一个配套的数据下载门户网站。
Flood sensor data present several opportunities for real-world application. One key application for the data is a real-time flood alerting system, triggered when measured flood depths reach a specified depth threshold. This alerting system has been identified as a critical feature by some of our stakeholders, including both community members and municipal employees involved in emergency management and response. During pilot testing of the alerting system during summer 2021, flood alerts were sent via email or messaging app to registered users when measured flood depths exceeded 7.6 cm. During Hurricane Ida (1 Sept 2021), the FloodNet system alerted NYC Emergency Management of flooding recorded by the sensors installed in the Gowanus neighborhood in Brooklyn 50 min before they received other notifications of the event.
洪水传感器数据为现实世界的应用提供了多个机会。数据的一个重要应用是实时洪水警报系统，当测量到的洪水深度达到指定深度阈值时就会触发该系统。我们的一些利益相关者，包括社区成员和参与应急管理和响应的市政人员，都认为该警报系统是一项重要功能。在 2021 年夏季对警报系统进行试点测试期间，当测量到的洪水深度超过 7.6 厘米时，就会通过电子邮件或消息应用程序向注册用户发送洪水警报。在伊达飓风（2021 年 9 月 1 日）期间，FloodNet 系统在纽约市应急管理部门收到其他事件通知前 50 分钟，就已将安装在布鲁克林戈瓦纳斯社区的传感器记录的洪水警报发送给了他们。
5 Results and Discussion
5 结果与讨论
FloodNet sensor installation and data collection are ongoing. To date, a total of 87 FloodNet sensors have been installed in flood-prone areas across the five boroughs of NYC (Figure 6). Sensor installations have been staggered, with some sensors installed for years and others installed more recently. We plan to expand the network across NYC in the coming years. Nonetheless, analysis of sensor data from flood events and static baseline measurements collected thus far provides insight about the functionality of ultrasonic sensors and the particularities of flood monitoring in general.
FloodNet 传感器的安装和数据收集工作正在进行中。迄今为止，纽约市五个区的洪水易发区共安装了 87 个 FloodNet 传感器（图 6）。传感器的安装是交错进行的，有些传感器已安装多年，有些则是最近才安装的。我们计划在未来几年将网络扩展到整个纽约市。不过，通过分析洪水事件中的传感器数据和迄今为止收集到的静态基线测量数据，我们可以深入了解超声波传感器的功能和洪水监测的一般特性。
FloodNet’s public facing data dashboard. The map view of the dashboard, shown here, includes the locations of the 87 sensors installed across the five boroughs of NYC as of February 2024. Each circle indicates a sensor location and is color coded corresponding to the real-time measured flood depth; the depth value in inches is indicated by the number located in the center of each symbol.
FloodNet 面向公众的数据仪表盘。此处显示的仪表板地图视图包括截至 2024 年 2 月纽约市五个区安装的 87 个传感器的位置。每个圆圈表示一个传感器的位置，并根据实时测量的洪水深度进行了颜色编码；每个符号中心的数字表示以英寸为单位的深度值。
5.1 Sources of Sensor Measurement Error
5.1 传感器测量误差的来源
The main sources of error in the sensor's depth measurements are related to the noise floor of the ultrasonic range finder, exposure to direct sunlight, shifts in ambient temperature, and the mount angle of the sensor. The noise floor of the ultrasonic range finder is a function of the manufacturer stated accuracy of 1% or better (i.e., ±15 mm at a typical mounting height of 3 m) and measurement deviations caused by variations in surface profile beneath the sensor. To assess the extent of the noise floor error, baseline data collected by 46 sensors from February 14 to 21, 2024 between the times of 10 p.m. and 5 a.m. were assessed for variability in distance readings; no floods occurred during this time and other sensors were not included due to recorded non-flood measurements (e.g., blips and boxes described in Section 10) that would have confounded analysis. Night time measurements were used to avoid the influence of direct sunlight on measurements. Of 98,013 data points where the water depth measurements should have been 0 mm, the minimum measurement was −14 mm and the maximum measurement was 19 mm, resulting in a error range of 33 mm (Table 3), which is within our accuracy tolerance of ±25 mm.
传感器深度测量的主要误差来源与超声波测距仪的本底噪声、阳光直射、环境温度变化和传感器的安装角度有关。超声波测距仪的本底噪声是制造商规定的 1%或更高精度（即在 3 米的典型安装高度下为 ±15 毫米）和传感器下方表面轮廓变化造成的测量偏差的函数。为评估噪声底限误差的程度，对 46 个传感器在 2024 年 2 月 14 日至 21 日晚上 10 点至凌晨 5 点期间收集的基线数据进行了评估，以了解距离读数的变化情况；在此期间未发生洪水，其他传感器也未包括在内，因为记录的非洪水测量值（如第 10 节所述的突波和方框）会影响分析。夜间测量是为了避免阳光直射对测量的影响。在水深测量值应为 0 毫米的 98,013 个数据点中，最小测量值为-14 毫米，最大测量值为 19 毫米，误差范围为 33 毫米（表 3），在我们的 ±25 毫米精度容差范围内。
表 3.传感器测量中的潜在误差来源及其对潜在测量不准确度的贡献统计摘要
| Source of error 错误来源 | Ultrasonic noise floor 超声波本底噪声 | Direct sunlight 阳光直射 | Ambient temperature 环境温度 | Mount angle 安装角度 |
|---|---|---|---|---|
| Number of sensors evaluated 评估的传感器数量 |
46 | 46 | 1 | 10 |
| Sample duration (days) 样本持续时间（天） | 7 | 7 | 365 | - |
| Number sensor measurements assessed (n) 评估的传感器测量次数（n） |
98,013 | 120,739 | 525,543 | 6 |
| Mean depth measurement (mm) 平均深度测量值（毫米） |
−0.4 -0.4 | −12.2 -12.2 | 0.7 | 5.7 |
| Standard deviation (mm) 标准偏差（毫米） | 2.6 | 10.7 | 4.2 | 4.3 |
| Median depth measurement (mm) 中间深度测量值（毫米） |
0 | −9.0 -9.0 | 1.7 | 5.7 |
| Minimum depth (mm) 最小深度（毫米） | −14.0 -14.0 | −73.0 -73.0 | −5.1 -5.1 | 0 |
| Maximum depth (mm) 最大深度（毫米） | 19.0 | 0 | 6.1 | 11.5 |
| Error range = maximum-minimum (mm) 误差范围 = 最大-最小（毫米） |
33.0 | 73.0 | 11.2 | 11.5 |
| % of measurements within error tolerance 误差在容许范围内的测量百分比 |
100% | 88.7% | 100% | 100% |
-
Note. For all data points evaluated, the water depth measurement should have been 0 mm (i.e., no flood was occurring). Data used for evaluation of error related to ultrasonic noise floor and ambient temperature were limited to measurements collected between 10 p.m. and 5 a.m.; data used to assess the impact of direct sunlight exposure were limited to those collected between 8 a.m. and 5 p.m.
注。对于所有评估数据点，水深测量值应为 0 毫米（即未发生洪水）。用于评估与超声波噪音底限和环境温度有关的误差的数据仅限于在晚上 10 点至凌晨 5 点之间收集的测量数据；用于评估阳光直射影响的数据仅限于在上午 8 点至下午 5 点之间收集的数据。
The largest contributor of measurement error was observed when direct sunlight was incident on the ultrasonic range finder. To isolate the effects of direct sunlight and calculate its impact, data collected by the same sensors and time period as above, but during the hours of 8 a.m. and 5 p.m., were assessed. Although a maximum error of 73 mm was found, this occurred only during times of peak sunlight intensity. Further, the mean (−12.2 ± 10.7 mm) and median (−9 mm) measurements were within our allowable accuracy tolerance, and fewer than 12% of daytime measurements exceeded the allowable error.
当阳光直射超声波测距仪时，测量误差最大。为了隔离阳光直射的影响并计算其影响，我们评估了由上述相同传感器和时间段收集的数据，但时间是上午 8 点和下午 5 点。虽然发现最大误差为 73 毫米，但这只发生在阳光最强烈的时候。此外，测量结果的平均值（-12.2 ± 10.7 毫米）和中位数（-9 毫米）都在允许误差范围内，只有不到 12% 的白天测量结果超出了允许误差。
As previously mentioned, there is a known effect of ambient weather conditions on ultrasonic distance measurements (Mousa et al., 2016). To evaluate this effect on FloodNet sensor readings, daily mean distances to the ground were calculated for data collected by a FloodNet sensor located at the intersection of Hoyt Street and Fifth Street in Brooklyn between 10 p.m. and 5 a.m. each day, over the course of a year. Night time values were evaluated alone to assess seasonal-dependence of sensor readings due to ambient temperature, without the confounding impact of direct sunlight on the sensor during the day. To understand baseline (i.e., non-flood) conditions, known flood and snow events were excluded from the dataset, as were absolute values greater than two standard deviations from the mean.
如前所述，已知环境天气条件会对超声波距离测量产生影响（Mousa 等人，2016 年）。为了评估这种对 FloodNet 传感器读数的影响，我们计算了位于布鲁克林霍伊特街与第五街交叉口的 FloodNet 传感器在一年中每天晚上 10 点到凌晨 5 点之间收集到的数据的日平均地面距离。仅对夜间值进行评估，以评估传感器读数因环境温度而产生的季节性变化，而不考虑白天阳光直射传感器的影响。为了解基线（即非洪水）条件，数据集中不包括已知的洪水和降雪事件，也不包括与平均值相差两个标准差以上的绝对值。
Despite no change in the physical height of the sensor as installed, there was a statistically significant shift in the baseline distance measurements made by the sensor over the course of a year (Figure 7), with greatest distances measured in January and smallest measured in August. Colder ambient temperatures in the winter decrease the velocity of ultrasonic pulses, resulting in longer times of return and exaggerated distance measurement if the temperature was not fully accounted for in the distance calculation. Conversely, warmer temperatures in summer would lead to calculation of shorter distances due to faster velocity and shorter time of return of the ultrasonic pulse. As such, it is possible that the internal temperature sensor housed within the casing of the ultrasonic sensor erroneously detected greater than ambient temperatures during the winter and lower than ambient during the summer. Nonetheless, the magnitude of difference between greatest and smallest distance measurement was only 11.2 mm, which is within the allowable error of the sensor. Additionally, the dynamic calibration procedure described in Section 9 accounts and corrects for this temperature-dependent phenomena.
尽管安装时传感器的物理高度没有变化，但在一年的时间里，传感器的基线距离测量值却发生了显著的变化（图 7），1 月份测量的距离最大，8 月份测量的距离最小。冬季较低的环境温度会降低超声波脉冲的速度，导致回波时间延长，如果在距离计算中没有充分考虑温度因素，则会夸大距离测量值。相反，夏季气温较高，超声波脉冲速度较快，回波时间较短，因此计算出的距离较短。因此，安装在超声波传感器外壳内的内部温度传感器有可能在冬季错误地检测到高于环境温度的温度，而在夏季则检测到低于环境温度的温度。不过，最大和最小距离测量值之间的差值仅为 11.2 毫米，在传感器的允许误差范围之内。此外，第 9 节所述的动态校准程序也考虑并纠正了这种与温度有关的现象。
Difference in daily nighttime mean distance measurements (under non-flooded conditions from 10 p.m. to 5 a.m.; solid line) from the annual mean for a flood sensor located at Hoyt Street and Fifth Avenue (Brooklyn) over 12 months from 2020 to 2021. Temperature data (dashed line) represent ambient conditions and were sourced from a nearby New York State Mesonet (Brotzge et al., 2020) managed weather station at the Brooklyn Navy Yard (ID: bknyrd), which is located ≈ 3 km away from the sensor.
从 2020 年到 2021 年的 12 个月中，位于 Hoyt 街和第五大道（布鲁克林）的洪水传感器的日夜间平均距离测量值（在晚上 10 点到凌晨 5 点的非洪水条件下；实线）与年平均值的差异。温度数据（虚线）代表环境条件，来自附近位于布鲁克林海军船坞（ID：bknyrd）的纽约州气象网（Brotzge 等人，2020 年）管理的气象站，该气象站距离传感器≈ 3 公里。
The ideal mounting angle between the ground and the ultrasonic range finder is 90°. If the mounting angle deviates from 90° by an angle of θ then measured flood depths will be exaggerated, given that the pathlength of the ultrasonic pulse would be along the hypotenuse, with a distance equal to the height of the sensor multiplied by cos−1(θ). Ten different ultrasonic range finders were evaluated in the laboratory to assess the maximum θ at which ultrasonic pulses reflected off the ground were still able to return to the sensor, which was found to be 5°. At angles greater than this, the majority of reflected ultrasonic pulses are not received, resulting in an invalid measurement. When mounting sensors, a two-axis spirit level is used to ensure the ultrasonic range finder is at 90° to the surface beneath, reducing the likelihood of this error. If, however, the mounting angle deviates from 90° by the 5° maximum θ, then the maximum error in flood depth measurement would be 11.5 mm at a typical sensor mount height of 3 m (Table 3), which is within our accuracy tolerance of ±25 mm.
地面与超声波测距仪之间的理想安装角度为 90°。如果安装角度偏离 90°，偏离角度为 θ，那么测得的洪水深度将被夸大，因为超声波脉冲的路径长度将沿着斜边，距离等于传感器高度乘以 cos −1 (θ)。在实验室中对十种不同的超声波测距仪进行了评估，以确定从地面反射的超声波脉冲仍能返回传感器的最大 θ，结果发现最大 θ 为 5°。大于这个角度时，大部分反射的超声波脉冲将无法接收，导致测量无效。在安装传感器时，可使用双轴水平仪确保超声波测距仪与下方表面成 90°，从而减少出现这种误差的可能性。但是，如果安装角度偏离 90° 的最大值 θ 为 5°，那么在传感器安装高度通常为 3 米的情况下，洪水深度测量的最大误差将为 11.5 毫米（表 3），在我们的精度公差 ±25 毫米范围内。
When assessing all data points collected by the sensors between February 14 and 21, 2024, significant majority (95.8%) of the sensor measurements (n = 329,124) were within our allowable range of error.
在评估 2024 年 2 月 14 日至 21 日期间传感器收集的所有数据点时，绝大多数（95.8%）传感器测量值（n = 329 124）都在我们允许的误差范围内。
5.2 Validation of Sensor Measurements
5.2 验证传感器测量结果
Sensor measurements were validated in three ways: (a) in-laboratory testing before deployment; (b) comparison with data from tide gauges operated by the National Oceanographic and Atmospheric Administration (NOAA); and (c) comparison with manually collected flood depth measurements during a street-level, tidal flood event.
传感器测量结果通过三种方式进行验证：(a) 部署前的实验室测试；(b) 与美国国家海洋和大气管理局 (NOAA) 运行的验潮仪数据进行比较；(c) 与街道潮汐洪水事件期间人工收集的洪水深度测量结果进行比较。
The flood sensor assembly process follows a detailed quality assurance (QA) procedure (FloodNet, 2024a) (https://floodnet-nyc.github.io). After the assembly, and before deployment, sensors undergo a data validation test to ensure accurate measurements. During this in-laboratory quality control (QC) testing, each sensor is tested for accuracy at known mounting heights that are similar to actual deployment scenarios. These heights are measured with a standard scale, and the mounts are aligned at a perpendicular angle to the ground surface using a spirit level. The duration of this test is 1 hour, and a median of seven measurements are collected every minute. The observed sensor measurements at all mounting heights must be within hi ± nallowed to pass the QC test, where hi is the known mounting height and nallowed is the acceptable noise floor, which is the ultrasonic range finder’s manufacturer stated 1% of the measured distance. All of the FloodNet sensors constructed thus far have met this QC criteria.
洪水传感器的装配过程遵循详细的质量保证 (QA) 程序（FloodNet，2024a）（https://floodnet-nyc.github.io）。组装完成后，在部署之前，传感器要经过数据验证测试，以确保测量的准确性。在实验室内的质量控制 (QC) 测试中，每个传感器都要在与实际部署情况类似的已知安装高度下进行精度测试。这些高度使用标准刻度进行测量，并使用水平仪以垂直于地表的角度对准支架。测试持续时间为 1 小时，每分钟收集 7 个测量值。在所有安装高度上观察到的传感器测量值必须在 h ± n allowed 范围内才能通过质量控制测试，其中 h 是已知的安装高度，n allowed 是可接受的本底噪声，即超声波测距仪制造商规定的测量距离的 1%。迄今为止制造的所有 FloodNet 传感器都符合这一质量控制标准。
To test the ability of FloodNet sensors to detect water level changes under field conditions, sensors were mounted over tidally influenced water bodies (Gowanus Canal in Brooklyn and the Jamaica Bay estuary in Queens). Changes in water level due to the daily tidal cycle measured by the FloodNet sensors were compared with measurements collected by nearby tide gauges operated by NOAA, and found to be similar with similar periodicity (Figure 8).
为了测试 FloodNet 传感器在现场条件下探测水位变化的能力，将传感器安装在受潮汐影响的水体（布鲁克林的戈瓦纳斯运河和皇后区的牙买加湾河口）上。将 FloodNet 传感器测量到的每日潮汐周期引起的水位变化与附近由 NOAA 运行的验潮仪收集到的测量值进行比较，发现两者具有相似的周期性（图 8）。
(a) Comparison of water level data collected by a FloodNet sensor installed above the Gowanus Canal in Brooklyn (black line), which is tidally influenced, with the tide level measured by a NOAA tide gauge located at the Battery, NY (NOAA station ID: 8,518,750; dashed blue line). The FloodNet sensor measurements were very similar to those from the NOAA tide gauge with a slight temporal delay due to being installed in different locations in New York Harbor. (b) FloodNet sensor measurements (black line) collected during a high tide flood in Hamilton Beach, Queens on 23 July 2021, compared with measurements collected intermittently by manually reading the depth value off of a standard ruler (blue data points and line).
(a) 受潮汐影响的布鲁克林戈阿纳斯运河（黑线）上方安装的 FloodNet 传感器收集的水位数据与位于纽约炮台的 NOAA 验潮仪（NOAA 站点 ID：8,518,750；蓝色虚线）测量的潮位数据的比较。FloodNet 传感器的测量结果与 NOAA 验潮仪的测量结果非常相似，只是由于安装在纽约港的不同位置而略有时间延迟。(b) FloodNet 传感器在 2021 年 7 月 23 日皇后区汉密尔顿海滩涨潮期间采集的测量值（黑线），与通过手动读取标准尺上的深度值间歇采集的测量值（蓝色数据点和线）进行比较。
Finally, the FloodNet sensors demonstrated accuracy in measuring water depths during an actual street-level flood event. A validation experiment was conducted during a high tide flood event in Hamilton Beach, Queens, on 23 July 2021. A standard measuring scale was installed on the same pole used to mount a FloodNet sensor, and flood depth measurements were manually recorded at regular time intervals during the flood, concurrent with flood sensor measurements. The FloodNet sensor readings were in agreement with the manual flood depth measurements, with an accuracy within a few millimeters (Figure 8).
最后，FloodNet 传感器展示了在实际街道洪水事件中测量水深的准确性。2021 年 7 月 23 日，在皇后区汉密尔顿海滩的一次涨潮洪水事件中进行了验证实验。在用于安装 FloodNet 传感器的同一根杆子上安装了一个标准测量秤，在洪水期间以固定时间间隔手动记录洪水深度测量值，同时进行洪水传感器测量。FloodNet 传感器读数与人工洪水深度测量结果一致，精度在几毫米之内（图 8）。
5.3 Network and Data Acquisition Performance
5.3 网络和数据采集性能
FloodNet sensors collect distance data every minute and transmit these data payloads to our servers via LoRaWAN. However, one factor affecting data transmission (also referred to as sensor uptime) is the signal strength between the sensor transmitting the data and the gateway receiving it. Low signal strength caused, for example, by a gateway being out of range, non-ideal topography between the sensor and gateway, or objects obstructing signal transmission (e.g., buildings, trees), can result in the loss of data payloads. The sensor design described herein does not store data on the device - a decision made to keep the sensor small and low-cost - and data can be lost if not received by the gateway.
FloodNet 传感器每分钟都会收集距离数据，并通过 LoRaWAN 将这些数据有效载荷传输到我们的服务器。然而，影响数据传输（也称为传感器正常运行时间）的一个因素是发送数据的传感器与接收数据的网关之间的信号强度。例如，由于网关超出范围、传感器和网关之间的地形不理想或物体（如建筑物、树木）阻碍信号传输等原因造成的信号强度低，都可能导致数据有效载荷丢失。本文所述的传感器设计不在设备上存储数据--这是为了保持传感器体积小、成本低而做出的决定--如果网关接收不到数据，数据就会丢失。
Much effort has been made in installing sensors and gateways in locations that ensure good signal strength. Nonetheless, FloodNet sensors deployed thus far have had transmission efficiencies of less than 100% (Figure 9), which is typical for LoRaWAN networks. For example, in analyzing the transmission efficiencies of 21 sensors that were installed and operating during the full six month period between October 2022 and April 2023, eight had transmission efficiencies that were greater than 75%, seven were between 50% and 75%, and the remaining six were between 35% and 46%. Given that the sensors collect and transmit data every minute, a transmission efficiency of greater than 50% is relatively good, and means that, on average, data packets sent by a sensor were received at a frequency of at least every other minute.
为了确保良好的信号强度，我们在安装传感器和网关方面做了大量工作。尽管如此，迄今为止部署的 FloodNet 传感器的传输效率仍低于 100%（图 9），这是 LoRaWAN 网络的典型情况。例如，在分析 2022 年 10 月至 2023 年 4 月整个 6 个月期间安装并运行的 21 个传感器的传输效率时，有 8 个传感器的传输效率超过 75%，7 个传感器的传输效率在 50% 至 75% 之间，其余 6 个传感器的传输效率在 35% 至 46% 之间。鉴于传感器每分钟收集和传输一次数据，传输效率超过 50%相对较好，这意味着平均至少每隔一分钟就能收到一个传感器发送的数据包。
Data transmission efficiencies of a subset of FloodNet sensors, showing percent of payload uploads per day, where 100% equals 60 successful uploads/hour over 24 hr. The percentage provided on the left hand side of the figure is the average daily transmission efficiency across all days in the period between October 2022 to April 2023. The 21 sensors included here are a subset of the total number of sensors installed thus far, and were selected because they were installed and operated during the full duration of the six month period from October 2022 to April 2023.
FloodNet 传感器子集的数据传输效率，显示每天有效载荷上传的百分比，其中 100% 相当于 24 小时内每小时成功上传 60 次。图中左侧提供的百分比是 2022 年 10 月至 2023 年 4 月期间所有天数的平均日传输效率。这里包含的 21 个传感器是迄今为止安装的传感器总数的一个子集，之所以选择这些传感器，是因为它们在 2022 年 10 月至 2023 年 4 月这六个月的整个期间都已安装并运行。
Poor sensor uptime can be caused by a variety of factors. First, if a sensor is observed to have consistently low upload rates, it is often caused by poor or inconsistent gateway coverage. This can be a function of pure distance between the sensor and the gateway - as seen for sensor 19 (Figure 9), which was located 1.8 km from the nearest gateway - or can be due to seasonal or environmental effects that could impede the signal, such as increased tree foliage in summer or high voltage electrical or heavy metal infrastructure located nearby. Sensor 18, for example, was mounted next to the underpass of a steel railway bridge, which impeded signal transmission.
传感器正常运行时间不足可能由多种因素造成。首先，如果观察到传感器的上传率一直很低，这通常是由于网关覆盖不佳或不一致造成的。这可能是传感器与网关之间的纯粹距离造成的，如传感器 19（图 9），它距离最近的网关有 1.8 千米，也可能是由于季节或环境影响可能会阻碍信号，如夏季树叶增多或附近有高压电或重金属基础设施。例如，传感器 18 就安装在一座钢结构铁路桥的地下通道旁，这阻碍了信号传输。
Second, downtime can manifest in an otherwise well-performing sensor as a sudden and complete loss of connectivity for a period of time, which can be caused by the temporary malfunction of a gateway or sensor. A gateway malfunction can cause an outage across multiple sensors that rely on that gateway for connectivity. For example, in February 2023, two gateway outages occurred, the first affecting sensors 8, 11, 12, and 14, and the second affecting sensors 9, 13, 16, and 18. During the first gateway outage, sensor 8 initially had a total outage, but was able to regain connectivity, albeit weaker in signal strength, with another gateway shortly after, and sensor 11 saw a drop in connectivity, but never completely lost connectivity due to proximity to another, further gateway. Improved gateway connectivity was gained through deployment of two new gateways in March 2023, first reconnecting sensors 8 and 12, and later reconnecting sensors 9, 11, 13, 14, 16, and 18. Additionally, vertical bands can be observed across all sensors in January and February 2023 related to short term system-wide outages caused by server infrastructure maintenance, when downtime was less than a day.
其次，宕机可能表现为原本性能良好的传感器在一段时间内突然完全失去连接，这可能是由网关或传感器的临时故障造成的。网关故障可导致依赖该网关进行连接的多个传感器出现故障。例如，2023 年 2 月发生了两次网关故障，第一次影响到传感器 8、11、12 和 14，第二次影响到传感器 9、13、16 和 18。在第一次网关断电期间，传感器 8 最初完全断电，但不久后又能与另一个网关恢复连接，尽管信号强度较弱；传感器 11 的连接性有所下降，但由于靠近另一个更远的网关，从未完全失去连接。通过在 2023 年 3 月部署两个新网关，网关连接得到改善，首先重新连接了 8 号和 12 号传感器，随后又重新连接了 9 号、11 号、13 号、14 号、16 号和 18 号传感器。此外，在 2023 年 1 月和 2 月，所有传感器都可以观察到垂直条带，这与服务器基础设施维护造成的短期全系统故障有关，当时停机时间不到一天。
To reduce the loss of data during transmission over LoRaWAN, one could use a stronger antenna on the sensor to improve signal strength and increase the packet success rate. For the next iteration of the FloodNet sensor, however, we have decided to pivot to the use of a cellular network for data transmission instead of LoRaWAN. In addition to improving the packet success rate, switching to a cellular network will alleviate other challenges associated with using LoRaWAN, including the need to identify and gain permission for gateway installation locations that have access to power, and gateway downtime due to power outages or equipment malfunctions. We also plan to implement onboard flash memory to store data on the sensor. In the event of a network disruption, the sensor will re-transmit stored data once the disruption is resolved, minimizing the loss of data packets.
为了减少数据在 LoRaWAN 传输过程中的丢失，我们可以在传感器上使用更强的天线，以提高信号强度和数据包成功率。不过，在 FloodNet 传感器的下一次迭代中，我们决定转而使用蜂窝网络进行数据传输，而不是 LoRaWAN。除了提高数据包成功率之外，改用蜂窝网络还能减轻使用 LoRaWAN 所面临的其他挑战，包括需要确定并获得许可才能在有电源的地方安装网关，以及网关因断电或设备故障而停机。我们还计划采用板载闪存来存储传感器上的数据。在网络中断的情况下，一旦中断解决，传感器将重新传输存储的数据，从而最大限度地减少数据包的丢失。
5.4 Data Filter Performance
5.4 数据过滤器性能
Between October 2020 and May 2023, a total of 6,641 events were recorded by the sensors (Table 4); each event is defined as a series of one or more depth readings greater than 10 mm. Of these, 360 were manually identified as flood events, 7 as snow events (i.e., snow accumulation during winter storms), and the remaining 6,274 as non-flood events. The recording of non-flood events (described in Section 10) resulted from the nature of the sensors being located over active streets and sidewalks in NYC, and included blips, boxes, pulse chains, and complex noise, the latter being a catch-all category for events that did not fit into the other three categories. Most of the non-flood events (74%) were identified as blips.
2020 年 10 月至 2023 年 5 月期间，传感器共记录了 6641 个事件（表 4）；每个事件的定义是一系列深度大于 10 毫米的一个或多个读数。其中，360 个事件被人工识别为洪水事件，7 个事件被识别为积雪事件（即冬季暴风雪期间的积雪），其余 6274 个事件被识别为非洪水事件。非洪水事件的记录（在第 10 节中描述）是由于传感器位于纽约市的活动街道和人行道上，包括突波、方框、脉冲链和复杂噪声，后者是不属于其他三个类别的事件的总括类别。大多数非洪水事件（74%）被识别为突波。
表 4.2020 年 10 月至 2023 年 5 月期间由 FloodNet 传感器记录的洪水、雪和非洪水事件计数，其中事件被定义为一系列一个或多个深度读数大于 10 毫米的事件
| Unfiltered data 未经过滤的数据 | After gradient filter (stage 2) 梯度滤波后（第 2 阶段） |
After all filter stages (stage 3) 在所有过滤阶段之后（第 3 阶段） |
|||
|---|---|---|---|---|---|
| Event type 活动类型 | Number events identified manually 人工识别的事件数量 |
Number non-flood events identified 确定的非洪水事件数量 |
Percent of total from unfiltered data 未过滤数据占总数的百分比 |
Number non-flood events identified 确定的非洪水事件数量 |
Percent of total from unfiltered data 未过滤数据占总数的百分比 |
| Flood | 360 | 0 | 0% | 0 | 0% |
| Snow | 7 | 0 | 0% | 0 | 0% |
| Blip | 4,641 | 990 | 21% | 4,298 | 93% |
| Box | 812 | 235 | 29% | 528 | 65% |
| Pulse Chain 脉冲链 | 597 | 58 | 10% | 193 | 32% |
| Complex Noise 复杂噪音 | 224 | 4 | 2% | 19 | 8% |
| Total Non-Flood Events 非洪水事件总数 | 6,634 | 1,287 | 19% | 5,038 | 76% |
-
Note. Total event numbers presented for the unfiltered data were manually counted. The numbers of non-flood events identified after the gradient filter stage and after application of all three filter stages are presented; the percentages provided are calculated as the number of events identified by the filters divided by the number of events identified in the unfiltered data. Zero non-flood events identified by the filters for flood and snow events means that there were no false negative results.
注。未过滤数据的事件总数由人工统计。在梯度滤波阶段和应用所有三个滤波阶段后识别出的非洪水事件数均已列出；所提供的百分比是以滤波识别出的事件数除以未滤波数据中识别出的事件数计算得出的。在洪水和雪灾事件中，过滤器识别出的非洪水事件为零，这意味着没有错误的负面结果。
As described in Section 10, we developed and implemented a series of automated data processing stages to filter non-flood events from the depth data, thereby removing false positive events from the data visualization and alerting platforms. As shown in Table 4, the use of all three filter stages described above resulted in an overall reduction in reported non-flood events of 76%, whereas only 19% of the non-flood events were captured after the gradient filter stage (i.e., using only the first two filter stages). Blip events were the most common non-flood events and the easiest to filter, with 92% of blips correctly classified. Boxes and pulse-chains occurred less frequently, but were more challenging to capture, as shown by the correct identification of 65% of boxes and 32% of pulse chains. All flood and snow events remained unfiltered, indicating that the filters did not result in false negative results.
如第 10 节所述，我们开发并实施了一系列自动数据处理阶段，以从深度数据中过滤非洪水事件，从而从数据可视化和警报平台中去除误报事件。如表 4 所示，使用上述全部三个过滤阶段后，报告的非洪水事件总体减少了 76%，而梯度过滤阶段（即仅使用前两个过滤阶段）后捕获的非洪水事件仅占 19%。闪烁事件是最常见的非洪水事件，也是最容易过滤的事件，92% 的闪烁事件被正确分类。方框和脉冲链出现的频率较低，但捕捉难度较大，65% 的方框和 32% 的脉冲链被正确识别。所有洪水和降雪事件均未被过滤，这表明过滤结果不会出现假阴性。
There are some challenges to automatically distinguishing flood events from non-flood events, including having sparse data points at times for some sensors. More specifically, when data are not received from the sensor regularly (e.g., every minute), it is difficult to distinguish a possible flood from a non-flood event. For example, with frequent data points, large gradients of flood depth per time can be readily identified as unlikely coming from a flood event. However if the data points are separated by long or irregular time intervals, a large increase in depth that is physically unlikely at a 1 min resolution could be possible at a five or 10 min resolution because the large instantaneous gradient is now spread over a longer period of time. In an extreme case, if only one measurement is received over the duration of a flood, it could appear as just a blip, which is why blip and box filters are both limited by the length of the time window they are allowed to operate on; this challenge will be alleviated with engineering improvements to increase data transmission.
自动区分洪水事件和非洪水事件存在一些挑战，包括某些传感器的数据点有时比较稀疏。更具体地说，当传感器没有定期（如每分钟）接收数据时，就很难区分可能发生的洪水和非洪水事件。例如，如果数据点的频率很高，那么每次洪水深度的较大梯度很容易被识别为不可能来自洪水事件。但是，如果数据点之间的时间间隔较长或不规则，那么在 1 分钟分辨率下物理上不可能出现的深度大幅增加，在 5 或 10 分钟分辨率下就有可能出现，因为此时较大的瞬时梯度分布在较长的时间段内。在极端情况下，如果在洪水持续时间内只接收到一次测量结果，那么它可能只是一个突波，这就是突波滤波器和方框滤波器都受到允许其工作的时间窗口长度限制的原因；随着工程技术的改进，数据传输量的增加，这一挑战将得到缓解。
Additionally, there are events that do not match the aforementioned non-flood or flood event patterns. For example, some non-flood events have two points that meet neither the characteristics of a box nor a blip, some boxes don’t start at zero and are not identified because the box filter requires the first point to have a depth of zero, and sometimes the tops of boxes have a large variance or have boxes or noise on top of boxes, both of which would bypass the box filter.
此外，还有一些事件不符合上述非洪水或洪水事件模式。例如，有些非洪水事件的两个点既不符合方框的特征，也不符合突波的特征；有些方框的起点不是零，没有被识别出来，因为方框过滤器要求第一个点的深度为零；有时方框的顶部差异很大，或者方框顶部有方框或噪声，这两种情况都会绕过方框过滤器。
Our goal is to release data to stakeholders as close to real-time as possible, ideally within 5–10 min of the start of a flood. Therefore, frequent data collection (≈1 min interval) is needed for the filters to be able to rapidly recognize flood events. Additionally, we are conducting ongoing research to improve the data filters' ability to distinguish floods from non-flood events. One avenue being explored is training machine learning models to identify the patterns of flood event data, with a goal of reducing the number of false positive floods reported for non-flood events, which is important for a reliable public flood alert system. Additionally this will help in post-hoc flood labeling and boundary regression allowing computation of flood event statistics.
我们的目标是向利益相关者发布尽可能接近实时的数据，最好是在洪水开始后 5-10 分钟内。因此，需要频繁收集数据（间隔时间≈1 分钟），以便过滤器能够快速识别洪水事件。此外，我们正在进行研究，以提高数据过滤器区分洪水和非洪水事件的能力。目前正在探索的一个途径是训练机器学习模型来识别洪水事件数据的模式，目的是减少非洪水事件中报告的假阳性洪水的数量，这对于建立可靠的公共洪水警报系统非常重要。此外，这将有助于事后洪水标记和边界回归，从而计算洪水事件统计数据。
5.5 Community Engagement With Flood Sensor Data
5.5 社区参与获取洪水传感器数据
Community engagement is an important component of the FloodNet project, given our goal to meaningfully collect and share data that can contribute to flood risk mitigation and community flood resilience. FloodNet's ongoing outreach and collaboration with local community members has assisted with project design and implementation. In addition, it is through community engagement activities that we have learned that flood sensor data have the potential to support community residents' decision making when faced with chronic floods; for example, a sensor placed on an important roadway that floods semi-monthly has helped residents know if it is passable in real time. In addition, the sensor data offer opportunities to assist in increasing residents' understanding of how flooding impacts their neighborhoods. For example, community members who participated in a focus group on FloodNet data and public engagement noted that sharing summary flood data with their neighbors could be a good way to communicate whether they live in a flood zone, in what time of year floods are most likely to occur, and what resources are available to prepare for flood events. Through presentations, community meetings, educational workshops, community walk-throughs, and community-collected feedback on flood sensor placement, we have also heard a desire among community members to use flood sensor data to validate community experiences of flooding in the eyes of elected officials and other people in power. This can serve to support action plans and advocacy that connect flooding to other relevant community issues in flood-prone areas. Given the various potential uses of flood data, a goal for future research is to assess how community residents actually use the data when it is available.
社区参与是 FloodNet 项目的一个重要组成部分，因为我们的目标是有意义地收集和共享有助于降低洪水风险和提高社区抗洪能力的数据。FloodNet 与当地社区成员的持续外联与合作有助于项目的设计和实施。此外，通过社区参与活动，我们了解到洪水传感器数据有可能在社区居民面临长期洪灾时为他们的决策提供支持；例如，在每半个月洪水泛滥一次的重要道路上安装传感器，可以帮助居民实时了解道路是否可以通行。此外，传感器数据还能帮助居民更好地了解洪水对社区的影响。例如，参加 FloodNet 数据和公众参与焦点小组的社区成员指出，与邻居分享洪水数据摘要是一种很好的方式，可以让他们了解自己是否居住在洪水区、洪水最有可能在一年中的哪个时间段发生以及有哪些资源可以为洪水事件做好准备。通过演讲、社区会议、教育研讨会、社区走访以及社区收集的有关洪水传感器安置的反馈意见，我们还听到社区成员希望使用洪水传感器数据来验证民选官员和其他当权者眼中的洪水社区经验。这有助于支持行动计划和宣传活动，将洪水与洪水易发地区的其他相关社区问题联系起来。鉴于洪水数据的各种潜在用途，未来研究的一个目标是评估社区居民在获得数据后的实际使用情况。
5.6 Flood Sensor Data Examples: Hurricane Ida and Tidal Flooding
5.6 洪水传感器数据示例：伊达飓风和潮汐洪水
Even with ongoing upgrades to the sensor design, network performance, and data analysis pipeline, the FloodNet sensors deployed thus far have been able to collect a rich dataset, capturing the profiles of pluvial and tidal floods in NYC neighborhoods. The arrival of the remnants of Hurricane Ida in NYC on 1 September 2021 was a landmark event, with record rainfall (up to 79 mm/hr recorded in Central Park) resulting in unprecedented flooding across NYC. All three FloodNet sensors deployed at the time in the Gowanus neighborhood in Brooklyn measured flood profiles during the storm, with the most extreme flooding occurring at the intersection of Carroll Street and 4th Avenue (Figure 10). A peak flood depth of 890 mm above the sidewalk was recorded (water depths in the roadway were deeper), as well as a rapid rate of onset: up to 91 mm/min for the first 5 min of the flood. See Silverman et al. (2022) for a description of flood data collected by the same sensor during Tropical Storm Henri on 21 August 2021.
即使传感器设计、网络性能和数据分析管道不断升级，迄今为止部署的 FloodNet 传感器仍能收集到丰富的数据集，捕捉到纽约市街区的冲积洪水和潮汐洪水概况。2021 年 9 月 1 日，飓风 "艾达 "的残余物抵达纽约市，这是一个具有里程碑意义的事件，创纪录的降雨量（中央公园的降雨量高达 79 毫米/小时）导致整个纽约市发生了前所未有的洪灾。当时部署在布鲁克林 Gowanus 社区的所有三个 FloodNet 传感器都测量到了暴风雨期间的洪水剖面，最严重的洪水发生在卡罗尔街和第四大道的交叉口（图 10）。记录到人行道上方的洪水峰值深度为 890 毫米（路面的水深更深），洪水发生的速度也很快：在洪水发生的前 5 分钟内，洪水速度高达 91 毫米/分钟。有关 2021 年 8 月 21 日热带风暴亨瑞期间同一传感器收集的洪水数据，请参见 Silverman 等人 ( 2022 年)。
Examples of flood data collected by FloodNet sensors: (a) Data collected by the sensor located at the intersection of Carroll Street and 4th Avenue in Brooklyn during Hurricane Ida (1 September 2021). The black line represents the flood depth data (secondary y-axis); blue bars represent rainfall intensity (primary y-axis; precipitation data was sourced from a rain gauge at a New York State Mesonet (Brotzge et al., 2020) managed weather station at the Brooklyn Navy Yard (ID: bknyrd) located ≈ 2.7 km from the sensor). The flood sensor is located over a sidewalk, therefore flood depths were greater in the adjacent roadway. (b) Data collected by the sensor located on Beach 84th Street in Rockaway, Queens over the course of 18 months (December 2021–May 2023). The Black line represents the flood depth data (secondary y-axis); the blue line (primary y-axis) represents the tide height above the mean higher high water (MHHW) level. Tide data are from NOAA for the North Channel, NY tide station (NOAA station ID: 8,517,201). (c) Same data as in Panel B, but zoomed in to a week period from September 5 to 14 September 2022, when high tide flooding caused by the full moon on September 10 resulted in 13 consecutive flood events, with floods occurring every 12–24 hr.
FloodNet 传感器收集的洪水数据示例：(a) 位于布鲁克林卡罗尔街与第四大道交叉口的传感器在飓风 "艾达"（2021 年 9 月 1 日）期间收集的数据。黑线代表洪水深度数据（二级 y 轴）；蓝条代表降雨强度（一级 y 轴；降雨数据来自布鲁克林海军船坞（ID：bknyrd）距离传感器 ≈ 2.7 千米处的纽约州中线网（Brotzge 等人，2020 年）管理的气象站的雨量计）。洪水传感器位于人行道上方，因此邻近道路的洪水深度更大。(b) 位于皇后区罗卡韦海滩第 84 街的传感器在 18 个月内（2021 年 12 月至 2023 年 5 月）收集的数据。黑线代表洪水深度数据（二级 y 轴）；蓝线（一级 y 轴）代表高于平均高潮水位 (MHHW) 的潮汐高度。潮汐数据来自 NOAA 的纽约州北海峡潮汐站（NOAA 站 ID：8,517,201）。(c) 与面板 B 中的数据相同，但放大到 2022 年 9 月 5 日至 9 月 14 日的一周内，9 月 10 日满月引起的高潮洪水导致 13 次连续洪水事件，每 12-24 小时发生一次洪水。
Flood events have been measured more frequently by sensors installed in coastal areas susceptible to regular tidal flooding. The northernmost block of Beach 84th Street on the Rockaway peninsula in Queens is one such example (Figure 10). In the 17 months between December 2021, when the sensor was installed, and April 2023, the sensor located on this block measured 121 distinct flood events, each between 29 and 560 mm in depth, during relatively high, high-tide events, which occur on a semi-monthly basis, typically coinciding with the full moon or new moon. For example, in September 2022, the FloodNet sensor recorded 13 distinct flood events on 7 consecutive days, following the approximately 12-hr period of the tidal cycle. Other FloodNet sensors located in additional coastal areas have measured a similar degree of flooding.
安装在易受定期潮汐洪水影响的沿海地区的传感器可以更频繁地测量洪水事件。皇后区洛克威半岛第 84 街海滩最北端的街区就是这样一个例子（图 10）。从 2021 年 12 月安装传感器到 2023 年 4 月的 17 个月中，位于该区块的传感器测量到 121 次不同的洪水事件，每次洪水深度在 29 毫米到 560 毫米之间，发生在相对较高的涨潮事件期间，涨潮事件每半月发生一次，通常与满月或新月同时发生。例如，在 2022 年 9 月，FloodNet 传感器连续 7 天记录到 13 次不同的洪水事件，潮汐周期约为 12 小时。位于其他沿海地区的其他 FloodNet 传感器也测量到了类似程度的洪水。
6 Conclusions and Future Work
6 结论和未来工作
The design of the FloodNet sensor allows accurate measurement of street-level floods as shallow as 10 mm, at a frequency of one measurement per minute under optimal conditions. The associated sensor network transmits measured data to our servers and the FloodNet data analysis pipeline, which enables flood data visualization and provision of alerts to users in near real-time.
FloodNet 传感器的设计允许在最佳条件下以每分钟一次的频率精确测量最浅 10 毫米的街道水位。相关的传感器网络将测量数据传输到我们的服务器和 FloodNet 数据分析管道，从而实现洪水数据的可视化，并近乎实时地向用户发出警报。
Given the success of FloodNet sensors in monitoring floods thus far, and the potential utility of flood sensor data (Silverman et al., 2022), we plan to expand the sensor network in NYC. To accomplish the expansion, ongoing research is being conducted to iterate the sensor design to make it more manufacturable for scale-up, to improve data transmission rates through the use of a cellular network, and to update the data analysis pipeline to improve the detection and logging of flood events. All sensor design files are located on the FloodNet Github page (FloodNet, 2024b) (https://github.com/floodnet-nyc/flood-sensor), which is updated as new designs are implemented.
鉴于 FloodNet 传感器迄今为止在监测洪水方面取得的成功，以及洪水传感器数据的潜在效用（Silverman 等人，2022 年），我们计划在纽约市扩展传感器网络。为了实现扩展，我们正在进行研究，以不断改进传感器的设计，使其更易于制造以扩大规模，通过使用蜂窝网络提高数据传输速率，并更新数据分析管道以改进洪水事件的检测和记录。所有传感器设计文件都位于 FloodNet Github 页面上（FloodNet, 2024b）（https://github.com/floodnet-nyc/flood-sensor），该页面会随着新设计的实施而更新。
A variety of stakeholders, including NYC residents, are being consulted as we iterate the design of the FloodNet web-based data dashboard and other data sharing tools, such as printable data sharing reports, to improve the meaningful sharing and communication of collected flood data. Additionally, given the thousands of locations that are at risk of flooding in NYC (New York City Stormwater Resiliency Plan, 2021), it is unlikely that a flood sensor network will be distributed enough to monitor every one. As such - and given the relevance of the data to municipalities for uses such as emergency management, urban planning, decision making for capital improvement projects and other resource allocation - an additional line of ongoing research is to create a risk and equity based framework to help prioritize flood sensor deployment locations in a systematic and equitable way. In addition, open questions exist related to how flood sensor data can be incorporated to hydrologic and hydraulic models to improve flood prediction and risk assessment.
在我们不断改进 FloodNet 网络数据仪表板和其他数据共享工具（如可打印的数据共享报告）的设计过程中，我们征求了包括纽约市居民在内的各利益相关方的意见，以改进所收集洪水数据的有意义的共享和交流。此外，鉴于纽约市面临洪水风险的地点数以千计（纽约市 2021 年雨水恢复计划），洪水传感器网络的分布不可能足以监测到每一个地点。因此，考虑到这些数据与市政当局在应急管理、城市规划、资本改善项目决策和其他资源分配等方面的相关性，目前正在进行的另一项研究是创建一个基于风险和公平的框架，以帮助以系统和公平的方式确定洪水传感器部署地点的优先次序。此外，在如何将洪水传感器数据纳入水文和水力模型以改进洪水预测和风险评估方面，还存在一些未决问题。
In conclusion, FloodNet sensors were specifically designed to overcome challenges related to measuring floods in a distributed manner across a complex urban environment, and have a demonstrated ability to collect street-level flood data, regardless of the flood typology (e.g., pluvial, fluvial, tidal, storm surge, infrastructure-related). While the FloodNet project is based in NYC, the sensor design is flexible for other locations and contexts, with an open-source design available for others who would like to build and deploy flood sensors in their own communities.
总之，FloodNet 传感器是专门为克服在复杂的城市环境中以分布式方式测量洪水所面临的挑战而设计的，并已证明有能力收集街道级洪水数据，无论洪水类型如何（如冲积、河川、潮汐、风暴潮、基础设施相关）。虽然 FloodNet 项目以纽约市为基地，但传感器的设计非常灵活，可适用于其他地点和环境，并为其他希望在自己的社区建立和部署洪水传感器的人提供开源设计。
Acknowledgments 致谢
Funding for this work was provided by the New York City Department of Environmental Protection, the Connected Cities for Smart Mobility towards Accessible and Resilient Transportation (C2SMART) University Center (U.S. Department of Transportation award number 69A3351747124), the Alfred P. Sloan Foundation, the New York State Empire State Development’s New York Smart Cities Innovation Partnership, and seed grants from the NYU Marron Institute of Urban Management and the CUNY Office of Research. Rainfall and temperature data were provided by the New York State Mesonet (NYSM). Original funding for NYSM buildup was provided by Federal Emergency Management Agency Grant FEMA-4085-DR-NY; the continued operation and maintenance of NYSM is supported by National Mesonet Program, University at Albany, Federal and private grants, and others. We thank colleagues at the City of New York for feedback and assistance in the development and installation of the sensor network, and our community partners for valuable insight regarding their experiences with flooding. We also thank student researchers at the high school, undergraduate, and graduate levels for assisting in aspects of this work. There are no conflicts of interest to declare.
这项工作的资金由纽约市环境保护局、"互联城市智能交通，实现无障碍和弹性交通（C2SMART）大学中心"（美国交通部奖励编号 69A3351747124）、Alfred P. Sloan 基金会、纽约州帝国发展部的纽约智能城市创新合作项目以及纽约大学马龙城市管理研究所和纽约市立大学研究办公室的种子基金提供。降雨量和温度数据由纽约州中间网（NYSM）提供。纽约州气象站的最初建设资金由联邦紧急事务管理局 FEMA-4085-DR-NY 拨款提供；纽约州气象站的持续运行和维护得到了国家气象站计划、奥尔巴尼大学、联邦和私人拨款以及其他方面的支持。我们感谢纽约市政府的同事在传感器网络的开发和安装过程中提供的反馈和帮助，并感谢我们的社区合作伙伴在洪灾经验方面提供的宝贵见解。我们还要感谢高中、本科和研究生阶段的学生研究人员对本研究工作提供的帮助。没有利益冲突需要声明。
