LAGOS-US RESERVOIR: Data module classifying conterminous U.S. lakes 4 hectares and larger as natural lakes or reservoirs. Environmental Data Initiative

This data package, LAGOS-US RESERVOIR, is one of the extension modules of the LAGOS-US platform for studying lakes in the United States. Although naturally-formed lakes and reservoirs are thought to differ in many properties, there is currently no data source that differentiates between lakes and reservoirs in the conterminous US. This absence of data stems from how challenging it is to identify reservoirs at broad scales -- there is a wide variety of dam types and sizes that results in various reservoir shapes and sizes, making a simple classification difficult. Furthermore, reservoirs are understudied compared to natural lakes. The LAGOS-US RESERVOIR data module fills these data and the resulting knowledge gaps by classifying lakes greater than or equal to 4 hectares in the conterminous US (137,465 lakes) into one of two classes: natural lakes (NL) or reservoirs (RSVR). We define RSVRs (using visual interpretation of imagery) as lakes that are likely to be either human-made or highly human-altered by the presence of a relatively large water control structure that significantly changes the flow of water. We define NLs (using visual interpretation of imagery) as lakes that are likely to be either naturally-formed or do not have a relatively large, apparently flow-altering structure on or near it. The RSVR and NL classification is based on high resolution imagery and model predictions. We trained machine learning models using 12,127 manually (i.e., visually) classified lakes. When then used these models to assign NL or RSVR predictions to the remaining 77,604 NLs and 59,861 RSVRs. RESERVOIR also includes model-based prediction probabilities and variables that are commonly used when studying reservoirs (e.g., lake shape). These data can be used for studying reservoirs at the regional to conterminous US scale.本数据集套件LAGOS-US RESERVOIR是面向美国湖泊研究的LAGOS-US平台的扩展模块之一。尽管天然形成的湖泊与水库在诸多属性上存在差异，但目前美国本土范围内尚无能够区分湖泊与水库的数据源。此类数据缺失源于大范围尺度下识别水库极具挑战性：水库的坝体类型与尺寸多样，进而造就了各异的水库形态与规模，使得简单分类难以实现。此外，相较于天然湖泊，水库的相关研究仍较为匮乏。LAGOS-US RESERVOIR数据模块填补了这一数据与认知空白：将美国本土范围内面积≥4公顷的137465个湖泊划分为两类：天然湖泊（Natural Lakes, NL）与水库（Reservoirs, RSVR）。本研究通过影像目视解译，将RSVR（水库）定义为极有可能为人工建造，或因存在可显著改变水流的大型控水设施而被高度人工改造的湖泊；将NL（天然湖泊）定义为极有可能为天然形成，或其本体及周边不存在可明显改变水流的大型设施的湖泊。该分类结果基于高分辨率影像与机器学习模型预测生成。研究团队利用12127个经人工（即目视）分类的湖泊训练机器学习模型，随后将训练好的模型应用于剩余的77604个天然湖泊与59861个水库，为其分配NL或RSVR分类标签。此外，LAGOS-US RESERVOIR数据集还包含基于模型生成的预测概率，以及水库研究中常用的各类变量（如湖泊形态参数）。该数据集可用于美国本土区域乃至全国尺度下的水库相关研究。