Landscape Change Monitoring System (LCMS) Alaska Annual QA Bits

This QA bit product is part of the Landscape Change Monitoring System (LCMS) data suite. It provides information about each pixel of the annual composites that are used as inputs to LandTrendr data used in the model. This information includes whether the data value is an observation or is interpolated, the Landsat sensor that observed that data value, and the Julian day of that observation. See additional information about QA in the Entity_and_Attribute_Information or Fields section below. LCMS is a remote sensing-based system for mapping and monitoring landscape change across the United States. Its objective is to develop a consistent approach using the latest technology and advancements in change detection to produce a "best available" map of landscape change. Because no algorithm performs best in all situations, LCMS uses an ensemble of models as predictors, which improves map accuracy across a range of ecosystems and change processes (Healey et al., 2018). The resulting suite of LCMS Change, Land Cover, and Land Use maps offer a holistic depiction of landscape change across the United States over the past four decades. Predictor layers for the LCMS model include outputs from the LandTrendr and CCDC change detection algorithms and terrain information. These components are all accessed and processed using Google Earth Engine (Gorelick et al., 2017). To produce annual composites, the cFmask (Zhu and Woodcock, 2012), cloudScore, Cloud Score + (Pasquarella et al., 2023), and TDOM (Chastain et al., 2019) cloud and cloud shadow masking methods are applied to Landsat Tier 1 and Sentinel 2a and 2b Level-1C top of atmosphere reflectance data. The annual medoid is then computed to summarize each year into a single composite. The composite time series is temporally segmented using LandTrendr (Kennedy et al., 2010; Kennedy et al., 2018; Cohen et al., 2018). All cloud and cloud shadow free values are also temporally segmented using the CCDC algorithm (Zhu and Woodcock, 2014). LandTrendr, CCDC and terrain predictors can be used as independent predictor variables in a Random Forest (Breiman, 2001) model. LandTrendr predictor variables include fitted values, pair-wise differences, segment duration, change magnitude, and slope. CCDC predictor variables include CCDC sine and cosine coefficients (first 3 harmonics), fitted values, and pairwise differences from the Julian Day of each pixel used in the annual composites and LandTrendr. Terrain predictor variables include elevation, slope, sine of aspect, cosine of aspect, and topographic position indices (Weiss, 2001) from the USGS 3D Elevation Program (3DEP) (U.S. Geological Survey, 2019). Reference data are collected using TimeSync, a web-based tool that helps analysts visualize and interpret the Landsat data record from 1984-present (Cohen et al., 2010). Outputs fall into three categories: Change, Land Cover, and Land Use. Change relates specifically to vegetation cover and includes slow loss (not included for PRUSVI), Fast Loss (which also includes hydrologic changes such as Inundation or Desiccation), and Gain. These values are predicted for each year of the time series and serve as the foundational products for LCMS.本QA位（QA bit）产品隶属于景观变化监测系统（Landscape Change Monitoring System, LCMS）数据集套件。其提供了作为模型中LandTrendr数据输入的年度合成影像各像素的相关信息，具体包括：数据值为实测观测值还是插值生成值、获取该数据值的Landsat传感器型号，以及该次观测的儒略日。有关QA的更多详情，请参见下文的Entity_and_Attribute_Information或Fields章节。LCMS是一套基于遥感技术的美国全境景观变化制图与监测系统，其目标是依托最新技术与变化检测领域的研究进展，构建一套「最优可用」的景观变化制图方案。由于不存在在所有场景下均表现最优的算法，LCMS采用多模型集成作为预测器，这一策略有效提升了不同生态系统与变化过程下的制图精度（Healey et al., 2018）。最终产出的LCMS变化、土地覆盖与土地利用系列地图，完整呈现了美国全境过去四十余年的景观变化全貌。LCMS模型的预测变量层包括LandTrendr与CCDC变化检测算法的输出结果，以及地形信息。所有这些组件均通过Google Earth Engine（Gorelick et al., 2017）进行调取与处理。为生成年度合成影像，研究团队将cFmask（Zhu and Woodcock, 2012）、cloudScore、Cloud Score +（Pasquarella et al., 2023）以及TDOM（Chastain et al., 2019）等云与云阴影掩膜方法，应用于Landsat Tier 1以及Sentinel 2a、2b Level-1C级大气顶部反射率数据。随后通过计算年度中值点（medoid），将每一年的观测数据汇总为单幅合成影像。合成影像时间序列通过LandTrendr算法进行时间分段（Kennedy et al., 2010; Kennedy et al., 2018; Cohen et al., 2018）。所有无云与云阴影的像素值也通过CCDC算法进行时间分段（Zhu and Woodcock, 2014）。LandTrendr、CCDC与地形预测变量均可作为独立变量应用于随机森林（Random Forest, Breiman, 2001）模型。其中LandTrendr预测变量包括拟合值、两两差值、分段时长、变化幅度与变化斜率；CCDC预测变量包括CCDC的正弦与余弦系数（前3次谐波）、拟合值，以及年度合成影像与LandTrendr所用各像素观测儒略日的两两差值。地形预测变量包括美国地质调查局3D高程计划（3D Elevation Program, 3DEP, U.S. Geological Survey, 2019）提供的高程、坡度、坡向正弦值、坡向余弦值，以及地形位置指数（Weiss, 2001）。参考数据通过TimeSync工具采集，这是一款基于网页的工具，可帮助分析人员可视化并解读1984年至今的Landsat数据序列（Cohen et al., 2010）。最终产出分为三大类：变化、土地覆盖与土地利用。其中变化类特指植被覆盖相关变化，包括缓慢退化（PRUSVI数据集未包含该类）、快速退化（该类同时涵盖洪涝、干涸等水文变化）与增益。上述数值针对时间序列的每一年进行预测，是LCMS的核心基础产品。