Exploring discrepancies between in situ phenology and remotely derived phenometrics at NEON sites
Exploring discrepancies between in situ phenology and remotely derived phenometrics at NEON sites
复制标题
DOI：
10.1002/ecs2.3912
复制
发表时间：
2022-01
期刊：
影响因子：
2.7
通讯作者：
Alison Donnelly;Rong Yu;Katherine D. Jones;M. Belitz;Bonan Li;Katharyn A. Duffy;Xiaoyang Zhang;Jianmin Wang;B. Seyednasrollah;Kathy Gerst;Daijiang Li;Y. Kaddoura;K. Zhu;J. Morisette;Colette A. Ramey;Kathleen Smith
中科院分区：
文献类型：
--
作者：
Alison Donnelly;Rong Yu;Katherine D. Jones;M. Belitz;Bonan Li;Katharyn A. Duffy;Xiaoyang Zhang;Jianmin Wang;B. Seyednasrollah;Kathy Gerst;Daijiang Li;Y. Kaddoura;K. Zhu;J. Morisette;Colette A. Ramey;Kathleen Smith
近几十年来，使用卫星传感器、近地表摄像机和其他远程方法来监测景观和更高尺度的植被物候已变得越来越普遍。这些技术提供了一种在不同空间分辨率下确定物候期时间和生长季节长度的方法；人类观察者无法达到的覆盖范围。然而，需要进行实地地面观测来验证远程导出的表观测量。尽管知识和专业知识有所增加，但将单个植物水平的地面观测与景观或更大规模的遥感（RS）表观测量相协调仍然存在持续的挑战。我们将原位物候估计（春季和秋季）的时间与美国 NEON 五个陆地地点（哈佛森林 [MA] [HARV]、奥纳基 [UT] [ONAQ]、艾比路 [WA] [ABBY]、迪士尼荒野保护区的一系列相应遥感（中分辨率成像光谱辐射计 [MODIS]、可见红外成像辐射计套件 [VIIRS]、PhenoCam）表观测量进行了比较[FL] [DSNY] 和 Ordway-Swisher 生物站 [FL] [OSBS]）重点关注 2017 年至 2019 年的 3 年期间。我们的主要目标是探索原位和遥感表象测量之间观测到的差异的潜在原因，并确定哪些技术能够更好地捕获地面观测结果。春季物候期的统计显着性关系最强（p< 0.001），而唯一与秋季物候期原位估计显着相关的RS表观数据是落叶（p< 0.01）和叶片（p< 0.000）。特别是，MODIS增强植被指数-2波段（EVI2）、VIIRS-EVI2和PhenoCam-green色度坐标（GCC）的均方根误差（RMSE）（平均偏差误差[MBE]）得出的早春过渡日期表明总体差异为21.7天（-4.6天）、28.4天（-1.2天）和早期抽叶日期的现场估计为 24.1 天（11.9 天）。秋季，与来自 MODIS（13.5 天/7.7 天）和 PhenoCam（GCC 绿度下降）的等效表观测量相比，原位落叶最新日期的 phenesse 估计值（第 95 个百分位数日期）和 VIIRS 得出的衰老结束之间的 RMSE/MBE 最小（10.9 天/−2.2 天）。 （13.8 天/−5.1 天）。总体而言，原位物候与遥感物候之间的差异与规模、物种可用性和时间序列的持续时间较短（3 年）有关。然而，随着 NEON 项目的进展，随着更多数据的可用，这些挑战预计将减少。
In recent decades, the use of satellite sensors, near‐surface cameras, and other remote methods for monitoring vegetation phenology at landscape and higher scales has become increasingly common. These technologies provide a means to determine the timing of phenophases and growing season length at different spatial resolutions; coverage that is not attainable by human observers. However, in situ ground observations are required to validate remotely derived phenometrics. Despite increased knowledge and expertise there still remains the persistent challenge of reconciling ground observations at the individual plant level with remotely sensed (RS) phenometrics at landscape or larger scales. We compared the timing of in situ phenophase estimates (spring and autumn) with a range of corresponding remote sensing (moderate resolution imaging spectroradiometer [MODIS], visible infrared imaging radiometer suite [VIIRS], PhenoCam) phenometrics across five terrestrial sites in the United States' NEON (Harvard Forest [MA] [HARV], Onaqui [UT] [ONAQ], Abby Road [WA] [ABBY], Disney Wilderness Preserve [FL] [DSNY], and Ordway‐Swisher Biological Station [FL] [OSBS]) focusing on the 3‐year period from 2017 to 2019. Our main objective was to explore potential reasons for the observed discrepancies between in situ and RS phenometrics and to determine which technologies were better able to capture ground observations. Statistically significant relationships were strongest (p< 0.001) for spring phenophases, while the only RS phenometrics significantly correlated with in situ estimates of autumn phenophases were leaf fall (p< 0.01) and leaves (p< 0.000). In particular, root mean square error (RMSE) (mean bias error [MBE]) for MODIS‐Enhanced Vegetation Index‐2 band (EVI2), VIIRS‐EVI2, and PhenoCam‐green chromatic coordinate (GCC) derived early spring transition dates indicated overall differences of 21.7 days (−4.6 days), 28.4 days (−1.2 days), and 24.1 days (11.9 days) from in situ estimates of early leaf‐out dates. In autumn, RMSE/MBE was smallest (10.9 days/−2.2 days) between phenesse estimates (95th percentile date) of the latest date of in situ leaf fall and VIIRS derived end of senescence, compared to the equivalent phenometric derived from MODIS (13.5 days/7.7 days) and PhenoCam (GCC greenness‐falling) (13.8 days/−5.1 days). Overall, discrepancies between in situ and RS phenology related to scale, species availability, and the short duration of the time series (3 years). However, as the NEON project progresses these challenges are expected to be reduced as more data become available.