Sentinel-1 and Sentinel-2 based frequency of open and vegetated water across the United States (2017-2021)

High-frequency observations of surface water at fine spatial scales are critical to effectively manage aquatic habitat, flood risk and water quality. We developed inundation algorithms for Sentinel-1 and Sentinel-2 across 12 sites within the conterminous United States (CONUS) covering greater than 536,000 km2 and representing diverse hydrologic and vegetation landscapes. These algorithms were trained on data from 13,412 points spread throughout the 12 sites. Each scene in the 5-year (2017-2021) time series was classified into open water, vegetated water, and non-water at 20 m resolution using variables not only from Sentinel-1 and Sentinel-2, but also variables derived from topographic and weather datasets. The Sentinel-1 model was developed distinct from the Sentinel-2 model to enable the two time series to be integrated into a single high-frequency time series, while open water and vegetated water were both mapped to retain mixed pixel inundation. Results were validated against 7,200 visually inspected points derived from WorldView and PlanetScope imagery. Classification accuracy for open water was high across the 5-year period, with an omission and commission error of only 3.1% and 0.9% for Sentinel-1 and 3.1% and 0.5% for Sentinel-2, respectively. Vegetated water accuracy was lower, as expected given that the class represents mixed pixels. Sentinel-2 showed higher accuracy (10.7% omission and 7.9% commission error) relative to Sentinel-1 (28.4% omission and 16.0% commission error). Our results demonstrated that Sentinel-1 and Sentinel-2 time series can be integrated to improve the temporal resolution when mapping open and vegetated waters, although sensor-specific differences, such as sensitivity to vegetation structure versus pixel color, complicate the data integration for subpixel, vegetated water compared with open water.

精细空间尺度下的高频地表水观测，对于有效管控水生栖息地、洪涝风险与水环境质量至关重要。本研究针对美国本土（conterminous United States, CONUS）内12个覆盖面积超53.6万平方千米、涵盖多样水文与植被景观的研究区域，开发了适用于哨兵-1号（Sentinel-1）与哨兵-2号（Sentinel-2）的水体淹没提取算法。上述算法以分布于12个研究区域内的13412个采样点数据作为训练集。本研究以20米空间分辨率，针对2017-2021年共5年的时间序列影像，不仅利用哨兵-1号与哨兵-2号的观测变量，还结合了地形与气象数据集衍生的特征变量，将每景影像划分为开阔水体、植被覆盖水体与非水体三类。为实现两类传感器时间序列的融合以生成单一高频时间序列，本研究分别构建了哨兵-1号与哨兵-2号专属模型；同时针对开阔水体与植被覆盖水体开展分类，以保留混合像元的淹没信息。研究结果通过7200个基于世界视图（WorldView）与行星观测（PlanetScope）影像目视解译得到的采样点完成验证。5年监测期间，开阔水体的分类精度表现优异：哨兵-1号模型的漏检率与误检率仅分别为3.1%与0.9%，哨兵-2号模型则分别为3.1%与0.5%。如预期所示，植被覆盖水体的分类精度相对更低——这是由于该类别对应混合像元。相较于哨兵-1号模型（漏检率28.4%、误检率16.0%），哨兵-2号模型的分类精度更高（漏检率10.7%、误检率7.9%）。本研究结果表明，尽管相较于开阔水体，植被覆盖水体的亚像元级数据融合会因传感器特异性差异（例如对植被结构的敏感度与对像元色彩的敏感度存在差异）而更为复杂，但融合哨兵-1号与哨兵-2号的时间序列，能够提升开阔水体与植被覆盖水体的制图时间分辨率。