Harmonized Chlorophyll-a Retrievals in Inland Lakes From Landsat-8/9 and Sentinel 2A/B Virtual Constellation Through Machine Learning

Moderate-high resolution satellite missions provide an opportunity to capture subtle spatial variability in lakes; however, the sparsity of time series for individual satellite instruments cannot monitor temporal variation in the lake environment. To date, studies on the joint observations of chlorophyll-a (Chl-a) in inland lakes from multiple missions have been poorly reported. Here, we generated a harmonized Chl-a dataset for the lakes in the Yunnan–Guizhou Plateau in China from 2013 to 2022 by the Landsat 8/9 and Sentinel-2A/B virtual constellation. This study first examined the performance of four atmospheric correction processors to derive remote sensing reflectance (R _rs ) from Landsat 8/9 Operational Land Imager (OLI) and Sentinel-2A/B multispectral instrument (MSI) images. We determined that the dark spectral fitting algorithm generated better R _rs than the other processors, e.g., R _rs (561) mean absolute percentage error (MAPE)=15.2%, R _rs (665) MAPE=27.5%, and R _rs (704) MAPE=25.7%. OLI-derived R _rs at five visible and near-infrared bands showed satisfactory agreement with MSI (slope=0.94, MAPE=11.8%). The mixed density network outperformed the six state-of-the-art algorithms and other two machine learning models in retrieving Chl-a [MSI: MAPE=31.4% (N=109), OLI: MAPE=38.0% (N=74)]. The satisfactory agreement of Chl-a retrievals between the synchronous MSI and OLI images (N=2,293,821, MAPE=34.6%) supported the establishment of the virtual constellation. MSI- and OLI- derived Chl-a in nine major lakes in the studied area exhibited apparent seasonal variability from 2013 to 2022, particularly after 2017. Results highlight a solution to establish the Landsat/Sentinel-2 virtual constellation for improving the spatial and temporal resolutions of a database of lake water quality.