Landsat observations of chlorophyll-a variations in Lake Taihu from 1984 to 2019

Long-term datasets of chlorophyll-a (Chla) are necessary to evaluate changes in eutrophication and to assist in lake management; however, current aquatic remote sensing datasets usually start after 2000. Here, a 36-year Chla dataset was assembled from Landsat Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), and Operational Landsat Imager (OLI) imagery for Lake Taihu (China) over the period from 1984 to 2019. TM, ETM+ and OLI reflectances were compared to those using the MODerate resolution Imaging Spectroradiometer (MODIS) on Terra, and agreement was found within a mean absolute difference of 15%. An algorithm for Chla retrieval developed by a machine learning approach (XGBoost) had good performance (mean absolute percentage error = 35%, mean absolute error = 9%) and outperformed random forest and support vector machine regressors and existing empirical algorithms for Lake Taihu. Landsat-derived mean Chla ranged between 12.8 µg L-1 and 32.3 µg L-1 and indicates that Lake Taihu has been eutrophic from 1984 to 2019. Chla in the northern region was higher than that in other areas over the 36 years. With the limited number of Landsat images each year, we found that the annual variation in Chla had high values during the periods of 1984–1992 and 1994–1997 and significant increases in 1999–2009 and from 2012 to 2019. The spatial and temporal variations in Chla for Lake Taihu were correlated with dissolved nutrients and air temperature. This research illustrates the use of machine learning approaches to generate long-term datasets of water quality from multiple Landsat instruments, for extending watercolor archives for lakes.