A machine learning approach to estimate chlorophyll-a from Landsat-8 measurements in inland lakes

Abstract

Landsat-8 Operational Land Imager (OLI) provides an opportunity to map chlorophyll-a (Chla) in lake waters at spatial scales not feasible with ocean color missions. Although state-of-the-art algorithms to estimate Chla in lakes from satellite-borne sensors have improved, there are no robust and reliable algorithms to generate Chla time series from OLI imageries in turbid lakes due to the absence of a red-edge band and issues with atmospheric correction. Here, a machine learning approach termed the extreme gradient boosting tree (BST) was employed to develop an algorithm for Chla estimation from OLI in turbid lakes. This model was developed and validated by linking Rayleigh-corrected reflectance to near-synchronous in situ Chla data available from eight lakes in eastern China (N = 225) and three coastal and inland waters in SeaWiFS Bio-optical Archive and Storage System (N = 97). The BST model performed well on a subset of data (N = 102, R2 = 0.79, root mean squared difference = 7.1 μg L−1, mean absolute percentage error = 24%, mean absolute error = 1.4, Bias = 0.9), and had better Chla retrievals than several band-ratio algorithms and a random forest approach. The performance of BST model was judged as appropriate only for the range of conditions in the training data. Given these limitations, spatial and temporal variations of Chla in hundreds of lakes larger than 1 km2 in eastern China for the period of 2013–2018 were mapped using the BST model. OLI-derived Chla indicated that small lakes (<50 km2) had greater Chla than the larger lakes. This research suggests that machine-learning models provide practical approaches to estimate Chla in turbid lakes via broadband instruments like OLI and that extending to other regions requires training with a representative dataset.