Evaluating and Optimizing VIIRS Retrievals of Chlorophyll-a and Suspended Particulate Matter in Turbid Lakes Using a Machine Learning Approach
Zhigang Cao, Ronghua Ma, Nima Pahlevan, Miao Liu, John M. Mélack, Hongtao Duan, Kun Xue, Ming Shen
Abstract
The Visible Infrared Imaging Radiometer Suite (VIIRS) instrument was launched to continue the legacy of the MODerate Resolution Imaging Spectroradiometer (MODIS). Despite recent studies demonstrating the use of VIIRS observations over inland waters, VIIRS has not been widely used to generate water quality products (e.g., chlorophyll-a (Chl- <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</i> ), suspended particulate matter (SPM)) in relatively large turbid lakes. This study examines the quality of VIIRS-derived remote sensing reflectance (R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rs</sub> ) from four different atmospheric-correction processors with matchups from 13 lakes sized between 107 km <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and 2573 km <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> across the eastern plain of China. NOAA’s operational R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rs</sub> outperforming R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rs</sub> retrieved by other state-of-the-art algorithms were shown to contain mean uncertainties of 57%, 33%, 20%, 28% for R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rs</sub> (486), R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rs</sub> (551), R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rs</sub> (671), and R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rs</sub> (745), respectively, which induced ~55% uncertainty in satellite-retrieved SPM and Chl- <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</i> from recently developed algorithms in the studied lakes. A deep neural network was developed for simultaneous retrievals of Chl- <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</i> and SPM from VIIRS Rayleigh-corrected reflectance to improve accuracy. The model with satisfactory accuracy (mean uncertainty of 28% for Chl- <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</i> and 20% for SPM) outperformed other machine learning approaches and nearly halved uncertainties compared to those obtained from satellite-derived R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rs</sub> products. Within the 2012-2020 period, high-quality VIIRS-derived Chl- <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</i> and SPM across 61 lakes in eastern China had evident interannual variability in SPM but insignificant temporal variations in Chl- <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</i> . This study provides validated, high-quality, basin-scale VIIRS-derived Chl- <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</i> and SPM products in eastern China during the past decade. Our results offer a strategy for improving regional water quality products from VIIRS data.