Machine Learning Based Long-Term Water Quality in the Turbid Pearl River Estuary, China

Total suspended solid (TSS) and chlorophyll-a (Chl-a) are critical indicators of water quality. Moderate-resolution Imaging Spectroradiometer (MODIS) on the Aqua satellite provides a critical tool to map TSS and Chl-a on a daily basis. However, robust algorithms are required to retrieve TSS and Chl-a from MODIS/Aqua images in the turbid Pearl River Estuary (PRE). Here, a new artificial neural network algorithm was developed to estimate TSS and Chl-a in the PRE using a large amount of in situ measurements spanning more than a decade. Cross-validation showed that the ANN-based algorithm presented good performance with R(2)s of 0.73 and 0.66 for TSS and Chl-a, respectively. The ANN-based algorithm outperformed existing empirical and semi-analytical algorithms. By implementing the proposed algorithm to MODIS/Aqua images, TSS and Chl-a from August 2002 to July 2020 over the PRE were obtained. The results show that TSS decreased from northwest to southeast in all seasons while peaked in winter and summer. Chl-a peaked in summer and the high Chl-a patch transferred from the upper reach to the lower reach of the PRE with the increase of river discharge. TSS decreased at an annual rate of similar to 0.13 mg L-1 while Chl-a increased at an annual rate of similar to 0.05 mg m(-3) in the last 18 years. Our findings suggest that machine learning provides a practical approach to estimate TSS and Chl-a in turbid estuaries via MODIS/Aqua images and the approach developed in this study can be applied to other turbid waters with similar optical properties.