Comparison of Machine Learning Algorithms for Retrieval of Water Quality Indicators in Case-II Waters: A Case Study of Hong Kong

Anthropogenic activities in coastal regions are endangering marine ecosystems. Coastal waters classified as case-II waters are especially complex due to the presence of different constituents. Recent advances in remote sensing technology have enabled to capture the spatiotemporal variability of the constituents in coastal waters. The present study evaluates the potential of remote sensing using machine learning techniques, for improving water quality estimation over the coastal waters of Hong Kong. Concentrations of suspended solids (SS), chlorophyll-a (Chl-a), and turbidity were estimated with several machine learning techniques including Artificial Neural Network (ANN), Random Forest (RF), Cubist regression (CB), and Support Vector Regression (SVR). Landsat (5,7,8) reflectance data were compared with in situ reflectance data to evaluate the performance of machine learning models. The highest accuracies of the water quality indicators were achieved by ANN for both, in situ reflectance data (89%-Chl-a, 93%-SS, and 82%-turbidity) and satellite data (91%-Chl-a, 92%-SS, and 85%-turbidity. The water quality parameters retrieved by the ANN model was further compared to those retrieved by “standard Case-2 Regional/Coast Colour” (C2RCC) processing chain model C2RCC-Nets. The root mean square errors (RMSEs) for estimating SS and Chl-a were 3.3 mg/L and 2.7 µg/L, respectively, using ANN, whereas RMSEs were 12.7 mg/L and 12.9 µg/L for suspended particulate matter (SPM) and Chl-a concentrations, respectively, when C2RCC was applied on Landsat-8 data. Relative variable importance was also conducted to investigate the consistency between in situ reflectance data and satellite data, and results show that both datasets are similar. The red band (wavelength ≈ 0.665 µm) and the product of red and green band (wavelength ≈ 0.560 µm) were influential inputs in both reflectance data sets for estimating SS and turbidity, and the ratio between red and blue band (wavelength ≈ 0.490 µm) as well as the ratio between infrared (wavelength ≈ 0.865 µm) and blue band and green band proved to be more useful for the estimation of Chl-a concentration, due to their sensitivity to high turbidity in the coastal waters. The results indicate that the NN based machine learning approaches perform better and, thus, can be used for improved water quality monitoring with satellite data in optically complex coastal waters.