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.

No abstract

Precise atmospheric correction is important for applications where small differences in surface reflectance (SR) are significant, such as biomass estimation, crop phenology, and retrieval of water quality parameters. It also enables direct comparison between different image dates and different sensors. As a precursor to monitoring different parameters of water quality around the coastline of Hong Kong using medium-resolution sensors Landsat TM/ETM, and HJ-1A/B, this study evaluated the performance of five atmospheric correction methods. The estimated SR of the first four reflective bands of Landsat 7 ETM+ and of the identical bands of the HJ-1A/B satellites was compared with in situ multispectral radiometer (MSR) SR measurements over sand, artificial turf, grass, and water surfaces for the five atmospheric correction methodssecond simulation of the satellite signal in the solar spectrum (6S), fast line-of-sight atmospheric analysis of spectral hypercubes (FLAASH), atmospheric correction (ATCOR), dark object subtraction (DOS), and the empirical line method (ELM). Among the five methods, 6S was observed to be consistently more precise for SR estimation, with significantly less difference from the in-situ-measured SR, especially over lower reflective water surfaces. Of the two image-based methods, DOS performed well over the darker surfaces of water and artificial turf, although still inferior to 6S, while ELM worked well for grass sites as compared to the DOS and equalled the good performance of 6S over the high reflective sand surfaces.The study also evaluated the new standard Landsat SR product Landsat ecosystem disturbance adaptive processing system (LEDAPS) using the in situ measured SR data for the three land surface types -sand, artificial turf, and grass. For the highly and moderately reflecting bright sand and artificial turf, LEDAPS performed poorly, while for the darker grass site it performed better, although still inferior to 6S and ELM methods. This is probably due to the variable aerosol types and atmospheric conditions of Hong Kong, as LEDAPS was mainly compiled with reference to larger continental landmass areas.