A machine learning-based strategy for estimating non-optically active water quality parameters using Sentinel-2 imagery

Water-quality monitoring for small urban waterbodies by remote sensing gets to be difficult due to the coarse spatial resolution of remote-sensing imagery. The recently launched Sentinel-2 produces imagery with a spatial resolution of 10 × 10 m and a temporal resolution of 5 days. It provides an opportunity to conduct high-frequency water-quality monitoring for small waterbodies. Since illegal discharges are an important issue for urban water management, total phosphorous (TP), total nitrogen (TN), and chemical oxygen demand (COD) were chosen as the target water-quality parameters. TP, TN and COD, however, are non-optically active parameters. There are fairly limited previous studies on retrieving these parameters in comparison with optically active parameters, e.g. Chlorophyll-a etc. Based on the fact that non-optically active parameters may be highly correlated with optically active parameters, this study compared 255 possible Sentinel-2 imagery band compositions to identify the most appropriate ones for TP, TN and COD retrieval. Three machine-learning models, namely Random Forest (RF), Support Vector Regression (SVR) and Neural Networks (NN), were compared to seek the most robust ones for retrieving the above non-optically active parameters. Results showed that the most appropriate band (hereafter termed as ‘Bindex’ for brevity) compositions for TP, TN, and COD retrieval were ‘B3+B4+B5+B6+B7+B8’, ‘B3+B4+B5\breAK+B6+B7+B8’, and ‘B2+B3+B5+B6+B7+B8’ respectively. The coefficient of determination (R2) of TP, TN, and COD estimations by NN, RF and SVR was 0.94, 0.88, and 0.86, respectively. The retrieval performances of these non-optically active parameters were hence significantly improved by the optimized machine-learning models and imagery band selection. The developed models have limitations in applying to other areas, thus band selection and tuning parameters with new data are necessary for different areas. The water-quality mapping obtained from Sentinel-2 imagery provided a full spatial coverage of the water-quality characterization for the entire water surface, and helped identify illegal discharges to urban waterbodies. This study provides a new practical and efficient water-quality monitoring strategy for managing small waterbodies.