Satellite retrievals of water quality for diverse inland waters from Sentinel-2 images: An example from Zhejiang Province, China

Abstract

Due to its advantages of high spatiotemporal resolution, long-term stable observation, and historical retrospective data, satellite remote sensing has extensive application in the dynamic monitoring of water quality in coastal and inland waters. However, constructing generic satellite inversion models for water quality indicators, especially non-optically active parameters such as nutrients, remains challenging due to the varying pollution sources in different inland waters. This study aims to address this challenge from a data-driven perspective. Based on Sentinel-2 satellite images and in situ data from 311 automatic monitoring stations in Zhejiang Province, China, a large volume of matchups between satellite-derived water spectra and in situ water quality parameter concentrations were constructed, including the permanganate index (CODMn, N=8760, from 0.08 to 9.88 mg/L), total nitrogen (TN, N=7434, from 0.01 to 9.30 mg/L), total phosphorus (TP, N=8845, from 0.001 to 0.991 mg/L), ammonia nitrogen (NH3-N, N=8642, from 0.001 to 1.980 mg/L) and water turbidity (TUB, N=7913, from 0.1 to 1994 NTU). Satellite retrieval models for the five water quality parameters were constructed utilizing the extreme gradient boosting tree algorithm (XGBoost), which can be applied to diverse inland waters. The models showed robust performance on the additional independent dataset, with correlation coefficients (r) of 0.74, 0.79, 0.84, 0.72, and 0.87, and root mean square errors (RMSEs) of 0.73 mg/L, 0.56 mg/L, 0.18 mg/L, 0.24 mg/L, and 0.29 NTU for CODMn, TN, TP, NH3-N, and TUB, respectively. Comparisons between satellite-retrieved and in situ values showed good consistency in both spatial and temporal distributions. By applying these models to inland waters in Zhejiang Province, the monthly distributions of major rivers and reservoirs with a 10 m resolution for the five water quality parameters were obtained. The patterns of the five water quality parameters across the rivers typically indicated elevated values in the downstream regions and diminished values in the upstream areas, except for those of the Yongjiang River and Cao’e River. Compared to rivers, reservoirs had good water quality. However, for some water quality parameters, the temporal variations in the average values over the whole reservoir and the values at specific sites had opposite trends. Our research provides a practical reference for the construction of satellite retrieval models for determining the water quality parameters of inland waters, as well as robust technical support for dynamic remote sensing surveillance of inland water quality.

Keywords

• Water quality
• Inland water
• Remote sensing
• Sentinel-2
• Machine learning