Monitoring water quality using proximal remote sensing technology

Abstract

Accurate, high spatial and temporal resolution water quality monitoring in inland waters is vital for environmental management. However, water quality monitoring in inland waters by satellite remote sensing remains challenging due to low signal-to-noise ratios (SNRs) and instrumental resolution limitations. We propose the concept of proximal remote sensing for monitoring water quality. The proximal hyperspectral imager, developed by Nanjing Institute of Geography and Limnology, Chinese Academy of Sciences (CAS) and Hikvision Digital Technology, Ltd., is a high spatial, temporal and spectral resolution (1 nm) sensor for continuous observation, allowing for effective and practical long-term monitoring of inland water quality. In this study, machine learning and empirical algorithms were developed and validated using in situ total nitrogen (TN), total phosphorus (TP), chemical oxygen demand (COD) concentrations and spectral reflectance from Lake Taihu (N = 171), the Liangxi River (N = 94) and the Fuchunjiang Reservoir (N = 109) covering different water quality. Our dataset includes a large range for three key water quality parameters of TN from 0.93 to 6.46 mg/L, TP from 0.04 to 0.62 mg/L, and COD from 1.32 to 15.41 mg/L. Overall, the back-propagation (BP) neural network model had an accuracy of over 80% for TN (R² = 0.84, RMSE = 0.33 mg/L, and MRE = 11.4%) and over 90% for TP (R² = 0.93, RMSE = 0.02 mg/L, and MRE = 12.4%) and COD (R² = 0.91, RMSE = 0.66 mg/L, and MRE = 9.3%). Our results show that proximal remote sensing combined with machine learning algorithms has great potential for monitoring water quality in inland waters.

Keywords: BP neural networks; Empirical algorithms; Machine learning algorithms; Proximal remote sensing; Water quality.