Surveillance of urban river environment by quantifying distributions of water quality parameters using hyperspectral...

Efficient inspection of variation in water quality is of paramount importance to environmental protection and management of urban rivers. With remote sensing techniques prevailing in environmental monitoring in recent years, unmanned aerial vehicle (UAV) remote sensing has been applied to retrieve water quality parameters (WQPs) involving hyperspectral data and posed a great opportunity for flexible and effective water quality monitoring. However, current methods usually entailed more cost-prohibitive water samples to predict a few WQPs without a unified framework. In this study, a hybrid feedback ripple net (HF-RN) was proposed to effectively retrieve concentrations of WQPs including total phosphorus (TP), total nitrogen (TN), chemical oxygen demand (COD), biochemical oxygen demand (BOD), chlorophyll a (Chl-a), and total suspended solids (TSS) from UAV hyperspectral data. HF-RN integrated deep learning, spatial distribution pattern analysis, and probabilistic statistical analysis, which quantified the spatial distribution of WQPs concentrations. Additionally, HF-RN reduced the quantity demand for water samples as training samples and dependence of prediction accuracy and stability on spatial continuity of water sampling. HF-RN correlated sampled region and unsampled region through information sharing and delivery in ripple propagation graph network, enhancing their spatial relatedness and continuity of predicted WQPs. Moreover, the proposed method has been adopted in practice, applied to monitoring water quality of Shiqi River, Zhongshan, Guangdong, China, which laid theoretical and technical foundation to formulate an efficient urban river management scheme and served as a visualization tool for variation of water quality. This study posed a great opportunity to optimize management scheme of domestic and industrial sewage discharge, rendering sewage discharge amount to be under control. The performance of HF-RN was evaluated on real-world datasets and experimental results showed that HF-RN outperformed its baseline models with resulting mean absolute percent error (MAPE) ranging from 7.54% to 12.38%.