Estimation of water quality variables based on machine learning model and cluster analysis-based empirical model using multi-source remote sensing data in inland reservoirs, South China

• Published in: Environmental pollution (1987) 2024-02, Vol.342, p.123104-123104, Article 123104
• Main Authors: Tian, Di, Zhao, Xinfeng, Gao, Lei, Liang, Zuobing, Yang, Zaizhi, Zhang, Pengcheng, Wu, Qirui, Ren, Kun, Li, Rui, Yang, Chenchen, Li, Shaoheng, Wang, Meng, He, Zhidong, Zhang, Zebin, Chen, Jianyao
• Format: Article
• Language: English
• Subjects: Band combination model; Clustering algorithm; Inland water quality monitoring; Machine learning; Multi-source remote sensing
• ISSN: 0269-7491; 1873-6424
• DOI: 10.1016/j.envpol.2023.123104

Reservoirs play important roles in the drinking water supply for urban residents, agricultural water provision, and the maintenance of ecosystem health. Satellite optical remote sensing of water quality variables in medium and micro-sized inland waters under oligotrophic and mesotrophic status is challenging in terms of the spatio-temporal resolution, weather conditions and frequent nutrient status changes in reservoirs, etc., especially when quantifying non-optically active components (non-OACs). This study was based on the surface reflectance products of unmanned aerial vehicle (UAV) multispectral images, Sentinel-2B Multispectral instrument (MSI) images and Landsat 7 Enhanced Thematic Mapper Plus (ETM+) by utilizing fuzzy C-means (FCM) clustering algorithm was combined with band combination (BC) model to construct the FCM-BC empirical model, and used mixed density network (MDN), extreme gradient boosting (XGBoost), deep neural network (DNN) and support vector regression (SVR) machine learning (ML) models to invert 12 kinds of optically active components (OACs) and non-OACs. Compared with the unclustered BC (UC) model, the mean coefficient of determination (MR) of the FCM-BC models was improved by at least 46.9%. MDN model showed best accuracy (R2 in the range of 0.60–0.98) and stability (R2 decreased by up to 13.2%). The accuracy of UAV was relatively higher in both empirical methods and machine learning methods. Additionally, the spatio-temporal distribution maps of four water quality variables were mapped based on the MDN model and UAV images, all platforms showed good consistency. An inversion strategy of water quality variables in various monitoring frequencies and weather conditions were proposed finally. The purpose of introducing the UAV platform was to cooperate with the satellite to improve the monitoring response ability of OACs and non-OACs in small and micro-sized oligotrophic and mesotrophic water bodies. [Display omitted] •Estimation of optically inactive and non-optically inactive components.•Clustering algorithm effectively improved the accuracy of the empirical models.•The mixture density network model had the best regression results.•Spatial distribution of four water quality variables map was generated.