DWE: Research on Water Quality Inversion and Scalability in Small Watersheds Using UAV Remote Sensing Data

Abstract

Urban rivers are faced with serious water quality deterioration, and an efficient, comprehensive and accurate dynamic monitoring method for urban water quality is urgently needed. Remote sensing water quality inversion model based on spectral features has been widely used in water quality monitoring. However, traditional water quality inversion models ignore the complex optical properties of water bodies, resulting in the performance of models that are often limited and cannot be reused across time and space. Therefore, this study applies the idea of integrated modeling to remote sensing inversion of water quality parameters (WQPs), and proposes an adaptive dynamic weighted ensemble learning (DWE) model, aiming to integrate the advantages of different machine learning models, improve the accuracy of WQP inversion, and adaptively assign dynamic weights to different machine learning models, thereby achieving the scalability of the model in time and space. In addition, to address the limitations of using drone data to infer small-scale water WQP due to spatial resolution limitations and limited input features, we propose an ASO-GSA feature selection method for selecting the optimal feature combination from potential feature set data. The results show that the DWE model based on optimal features has high accuracy and robustness in WQP prediction, with coefficients of determination ( $R^{2}$ ) of 0.8472, 0.8547, 0.8973, and 0.8393 for chlorophyll a (Chl-a), chemical oxygen demand (COD), total nitrogen (TN), and total phosphorus (TP), respectively, which improved the prediction accuracy by nearly 30% compared with the traditional inversion model. In order to evaluate the spatiotemporal scalability of the model, we applied the model to unmanned aerial vehicle (UAV) remote sensing data acquired from the Nanxintang River during the first half of the year from 2024 to 2025 (nearly one year later) and the Shangheng River between 2023 and 2024 to assess the scalability of the model. The results show that our model can be extended across time and space( $R^{2}$ of 0.8012, 0.8387,0.8462 and 0.7863; $R^{2}$ of 0.7892, 0.8153,0.7587 and 0.7664). The results of this paper can further inform water quality monitoring and water resource management in urban rivers.

Keywords

• UAV remote sensing
• ensemble learning
• machine learning
• water quality inversion