Integrated retrieval of water quality parameters using UAV hyperspectral images and satellite imagery: Leveraging deep learning and attention mechanisms for precision

Integrated retrieval of water quality parameters using UAV hyperspectral images and satellite imagery: Leveraging deep learning and attention mechanisms for precision • A novel DL model with residuals and attention mechanisms achieved R 2 >0.85 for some WQPs from UAV hyperspectral images. • Attention weights identified key spectral bands in retrieving WQPs. • Mapping the attention weights of UAV images to Planet images. • The framework of integrating ground observations, UAV and satellite images improved R 2 for TN and COD Mn by 0.16–0.18 than using single Planet images. Real-time and high-precision monitoring of water quality is essential for effective water management. Despite challenges in narrow waterways and intricate spectral characteristics, the integration of the unmanned aerial vehicle (UAV) hyperspectral images and deep learning (DL) shows promise for monitoring, though issues like small spatial coverage and poor interpretability must be addressed. This paper focused on retrieving water quality parameters (WQPs) in urban rivers at Guangzhou City, China, utilizing synchronously collected water quality data, water surface reflectance, UAV hyperspectral images, and multispectral PlanetScope images. A novel CNN-Attention-ResBlock (CAR) model was developed by combining attention mechanism, residual blocks, and neural networks to retrieve 16 WQPs such as the suspended solids (SS), ammonia nitrogen (NH 3 -N), total phosphorous (TP). Attention weights were applied to quantify the significance of each spectral band in retrieving a certain WQP. The CAR demonstrated good regression performance for SS (R 2 =0.85), NH 3 -N (R 2 =0.93), TP (R 2 =0.85), chemical oxygen demand (R 2 =0.87) and permanganate index (COD Mn , R 2 =0.96). A framework of integrating UAV and PlanetScope images improved the prediction accuracy based on PlanetScope images, with R 2 exceeding 0.7 for total nitrogen and COD Mn . Spatial distribution of WQPs in Guangzhou’s main urban area revealed poorer water quality in densely populated and agriculturally active sections, though an overall improving trend was observed. This paper not only develops a high-precision DL model for retrieving WQPs and identifying sensitive bands, but also presents a ground–UAV–satellite framework for monitoring spatio-temporal variations on a larger regional scale.