Deep-learning-based remote sensing retrieval of coastal water quality in the Pearl River Estuary

In recent years, climate change, industrial and agricultural production, and human activities have intensified eutrophication in coastal waters, underscoring the urgent need for comprehensive monitoring and remediation. Remote sensing retrieval has become essential for water quality monitoring due to its broad spatial coverage, timely data acquisition, and rapid response. However, the optical complexity of coastal waters—caused by terrestrial inputs, marine dynamics, and anthropogenic interference—limits the performance of traditional empirical and machine learning models. Deep learning, with strong feature extraction and nonlinear modeling capabilities, offers a promising alternative. We propose a multispectral coastal water quality retrieval network (MS-CoastNet) to retrieve key water quality parameters (WQPs)—chlorophyll-a (CHL-a), turbidity (TUB), total nitrogen (TN), total phosphorus (TP), chemical oxygen demand (COD), and dissolved inorganic nitrogen (DIN)—based on 1598 Sentinel-3 OLCI images and in situ data collected between 2016 and 2024 (R2=0.58 to 0.72). We further analyze the spatiotemporal characteristics of WQPs in coastal waters of the Pearl River Estuary (PRE). Results demonstrate that MS-CoastNet outperforms conventional machine learning algorithms for most WQPs. We present the potential of an AI-based remote sensing approach for monitoring water quality and supporting environmental management in the coastal waters of the PRE.