Coastal water quality monitoring is pivotal for marine ecosystem management. However, current monitoring capabilities are constrained by a trade-off between the high accuracy of sparse in-situ sampling and the broad coverage but coarse resolution of traditional ocean color satellites. To bridge this gap, this study establishes a cloud-native framework coupling multi-source satellite fusion with machine learning to reconstruct a high-resolution (30 m) water quality dataset for the coastal waters of Guangdong, China, spanning a 40-year period. Leveraging the Google Earth Engine (GEE) platform, we harmonized Landsat-5/7/8 and Sentinel-2 MSI, processing over 22,172 valid scenes via a hierarchical regression strategy to overcome cross-sensor spectral inconsistencies. A Random Forest ensemble model was developed to retrieve Chemical Oxygen Demand (COD, R 2 = 0.76), Total Nitrogen (TN, R 2 = 0.86), and Total Phosphorus (TP, R 2 = 0.88). The time series unveils a significant upward trajectory in COD concentrations (+4.08 × 10 −3 mg/L/yr) with a regime shift around 2013. Spatially, COD plumes exhibited a “Westward Expansion” pattern during the wet season, contrasting with winter maxima for TN and TP. This parameter‑specific divergence demonstrates that organic and nutrient pollutants are governed by different source and transport mechanisms – a finding that single‑parameter studies would miss. We further explored the associations between water quality variations and environmental drivers, revealing a “Human-Climate Coupled” control pattern. Beyond this qualitative mechanism, we identify a dual‑process pathway and quantify a 6‑month lag between ENSO events and peak COD anomalies, moving the discussion from simple correlation to potentially predictive understanding. This study provides the first multi-decadal, high-resolution benchmark for the Greater Bay Area and demonstrates the value of cloud-based multi-sensor fusion for precision coastal management.
Chen et al. (2026) studied this question.