Revealing the spatiotemporal drivers of coastal deoxygenation with a novel AI reconstruction framework

Reconstructing high-frequency dissolved oxygen (DO) data is crucial for understanding and managing coastal hypoxia, particularly in regions constrained by sparse, low-frequency monitoring. This study presents a novel framework that uses transfer learning (TL) to reconstruct daily DO time series for stratified coastal waters, followed by Structural Equation Modeling (SEM) to analyze the reconstructed results. The model was developed using a 30-year dataset (1993-2023) from Hong Kong, China integrating monthly DO measurements from 76 stations across ten water control zones with 13 daily hydrometeorological drivers. The TL model demonstrated robust superiority over a baseline deep learning approach, achieving median Nash-Sutcliffe Efficiency (NSE) values of 0.63 (surface), 0.72 (middle) and 0.78 (bottom), effectively transforming monthly data into continuous daily reconstructed records. This daily reconstruction capability was further corroborated by independent continuous buoy records, which verified that the model reliably captures intra-monthly DO dynamics. Analysis of the reconstructed series revealed that, despite a generally healthy DO status with an overall hypoxic event rate below 5%, distinct spatiotemporal patterns of hypoxia risk emerge in semi-enclosed bays and densely populated estuaries. SEM interpretability analysis identified air temperature as the dominant negative statistical predictor of DO variation across the three layers, while mean sea level height showed an additional significant negative association with surface DO. A corresponding long-term trend analysis confirmed a gradual decline in coastal DO concentrations concurrent with regional warming and sea-level rise over recent decades. This study validates that TL-based reconstruction can unlock nuanced, high-frequency dynamics and climatic trends from sparse observations. The framework provides a critical tool for converting routine monitoring data into actionable insights for stratified hypoxia management and climate adaptation planning.