A comparative analysis of empirical, machine learning, and deep learning models for water quality retrieval in Wenzhou Bay with sparse in situ measurements

Wenzhou Bay, Zhejiang province, China High-resolution water quality monitoring is crucial for environmental management, yet the comparative effectiveness of deep learning versus conventional retrieval methods remains insufficiently explored. This study systematically evaluates the performance of multiple approaches—including empirical models, machine learning methods, and a novel spatially aware deep learning architecture (DeepSeaWQ)—for retrieving water quality parameters, including chlorophyll-a and non-optically active parameters (total nitrogen, total phosphorus, and dissolved oxygen). In model training and validation, a random split of samples was adopted, and spatial block cross-validation was employed to confirm the model’s generalization capability. The water quality monitoring data used in this study consist of in situ ground-truth samples rather than processed data, ensuring reliability. However, due to logistical constraints, there is an average 3-day match-up window between the ground sampling data and the satellite imagery, which may introduce certain uncertainties, even though the water environment in this region remains relatively stable. Results indicate that machine learning and spatially explicit deep learning models are effective in capturing the bay's dynamic and interconnected hydrological processes. The proposed DeepSeaWQ model achieved superior performance, with R² values exceeding 0.9 for all four indicators. The derived high-resolution water quality maps successfully reveal distinct spatiotemporal patterns, supporting targeted environmental monitoring. • Comparative analysis of multiple water quality retrieval models conducted. • Proposed novel spatial-aware deep learning model for water retrieval. • Proposed model excels in TN, TP, chl-a, and DO retrieval. • Generated long-term high-resolution water quality maps for Wenzhou Bay.