Spectral Similarity-Driven Transfer Learning for Cross-Regional Retrieval of Nonoptically Active Water Quality Parameters

The diversity and variability of water environments in different regions pose a challenge to the cross-regional transferability of water quality retrieval models using remote sensing. This study proposed a spectral similarity-driven transfer learning (SSDTL) framework for cross-regional retrieval of non-optically active water quality parameters (NAWQPs). By integrating spectral similarity analysis and multi-task neural networks, the method addressed data scarcity in new target regions. Spectral reflectance curves from an original region were clustered using the k-means clustering algorithm, and the transfer threshold was defined using a spectral similarity index (SSI). For the new region, new samples were assigned to original clusters or new separate clusters via SSI matching. Key input features, including spectral bands, environmental factors, and optically active water quality parameters (OAWQPs), were selected through correlation analysis. A pre-trained multi-task model was fine-tuned with very few samples from the new region. Using in situ data from Shanghai and Nanning, China, results showed that spectra with higher similarity corresponded to similar total phosphorus (TP) and chemical oxygen demand (COD) concentrations, and over 70% of cross-regional samples fell within the statistical discrete range. The SSDTL model outperformed many common machine learning methods in transfer tasks, generally improving error metrics by at least 20%. The real-scene test with airborne hyperspectral images achieved a coefficient of determination (R²) of 0.31 and a mean absolute error (MAE) of 0.09 for TP, as well as R² of 0.47 and MAE of 8.11 for COD, demonstrating the applicability. The approach reduces field sampling costs by over 90%, offering a scalable solution for cross-regional NAWQPs monitoring.