Physically-informed modular transfer learning for cross-water-body water quality inversion using underwater RGB...

Cross-water-body domain shifts and implicit data leakage remain significant bottlenecks for the reliable inversion of water quality parameters from underwater imagery. This study develops a physically-informed modular transfer learning framework to enhance the robustness of RGB-based water quality estimation across heterogeneous aquatic environments. Using an underwater robotic platform, a synchronized dataset of RGB images and in-situ water quality parameters was collected from both river (source) and lake (target) domains. We first demonstrate that conventional random data partitioning suffers from severe spatial data leakage, which artificially inflates the average coefficient of determination (R²) of baseline convolutional neural networks by 0.594. To establish a credible evaluation boundary, we implemented a strict site-independent partitioning protocol and proposed an enhanced ResNet-18 architecture. This model incorporates multi-scale feature enhancement, domain adaptation, and a physically-informed consistency constraint based on optical attenuation laws. Multidimensional scaling analysis suggests that cross-domain robustness is strongly associated with the geometric consistency of learned representations rather than mere architectural complexity. Results indicate that the ResNet-18-DP configuration provides the optimal balance between accuracy and stability, achieving an R² of 0.965 (MAE = 0.257) for turbidity inversion. Furthermore, the model exhibits exceptional zero-shot transferability to suspended solids inversion (R² = 0.953) and maintains robust performance under independent cross-seasonal testing (R² = 0.846). These findings provide a rigorous methodological foundation for scalable, low-cost underwater water quality monitoring in complex, real-world scenarios.