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

Area of Science:

• Environmental Science
• Computer Vision
• Robotics

Background:

• Underwater water quality monitoring using RGB imagery faces challenges from domain shifts and data leakage.
• Existing methods often overestimate performance due to inadequate data partitioning, hindering real-world applicability.
• Developing robust models for heterogeneous aquatic environments is crucial for accurate water quality assessment.

Purpose of the Study:

• To develop a physically-informed transfer learning framework for robust RGB-based water quality estimation.
• To address cross-water-body domain shifts and implicit data leakage in underwater imagery analysis.
• To establish a credible evaluation protocol for assessing model performance across diverse aquatic environments.

Main Methods:

• Collected synchronized RGB images and in-situ water quality data from river and lake domains using an underwater robotic platform.
• Implemented a strict site-independent partitioning protocol to prevent data leakage and ensure reliable evaluation.
• Proposed an enhanced ResNet-18 architecture incorporating multi-scale feature enhancement, domain adaptation, and physically-informed consistency constraints based on optical attenuation laws.

Main Results:

• Demonstrated that conventional data partitioning inflates model performance (R² by 0.594) due to severe spatial data leakage.
• The proposed ResNet-18-DP model achieved high accuracy for turbidity inversion (R² = 0.965, MAE = 0.257).
• Showcased excellent zero-shot transferability to suspended solids inversion (R² = 0.953) and robust cross-seasonal performance (R² = 0.846).

Conclusions:

• The physically-informed transfer learning framework significantly enhances the robustness of water quality estimation across heterogeneous aquatic environments.
• Geometric consistency of learned representations is key for cross-domain robustness, surpassing mere architectural complexity.
• The developed methodology provides a foundation for scalable, low-cost underwater water quality monitoring in complex real-world scenarios.