Deep Learning-Enabled Dissolved Oxygen Sensing in Biofouling Environments for Ocean Monitoring
A physics-informed neural network cuts MAE to ~2 μmol/L even with algae buildup.
A team led by Nikolaos Salaris at arXiv introduced a deep learning approach to combat biofouling in dissolved oxygen sensors for ocean monitoring. Their system pairs low-cost camera-based optoelectronic sensors (using phosphorescent polymer films) with a visual transformer-based physics-informed neural network (ViT-PINN). The PINN embeds the Stern-Volmer equation directly into its loss function, ensuring physical consistency even as sensor drift and algal growth degrade raw readings.
Tested over 14 days in a controlled algae-laden tank, the ViT-PINN slashed mean average error by 92% compared to classical statistical methods and 89% over standard ML approaches, achieving ~2 μmol/L absolute error. A deep ensemble further quantifies predictive uncertainty, enabling self-diagnostic sensing. This breakthrough makes inexpensive DO sensors viable for long-term, robust climate monitoring in real-world marine environments.
- ViT-PINN reduces mean average error by 92% vs. classical statistical methods.
- Achieves ~2 μmol/L absolute error even under accelerated biofouling conditions.
- Deep ensemble quantifies predictive uncertainty for self-diagnostic sensing.
Why It Matters
Enables cheap, resilient ocean sensors for long-term climate tipping point monitoring.