Depth-Resolved Coral Reef Thermal Fields from Satellite SST and Sparse In-Situ Loggers Using Physics-Informed Neural Networks
A new physics-informed AI model predicts underwater coral temperatures with 0.27°C accuracy using just three sensors.
Researchers Alzayat Saleh and Mostafa Rahimi Azghadi have developed a novel physics-informed neural network (PINN) that significantly improves the accuracy of coral reef temperature monitoring. The model addresses a critical flaw in current satellite-based systems, which only measure sea surface temperature (SST) but apply it uniformly to all depths, overestimating thermal stress for corals living in cooler subsurface waters. By fusing NOAA Coral Reef Watch SST data with sparse, real-world temperature loggers, the PINN enforces the laws of physics—specifically the one-dimensional vertical heat equation—as a hard constraint. This allows it to jointly learn key environmental parameters like thermal diffusivity and light attenuation, creating a complete 3D thermal field from limited 2D surface data.
Validated across four sites on Australia's Great Barrier Reef, the model's performance is striking. It achieves a root mean square error (RMSE) between 0.25°C and 1.38°C at depths where it has no training data. Most impressively, under extreme data sparsity with only three training depth measurements, it maintains an RMSE of just 0.27°C at a 5-meter holdout depth and 0.32°C at 9.1 meters. In these sparse-data scenarios, standard statistical models fail, with errors ballooning beyond 1.8°C. The PINN outperformed a traditional physics-only finite-difference model in 90% of experiments.
The practical impact is a revolution in bleaching risk assessment. The model-generated depth-resolved Degree Heating Day (DHD) profiles reveal that thermal stress attenuates rapidly with depth. At Davies Reef, for example, DHD—a key metric for bleaching—dropped from 0.29 at the surface to zero by 10.7 meters, which matched logger observations. In contrast, satellite-derived DHD remained a constant, overestimated 0.31 across all depths. The researchers note the PINN's predictions are conservative, acting as a lower bound for stress, as its smooth outputs can underestimate short-duration temperature peaks. Nonetheless, this work demonstrates that physics-constrained AI can unlock the third dimension for coral reef monitoring using existing, sparse observational infrastructure, providing a more accurate and actionable picture of ecosystem health.
- PINN model achieves 0.27°C RMSE at 5m depth with only three training sensors, where baselines fail (>1.8°C error).
- Reveals thermal stress (Degree Heating Days) drops to zero by 10.7m depth, while satellite data overestimates it as constant.
- Outperforms a physics-only finite-difference baseline in 90% of experiments by fusing satellite data with sparse in-situ loggers.
Why It Matters
Enables accurate, depth-aware coral bleaching alerts using existing sensor networks, crucial for targeted conservation efforts.