PINNs Generate Depth-Resolved Coral Reef Thermal Maps from Satellite SST and Sparse Loggers
Physics-informed neural networks fuse satellite SST and sparse loggers to map subsurface reef temperatures, correcting surface-only overestimates of thermal stress with RMSE under 0.4 °C even under extreme data sparsity.
Lede: Researchers deployed a physics-informed neural network that fuses NOAA Coral Reef Watch satellite SST with sparse in-situ temperature loggers to solve the one-dimensional vertical heat equation and produce subsurface thermal profiles.
The arXiv:2604.13131 preprint enforces SST as a hard surface boundary condition while jointly optimizing effective thermal diffusivity and light attenuation Kd; validation on four Great Barrier Reef sites across 30 holdout experiments produced 0.25-1.38°C RMSE at unseen depths and maintained 0.27°C RMSE at 5 m and 0.32°C RMSE at 9.1 m under three-depth training regimes where linear statistical baselines exceeded 1.8°C RMSE (https://arxiv.org/abs/2604.13131). The PINN outperformed a physics-only finite-difference baseline in 90 % of trials.
Raissi et al. (arXiv:1711.10561) originated the PINN framework for PDE-constrained learning; Hughes et al. (Nature, 2018, https://www.nature.com/articles/s41586-018-0041-2) established that mass-bleaching frequency has risen with SST anomalies yet relied on surface-only metrics. Depth-resolved Degree Heating Day profiles from the new model show stress attenuating from 0.29 at the surface to zero by 10.7 m at Davies Reef, while satellite DHD remains fixed at 0.31 across depths.
Mainstream coral-bleaching coverage has uniformly applied satellite SST without depth correction; the PINN results indicate that reported thermal stress constitutes an overestimate below the surface, although its smoothed predictions yield conservative lower-bound DHD values at shallow sites because short-duration peaks are attenuated.
AXIOM: PINNs accurately reconstruct cooler subsurface temperatures that satellite SST misses, delivering depth-specific Degree Heating Day profiles that identify potential deep refugia and supply more precise inputs for coral bleaching forecasts.
Sources (3)
- [1]Depth-Resolved Coral Reef Thermal Fields from Satellite SST and Sparse In-Situ Loggers Using Physics-Informed Neural Networks(https://arxiv.org/abs/2604.13131)
- [2]Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations(https://arxiv.org/abs/1711.10561)
- [3]Global warming and recurrent mass bleaching of corals(https://www.nature.com/articles/s41586-018-0041-2)