arxiv: 2604.13131 · v1 · submitted 2026-04-14 · 💻 cs.LG · cs.CV

Recognition: unknown

Depth-Resolved Coral Reef Thermal Fields from Satellite SST and Sparse In-Situ Loggers Using Physics-Informed Neural Networks

Alzayat Saleh , Mostafa Rahimi Azghadi

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:06 UTC · model grok-4.3

classification 💻 cs.LG cs.CV

keywords physics-informed neural networkscoral reefssea surface temperaturethermal stressdepth-resolved modelingGreat Barrier ReefDegree Heating Daysdata fusion

0 comments

The pith

A physics-informed neural network reconstructs accurate depth-resolved temperatures in coral reefs by fusing satellite SST with sparse in-situ loggers via the vertical heat equation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that physics-informed neural networks can extend surface satellite temperature measurements to the depths where corals live by incorporating the physics of vertical heat transport and learning a few site-specific parameters. This is important because corals experience cooler temperatures and less thermal stress at greater depths, so using surface data alone overestimates bleaching risk. The approach uses the one-dimensional heat equation as a constraint in the neural network, with satellite data as the surface boundary condition, while jointly estimating thermal diffusivity and light attenuation from a small number of logger measurements. Validation on Great Barrier Reef sites demonstrates low prediction errors at holdout depths, even when only three depths are used for training, and reveals that degree heating days decrease rapidly with depth. The model provides conservative estimates because its smooth outputs miss some peak temperature events.

Core claim

By embedding the one-dimensional vertical heat equation into a neural network and enforcing satellite SST as a hard boundary condition while learning effective thermal diffusivity and light attenuation, the PINN can predict temperatures at unseen depths with 0.25-1.38°C RMSE across four sites, outperforming baselines especially under sparse data conditions, and produces depth profiles of thermal stress that attenuate with depth.

What carries the argument

The physics-informed neural network that solves the 1D vertical heat equation with satellite SST as a hard surface boundary condition and jointly learns constant thermal diffusivity kappa and light attenuation Kd from sparse logger data.

If this is right

Depth-resolved temperature and thermal stress assessments become feasible using only existing satellite products and a minimal set of in-situ loggers.
Coral bleaching monitoring can account for reduced stress at depth rather than assuming uniform conditions from the surface.
The method remains effective with as few as three training depths, enabling application in regions with limited observational data.
PINN-based Degree Heating Day estimates serve as conservative lower bounds on actual thermal stress due to smoothing of temperature peaks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Applying this model globally could improve coral reef management by providing depth-specific risk maps without requiring dense sensor networks.
Extensions incorporating horizontal flows or variable coefficients might address sites where the constant-parameter assumption breaks down.
Comparing PINN outputs with high-resolution 3D hydrodynamic models could test the validity of the 1D approximation.
The conservative DHD values suggest the need for ensemble methods or peak-capturing adjustments for more complete stress quantification.

Load-bearing premise

The vertical temperature dynamics at these coral reef sites are adequately captured by the one-dimensional heat equation using constant effective thermal diffusivity and light attenuation coefficients.

What would settle it

A direct comparison of PINN-predicted temperatures against new in-situ logger data at previously unseen depths and sites, where an RMSE exceeding 1.5°C or failure to show depth-attenuation in DHD would indicate the model does not generalize as claimed.

Figures

Figures reproduced from arXiv: 2604.13131 by Alzayat Saleh, Mostafa Rahimi Azghadi.

**Figure 2.** Figure 2: Study area showing the four Great Barrier Reef sites in the central GBR. Inset shows location on the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Holdout validation results across four reefs. (a) Holdout RMSE ( [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Figure 4: Reconstruction accuracy as a function of training data density for two Davies Reef holdout depths. The [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: Time series comparison at four holdout depths: (a) Davies Reef [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

**Figure 6.** Figure 6: Depth-resolved thermal stress profiles. (a) Davies Reef: cumulative Degree Heating Days (DHD, [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

**Figure 7.** Figure 7: Learned physical parameters across reefs and holdout experiments (5 random seeds per configuration; each [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

read the original abstract

Satellite sea surface temperature (SST) products underpin global coral bleaching monitoring, yet they measure only the ocean skin. Corals inhabit depths from the shallows to beyond 20 metres, where temperatures can be 1-3{\deg}C cooler than the surface; applying satellite SST uniformly to all depths therefore overestimates subsurface thermal stress. We present a physics-informed neural network (PINN) that fuses NOAA Coral Reef Watch SST with sparse in-situ temperature loggers within the one-dimensional vertical heat equation, enforcing SST as a hard surface boundary condition and jointly learning effective thermal diffusivity (\k{appa}) and light attenuation (Kd). Validated across four Great Barrier Reef sites (30 holdout experiments), the PINN achieves 0.25-1.38{\deg}C RMSE at unseen depths. Under extreme sparsity (three training depths), the PINN maintains 0.27{\deg}C RMSE at the 5 metre holdout and 0.32{\deg}C at the 9.1 metre holdout, where statistical baselines collapse to >1.8{\deg}C; it outperforms a physics-only finite-difference baseline in 90% of experiments. Depth-resolved Degree Heating Day (DHD) profiles show that thermal stress attenuates with depth: at Davies Reef, DHD drops from 0.29 at the surface to zero by 10.7 metres, consistent with logger observations, while satellite DHD remains constant at 0.31 across all depths. However, the PINN underestimates absolute DHD at shallow depths because its smooth predictions attenuate the short-duration peaks that drive threshold exceedances; PINN DHD values should be interpreted as conservative lower bounds on depth-resolved stress. These results demonstrate that physics-constrained fusion of satellite SST with sparse loggers can extend bleaching assessment to the depth dimension using existing observational infrastructure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PINN depth extrapolation for reef temps works on sparse data but the 1D constant physics may be too limited.

read the letter

The paper's main result is that a physics-informed network can turn sparse logger readings plus satellite SST into depth profiles of temperature and thermal stress for reefs, with decent accuracy even when data is very thin. What stands out is the performance under extreme sparsity. With only three training depths, the model keeps RMSE around 0.3 °C at holdout depths where statistical methods jump above 1.8 °C. It also learns the effective diffusivity and light attenuation jointly rather than fixing them. The DHD profiles are a practical output that shows stress dropping with depth, matching the loggers at the sites they tested. The soft spot is the underlying 1D heat equation with fixed coefficients. Reefs have horizontal currents, tidal mixing, and variable properties that this leaves out. Because the finite-difference baseline solves the same equation, the comparison shows the network is better at the numerics or fitting but does not test if the physics is sufficient. The underestimation of DHD peaks is another sign that short-term variability is not fully captured, so the depth-resolved stress numbers are lower bounds. This is for applied researchers in marine ecology or remote sensing who need subsurface temperature estimates without dense sensor arrays. A reader working on physics-informed methods for environmental monitoring would get concrete numbers on how well it works with real sparse data. It deserves a serious referee because the holdout tests are clear and the problem is well-motivated, even with the modeling assumptions. I would recommend peer review, with the referees asked to look closely at how sensitive the results are to the constant-coefficient assumption and whether more complex physics would change the conclusions.

Referee Report

1 major / 2 minor

Summary. The paper claims that a physics-informed neural network (PINN) fusing NOAA Coral Reef Watch satellite SST with sparse in-situ temperature loggers, while enforcing the one-dimensional vertical heat equation as a hard constraint and jointly learning effective thermal diffusivity (kappa) and light attenuation (Kd), can accurately extrapolate temperatures to unseen depths. Validated on four Great Barrier Reef sites via 30 holdout experiments, it reports RMSE of 0.25-1.38°C at unseen depths; under extreme sparsity (three training depths) it achieves 0.27°C RMSE at the 5 m holdout and 0.32°C at 9.1 m (versus >1.8°C for statistical baselines) and outperforms a physics-only finite-difference baseline in 90% of cases. Depth-resolved Degree Heating Day (DHD) profiles are derived showing attenuation of thermal stress with depth, though the PINN underestimates absolute DHD at shallow depths due to smoothing of short-duration peaks and should be treated as conservative lower bounds.

Significance. If the results hold, the work provides a practical, data-efficient method to extend satellite SST-based coral bleaching monitoring into the subsurface using existing sparse logger infrastructure and physics constraints, addressing a key limitation in current global assessments. Notable strengths include the quantitative multi-site holdout validation across sparsity regimes, direct comparison to both statistical and physics-only baselines, and explicit acknowledgment of limitations for threshold-based metrics.

major comments (1)

[Abstract] Abstract: The reported outperformance versus the physics-only finite-difference baseline (in 90% of experiments) employs the identical 1D vertical heat equation with constant effective kappa and Kd; this comparison therefore does not test the sufficiency of the model form itself for reef dynamics. Omitted processes (horizontal advection, tidal mixing, vertical velocity, depth/time-varying properties) could affect generalizability, consistent with the acknowledged underestimation of short-duration DHD peaks that drive bleaching thresholds.

minor comments (2)

[Abstract] Abstract: The thermal diffusivity is denoted as (k{appa}); adopt standard mathematical notation (kappa) for clarity and consistency with the heat equation.
[Abstract] Abstract: Temperature differences are written as 1-3{deg}C; ensure uniform degree-symbol formatting and check for similar LaTeX artifacts elsewhere.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive comments. We address the major comment below and have revised the manuscript to clarify the scope of the baseline comparison and the limitations of the 1D model.

read point-by-point responses

Referee: The reported outperformance versus the physics-only finite-difference baseline (in 90% of experiments) employs the identical 1D vertical heat equation with constant effective kappa and Kd; this comparison therefore does not test the sufficiency of the model form itself for reef dynamics. Omitted processes (horizontal advection, tidal mixing, vertical velocity, depth/time-varying properties) could affect generalizability, consistent with the acknowledged underestimation of short-duration DHD peaks that drive bleaching thresholds.

Authors: We agree that the comparison to the finite-difference baseline does not test the sufficiency of the 1D heat equation model itself, as both approaches use the same governing equation (with the baseline employing constant kappa and Kd). The PINN's outperformance in 90% of experiments stems from its ability to jointly learn effective, potentially varying kappa and Kd values from the sparse data while enforcing the physics constraint. We acknowledge that omitted processes such as horizontal advection, tidal mixing, vertical velocity, and depth/time-varying properties could affect generalizability, consistent with the underestimation of short-duration DHD peaks. This limitation is already noted in the manuscript, where PINN DHD values are described as conservative lower bounds. We have revised the abstract to explicitly state that the baseline comparison evaluates the benefit of data-informed parameter learning within the physics model rather than validating the model form. We have also expanded the discussion to highlight the potential impact of unmodeled processes on generalizability. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the PINN derivation or predictions

full rationale

The paper's central derivation enforces the 1D vertical heat equation as a hard constraint in the PINN loss while learning effective constant parameters κ and Kd from the combination of satellite SST (hard surface BC) and sparse in-situ logger data at a subset of depths. Predictions at holdout depths are obtained by solving the PDE forward with the learned parameters, not by direct regression on the target temperatures. This is confirmed by the 30 holdout experiments, extreme sparsity tests (three training depths), and outperformance versus both statistical baselines and a physics-only finite-difference solver using the identical equation. No quoted step reduces a claimed prediction to a fitted input by construction, no self-citation chain is load-bearing for the extrapolation result, and the method remains self-contained against the external holdout benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim depends on the 1D vertical heat diffusion equation being a sufficient model and on the learned effective parameters generalizing across sites and conditions.

free parameters (2)

effective thermal diffusivity kappa
Learned jointly during PINN training from data and physics loss; not a fixed physical constant.
light attenuation coefficient Kd
Learned jointly during PINN training; effective value for the water column.

axioms (1)

domain assumption Temperature evolution obeys the one-dimensional vertical heat equation with constant effective diffusivity and attenuation.
Enforced as hard constraint in the PINN; invoked throughout the method description.

pith-pipeline@v0.9.0 · 5655 in / 1392 out tokens · 43567 ms · 2026-05-10T16:06:20.016586+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 18 canonical work pages

[1]

T. P. Hughes, J. T. Kerry, M. Álvarez-Noriega, J. G. Álvarez-Romero, K. D. Anderson, A. H. Baird, R. C. Babcock, M. Beger, D. R. Bellwood, R. Berkelmans, et al., Global warming and recurrent mass bleaching of corals, Nature 543 (7645) (2017) 373–377.doi:10.1038/ nature21707

2017
[2]

T. P. Hughes, K. D. Anderson, S. R. Connolly, S. F. Heron, J. T. Kerry, J. M. Lough, A. H. Baird, J. K. Baum, M. L. Berumen, T. C. Bridge, et al., Spatial and temporal patterns of mass bleaching of corals in the Anthropocene, Science 359 (6371) (2018) 80–83.doi: 10.1126/science.aan8048

work page doi:10.1126/science.aan8048 2018
[3]

URLhttps://www.aims.gov.au/monitoring-great-barrier-reef/ gbr-condition-summary-2024-25 21

Australian Institute of Marine Science, Annual summary report of coral reef condition 2024/25, accessed 2026-03-16 (2025). URLhttps://www.aims.gov.au/monitoring-great-barrier-reef/ gbr-condition-summary-2024-25 21

2024
[4]

G. Liu, S. F. Heron, C. M. Eakin, F. E. Muller-Karger, M. Vega-Rodriguez, L. S. Guild, J. L. De La Cour, E. F. Geiger, W. J. Skirving, T. F. D. Burgess, et al., Reef-scale thermal stress monitoring of coral ecosystems: New 5-km global products from NOAA Coral Reef Watch, Remote Sensing 6 (11) (2014) 11579–11606.doi:10.3390/rs61111579

work page doi:10.3390/rs61111579 2014
[5]

W. J. Skirving, S. F. Heron, B. L. Marsh, G. Liu, J. L. De La Cour, E. F. Geiger, C. M. Eakin, The relentless march of mass coral bleaching: a global perspective of changing heat stress, Coral Reefs 38 (3) (2019) 547–557.doi:10.1007/s00338-019-01799-4

work page doi:10.1007/s00338-019-01799-4 2019
[6]

J. J. Leichter, B. Helmuth, A. M. Fischer, Variation beneath the surface: Quantifying complex thermal environments on coral reefs in the Caribbean, Bahamas and Florida, Journal of Marine Research 64 (4) (2006) 563–588.doi:10.1357/002224006778715711

work page doi:10.1357/002224006778715711 2006
[7]

A. H. Baird, J. S. Madin, M. Álvarez-Noriega, L. Fontoura, J. T. Kerry, C.-Y. Kuo, K. Precoda, D. Torres-Pulliza, R. M. Woods, K. J. A. Zawada, T. P. Hughes, A decline in bleaching suggests that depth can provide a refuge from global warming in most coral taxa, Marine Ecology Progress Series 603 (2018) 257–264.doi:10.3354/meps12732

work page doi:10.3354/meps12732 2018
[8]

T. C. L. Bridge, A. S. Hoey, S. J. Campbell, E. Muttaqin, R. M. Bonaldo, A. H. Baird, Depth- dependent mortality of reef corals following a severe bleaching event: implications for thermal refuges and population recovery, F1000Research 2 (2013) 187.doi:10.12688/f1000research. 2-187.v3

work page doi:10.12688/f1000research 2013
[9]

P. R. Frade, P. Bongaerts, N. Englebert, A. Rogers, M. Gonzalez-Rivero, O. Hoegh-Guldberg, Deep reefs of the Great Barrier Reef offer limited thermal refuge during mass coral bleaching, Nature Communications 9 (1) (2018) 3447.doi:10.1038/s41467-018-05741-0

work page doi:10.1038/s41467-018-05741-0 2018
[10]

deep reef refugia

P. Bongaerts, T. Ridgway, E. M. Sampayo, O. Hoegh-Guldberg, Assessing the “deep reef refugia” hypothesis: focus on Caribbean reefs, Coral Reefs 29 (2) (2010) 309–327.doi:10. 1007/s00338-009-0581-x

2010
[11]

S. J. Bainbridge, Temperature and light patterns at four reefs along the Great Barrier Reef during the 2015–2016 austral summer: Understanding patterns of observed coral bleaching, Journal of Operational Oceanography 10 (1) (2017) 16–29.doi:10.1080/1755876X.2017. 1290863

work page doi:10.1080/1755876x.2017 2015
[12]

Raissi, P

M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics 378 (2019) 686–707.doi:10.1016/j.jcp.2018. 10.045

work page doi:10.1016/j.jcp.2018 2019
[13]

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, Physics- informed machine learning, Nature Reviews Physics 3 (6) (2021) 422–440.doi:10.1038/ s42254-021-00314-5

2021
[14]

Kashinath, M

K. Kashinath, M. Mustafa, A. Albert, J. Wu, C. Jiang, S. Esmaeilzadeh, K. Azizzadenesheli, R. Wang, A. Chattopadhyay, A. Singh, et al., Physics-informed machine learning: case stud- ies for weather and climate modelling, Philosophical Transactions of the Royal Society A 379 (2194) (2021) 20200093.doi:10.1098/rsta.2020.0093. 22

work page doi:10.1098/rsta.2020.0093 2021
[15]

J. D. Willard, X. Jia, S. Xu, M. Steinbach, V. Kumar, Integrating scientific knowledge with machine learning for engineering and environmental systems, ACM Computing Surveys 55 (4) (2023) 1–37.doi:10.1145/3514228

work page doi:10.1145/3514228 2023
[16]

Y. Xiao, Y. Tang, Y. Li, Observation-guided physics-informed neural network (OG-PINN): Application to subsurface ocean temperature and salinity structure reconstructionPreprint, submitted to Climate Dynamics (2026).doi:10.21203/rs.3.rs-8879567/v1

work page doi:10.21203/rs.3.rs-8879567/v1 2026
[17]

L. Han, C. Dong, Y. Liu, H. Xie, H. Zhang, W. Zhu, Application of physics-informed neural networks in solving temperature diffusion equation of seawater, Journal of Oceanology and Limnology 44 (2026) 1–18.doi:10.1007/s00343-025-4348-1

work page doi:10.1007/s00343-025-4348-1 2026
[18]

S. G. Monismith, Hydrodynamics of coral reefs, Annual Review of Fluid Mechanics 39 (2007) 37–55.doi:10.1146/annurev.fluid.38.050304.092125

work page doi:10.1146/annurev.fluid.38.050304.092125 2007
[19]

R. J. Lowe, J. L. Falter, Oceanic forcing of coral reefs, Annual Review of Marine Science 7 (2015) 43–66.doi:10.1146/annurev-marine-010814-015834

work page doi:10.1146/annurev-marine-010814-015834 2015
[20]

S. Wang, Y. Teng, P. Perdikaris, Understanding and mitigating gradient flow pathologies in physics-informed neural networks, SIAM Journal on Scientific Computing 43 (5) (2021) A3055– A3081.doi:10.1137/20M1318043

work page doi:10.1137/20m1318043 2021
[21]

Bradbury, R

J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne, Q. Zhang, JAX: Composable transformations of Python+NumPy programs (2018). URLhttp://github.com/jax-ml/jax

2018
[22]

G. Liu, A. E. Strong, W. J. Skirving, F. Arzayus, Overview of NOAA Coral Reef Watch program’s near-real-time satellite global coral bleaching monitoring activities, in: Proceedings of the 10th International Coral Reef Symposium, 2006, pp. 1783–1793

2006
[23]

J. T. O. Kirk, Light and Photosynthesis in Aquatic Ecosystems, 3rd Edition, Cambridge University Press, Cambridge, 2011.doi:10.1017/CBO9781139168212

work page doi:10.1017/cbo9781139168212 2011
[24]

Donlon, B

C. Donlon, B. Berruti, A. Buongiorno, M.-H. Ferreira, P. Féménias, J. Frerick, P. Goryl, U. Klein, H. Laur, C. Mavrocordatos, et al., The Global Monitoring for Environment and Security (GMES) Sentinel-3 mission, Remote Sensing of Environment 120 (2012) 37–57.doi: 10.1016/j.rse.2011.07.024. 23

work page doi:10.1016/j.rse.2011.07.024 2012