Depth-Resolved Coral Reef Thermal Fields from Satellite SST and Sparse In-Situ Loggers Using Physics-Informed Neural Networks Journal: Remote Sensing of Environment

Alzayat Saleh, Mostafa Rahimi Azghadi · 2026

Read Depth-Resolved Coral Reef Thermal Fields from Satellite SST and Sparse In-Situ Loggers Using Physics-Informed Neural Networks Journal: Remote Sensing of Environment by Alzayat Saleh, Mostafa Rahimi Azghadi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-13131.txt.

Introduction

Saleh and Azghadi build a physics-informed neural network that maps coral reef temperatures below the surface, where satellite readings alone mislead. Fusing satellite sea surface temperature with sparse underwater loggers inside a vertical heat equation, the model learns thermal diffusivity and light attenuation, hitting 0.25 to 1.38 degrees Celsius error at unseen depths. Depth-resolved heat stress falls with depth at Davies Reef, and the model's smooth predictions make its heat-stress values conservative lower bounds.

Catalog record

CREATORAlzayat Saleh, Mostafa Rahimi Azghadi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62605
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.13131

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Depth-Resolved Coral Reef Thermal Fields from Satellite SST and Sparse In-Situ Loggers Using Physics-Informed Neural Networks Journal: Remote Sensing of Environment Alzayat Saleh, Mostafa Rahimi Azghadi 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 m 20\text{\,}\mathrm{m} , where temperatures can be 1– 3 °C 3\text{\,}\mathrm{\SIUnitSymbolCelsius} 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 ( κ \kappa ) and light attenuation ( K d K_{d} ). Validated across four Great Barrier Reef sites (30 holdout experiments), the PINN achieves 0.25– 1.38 °C 1.38\text{\,}\mathrm{\SIUnitSymbolCelsius} RMSE at unseen depths. Under extreme sparsity (three training depths), the PINN maintains 0.27 °C 0.27\text{\,}\mathrm{\SIUnitSymbolCelsius} RMSE at the 5 m 5\text{\,}\mathrm{m} holdout and 0.32 °C 0.32\text{\,}\mathrm{\SIUnitSymbolCelsius} at the 9.1 m 9.1\text{\,}\mathrm{m} holdout, where statistical baselines collapse to > > 1.8 °C 1.8\text{\,}\mathrm{\SIUnitSymbolCelsius} ; 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 m 10.7\text{\,}\mathrm{m} , 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. 1 Introduction Mass coral bleaching has shifted from a rare disturbance to a recurring crisis. Since 2016, the Great Barrier Reef (GBR) alone has experienced six mass bleaching events (2016, 2017, 2020, 2022, 2024, and 2025), each triggered by marine heatwaves that push reef temperatures above coral thermal tolerance thresholds for sustained periods [ 11 , 10 , 1 ] . Satellite sea surface temperature (SST) products, principally the NOAA Coral Reef Watch (CRW) system, are the backbone of global bleaching monitoring: CRW delivers daily 5 km SST and cumulative thermal stress metrics (Degree Heating Weeks) for every reef area on Earth [ 16 , 21 ] . These products have transformed our ability to detect and forecast bleaching events at planetary scale. Satellite SST, however, measures only the skin temperature of the ocean surface. Corals inhabit a range of depths, from reef flats at less than 1\text{\,}\mathrm{m} to reef slopes and walls extending beyond 20\text{\,}\mathrm{m} . In-situ measurements consistently reveal that temperature decreases with depth on coral reefs, with subsurface temperatures 1\text{\,}\mathrm{\SIUnitSymbolCelsius}3\text{\,}\mathrm{\SIUnitSymbolCelsius} cooler than the surface during thermal stress events [ 15 , 3 ] . This vertical thermal gradient means that corals at 10\text{\,}\mathrm{m}20\text{\,}\mathrm{m} may experience far less thermal stress than satellite observations suggest. Bleaching alerts derived from satellite SST alone therefore overestimate thermal stress at depth, a bias that affects management decisions about depth refugia, reef prioritization, and the interpretation of bleaching surveys [ 6 , 8 , 4 ] . Figure 1: Conceptual overview of the PINN framework. Left : Satellite SST provides a surface boundary condition while sparse AIMS temperature loggers supply subsurface observations at discrete depths, leaving gaps in between. Center : A physics-informed neural network embeds the 1D vertical heat equation as a soft constraint, enforces the satellite SST as a hard boundary condition at the surface, and jointly learns effective thermal diffusivity ( \kappa ) and light attenuation ( K_{d} ). Right : The trained PINN produces continuous depth-resolved temperature and Degree Heating Day (DHD) profiles, revealing that thermal stress attenuates with depth; this vertical structure is invisible to satellite-only monitoring. Resolving this depth bias requires depth-resolved thermal fields, but constructing them from observations alone is difficult. In-situ temperature loggers provide point measurements at specific depths, yet even well-instrumented networks such as the AIMS temperature logger program on the GBR cover only tens of sites with varying depth coverage [ 2 ] . Reconstructing continuous thermal fields across the full depth range from these sparse observations demands a method that can honor the measurements, enforce physical consistency between depths, and use satellite SST as a reliable surface anchor. Standard interpolation methods (inverse distance weighting, nearest neighbor, Gaussian process regression) can fill gaps within the observed depth range but lack physical constraints; they degrade rapidly when asked to extrapolate beyond training data or when observations become sparse. Physics-informed neural networks (PINNs) offer a framework for this problem. PINNs embed governing physical equations directly into the neural network training process, producing solutions that simultaneously fit observed data and approximately satisfy the underlying physics [ 20 , 12 ] . For depth-resolved temperature reconstruction, the relevant physics is the vertical heat equation: temperature evolves through a balance of vertical diffusive mixing and depth-attenuated solar heating. By encoding this equation as a soft constraint, a PINN can interpolate and extrapolate temperature fields that are physically plausible even where no observations exist. This is particularly valuable for coral reef applications, where the physics is well understood but observations are sparse. PINNs have demonstrated success in related geophysical inverse problems, including subsurface flow reconstruction and ocean state estimation [ 13 , 23 ] . Recently, PINNs have been applied to subsurface ocean temperature reconstruction from surface observations [ 24 , 9 ] , though none have targeted coral reef environments or the specific challenge of depth-resolved thermal stress assessment. Figure 1 illustrates the approach we develop here: satellite SST and sparse in-situ loggers are fused through a PINN constrained by the vertical heat equation, producing continuous depth-resolved temperature and thermal stress profiles. This paper makes three contributions: 1. A PINN framework for reconstructing depth-resolved thermal fields from satellite SST and sparse temperature loggers, enforcing the one-dimensional heat equation with depth-dependent solar heating and jointly learning effective thermal diffusivity ( \kappa ) and light attenuation ( K_{d} ) from data. 2. Holdout validation across four GBR sites, demonstrating reconstruction accuracy ranging from 0.25 to 1.38\text{\,}\mathrm{\SIUnitSymbolCelsius} RMSE and stability under extreme data sparsity where statistical baselines collapse. 3. Depth-resolved Degree Heating Day profiles that reveal vertical attenuation of thermal stress, with matched-window logger validation demonstrating that the PINN captures the qualitative depth structure of DHD while providing conservative absolute estimates. The remainder of this paper is organized as follows. Section 2 describes the four GBR study sites and the satellite and in-situ data sources. Section 3 details the PINN formulation, network architecture, baseline methods, and experimental design. Section 4 presents holdout validation, sparsity robustness, depth-resolved thermal stress comparisons, and learned physical parameters. Section 5 discusses implications for bleaching monitoring, the PINN’s advantages and limitations, and future directions. Section 6 summarizes the findings. 2 Study area and data 2.1 Great Barrier Reef study sites We selected four reefs in the central GBR that span a range of shelf positions, depth ranges, and bleaching histories (Fig. 2 ; Table 1 ). Davies Reef ( 18.83\text{\,}\mathrm{\SIUnitSymbolDegree} S, 147.63\text{\,}\mathrm{\SIUnitSymbolDegree}

Read the complete text