🤖 machine learning

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

This paper introduces a physics-informed neural network that fuses satellite sea surface temperature data with sparse in-situ logger measurements to accurately reconstruct depth-resolved thermal fields and coral bleaching stress across the Great Barrier Reef, significantly outperforming statistical and physics-only baselines even under extreme data sparsity.

Original authors: Alzayat Saleh, Mostafa Rahimi Azghadi

Published 2026-04-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alzayat Saleh, Mostafa Rahimi Azghadi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Surface Illusion"

Imagine you are trying to understand the temperature of a swimming pool, but you are only allowed to look at the very top layer of water. You see the sun baking the surface, making it feel like a scorching hot tub. You assume the water at the bottom is just as hot.

But it's not.

In the ocean, specifically on coral reefs, the surface water (measured by satellites) can be 1 to 3 degrees Celsius hotter than the water just a few meters down. Corals live at all these different depths. If we only look at the satellite data, we think the deep corals are baking in heat stress when, in reality, they might be in a cool, safe "refuge."

This is like judging the temperature of a whole house by sticking a thermometer in the attic on a sunny day. You'd think the basement is on fire, but it's actually quite comfortable.

The Solution: The "Physics Detective" (PINN)

The researchers created a smart computer program called a Physics-Informed Neural Network (PINN). Think of this PINN as a detective who solves a mystery using two types of clues:

The Satellite Clue (The Surface): It knows exactly how hot the ocean skin is every day.
The Logger Clue (The Sparse Points): It has a few scattered temperature sensors (loggers) dropped at specific depths, like a few people standing at different floors of a skyscraper reporting the temperature.

The Problem: The loggers are too far apart. There are huge gaps between them where we don't know the temperature.

The Magic Trick: Instead of just guessing the temperature in the gaps (which usually fails), the PINN uses the laws of physics as a rulebook. It knows that heat moves slowly through water and that sunlight gets weaker the deeper you go.

Analogy: Imagine trying to draw a smooth curve connecting a few dots on a piece of paper.
- A statistical guess might draw a jagged, crazy line that fits the dots but makes no sense physically.
- The PINN is like a ruler that must follow the laws of gravity and fluid dynamics. It draws a smooth, realistic curve that connects the dots and obeys the rules of how heat actually behaves.

How It Works (The "Recipe")

The Anchor: The PINN is "anchored" to the satellite data at the very top (the surface). It knows, "At depth 0, the temperature must be exactly what the satellite says."
The Learning: It then looks at the few deep sensors it has. It asks, "How much cooler is it down here? How fast does the heat disappear?"
The Prediction: It fills in the blanks for every single meter of depth, creating a continuous 3D picture of the reef's temperature.

Why This Matters: The "Deep Dive" Results

The researchers tested this on four different reefs in the Great Barrier Reef. Here is what they found:

The "Collapse" of Old Methods: When they removed most of the deep sensors (making the data very sparse), old computer methods (statistical guesses) completely failed. They started predicting wild, wrong temperatures.
The PINN's Superpower: Even with only three sensors, the PINN stayed accurate. It didn't panic because it was relying on the "rulebook" of physics to guide it.
The "Heat Stress" Map: They calculated "Degree Heating Days" (a measure of how much heat stress corals feel).
- Satellite View: Shows high stress everywhere, from the surface to the deep.
- PINN View: Shows that the stress drops off as you go deeper. At 10 meters, the stress might be zero, even if the surface is boiling.

The Catch: The "Smoothie" Effect

There is one small limitation. The PINN is very good at showing the average temperature and the trend of cooling with depth. However, it tends to "smooth out" the data.

Analogy: Imagine a coral reef experiencing a sudden, short burst of extreme heat for just one hour.
- The real sensor sees this spike and says, "Wow, that was dangerous!"
- The PINN sees the average and says, "It was a bit warm, but not a crisis."
Why? Because the satellite data it uses is an average (daily), so the PINN can't "see" those tiny, dangerous spikes.
The Verdict: The PINN is a conservative estimate. It tells us, "The stress is at least this low." It might miss the very worst moments, but it is excellent at telling us that the deep water is generally a safe haven.

The Bottom Line

This paper introduces a new way to save coral reefs. By combining satellite photos with a few cheap underwater sensors and a smart "physics detective," we can finally see the 3D temperature map of a reef.

This helps scientists and managers realize that deep water is a lifeboat. While the surface might be too hot for corals, the deep corals might be surviving. This changes how we protect reefs, allowing us to focus on the "safe zones" that satellites were previously ignoring.

In short: We stopped guessing the temperature of the deep ocean and started calculating it using the laws of nature, revealing that the deep reef is cooler and safer than we thought.

1. Problem Statement

Global coral bleaching monitoring relies heavily on satellite Sea Surface Temperature (SST) products (e.g., NOAA Coral Reef Watch). However, these products measure only the ocean's "skin" temperature. Corals inhabit a vertical range from shallow reef flats to depths exceeding 20 meters, where temperatures can be 1–3°C cooler than the surface during thermal stress events.

The Bias: Applying surface SST uniformly to all depths leads to a systematic overestimation of subsurface thermal stress.
The Data Gap: While in-situ temperature loggers exist (e.g., AIMS network), they provide only sparse, point measurements at discrete depths. Standard interpolation methods (e.g., Gaussian Processes, Inverse Distance Weighting) fail to extrapolate accurately when data is sparse or when predicting at unobserved depths because they lack physical constraints.
The Goal: To reconstruct continuous, depth-resolved thermal fields that accurately capture the vertical attenuation of thermal stress using a fusion of satellite SST and sparse in-situ loggers.

2. Methodology: Physics-Informed Neural Network (PINN)

The authors propose a PINN framework that embeds the governing physics of heat transfer directly into the neural network training process.

A. Governing Physics

The model solves the one-dimensional vertical heat equation:
$\frac{\partial T}{\partial t} = \kappa \frac{\partial^2 T}{\partial z^2} + \frac{Q_{solar}(z, t)}{\rho c_p}$
Where:

$T(z, t)$ is temperature at depth $z$ and time $t$ .
$\kappa$ is the effective thermal diffusivity (learnable).
$Q_{solar}$ is depth-attenuated solar heating derived from Beer's Law, dependent on the light attenuation coefficient $K_d$ (learnable).
The model assumes vertical mixing dominates over lateral advection for reef-scale profiles.

B. Network Architecture & Constraints

Hard Boundary Condition: The satellite SST is enforced as an exact boundary condition at the surface ( $z=0$ ). The network output is structured as $T(z,t) = SST(t) + \frac{z}{z_{max}} T_{scale} f_{nn}(z,t)$ , ensuring the prediction matches the satellite data exactly at the surface regardless of network weights.
Learnable Parameters: The network jointly learns the effective thermal diffusivity ( $\kappa$ ) and light attenuation coefficient ( $K_d$ ) alongside the network weights. This allows the model to adapt to specific reef conditions (e.g., wave energy, water clarity).
Architecture: A modified Multilayer Perceptron (MLP) with multiplicative gating and Fourier feature encoding for time (diurnal and annual cycles) to capture periodic patterns.
Loss Function: A composite loss minimizes:
1. Data Loss: MSE between predictions and logger observations.
2. PDE Loss: Residual of the heat equation at collocation points (enforcing physical consistency).
3. Boundary Loss: Neumann (zero-flux) condition at the bottom depth.
- Curriculum Learning: The weight of the PDE loss is gradually decreased during training to allow the network to first learn the physical structure and then fine-tune to specific data points.

C. Experimental Design

Study Sites: Four Great Barrier Reef (GBR) sites (Davies, Myrmidon, Rib, Kelso) with varying depths, shelf positions, and historical bleaching events.
Validation Strategy: "Holdout" experiments where one depth is excluded from training and used for testing.
Sparsity Tests: Experiments reducing training depths to as few as 2–3 to test robustness.
Baselines: Compared against statistical methods (Gaussian Process, IDW, Nearest Neighbor, Random Forest), a physics-only Finite Difference (FD) solver, and the standard Satellite-only approach.

3. Key Contributions

Novel Framework: A PINN architecture that fuses satellite SST (surface boundary) with sparse loggers to reconstruct continuous vertical temperature profiles.
Joint Parameter Learning: The ability to simultaneously estimate effective thermal diffusivity ( $\kappa$ ) and light attenuation ( $K_d$ ) from data, providing physical consistency checks.
Robustness to Sparsity: Demonstrated that the PINN maintains accuracy even with extremely sparse data (3 training depths), a regime where statistical baselines fail catastrophically.
Depth-Resolved Stress Metrics: Generation of depth-resolved Degree Heating Days (DHD) profiles, revealing that thermal stress attenuates significantly with depth.

4. Key Results

A. Reconstruction Accuracy

Overall Performance: The PINN achieved Root Mean Square Error (RMSE) between 0.25°C and 1.38°C across 30 holdout experiments.
Sparsity Robustness:
- At Davies Reef (9.1 m holdout) with only 3 training depths, the PINN maintained an RMSE of 0.32°C.
- In contrast, statistical baselines (NN, IDW) collapsed to >1.8°C RMSE under the same conditions.
- The PINN outperformed the physics-only Finite Difference baseline in 90% of experiments, proving that data-driven adaptation of parameters adds value over fixed-parameter physics models.

B. Thermal Stress Attenuation (DHD)

Vertical Gradient: The study confirmed that thermal stress decreases with depth.
- Example (Davies Reef): Satellite DHD was constant at 0.31 across all depths. The PINN and loggers showed DHD dropping to zero by 10.7 m, identifying a "depth refugia" invisible to satellites.
- Example (Rib Reef): DHD dropped from 3.33 at 1 m to 0.14 at 9 m (PINN), whereas satellite data remained constant at 3.08.
Systematic Limitation: The PINN underestimates absolute DHD at shallow/mid depths compared to loggers.
- Cause: The PINN is anchored to daily satellite SST, which lacks the sub-hourly temporal resolution to capture short-duration thermal peaks that drive threshold exceedances.
- Interpretation: PINN DHD values should be treated as conservative lower bounds. While the absolute magnitude may be underestimated, the shape of the depth-stress profile (attenuation) is reliable.

C. Learned Physical Parameters

Thermal Diffusivity ( $\kappa$ ): Learned values ranged from $0.6 \times 10^{-4}$ to $5.0 \times 10^{-4} \, m^2 s^{-1}$ , consistent with literature for turbulent mixing on coral reefs.
Light Attenuation ( $K_d$ ): Values varied by reef, reflecting differences in water clarity (e.g., lower $K_d$ for offshore clear waters, higher for turbid mid-shelf reefs).

5. Significance and Implications

Management Impact: Current bleaching alerts based on satellite SST likely overestimate stress at depth, potentially misidentifying "refugia" (safe zones) or misprioritizing reefs. This framework provides the necessary depth-resolved data to correct these biases.
Infrastructure Efficiency: The method requires no new hardware. It utilizes existing satellite data and sparse logger networks (e.g., AIMS), adding a physics-constrained computational layer to convert point data into 3D fields.
Scalability: Training a PINN for a reef takes ~4 minutes on a consumer GPU. The framework is scalable to the ~3,000 reefs in the GBR and potentially globally, provided 2–3 logger depths exist per site.
Future Directions: The authors suggest future work on peak-preserving loss functions to improve DHD accuracy, 2D extensions for lateral gradients, and transfer learning to apply models to reefs with minimal data.

Conclusion

This paper demonstrates that Physics-Informed Neural Networks can effectively bridge the gap between surface satellite observations and sparse subsurface measurements. By enforcing physical laws of heat transfer, the PINN reconstructs depth-resolved thermal fields that reveal significant vertical attenuation of thermal stress, offering a critical tool for more accurate, three-dimensional coral bleaching assessment in a warming climate.