Evaluating satellite and modeled lake surface water temperature across the contiguous United States.

This study developed and validated random forest models using Sentinel-3, Landsat, and in situ data to accurately predict surface water temperatures across U.S. lakes, thereby providing a critical tool for improving cyanobacterial harmful algal bloom forecasting and water resource management.

The Gist

The Big Picture: Why Do We Care About Lake Temperatures?

Imagine a lake as a giant, natural swimming pool. Just like in a pool, the water temperature dictates what can live there. When the water gets too warm, a specific type of algae called cyanobacteria (often called "blue-green algae") throws a wild party. They multiply rapidly, forming harmful algal blooms (cyanoHABs).

These blooms are like a toxic fog over the water. They can make people and pets sick, kill fish, and force beaches to close. To stop these parties before they start, scientists need to predict exactly when and where the water will get warm enough for the algae to thrive.

The Problem: You can't put a thermometer in every single lake in the United States. There are too many, and they are too far apart. It's like trying to check the temperature of every puddle in a city by walking to each one; you'd miss most of them.

The Solution: Two Different "Thermometers"

The researchers tried to solve this by building a "super-predictor" using two different sources of data, kind of like trying to guess the weather using two different tools:

1. The "Boots on the Ground" Team (In Situ Data): These are real thermometers dropped into lakes by scientists and volunteers.
   • Pros: They are incredibly accurate.
   • Cons: They are sparse. You only get a reading if someone happened to be there that day. It's like having a weather report for only three cities in the whole country.
2. The "Eye in the Sky" Team (Satellite Data): These are satellites (like Landsat) that fly overhead and take pictures of the water's "skin" temperature.
   • Pros: They cover almost every lake, every day.
   • Cons: They are easily confused by clouds. If a cloud passes over, the satellite can't see the water, or the cloud makes the water look colder than it is. It's like trying to take a selfie with a friend, but a cloud keeps blocking the camera lens.

The Experiment: Teaching a Robot to Predict

The team used a powerful computer algorithm called a Random Forest. Think of this algorithm as a very smart student who learns by looking at thousands of examples. They trained this student in two different ways:

• Class A: The student studied only the "Boots on the Ground" data (real thermometers).
• Class B: The student studied the "Eye in the Sky" data (satellite photos).

They then tested both students to see who could predict the temperature of lakes they had never seen before.

The Results: Who Won?

The Winner: The "Boots on the Ground" Student (In Situ Model)

• Performance: This student was the most accurate. It made very few mistakes.
• The Catch: Because it only learned from the few lakes where real thermometers existed, it struggled to guess temperatures for lakes in places where no one had ever measured the water. It was a great student, but it had a very limited library of books to study from.

The Runner-Up: The "Eye in the Sky" Student (Satellite Model)

• Performance: This student had a much bigger library (data from thousands of lakes), but it made more mistakes.
• The Issue: The satellite data was "noisy." Clouds and shadows confused the satellite, making the water look colder or warmer than it really was. Even though the student studied more examples, the examples were sometimes flawed.
• The Cloud Problem: The researchers found that even if a cloud was far away from the lake they were looking at, it could still mess up the temperature reading. It's like how a storm in the next town can change the wind direction where you are standing.

The "Secret Sauce": What Makes a Lake Warm?

The computer model figured out that to predict lake temperature, you don't just need the current air temperature. You need to know:

• Where the lake is: Lakes in the south are warmer; lakes in the north are cooler.
• How high up it is: High mountain lakes are colder.
• How big it is: Small ponds heat up and cool down faster than giant lakes.
• The weather history: What was the temperature 30 days ago? (Lakes are slow to change, like a giant bathtub that takes a long time to warm up).

The Takeaway: Why This Matters

The researchers successfully built a system that can predict lake temperatures for 2,192 lakes across the entire US, filling in the gaps where no thermometers exist.

• For the Public: This means better warnings. If the model predicts a lake is getting warm enough for toxic algae, health officials can close the beach before people get sick.
• For the Future: While the satellite data wasn't perfect on its own, combining it with the real thermometer data creates a powerful tool. It's like having a weather forecast that combines local reports with satellite imagery to give you the most accurate prediction possible.

In a nutshell: The scientists built a digital crystal ball that uses real measurements and satellite photos to tell us which lakes are about to get "too hot to handle," helping us keep our water safe and our beaches open.

Technical Summary

1. Problem Statement

Routine monitoring of lake surface water temperature (LSWT) is critical for predicting cyanobacterial harmful algal blooms (cyanoHABs), which pose risks to human health, aquatic ecosystems, and the economy. However, existing monitoring faces significant challenges:

• Data Gaps: In situ measurements are sparse, inconsistent in methodology, and lack spatial/temporal continuity. Over 95% of U.S. waterbodies in the National Water Information System lack temperature data.
• Satellite Limitations: While satellite remote sensing (e.g., Landsat) offers broad coverage, it is susceptible to cloud cover, cloud shadows, and atmospheric interference, leading to data gaps and retrieval errors. Previous attempts to use Landsat data for cyanoHAB forecasting (e.g., in Florida) required supplementing gaps with climatology due to these limitations.
• Modeling Needs: There is a need for a continuous, daily, high-resolution temperature dataset for all Sentinel-3 Ocean and Land Colour Instrument (OLCI) resolved lakes (2,192 lakes in the contiguous U.S.) to support national-scale cyanoHAB forecasting.

2. Methodology

The study employed a comparative modeling approach using Random Forest algorithms to generate daily LSWT predictions.

Data Sources:

• Target Lakes: 2,192 lakes in the contiguous U.S. (CONUS) resolved by the 300-m Sentinel-3 OLCI sensor.
• In Situ Data: High-quality discrete measurements from the National Water Information System (NWIS, 2016–2022) and National Lakes Assessment (NLA, 2007, 2012, 2017). Samples were filtered to the upper 2m of the water column.
• Satellite Data: Landsat Analysis Ready Data (ARD) from Landsat 7, 8, and 9 (2016–2022).
  • Filtering Strategy: Two distinct Landsat datasets were created to test cloud sensitivity:
    1. Landsat(LakeCloudFree): Pixels within the lake were clear, but the scene could contain clouds/shadows outside the lake.
    2. Landsat(SceneCloudFree): Scenes with <1% cloud cover and shadow across the entire image.

Modeling Approach:

• Algorithm: Random Forest models (100 trees) were trained using the randomForest R package.
• Input Variables: Lake centroid coordinates (Northing/Easting), elevation, static shoreline length, surface area, day of year, daily air temperature, and rolling 30-day average air temperature (from PRISM data).
• Three Models Developed:
  1. Trained on Landsat(LakeCloudFree) data.
  2. Trained on Landsat(SceneCloudFree) data.
  3. Trained on In Situ data (80% training, 20% validation).
• Validation: Models were validated against independent in situ data (same calendar day matchups). Performance metrics included RMSE, MAE, Bias, and $R^2$.

3. Key Contributions

• Comprehensive Validation: The study provides a rigorous validation of Landsat-derived LSWT under varying cloud conditions, quantifying how cloud cover and shadows in the broader scene affect retrieval accuracy even when the lake itself appears clear.
• National-Scale Dataset: Generated a continuous, daily predicted LSWT dataset for all 2,192 OLCI-resolved lakes across the CONUS for the entire calendar year.
• Methodological Comparison: Directly compared the efficacy of training predictive models on raw satellite data versus sparse in situ data, highlighting the trade-offs between data volume and data quality.
• CyanoHAB Support: Created a foundational temperature layer specifically designed to improve the accuracy of national cyanoHAB forecasting models.

4. Key Results

Satellite Validation:

• Cloud Impact: Cloud cover significantly degrades Landsat accuracy.
  • Unfiltered (LakeCloudFree): MAE = 2.99°C, Bias = -1.3°C.
  • Strictly Cloud-Free (<1% SceneCloudFree): MAE = 0.94°C, Bias = +0.39°C.
• Conclusion: Even if the lake pixel is clear, clouds elsewhere in the scene introduce error, likely due to atmospheric correction uncertainties.

Model Performance:

• In Situ Model: Outperformed both satellite-trained models.
  • RMSE: 1.11°C (Applied Validation)
  • MAE: 0.77°C
  • Bias: 0.01°C
  • It showed no spatial bias and captured seasonal trends accurately.
• Landsat(LakeCloudFree) Model: Consistently underestimated temperatures (Bias = -1.97°C). While it had broad spatial coverage, the error was high (MAEapplied ~2.98°C).
• Landsat(SceneCloudFree) Model: Showed a positive bias (overestimation, Bias = +0.81°C) and high error (MAEapplied ~2.48°C).
  • Spatial Bias: Overestimated temperatures in the western U.S. (high elevation/dry regions) and underestimated in the eastern U.S.
  • Seasonal Bias: Poor performance in winter months due to a lack of training data (clouds prevent winter satellite observations), leading to unrealistic warm predictions in winter.

Variable Importance:

• In Situ Model: Most driven by air temperature (daily and 30-day rolling) and day of year.
• Landsat Models: Heavily influenced by spatial coordinates (Easting/Northing) and elevation, suggesting the models relied on geographic proxies to compensate for the lack of direct thermal observations in certain conditions.

5. Significance and Implications

• Water Quality Management: The developed In Situ-based Random Forest model provides the most reliable continuous temperature dataset for national management. It fills critical spatio-temporal gaps, enabling daily to weekly forecasting of cyanoHABs.
• Limitations of Satellite-Only Approaches: The study demonstrates that training models directly on satellite data (even with strict cloud filtering) is inferior to training on high-quality in situ data due to atmospheric correction errors and seasonal sampling biases (e.g., lack of winter data).
• Future Directions: The authors suggest that while Random Forests are effective within observed ranges, future work should explore hybrid approaches (e.g., combining Random Forests with Long Short-Term Memory networks) and the inclusion of wind speed data to better model physical mixing processes.
• Public Health: By improving the accuracy of temperature predictions, this tool directly supports early warning systems for harmful algal blooms, protecting drinking water supplies and recreational areas.

In summary, while satellite data offers necessary spatial coverage, the study concludes that models trained on high-quality in situ data yield superior predictive accuracy for lake surface temperatures, serving as a robust tool for national ecosystem protection and public health forecasting.