The ocean absorbs about 90% of the extra heat trapped by greenhouse gases. This has caused ocean temperatures to rise steadily, with marine heat waves becoming more frequent and intense. For corals living near their thermal limits, this is devastating.
Scientists use Degree Heating Weeks (DHW) to measure cumulative heat stress on coral reefs. DHW tracks how long temperatures stay above the local bleaching threshold (typically 1°C above the warmest monthly average).
Image classification is a machine learning task where a computer system learns to assign categories to images. The classic example is teaching a computer to distinguish cats from dogs: it's trained on thousands of labeled cat and dog photos, learns the visual patterns that distinguish them, and then can classify new photos it hasn't seen before.
For coral bleaching, the classification task is similar: train a model on satellite images labeled "healthy reef," "partially bleached reef," or "severely bleached reef," and the model learns to classify new satellite images into these categories. But instead of learning cat vs. dog features, the model learns the color, texture, and spectral patterns that distinguish healthy from stressed corals.
A healthy reef has distinct colors depending on depth and composition: deep water is dark blue, shallow water over sand is lighter blue or turquoise, and coral-rich areas show browns, greens, or reds reflecting the colors of corals and algae. A bleached reef looks whiter or lighter in the same location—the lack of zooxanthellae causes corals to reflect more light, making the reef appear washed out.
Image classification models work by analyzing the pixel values of an image. Each pixel contains measurements in multiple spectral bands: red, green, blue (visible light), and infrared. The model learns which combinations of values are associated with which classifications. If healthy reefs tend to have high green and low blue pixel values in specific locations, while bleached reefs have low green and high blue, the model learns these patterns and uses them to classify new images.
Building a working classification model requires five steps: data collection, labeling, preprocessing, training, and validation.
Data collection: Scientists gather satellite imagery from the region of interest—perhaps several years of MODIS or Landsat imagery covering a particular reef system. The imagery might be downloaded from NOAA, NASA, or a similar data provider.
Labeling: A subset of the collected images—typically 5-10% of the full dataset—is manually labeled by expert scientists who have conducted field surveys or have deep knowledge of the reefs. Each image is marked as "healthy," "stressed," or "bleached" based on expert judgment and field validation. This labeled dataset becomes the "training data" that teaches the model.
Preprocessing: Raw satellite images contain noise and artifacts. Scientists standardize the images, removing clouds, correcting for atmospheric effects, and normalizing brightness and contrast. They also resize images to a consistent size so all images are analyzed at the same scale.
Training: The model ingests thousands of labeled training images and learns the pixel patterns associated with each classification. Modern image classification uses deep neural networks—specifically, convolutional neural networks (CNNs) that are designed to recognize spatial patterns and textures in images. During training, the network adjusts internal parameters thousands of times, progressively improving its ability to classify the training images correctly.
Validation: Once training is complete, the model is tested on images it has never seen before—the remaining 90% of the collected data that was held back from training. If the model correctly classifies 85-90% of validation images, it's ready for deployment. If accuracy is lower, scientists debug the problem: maybe more training data is needed, maybe the image preprocessing needs adjustment, or maybe the spectral bands used don't contain enough information to distinguish categories reliably.
No classification model is 100% accurate. It will sometimes incorrectly classify a healthy reef as bleached (a false positive) or fail to detect actual bleaching (a false negative). Understanding these errors is critical for interpreting AI predictions.
False positives occur when the model flags a reef as bleached when it's actually healthy. This might happen if cloudy water, sediment, or algae blooms create color patterns similar to bleaching. False positives lead to false alarms—scientists dispatch expensive field surveys only to discover the reef is fine.
False negatives occur when the model misses actual bleaching. This is more dangerous than false positives because it means scientists don't send help to reefs that need it. A model with 95% accuracy might still miss 5% of actual bleaching events, leaving some reefs unmonitored.
Scientists measure accuracy using several metrics: - Overall accuracy: Percentage of all predictions that are correct - Sensitivity (or recall): What percentage of actual bleached reefs does the model correctly identify? (1 - false negative rate) - Specificity: What percentage of healthy reefs does the model correctly identify? (1 - false positive rate) - Precision: Of all the reefs the model flags as bleached, what percentage are actually bleached?
A model with 95% overall accuracy, 88% sensitivity, and 98% specificity might be considered adequate for monitoring because it catches most real bleaching events (88%) while keeping false alarms at a low rate (2%).
Scientists also consider the cost of errors. For reef conservation, false negatives (missing real bleaching) are typically more costly than false positives (unnecessary surveys), so they often tune models to be more sensitive—catching nearly all real bleaching even if it means more false alarms.
In 2016, a severe global bleaching event affected two-thirds of the Great Barrier Reef. Detecting and mapping this event required data from multiple sources: thermal monitoring showed extremely high DHW values, field surveys conducted by boat and helicopter documented massive bleaching, aerial photographs showed the color change across reef systems, and satellite imagery confirmed the spatial extent.
Scientists trained a new image classification model on a combination of historical Landsat and MODIS imagery from the Great Barrier Reef, labeled with field survey data from previous years. The model learned the spectral signatures of healthy reef, moderately stressed reef, and severely bleached reef in this specific region.
When applied to 2016 satellite imagery, the model automatically flagged the affected areas, providing a rapid assessment of the spatial extent of bleaching across thousands of square kilometers. The AI predictions aligned well with field surveys—where the model predicted severe bleaching, divers found widespread coral death; where the model predicted moderate bleaching, surveys found some coral survival and recovery.
However, the model also made mistakes. In some areas with heavy cloud cover or fresh water plumes, it miscategorized conditions. Scientists had to manually review the model's predictions against other data sources before issuing final assessments. But the AI model provided a first-pass analysis that guided where field surveys would be most valuable, dramatically reducing the time and cost required to characterize the bleaching event.
The 2016 case demonstrates the value of AI monitoring: within days of acquiring satellite imagery, scientists had automatically processed imagery covering tens of thousands of square kilometers and identified the most severely affected regions. Without automation, conducting manual visual analysis of thousands of satellite images would have taken weeks or months.
In this activity, you'll take on the role of an AI training data labeler and a model validator.
Part 1 - Manual classification: Your teacher will provide 10-15 satellite image patches showing different reef regions under various conditions. For each image, classify the reef as "healthy," "stressed," or "bleached" based on color and visual patterns. Note the difficulty of distinguishing categories in edge cases—some images might be ambiguous.
Part 2 - Compare to model predictions: Your teacher will then reveal the classifications made by a trained AI model on the same images. In how many cases did you and the model agree? Where did you disagree? Were the disagreements in cases you found ambiguous as well?
Part 3 - Assess real-world implications: Imagine these predictions drive field survey decisions. For one of the images where you and the model disagreed, discuss what would happen: if the model was wrong, scientists might survey a healthy reef (wasting time) or miss a bleached reef (harmful outcome). What information would you need to verify which classification is correct?
You now understand how neural networks learn to recognize bleaching in satellite imagery and how to interpret AI predictions, including their limitations. In Level 4, you'll explore multi-factor predictive models that combine temperature data, satellite imagery, and other variables to forecast which reefs are most at risk before bleaching occurs.