What condition best predicts a possible bleaching event?

Summary

Rising sea surface temperatures are a primary driver of coral reef bleaching, but predicting these events globally is difficult. What counts as "hot" to corals would vary drastically from region to region. This project built a supervised machine learning pipeline to predict coral bleaching events. By pivoting from a struggling global baseline model to a regionally-tuned model, and by engineering a Cumulative Heat Stress feature, the final Random Forest classifier successfully prioritized "early warnings" (Recall) to identify possible bleaching events. 

The Problem with a Simple Global Model 

An initial exploratory data analysis revealed that corals do not bleach simply because the water reaches a specific temperature. They bleach due to prolonged thermal stress. Feeding raw Sea Surface Temperature (SST) into a baseline global Random Forest resulted in a poor recall rate (30%), meaning the model missed 70% of bleaching events. The model was confused by the varying biological thresholds of different oceans. 

Feature Engineering & Model Iteration  

To solve this, I implemented an iterative testing approach to model optimization in three different ways:

• Feature Engineering: I engineered a Heat_Stress_12Wk feature. By calculating the 12-week rolling sum of temperature anomalies for each region, similar to NOAA's "Degree Heating Weeks" metric. This proved to be a much better predictor than daily temperatures.

• Hyperparameter Tuning: I used GridSearchCV to constrain tree depth, preventing the model from overfitting.

• Regionalization: I isolated the data to a single area (The Great Barrier Reef). This allowed the algorithm to learn the specific biological rules of one environment rather than searching for an impossible "global" rule.

• Threshold Adjustment: I lowered the decision threshold to 0.40. In a conservation context, a False Positive is far less costly than a False Negative (missing a mass bleaching event entirely).

Limitations & Next Steps

Because this model was restricted to the Great Barrier Reef to improve accuracy, the test set size was reduced to just 18 samples. While the predictive patterns are mathematically sound, the model requires a larger volume of localized historical data before it can be realistically functional.