Ensemble-Based Detection of Global Climate Anomalies Using Surface Temperature and CO₂ Emission Indicators: A Majority-Voting Unsupervised Framework with SHAP Interpretability

Significant Research Contributions:

1. First ensemble application at this scale: To the best of the authors’ knowledge, this is the first study applying a majority-voting ensemble of isolation-based, boundary-based, and density-based unsupervised detectors to the joint global temperature and CO₂ record spanning 171 years (1854–2024).

2. Quantitative evaluation without labels: A proxy ground truth of twelve independently documented climate events is used to compute precision, recall, and F1 — achieving Precision = 0.80 with only one false positive, outperforming individual detectors and a z-score baseline by a factor of 3.8 on precision.

3. SHAP-based interpretability: SHAP analysis reveals that rate-of-change features (CO₂_YoY and Temp_YoY) drive anomaly detection more than absolute levels — a physically meaningful and novel finding not recoverable from raw data alone.

4. Ablation-verified feature engineering: A systematic ablation study confirms the full six-feature set is necessary, with precision dropping from 0.80 to 0.38 when only base features are used.

5. Quantified climate acceleration: The anomaly score exhibits a statistically significant downward trend (R² = 0.232, p < 0.001), and the 2020s already register a 40 percent anomaly density against a historical average of 3 percent per decade.