Integrating Explainable Machine Learning, Extreme Value Analysis, and CMIP6 Projections for Extreme Precipitation Prediction in the Eastern Mediterranean Coastal Region of Türkiye

Technical Summary: Integrating Explainable Machine Learning, Extreme Value Analysis, and CMIP6 Projections for Extreme Precipitation Prediction in the Eastern Mediterranean Coastal Region of Türkiye

Problem Statement
The eastern Mediterranean coastal region of Türkiye (encompassing Mersin, Adana, Osmaniye, and Hatay) faces severe risks from extreme precipitation events, which threaten densely populated urban centers and agricultural lands. Despite this vulnerability, previous research has lacked a unified analytical framework that integrates physical drivers, statistical recurrence properties, and future climate trajectories for this specific subregion. Existing studies often rely on isolated station analyses, fail to correlate recurrence period estimates with machine learning (ML) prediction models, or do not adequately account for the heterogeneity of precipitation regimes caused by complex topography (e.g., the Taurus Mountains). Furthermore, while CMIP6 models predict significant warming and summer drought, the specific future behavior of extreme precipitation frequency versus intensity remains unclear at the local level.

Methodology
This study presents a hybrid machine learning and statistical pipeline designed to address these gaps through eight integrated stages:

1. Data Acquisition and Quality Control: The study utilized 24 Turkish State Meteorological Service (TSMS) ground stations (2010–2024), augmented by CHIRPS satellite precipitation data, ERA5-Land reanalysis, and NOAA teleconnection indices. A rigorous two-stage quality control process excluded stations with >5% missing data and applied homogeneity tests (SNHT, Pettitt, Buishand) to identify regime shifts.
2. Multi-Source Data Fusion: A new daily dataset was created by fusing ground observations, satellite data, and reanalysis products. This involved spatial inference, sea-masking corrections for coastal stations, bias correction for temperature variables, and gap filling.
3. Spatial Clustering: K-Means clustering partitioned the 24 stations into three physically interpretable precipitation regime zones: Coastal, Inland, and Mountainous, based on elevation, precipitation totals, seasonal ratios, and temperature.
4. Feature Engineering: A 40-attribute predictor matrix was constructed, including meteorological variables, topographic data, temporal encodings, cluster labels, and 30 teleconnection features (lagged indices for ONI, NAO, AO, AMO, EAWR).
5. Machine Learning Classification: An XGBoost binary classifier was trained to predict extreme precipitation (defined as daily rainfall ≥95th percentile). The model addressed class imbalance (1:18 ratio) using inverse frequency weighting. Performance was evaluated using temporal cross-validation and an independent test set (2023–2024).
6. Explainability (XAI): SHAP (SHapley Additive exPlanations) analysis was employed to interpret the XGBoost model, identifying dominant predictors and their nonlinear interactions across different spatial clusters.
7. Extreme Value Analysis (EVA): Return periods were estimated using 44 years of CHIRPS data (1981–2024) via two complementary methods: Generalized Extreme Value (GEV) distribution for annual maxima and Generalized Pareto Distribution (GPD) for peaks over threshold (POT).
8. Future Projections: CMIP6 projections from the MPI-ESM1-2-LR model under SSP2-4.5 and SSP5-8.5 scenarios were bias-corrected and analyzed for the 2041–2060 horizon to assess changes in extreme event frequency and wet-day intensity.

Key Results

• Regime Shift: Homogeneity testing identified a statistically significant precipitation regime shift beginning in 2020. The domain-mean anomaly relative to the preceding decade was −20.5%, with the most severe decreases observed in inland stations (e.g., Osmaniye: −35.9%). This shift coincided with a change in the target variable distribution between the training (2010–2022) and test (2023–2024) periods.
• Model Performance: The XGBoost classifier achieved an AUC-ROC of 0.835 on the independent test set, demonstrating robust discriminatory ability despite the non-stationary conditions. The F1 score on the test set was 0.288, lower than the cross-validation score (0.403), attributed to the reduced frequency of extreme events in the post-2020 drought period.
• Feature Importance: SHAP analysis revealed near-surface relative humidity as the overwhelmingly dominant predictor (mean |SHAP| = 0.557). Its importance intensified with elevation (0.469 in Coastal vs. 0.642 in Mountainous clusters), reflecting the orographic precipitation mechanism. A nonlinear activation threshold was identified near 65–70% relative humidity. Mean temperature was the second most important feature, showing a negative contribution above 15–20°C.
• Teleconnections: Large-scale indices AMO (negative correlation, r = −0.335) and EAWR (positive correlation, r = +0.299) were identified as significant drivers of interannual variability.
• Return Levels: GEV fitting yielded a domain-mean 100-year return level of 102.4 mm day⁻¹. The highest hazard was concentrated in the eastern inland corridor (e.g., Düzici: 153.1 mm day⁻¹), where moist air interacts with the Taurus foothills. Tufanbeyli, despite high elevation, showed the lowest return level (54.9 mm day⁻¹) due to rain shadow effects.
• Future Projections: CMIP6 projections indicate a consistent decrease in extreme event frequency by 2041–2060 under both scenarios (−19.5% for SSP2-4.5 and −12.1% for SSP5-8.5). However, wet-day intensity is projected to decrease under SSP2-4.5 (−2.2%) but increase under SSP5-8.5 (+1.2%), suggesting a "fewer but potentially more intense" future under high-emission scenarios.

Significance and Contributions
The authors claim this study provides the first integrated ML–extreme value–climate projection pipeline specifically for the eastern Mediterranean coastal region of Türkiye. The significance of the work lies in:

1. Unified Framework: Successfully combining multi-source data fusion, spatial clustering, explainable ML, and extreme value analysis to characterize extreme precipitation hazards.
2. Physical Interpretability: Using SHAP to empirically recover physical relationships (such as the Clausius–Clapeyron relationship and orographic lifting mechanisms) without explicit physical equation inputs, validating the model's alignment with atmospheric physics.
3. Regime Change Detection: Identifying a genuine hydroclimatological shift in 2020 that impacts model transferability, highlighting the need for adaptive retraining protocols in operational forecasting.
4. Infrastructure Implications: Demonstrating that estimated 100-year return levels in the eastern corridor substantially exceed current urban drainage design thresholds (50–100 mm day⁻¹), indicating structural vulnerability.
5. Future Risk Characterization: Clarifying the divergence between frequency and intensity trends under different emission scenarios, suggesting that while extreme days may become less frequent, their intensity could increase under high-emission pathways.

The study concludes that while the region faces a drying trend in terms of event frequency, the potential for intensified rainfall events under high-emission scenarios necessitates a revision of infrastructure design standards and the development of adaptive early warning systems that account for non-stationary hydroclimatological conditions.