Unveiling teleconnection drivers for heatwave prediction in South Korea using explainable artificial intelligence

This study investigates how remote land and ocean variability (teleconnections) influence heatwaves in South Korea and whether these can be leveraged for improved long-lead prediction of annual heatwave frequency (HF). Physically based general circulation models (GCMs) struggle to represent complex land–atmosphere–ocean interactions involving snow depth (SD), soil moisture (SM), sea surface temperature (SST), and sea-ice concentration (SIC). Data-driven models can capture nonlinear relationships among these variables and local climate extremes, potentially improving predictive skill. The research aims to (1) identify statistically robust teleconnection drivers for Korean HF, (2) build an ML model to predict annual HF using these drivers, and (3) apply explainable AI (SHAP) to interpret driver contributions and physical linkages.

Prior work shows that land and ocean anomalies affect upper-atmospheric circulation and regional extremes: snow depth and cover over Eurasia and the Tibetan Plateau, soil moisture anomalies, ENSO-related SST variability in the tropical Pacific and Atlantic, and Arctic sea-ice concentration changes. Previous studies often identify teleconnection regions via correlation analyses (e.g., global SST anomalies) and then relate them to regional climate variability. Challenges persist for physics-based models to resolve these interactions, motivating data-driven approaches that have shown improved skill for extremes (e.g., heatwaves and precipitation) using regression, SVMs, and neural networks. The East Asian summer climate is influenced by patterns such as ENSO, the North Atlantic Oscillation, the Western North Pacific Subtropical High, the Tibetan Plateau high, and the East Asian jet, with decadal modulation reported. These provide a basis to search for predictors over the Indo-Pacific warm pool, central/eastern Pacific, North Atlantic, Barents–Kara seas, and Eurasian land (snow and soil moisture).

Data: Global monthly SST and SIC (HadISST, 1°×1°); ERA5 monthly fields for snow depth (SD), soil moisture (SM), and atmospheric variables (Z850, Z500, Z200, U200, T2M); precipitation from MSWEP; in situ daily maximum temperature at 103 KMA stations to compute annual HF (number of days in July–August with Tmax ≥ 33 °C) for 1960–2022. Teleconnection driver selection: Predictor regions and seasons were compiled from prior literature for SST, SIC, SD, and SM. For each grid point within these regions, Spearman rank correlations with annual HF were computed; points with |R| > 0.3 and p < 0.05 were retained. Adjacent significant points were clustered into teleconnection drivers using DBSCAN. Within each cluster, anomalies were averaged and detrended to form driver indices. Models and evaluation: A Light Gradient Boosting Machine (LGBM) regression model was trained to predict annual HF from selected drivers. Hyperparameters (num_leaves, learning_rate, n_estimators, bagging_fraction, lambda_l1, lambda_l2) were tuned via GridSearchCV to minimize mean squared error. Predictive skill was assessed by leave-one-year-out cross-validation (LOOCV) and a hindcast/backward validation setup, using RMSE, correlation (R), and mean square skill score (MSSS). Benchmark models included multiple linear regression (MLR) using the same inputs and the operational PNU Coupled CGCM (PNU CGCM). Categorical predictability was assessed with a contingency table based on ±1 standard deviation thresholds of observed HF (less normal, normal, above normal). Explainable AI: SHAP and SHAP interaction values were computed to quantify individual and pairwise contributions of drivers to HF predictions over time. Scatter plots of driver anomalies versus SHAP values assessed temporal dependence and nonlinearity of effects. Composite and pattern similarity analyses: To connect top drivers to summer circulation, composites contrasted negative vs positive phases (thresholds ±0.5σ) of key SD drivers (MAM Gobi Desert SD and DJF Tianshan Mountains SD) for July–August fields (Z850, Z500, Z200, U200, T2M, precipitation). Structural Similarity Index Measure (SSIM) compared Z200 teleconnection patterns of MAM GD SD with the SCAND pattern, and DJF TM SD with a double-trajectory type circulation index, to assess pattern resemblance.

- Sixteen teleconnection drivers were identified across SST, SIC, SD, and SM. Notable regions include: SD over the Tianshan Mountains (DJF TM SD) and eastern Tibetan Plateau (DJF ET SD); multiple MAM SD/SM regions over East/Central Asia; SST regions tied to ENSO (DJF Indo-Pacific warm pool, MAM warm pool, central Pacific, and eastern Pacific), and DJF North Atlantic SST associated with NAO; SIC variability over the Barents–Kara seas (DJF and MAM). - LGBM outperformed baselines (Table 1): LGBM RMSE 3.151 days, MSSS 0.804, R 0.644; MLR RMSE 3.740, MSSS 0.731, R 0.500; PNU CGCM RMSE 6.586, MSSS 0.264, R −0.200. - Categorical predictability (±1σ thresholds): Above normal precision 0.833 and recall 0.714; Normal precision 0.896 and recall 0.876; Less normal precision 0.444 and recall 0.571. - Years with significant or recent heatwaves (e.g., 2012, 2013, 2016, 2018, 2020, 2021) were better captured by LGBM than PNU CGCM. - SHAP analyses identified two dominant drivers: MAM Gobi Desert snow depth (MAM GD SD; negative anomalies increase HF) and DJF Tianshan Mountains snow depth (DJF TM SD; positive phase linked to hot/dry summer conditions). Time-varying SHAP values highlighted strong, nonlinear contributions and interactions. - Ablation tests: Removing MAM GD SD reduced R from 0.804 to 0.687 and increased RMSE from 3.15 to 3.96 days. Removing DJF TM SD reduced R to 0.756 and increased RMSE to 3.48 days. Removing both yielded metrics similar to removing MAM GD SD alone, underscoring MAM GD SD’s dominant role and DJF TM SD’s secondary but important impact. - Composite analyses showed vertically coherent ridges, positive T2M anomalies, and reduced precipitation over Korea during the negative phase of MAM GD SD and positive phase of DJF TM SD, consistent with conditions favoring heatwaves (northward-shifted EAS jet, altered moisture transport, subsidence/clear skies). - Pattern similarity: MAM GD SD’s Z200 pattern resembled SCAND with high SSIM (~0.859), linking it to stationary wave patterns associated with prolonged East Asian heatwaves. DJF TM SD showed similarity to a double-trajectory synoptic circulation type, suggesting pathways for influencing Eurasian summer circulation.

Findings address the core question of whether teleconnections can provide skillful, interpretable prediction of Korean heatwaves. LGBM, trained on statistically clustered teleconnection drivers, delivered substantially lower errors and higher correlations than both linear statistical modeling and an operational coupled climate model. Explainable AI indicates that Eurasian spring and winter snow depth anomalies—especially decreased MAM Gobi Desert SD and increased DJF Tianshan SD—are pivotal, shaping summertime ridging, thermal and precipitation anomalies over Korea. The SHAP-based interpretation, supported by composites and pattern similarity to known teleconnections (SCAND, double-trajectory), links data-driven importance to plausible dynamics (stationary waves, jet shifts, modulation of the WNPSH and Tibetan High), thereby enhancing physical credibility. Ablation experiments confirm these drivers’ causal relevance within the model, particularly MAM GD SD. Overall, the approach demonstrates how combining targeted variable selection, modern ML, and XAI can reveal actionable, physically consistent teleconnection drivers for heatwave prediction.

This study demonstrates that a teleconnection-informed, explainable ML framework can skillfully predict annual heatwave frequency in South Korea and elucidate the key remote drivers. LGBM outperforms MLR and an operational coupled climate model. XAI identifies two dominant and physically interpretable drivers—MAM Gobi Desert SD and DJF Tianshan Mountains SD—whose phases align with circulation patterns favoring Korean heatwaves. Composites and SSIM analyses link these drivers to known teleconnection structures (e.g., SCAND-like patterns and double-trajectory types). These insights can inform early-warning systems and targeted monitoring of key regions. Future work should (1) further unravel mechanisms connecting Eurasian snow anomalies to East Asian summer circulation, (2) expand driver discovery beyond pre-defined regions to capture emerging predictors, (3) increase training samples or leverage data augmentation/transfer learning, and (4) explore multi-model ensembles and hybrid physics–ML approaches to enhance robustness.

- Limited sample size: Annual HF yields relatively few samples (1960–2022), increasing uncertainty. Bootstrapping (100 runs with 70–90% subsampling) showed stable mean predictions but increased variance with smaller samples. - Hyperparameter sensitivity: LGBM performance depends on extensive tuning (grid search with cross-validation), which is computationally intensive and may not guarantee globally optimal settings. - Predictor discovery constraints: The DBSCAN-based clustering relied on pre-defined regions/seasons from literature, potentially missing newly emerging or nontraditional drivers. Alternative global search and selection methods could uncover additional predictors.