Seasonal-Imbalance Mitigation and Shapley Additive Explanations (SHAP)-Based Interpretation for Sporadic-E Prediction at Jeju

Sporadic-E (Es) layers occur irregularly and show strong seasonal concentration. At mid-latitudes, occurrence rates are high in summer and low in the rest of the year. Such seasonal bias can cause machine-learning predictors to overpredict in summer (May–Aug) and underpredict in non-summer (Sep–Apr). Using Jeju ionosonde data for 2009–2024, we compare three strategies to mitigate this bias: (1) monthly oversampling, (2) threshold adjustment, and (3) probability calibration. Performance is evaluated with the true skill statistic (TSS), geometric mean (G-mean), and count-based symmetry index (C-SI). Experiments show that combining monthly oversampling with threshold adjustment yields the most stable and consistent improvements, significantly reducing summer overprediction and non-summer underprediction. Shapley additive explanations (SHAP) analysis indicates that foEs (the Es critical frequency), day of year (DOY), and vertical ion convergence (VIC) are the key predictors in the model, consistent with known Es formation mechanisms. This study underscores the importance of training strategies that account for seasonal data bias in mid-latitude Es prediction and points to future multi-station extension and model generalization.