فیزیک زمین و فضا

فیزیک زمین و فضا

Cross-Validated Aftershock Pattern Forecasting: Deep Learning vs. Gradient Boosting

نوع مقاله : مقاله پژوهشی

نویسندگان
1 هیات علمی دانشگاه تبریز
2 Surveying Engineering Department, Ahar Azad University, Ahar, Iran
10.22059/jesphys.2026.415915.1007790
چکیده
Accurate spatial forecasting of aftershocks remains a critical challenge in seismology. The Deep Learning approach of DeVries et al. (2018) reported an AUC of 0.849 on a single held-out test set and has been widely adopted as a benchmark. We compare the DeepMLP (DeVries18 architecture) against four gradient-boosting methods—AdaBoost, XGBoost, CatBoost, and LightGBM—for binary classification of 5 km grid cells on the same SRCMOD and ISC dataset. Our central finding is a systematic rank reversal. Spearman rank correlations between test and 5-fold cross-validation (CV) rankings are strongly negative for Accuracy (ρ=−1.00), Average Precision (ρ=−0.90), and AUC (ρ=−0.70). The MLP achieves the highest test AUC (0.851) but ranks last in CV (0.751±0.019), while LightGBM and XGBoost achieve the highest CV AUC (0.798), with gaps of only +0.015 to 0.018. The MLP exhibits four pathologies: extreme data-hunger (CV AUC = 0.516 at 10% data), training-budget insensitivity (ΔAUC = +0.008 from 30 to 200 epochs), high variance with calibration reversal (CV std = 0.019; Brier 0.182 → 0.227), and unreliable uncertainty (r = −0.31 between uncertainty and error). We further report bootstrap 95% confidence intervals, F1-optimal thresholds, DeLong and McNemar tests, calibration analysis, cross-model SHAP agreement, and uncertainty quantification. LightGBM and CatBoost are statistically indistinguishable (DeLong p=0.465). Cross-model SHAP agreement is weak (minimum ρ=0.18). The MLP is 4.1× more uncertain than LightGBM and is the only model whose uncertainty is anti-correlated with error (r=−0.31). We recommend LightGBM for operational deployment due to its superior CV performance (0.798) and low training cost (4.8 s).
کلیدواژه‌ها
موضوعات


مقالات آماده انتشار، پذیرفته شده
انتشار آنلاین از 26 آبان 1405