Explainable Deep Ensemble Bias Correction of GloFAS-ERA5 Streamflow Across Snow-Influenced Transboundary Basins of Central Asia

Global streamflow reanalyses such as GloFAS-ERA5 are available everywhere yet lose fidelity in small, snow- and glacier-fed headwaters that feed Central Asia’s transboundary rivers. We present an explainable, calibrated deep ensemble framework that corrects the GloFAS-ERA5 log-residual at gauges of the Syr Darya and Amu Darya systems using the CA-discharge archive. An entity-aware long short-term memory (LSTM) backbone drives a regime-gated mixture of experts trained under a closed-form mixture continuous ranked probability score (CRPS) and augmented with snow physics constraints and a regime-conditional (Mondrian) conformal layer; skill was assessed under temporal holdout, leave-one-basin-out and prediction in ungauged region protocols, with grouped Shapley value attribution. Correction rendered all 74 gauges skillful, raising the median modified Kling–Gupta efficiency (KGE′) from 0.386 (raw) to 0.825 (flagship); on temporal point skill the framework is statistically tied with gradient boosting (paired Wilcoxon p = 0.49). Under-dispersed raw intervals (90% coverage 0.68) were recalibrated to near-nominal coverage (~0.90), and high-flow exceedance decision skill was moderate (Q90 Brier skill score 0.34, ROC-AUC 0.92), while low-flow (Q10) exceedance showed no skill over climatology. Transfer to ungauged, more glacierized catchments was a measured limit that degraded with glacier fraction and basin area. The framework’s value is calibration, an inspectable (supervised) regime structure and regional physical insight, not point skill superiority.