Answers are grounded in this paper and cite their sources.
Accurate daily temperature and precipitation are most needed over mountains, where observing networks are thin and reanalysis errors are largest. A growing literature corrects ERA5-Land with machine learning, but these corrections are typically validated against stations co-located with those used for training. They rarely attribute the bias to physical drivers or calibrate their uncertainty, and their transfer to regions absent from training remains untested. We address this with TRACE (Transferability-Aware Conformal Explainable bias correction), an interpretable and uncertainty-aware framework for 2-m temperature and precipitation over Türkiye. TRACE harmonizes 67 open stations, engineers 22 physically motivated predictors, and trains gradient-boosted corrections that are explained with Shapley additive explanations and bounded by distribution-free conformal prediction intervals. Skill is assessed under four cross-validation regimes of increasing spatial difficulty and, in addition, in both directions of transfer. Random cross-validation reports a 47% reduction of the daily temperature error, whereas withholding whole stations returns 26%, transfer from the western-central network into the eastern mountains 23%, and transfer across climate regimes 17%. The advantage over an elevation-only correction falls from 22 to 2 percentage points and disappears when the training network is thinned to the size of the target network. Trained on the mountains instead, the learner falls below the elevation-only correction. Precipitation declines from 10% to 4%. Conformal coverage falls from 90% to 62% under transfer, and five local stations restore it to near-nominal. The correction degrades the cold tail under transfer, which bounds its operational use.
No takes yet. Share an insight, caveat, or question.
Ertuğrul Gül (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: