Spatiotemporally Consistent Multivariate Bias Correction for Climate Projections via Nested Vine Copulas

Climate models are essential for understanding large-scale climate dynamics and long-term climate change, yet they exhibit systematic biases when compared with historical observations. Existing multivariate bias correction (MBC) approaches do not explicitly handel spatiotemporal dependence. However, preserving both spatiotemporal and inter-variable consistency is essential for realistic climate dynamics and reliable regional impact assessments. To address this gap, we propose a novel MBC method called GN-VBC that uses generalized additive models (GAMs) to disentangle spatiotemporal deterministic effects from stochastic residuals. To model joint distributions and dependencies across variables and locations, we introduce nsted vine copulas (NVCs), a hierarchical vine merging strategy. NVC in the context of MBC combines two dependence levels: (i) spatial dependence across locations, modeled separately for each variable, and (ii) inter-variable dependence modeled at a selected reference location, which links the spatial models into a coherent multivariate and spatial structure. An application to Switzerland shows improvements in preserving inter-variable, spatial and temporal dependence across a wide range of evaluation metrics.

💡 Research Summary

The paper introduces GN‑VBC, a novel multivariate bias‑correction framework that simultaneously preserves spatial, temporal, and inter‑variable dependencies in climate model outputs. The method consists of two main components. First, generalized additive models (GAMs) are fitted for each climate variable using covariates such as absolute time, latitude, longitude, elevation, and their interactions. GAMs capture deterministic spatio‑temporal patterns, and the residuals are transformed to uniform margins via the probability integral transform (PIT). This step avoids the extrapolation problems of traditional quantile mapping and retains the modeled climate‑change signal. Second, a nested vine copula (NVC) is constructed to model the joint distribution of the standardized residuals. The NVC has a hierarchical structure: (i) a spatial layer builds a separate R‑vine for each variable across all locations, using distance‑based or elevation‑based covariates to capture spatial dependence; (ii) an inter‑variable layer is fitted at a chosen reference (bridge) location, linking the variable‑specific vines into a coherent multivariate vine. Pair‑copulas are selected by BIC from families such as Gaussian, t, Clayton, and Gumbel.

Bias correction proceeds as follows: (1) apply the fitted GAMs to model simulations to obtain uniform residuals U⁽ᵐ⁾; (2) map U⁽ᵐ⁾ to the reference distribution using the inverse Rosenblatt transform of the NVC, i.e., U* = C⁻¹_NVC(F_{U⁽ʳ⁾}(U⁽ᵐ⁾)); (3) back‑transform U* through the inverse GAM to obtain the corrected climate field Y*. This pipeline preserves marginal distributions, spatial coherence, temporal autocorrelation, and the full inter‑variable dependence structure.

The authors evaluate GN‑VBC on a Swiss case study involving five key atmospheric variables (2‑m temperature, total precipitation, relative humidity, 10‑m wind speed, surface pressure) across 22 grid points in the Alpine region. They compare against state‑of‑the‑art multivariate bias‑correction methods, including MBCn, R2D2, dOTC, and the earlier vine‑bias‑correction (VBC) approach. Evaluation metrics cover marginal fit (Kolmogorov‑Smirnov, CRPS), spatial dependence (Kendall’s τ, variograms), temporal autocorrelation, multivariate rank histograms, and extreme‑event reproduction. GN‑VBC consistently outperforms the competitors, achieving 10–15 % improvements across most metrics and notably excelling at reproducing inter‑variable correlations and spatial variability. Sensitivity analysis shows that the choice of the reference location has limited impact on performance, indicating robustness.

Key contributions are: (1) explicit separation of deterministic spatio‑temporal effects via GAMs, preserving climate‑change trends; (2) introduction of the nested vine copula framework that merges variable‑specific spatial vines with an inter‑variable vine, enabling flexible, high‑dimensional dependence modeling; (3) a modular estimation procedure that is computationally tractable for moderate‑size multivariate, multi‑site datasets; and (4) a practical demonstration that the method can improve operational climate‑scenario production, especially for regions with complex topography like Switzerland. The paper suggests future extensions to non‑stationary dependence structures, higher‑dimensional variable sets (e.g., coupling atmospheric, hydrological, and ecological variables), and integration into real‑time climate‑service pipelines.