Scholarlink

Abstract <p>Accurate, high-resolution local climate projections are essential for climate adaptation and impact assessment. Most statistical downscaling approaches rely on regression-based models that produce single deterministic estimates, thereby limiting their ability to capture the inherent variability and multivariate dependencies of the climate system. To address this limitation, ClimDiT is introduced, a generative multivariate statistical downscaling framework based on latent diffusion transformer models. Thus, instead of a single expected result, it estimates the full conditional distribution of high-resolution climate fields. The proposed system operates within a compressed, autoencoded latent space, enabling computationally efficient sampling while implicitly capturing inter-variable dependencies. In this work, ClimDiT simultaneously downscales daily maximum temperature, minimum temperature, and accumulated precipitation to approximately 5 km resolution over the Iberian Peninsula. It is trained using ERA5 predictors and ROCIO-IBEB high-resolution targets and its performance is evaluated using metrics aligned with the VALUE framework. ClimDiT attains deterministic accuracy comparable to, or moderately superior to, state-of-the-art regression networks, while improving spatial coherence. Although specialised stochastic baselines, which suffer from high deterministic errors, retain an edge in univariate and multivariate probabilistic scores, a ClimDiT ensemble generated from 50 samples demonstrates reasonable probabilistic skill and joint calibration of the three target variables (evaluated with CRPS, Brier, Energy and Variogram scores). Overall, ClimDiT demonstrates that latent diffusion models enable a generative pathway toward coherent, probabilistic, and physically consistent regional climate downscaling, advancing beyond the limitations of traditional regression-based AI systems.</p> Graphical Abstract <p></p> <p>This graphical abstract provides a concise visual overview of ClimDiT: A Generative Latent Diffusion Transformer Framework for Multivariate Climate Downscaling. It summarizes the study objectives, methods, and main findings. At the top, the Data section illustrates the statistical downscaling approach, by linking coarse-resolution ERA5 predictors with high-resolution ROCIO-IBEB climate targets over the Iberian Peninsula. It highlights the three downscaled variables—maximum temperature, minimum temperature, and precipitation—as well as the corresponding training and testing periods. The Methods panel presents the core innovation: the ClimDiT model. It combines an autoencoder, which compresses climate fields into a latent space, with a latent diffusion transformer that progressively reconstructs high-resolution fields from Gaussian noise, conditioned on large-scale predictors. This reflects the proposed use of generative models, rather than the widely used regression-based AI approaches, as they allow the generation of multiple outputs for the same synoptic state. A benchmark comparison box displays the univariate state-of-the-art (SOTA) methods (DeepESD and U-Net) used as references, along with the tested loss functions (MSE, ASYM, and STO). STO refers to stochastic models, which use a Bernoulli–Gamma distribution for precipitation and a Gaussian distribution for temperature, trained using the negative log-likelihood loss. Additionally, a deterministic multivariate baseline (DeepESD-MULT, trained with MSE) is included to directly evaluate the impact of joint variable modeling within a traditional regression framework. The Results section visualizes performance across spectral, deterministic, and probabilistic metrics (RAPSD, MAE, CRPS, Variogram Score), emphasizing ClimDiT superior spatial coherence and balanced probabilistic skill. The Highlights box succinctly reinforces these contributions: a generative, latent-space-based approach that enables coherent multivariate downscaling.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ClimDiT: A Generative Latent Diffusion Transformer Framework for Multivariate Climate Downscaling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Marcos Esquivel-González,
  • Albano González,
  • Juan Pedro Díaz,
  • Juan Carlos Pérez,
  • Jose González-Abad,
  • Samuel Márquez-Cubas

摘要

Abstract

Accurate, high-resolution local climate projections are essential for climate adaptation and impact assessment. Most statistical downscaling approaches rely on regression-based models that produce single deterministic estimates, thereby limiting their ability to capture the inherent variability and multivariate dependencies of the climate system. To address this limitation, ClimDiT is introduced, a generative multivariate statistical downscaling framework based on latent diffusion transformer models. Thus, instead of a single expected result, it estimates the full conditional distribution of high-resolution climate fields. The proposed system operates within a compressed, autoencoded latent space, enabling computationally efficient sampling while implicitly capturing inter-variable dependencies. In this work, ClimDiT simultaneously downscales daily maximum temperature, minimum temperature, and accumulated precipitation to approximately 5 km resolution over the Iberian Peninsula. It is trained using ERA5 predictors and ROCIO-IBEB high-resolution targets and its performance is evaluated using metrics aligned with the VALUE framework. ClimDiT attains deterministic accuracy comparable to, or moderately superior to, state-of-the-art regression networks, while improving spatial coherence. Although specialised stochastic baselines, which suffer from high deterministic errors, retain an edge in univariate and multivariate probabilistic scores, a ClimDiT ensemble generated from 50 samples demonstrates reasonable probabilistic skill and joint calibration of the three target variables (evaluated with CRPS, Brier, Energy and Variogram scores). Overall, ClimDiT demonstrates that latent diffusion models enable a generative pathway toward coherent, probabilistic, and physically consistent regional climate downscaling, advancing beyond the limitations of traditional regression-based AI systems.

Graphical Abstract

This graphical abstract provides a concise visual overview of ClimDiT: A Generative Latent Diffusion Transformer Framework for Multivariate Climate Downscaling. It summarizes the study objectives, methods, and main findings. At the top, the Data section illustrates the statistical downscaling approach, by linking coarse-resolution ERA5 predictors with high-resolution ROCIO-IBEB climate targets over the Iberian Peninsula. It highlights the three downscaled variables—maximum temperature, minimum temperature, and precipitation—as well as the corresponding training and testing periods. The Methods panel presents the core innovation: the ClimDiT model. It combines an autoencoder, which compresses climate fields into a latent space, with a latent diffusion transformer that progressively reconstructs high-resolution fields from Gaussian noise, conditioned on large-scale predictors. This reflects the proposed use of generative models, rather than the widely used regression-based AI approaches, as they allow the generation of multiple outputs for the same synoptic state. A benchmark comparison box displays the univariate state-of-the-art (SOTA) methods (DeepESD and U-Net) used as references, along with the tested loss functions (MSE, ASYM, and STO). STO refers to stochastic models, which use a Bernoulli–Gamma distribution for precipitation and a Gaussian distribution for temperature, trained using the negative log-likelihood loss. Additionally, a deterministic multivariate baseline (DeepESD-MULT, trained with MSE) is included to directly evaluate the impact of joint variable modeling within a traditional regression framework. The Results section visualizes performance across spectral, deterministic, and probabilistic metrics (RAPSD, MAE, CRPS, Variogram Score), emphasizing ClimDiT superior spatial coherence and balanced probabilistic skill. The Highlights box succinctly reinforces these contributions: a generative, latent-space-based approach that enables coherent multivariate downscaling.