An intercomparison of generative machine learning methods for downscaling precipitation at fine spatial scales

Machine learning (ML) offers a computationally efficient approach for generating large ensembles of high-resolution climate projections, but deterministic ML methods often smooth fine-scale structures and underestimate extremes. While stochastic generative models show promise for predicting fine-scale weather and extremes, few studies have compared their performance under present-day and future climates. This study compares a previously developed conditional Generative Adversarial Network (cGAN) with an intensity constraint against different configurations of diffusion models for downscaling daily precipitation from a regional climate model (RCM) over Aotearoa New Zealand. Model skill is comprehensively assessed across spatial structure, distributional metrics, means, extremes, and their respective climate change signals. Both generative approaches outperform the deterministic baseline across most metrics and exhibit similar overall skill. Diffusion models better predict the fine-scale spatial structure of precipitation and the length of dry spells, but underestimate climate change signals for extreme precipitation compared to the ground truth RCMs. In contrast, cGANs achieve comparable skill for most metrics while better predicting the overall precipitation distribution and climate change responses for extremes at a fraction of the computational cost. These results demonstrate that while diffusion models can readily generate predictions with greater visual “realism”, they do not necessarily better preserve climate change responses compared to cGANs with intensity constraints. At present, incorporating constraints into diffusion models remains challenging compared to cGANs, but may represent an opportunity to further improve skill for predicting climate change responses.

💡 Research Summary

This study presents a comprehensive intercomparison of two leading generative machine learning methods—conditional Generative Adversarial Networks (cGANs) and Diffusion Models—for the task of statistical downscaling of daily precipitation. Conducted over the complex terrain of Aotearoa New Zealand, the research aims to evaluate which approach better emulates high-resolution Regional Climate Model (RCM) output, with a particular focus on capturing spatial patterns, extreme events, and critically, the climate change signal.

The core methodology employs a “residual correction” framework. A deterministic U-Net model first predicts a smoothed, conditional mean high-resolution precipitation field from coarse-resolution inputs (winds, temperature, humidity, and high-res topography). Subsequently, a generative model (either a cGAN or a Diffusion model) is trained to learn and generate the stochastic residual—the difference between this deterministic prediction and the ground-truth high-resolution RCM output. The final downscaled product is the sum of the deterministic mean and the generated residual. This setup ensures the generative model focuses on adding the missing fine-scale variability and extremes.

Models were trained on 12km RCM simulations driven by the ACCESS-CM2 global climate model and evaluated on independent simulations driven by EC-Earth3 and NorESM2-MM, assessing performance for both historical (1985-2014) and future (2070-2099) periods. Evaluation metrics were extensive, covering spatial correlation structure, precipitation intensity distribution, mean values, extreme precipitation (99th percentile), dry spell length, and the preservation of the climate change signal (the difference between future and historical periods for these metrics).

The key findings reveal a nuanced performance landscape. Both generative models significantly outperformed the deterministic baseline across most metrics. In a head-to-head comparison, the overall skill was similar, but each method exhibited distinct strengths and weaknesses. Diffusion models, particularly those with more denoising steps, generated outputs with superior visual realism, better captured fine-scale spatial autocorrelation structures, and more accurately reproduced the distribution of dry spell lengths. However, a critical shortcoming was identified: Diffusion models systematically underestimated the climate change signal for extreme precipitation compared to the ground-truth RCMs. They produced a weaker intensification of heavy rainfall under future warming.

In contrast, the cGAN—specifically one enhanced with an intensity-constrained loss function—achieved comparable skill on most standard metrics while demonstrating a stronger ability to preserve the climate change response of precipitation extremes. Furthermore, the cGAN accomplished this at a fraction of the computational cost required for training and inference by the Diffusion models.

The study concludes that while diffusion models excel in generating spatially realistic fields, their current formulations may not reliably capture crucial climate change responses, potentially due to an over-reliance on learning the historical data distribution. The cGAN, with its explicit constraint on precipitation intensity, offers a more computationally efficient and, in this context, more reliable tool for downscaling future climate projections where preserving change signals is paramount. The results underscore that visual “realism” in generated climate fields should not be conflated with fidelity in projecting climate change impacts. The challenge of incorporating physical constraints into diffusion models is highlighted as a promising avenue for future improvement.