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

TL;DR

This study benchmarks residual generative downscaling methods for daily precipitation over New Zealand, comparing a residual conditional GAN with an intensity constraint to residual diffusion models against a deterministic baseline within a perfect-model framework. Using a 12 km CCAM-RCM emulator trained on CMIP6 data, it evaluates performance across historical and future climates with metrics capturing mean climatologies, extremes, spatial structure, and climate-change signals. Diffusion models deliver stronger fine-scale spatial realism and dry-spell representation but tend to underpredict extreme precipitation changes, while the cGAN matches or surpasses in most metrics, including extremes and climate-change responses, at substantially lower computational cost. The results highlight that appearance of spatial realism does not guarantee reliable climate-change extrapolation, and they suggest that incorporating physics-based constraints into diffusion models or further tuning could yield robust, efficient downscaling suitable for climate risk assessment and ensemble projections.

Abstract

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.

Figures (7)

• Figure 1: Full model architecture, showing the use of a GAN (blue) and Diffusion model (Green) for residual correction. Note that the GAN and Diffusion models are mutually exclusive in our experiment configurations, i.e., only one is used in a given configuration.
• Figure 2: (a) Architecture of the deterministic U-net used for downscaling coarse-resolution predictors into the high-resolution precipitation field. The network takes a static high-resolution topography field and concatenated coarse predictors as inputs $\mathbf{x}$, which are processed through successive encoding layers (residual and pooling blocks) before being passed through a bottleneck, where the encoded topography and predictor features are concatenated. The decoder progressively upsamples the features using bicubic interpolation and residual blocks, and the final convolutional layers generate the deterministic prediction $\hat{\mathbf{y}}_\text{det}$ for the high-resolution precipitation field. (b) Architecture of the residual generator network (GAN) for predicting a residual to correct $\hat{\mathbf{y}}_\text{det}$. Similar to (a), but takes additional inputs: $\hat{\mathbf{y}}_\text{det}$ as the conditional mean and noise vectors. High-resolution inputs are processed through encoding layers, while coarse fields and noise pass through residual blocks without pooling. After concatenation in the bottleneck, features are upsampled with skip connections from the encoder. Additional noise is injected after the first upsample. Outputs residual prediction $\hat{\mathbf{r}}_\text{GAN}$ for small-scale corrections ($\hat{\mathbf{y}}_\text{GAN}=\hat{\mathbf{r}}_\text{GAN}+\hat{\mathbf{y}}_\text{det}$). (c) Architecture of the residual noise predictor (diffusion model) for correcting $\hat{\mathbf{y}}_\text{det}$. Takes diffusion step $t$ and noisy residual $\hat{\mathbf{r}}_t$, conditioned by $\hat{\mathbf{y}}_\text{det}$ and inputs $\mathbf{x}$. Similar encoder-decoder structure as (a), but with adaptive group normalization blocks. Outputs noise prediction $\hat{\boldsymbol{\epsilon}}$ to obtain denoised residual $\hat{\mathbf{r}}_0\gets \hat{\mathbf{r}}_t-\hat{\boldsymbol{\epsilon}}$. (d) Definitions of network components used in (a), (b), and (c).
• Figure 3: A representation of the reverse diffusion process from time T=999 to t=0. The top row shows residuals (relative to the deterministic baseline) evolving from white noise at T=999 to the final residual prediction at t=0. The middle row shows the residuals added to the deterministic baseline in logarithmic space (because precipitation is log-normalized, as described in Section \ref{['section:training-and-evaluation']}). The bottom row shows the final precipitation field after reversing the normalization and taking the exponential of the values.
• Figure 4: Spatial comparison of predicted daily precipitation fields for an ex-tropical cyclone event over New Zealand (NorESM2-MM historical period; 1987-03-26). The top row shows predictions from the diffusion model with 1000 steps (left), 100 steps (center), and the cGAN (right). The bottom row shows the deterministic U-Net baseline (left) and the ground truth CCAM simulation downscaled from NorESM2-MM (right). Maximum precipitation values are indicated for each panel.
• Figure 5: Precipitation intensity distributions (top) and radially integrated power spectral densities (bottom) for the historical (1986–2005, left) and future (2080–2099, right) periods. The precipitation intensity histograms are shown for days exceeding 1 mm/day, using a logarithmic scale for histogram counts. Bottom panels show radially-averaged power spectra as a function of radial wavenumber (km$^{-1}$).