SpateGAN-ERA5: Global Precipitation Downscaling

• SpateGAN-ERA5 is a conditional generative adversarial network that performs global spatio-temporal downscaling of ERA5 precipitation from 24 km/1 hr to 2 km/10 min.
• It employs a fully 3D CNN with a UNet-like architecture and dropout-based ensemble generation to realistically capture fine-scale structures and extreme intensities.
• Evaluated in Germany, the US, and Australia, the method outperforms traditional techniques like RainFARM by preserving detailed variability and extreme rainfall statistics.

Searching arXiv for the primary and closely related papers to ground the article. arXiv search: spateGAN-ERA5 precipitation global spatio-temporal downscaling generative AI SpateGAN-ERA5 is a conditional generative adversarial neural network for global spatio-temporal downscaling of ERA5 precipitation from about 24 km and 1 hour to 2 km and 10 minutes. It is described as the first deep learning based spatio-temporal downscaling of precipitation data on a global scale, and is designed to generate high-resolution rainfall fields with realistic spatio-temporal patterns and an accurate rain rate distribution including extremes while remaining computationally cheap enough to produce large ensembles. The system is trained solely on data from Germany and evaluated in the United States and Australia, where it is reported to generalize across diverse climate zones (Glawion et al., 2024).

1. Definition and problem setting

SpateGAN-ERA5 addresses the mismatch between the coarse spatial and temporal support of ERA5 precipitation and the finer scales required by hydrology, flood risk analysis, agriculture, and impact modelling. ERA5 precipitation is provided at about 0.25° and hourly resolution, with convective processes parameterized rather than explicitly resolved. At that scale, local convective cells, sharp rain-rate gradients, and short-duration extremes are smoothed or missed entirely. The method therefore targets simultaneous spatial and temporal super-resolution rather than static bias correction alone (Glawion et al., 2024).

The model takes ERA5 convective precipitation and large-scale precipitation as conditioning inputs and produces radar-like precipitation sequences at much higher resolution. Its nominal enhancement factors are 12× in space, from 24 km to 2 km, and 6× in time, from 1 hour to 10 minutes. A central design principle is that the fine-scale output should inherit large-scale consistency from ERA5 while recovering the intermittency, anisotropy, and high-intensity tail absent from the coarse input (Glawion et al., 2024).

A concise summary of the system’s scope is given below.

Aspect	Specification	Notes
Input	ERA5 convective precipitation and large-scale precipitation	24 km, 1 h
Output	Radar-like precipitation	2 km, 10 min
Training region	Germany	RADKLIM-YW target
Evaluation regions	Germany, United States, Australia	Independent radar products
Generative mechanism	Conditional GAN with ensemble generation	Dropout-based stochasticity

2. Architecture and downscaling mechanism

SpateGAN-ERA5 is a cGAN with a fully 3D spatio-temporal formulation. The conditioning tensor consists of ERA5 convective precipitation and large-scale precipitation on a coarse 28×28 grid over 16 hourly time steps, corresponding to a domain of about 672 km × 672 km. The target is the central 8 hours over 336 km × 336 km at 2 km and 10 minute resolution. The larger conditioning window supplies contextual information for advection and mesoscale organization, while the central crop suppresses boundary artefacts (Glawion et al., 2024).

The generator is a 3D-CNN with UNet-like multiscale structure. It begins with coarse-resolution feature extraction using residual 3D convolution blocks, followed by a branch with strided 3D downsampling and a parallel center-cropping branch. These are merged at reduced space-time resolution, after which four upsampling blocks use bilinear interpolation in space and linear interpolation in time, each followed by additional residual 3D blocks. Three further full-resolution residual blocks refine the 2 km, 10 minute output, and a final 3D convolution with Softplus enforces non-negative rain rates. The discriminator is also 3D and conditional: it processes the high-resolution sequence and the low-resolution ERA5 context in separate paths, then concatenates the resulting features and outputs a scalar real/fake score (Glawion et al., 2024).

The training objective combines adversarial discrimination with an ensemble-mean reconstruction term. With $X$ denoting the ERA5 input, $Y$ the radar target, and $\hat{Y}_i = G(X; z_i)$ the $i$-th ensemble member, the discriminator and generator terms are

$$\mathcal{L}_D = -\mathbb{E}_{X,Y}[\log D(X,Y)] - \mathbb{E}_{X,\hat{Y}}[\log(1 - D(X,\hat{Y}))]$$

$$\mathcal{L}_{\text{GAN}(G)} = -\mathbb{E}_{X}[\log D(X, G(X))]$$

$$\mathcal{L}_{\text{L1}(G)} = \overline{\left| Y - \frac{1}{3}\sum_{i=1}^{3} \hat{Y}_i \right|}$$

$$\mathcal{L}_G = \mathcal{L}_{\text{GAN}(G)} + \mathcal{L}_{\text{L1}(G)}.$$

The use of the ensemble mean in the $L_1$ term is intended to mitigate the double-penalty problem associated with small phase errors in localized convective structures. A separate physical consistency device is applied at inference: the mean precipitation in each high-resolution patch is multiplicatively adjusted to match the mean ERA5 precipitation for that patch, after discarding ERA5 values below 0.1 mm/h over oceans for stability. This constrains patch-wise mean rainfall while leaving sub-grid structure to the generator (Glawion et al., 2024).

3. Training data, sampling, and inference workflow

The training target is RADKLIM-YW, a gauge-adjusted, climatologically corrected German radar composite originally at 1 km and 5 minute resolution, here downsampled to 2 km and 10 minutes. The training period is 2009–2020, model selection uses January–June 2021, and evaluation uses July–December 2021. ERA5 precipitation is regridded to the radar projection and then downsampled to a 24 km coarse km-grid. Independent evaluation datasets are MRMS in the United States and AURA Rainfields 3 in Australia, both downsampled to 2 km and 10 minutes for comparison (Glawion et al., 2024).

Training samples are not drawn uniformly. Because dry and weakly precipitating scenes dominate the climatology, a stochastic sampling scheme preferentially selects informative rainfall episodes. For each candidate sample, both the ERA5 input $X$ and the radar target $Y$0 must exceed randomized total-rainfall and upper-quantile thresholds:

$Y$1

with $Y$2, $Y$3, and an additional 66th-percentile threshold with

$Y$4

This procedure yields about 20,000 training samples, corresponding to roughly 850 GB of data. Optimization uses AdamW, with learning rates $Y$5 for the generator and $Y$6 for the discriminator, $Y$7, $Y$8 for the generator, $Y$9, $\hat{Y}_i = G(X; z_i)$0 for the discriminator, a batch size of 9, and $\hat{Y}_i = G(X; z_i)$1 adversarial updates. Training is reported on 3 NVIDIA A100 GPUs and takes about 4 days. Checkpoints are selected by maximizing the mean ensemble Fractions Skill Score over thresholds $\hat{Y}_i = G(X; z_i)$2 mm/h and spatial scales $\hat{Y}_i = G(X; z_i)$3 km at $\hat{Y}_i = G(X; z_i)$4 h (Glawion et al., 2024).

At inference, stochasticity enters through temporally constant dropout applied at three depths in the generator. Each ensemble member uses a distinct dropout realization, fixed across time within the sample, so the perturbations are temporally coherent rather than framewise independent. The model downscales overlapping patches in space and time, discards the outermost ~24 km and 1 hour of each patch to suppress edge effects, and blends neighboring outputs linearly. This is the basis for global tiling onto a 0.018° latitude-longitude grid, approximately 2 km at the equator (Glawion et al., 2024).

4. Evaluation methodology and reported performance

The reported evaluation is deliberately multi-criteria. It includes MAE, RMSE, BIAS, CRPS, FSS, mean FSS, mean ensemble FSS, temporal power spectral density, radially averaged power spectral density, rank histograms, and anisotropy diagnostics derived from autocorrelation ellipses. This metric design reflects the fact that simple pointwise errors reward smooth fields and therefore do not adequately measure the realism of intermittent precipitation (Glawion et al., 2024).

On the 2 km, 10 minute grid, aggregated to 1 hour where required, SpateGAN-ERA5 outperforms RainFARM and interpolation in the probabilistic and scale-aware metrics most directly tied to high-resolution rainfall structure. The reported CRPS values are 0.018 in Germany, 0.020 in the United States, and 0.013 in Australia, compared with 0.023, 0.024, and 0.015 for RainFARM. The reported mean ensemble FSS values are 0.34, 0.25, and 0.19, compared with RainFARM mean FSS values of 0.24, 0.20, and 0.13, and interpolation mean FSS values of 0.22, 0.18, and 0.12. The relative improvement in mean FSS over interpolation is reported as +54% in Germany, +38% in the United States, and +58% in Australia (Glawion et al., 2024).

By contrast, smoother baselines achieve lower MAE and RMSE: for Germany, SpateGAN-ERA5 yields MAE 0.027 and RMSE 0.145, while RainFARM and interpolation are around 0.024–0.025 for MAE and about 0.102–0.103 for RMSE. The same pattern appears in the United States and Australia. This is explicitly attributed to the fact that interpolation and RainFARM under-represent extremes and therefore incur smaller pointwise penalties when convective cells are displaced or smoothed. The reported interpretation is therefore not that SpateGAN-ERA5 minimizes average squared error, but that it recovers physically relevant fine-scale variability and threshold exceedances more successfully than the smoother baselines (Glawion et al., 2024).

Distributional diagnostics further support that reading. In Germany and Australia, the hourly rain-rate distributions of SpateGAN-ERA5 ensemble members closely follow the observed log-normal-like shape and capture the heavy tail much better than ERA5 interpolation or RainFARM. Spatial RAPSD and temporal PSD show that the model reproduces the observed scale dependence of variability; after normalization, the spectral curves of SpateGAN-ERA5 and radar nearly coincide across regions and wavelengths, whereas interpolation remains too smooth and RainFARM exhibits an artificial kink near the coarse grid scale of about 24 km (Glawion et al., 2024).

5. Generalization, ensemble behaviour, and uncertainty

A defining feature of SpateGAN-ERA5 is that it is trained solely on German radar data and then validated in the United States and Australia. The reported performance indicates strong transfer across diverse climates, despite differences in radar products and quality-control regimes. The normalized spatial and temporal spectra, the anisotropy metrics, and the threshold-based skill scores all suggest that the generator learns a structural mapping from coarse precipitation organization to fine-scale variability that is not narrowly tied to one national radar archive (Glawion et al., 2024).

The ensemble behaviour is central rather than incidental. Rank histograms computed with 100 members are nearly flat in all three regions, with a slight under-dispersive tendency. RainFARM, by comparison, is described as strongly under-dispersive. CRPS is consistently lower for SpateGAN-ERA5 than for RainFARM, which indicates better probabilistic sharpness-reliability trade-offs at the gridpoint level. In the case studies, ensemble members differ mainly in local intensity realization while preserving the mesoscale placement and evolution implied by the ERA5 conditioning. This suggests that the stochastic component is being used primarily to sample sub-grid uncertainty rather than to randomize the large-scale event itself (Glawion et al., 2024).

The anisotropy analysis is particularly important because it bears on dynamical plausibility. ERA5 interpolation and RainFARM produce fields that are too isotropic and too spatially smooth. SpateGAN-ERA5, in contrast, reproduces elongated precipitation structures with realistic orientation and size, closer to the observed autocorrelation ellipse statistics. A plausible implication is that the 3D convolutions and large contextual window allow the generator to internalize advection-like organization without any explicit wind-field conditioning, although this remains an implicit rather than hard-constrained form of dynamics (Glawion et al., 2024).

6. Position in the literature, applications, and limitations

SpateGAN-ERA5 sits within a broader arXiv trajectory of generative modelling for atmospheric data, but its scope is distinct. SPATE-GAN introduced an autoregressive embedding loss for complex dynamic spatio-temporal patterns in a COT-GAN framework (Klemmer et al., 2021). Conditional GANs had already been used for rainfall post-processing and downscaling of ERA5 to 1 km over regional domains (Cooper et al., 2023), and related ERA5-driven super-resolution pipelines had been developed for wind in the Sup3rWind framework (Benton et al., 2024). SpateGAN-ERA5 differs from those systems in combining global applicability, joint spatial and temporal precipitation downscaling, and ensemble generation at 2 km and 10 minute resolution from ERA5 precipitation alone (Glawion et al., 2024).

The paper identifies several limitations. First, the model preserves ERA5 mean precipitation at patch scale, so regional mean biases in ERA5 are inherited rather than removed. Second, the training domain is limited to Germany, which leaves some extreme regimes, especially tropical and strongly monsoonal precipitation, underrepresented in the training climatology. Third, detailed validation is restricted to Germany, the United States, and Australia because globally homogeneous, gauge-adjusted radar references are not available. Fourth, no explicit physical conservation laws are embedded beyond the patch-mean precipitation adjustment. Fifth, the method does not explicitly address non-stationarity under climate change, although it is proposed as a tool for long historical reconstructions and potentially for AI-enhanced weather forecasting and impact modelling (Glawion et al., 2024).

The principal applications are high-resolution hydrological forcing, flood risk assessment, urban drainage analysis, climate-impact modelling, and post-processing of AI weather forecasts. Because the system is computationally efficient—one 672×672 km, 16 hour patch requires about 0.04 s on a single NVIDIA A100 GPU and about 10 GB of GPU memory—it is feasible to generate ensembles at continental or global scale. This places SpateGAN-ERA5 in a specific methodological niche: not a replacement for dynamical reanalysis, but a learned stochastic downscaler that preserves large-scale ERA5 precipitation while reconstructing the fine-scale spatial and temporal statistics that matter for rainfall-driven hazards (Glawion et al., 2024).