RainAtlas: A Multi-Continental Dataset for Precipitation Downscaling

Pierre-Louis Lemaire; Luca Schmidt; Wietze Suijker; Alex Hernandez-Garcia; David Rolnick

cs.LGSep 30, 2026

RainAtlas: A Multi-Continental Dataset for Precipitation Downscaling

Authors: Pierre-Louis Lemaire, Luca Schmidt, Wietze Suijker, Alex Hernandez-Garcia, David Rolnick

Organizations: connAIx Research School, Tübingen AI Center, University of Tübingen · Mila - Quebec AI Institute · Cluster of Excellence Machine Learning, University of Tübingen · IVADO · Universit´e de Montr´eal · McGill University

Abstract

Extreme rainfall events are increasing in intensity and frequency as climate change accelerates. While kilometer-scale precipitation forecasts are critical for supporting local decision-making, the limited availability of high-resolution precipitation observations hinders their accuracy, especially in under-resourced regions. Machine learning models are widely used to downscale precipitation data to km-scale, but their application to unseen geographies presents challenges. First, processing raw high-resolution precipitation datasets across regions requires significant engineering and domain expertise. Second, generalization across regions remains difficult. To help overcome these barriers, we release RainAtlas, a large-scale, ML-ready and multi-continental dataset for precipitation downscaling. Covering three continents, RainAtlas harmonizes heterogeneous hourly km-scale observations to a common 2-km grid. Each regional partition contains around 210,000 aligned low- and high-resolution precipitation pairs, respectively from ERA5 reanalysis and direct observations. We benchmark state-of-the-art ML-based downscaling models across RainAtlas using a wide range of metrics. Our evaluation reveals substantial variance in out-of-domain generalization depending on the training regions. This underscores the need for cross-regional, multi-source km-scale evaluation, establishing RainAtlas as a well-positioned benchmark for precipitation downscaling research.

Figures & tables

Figure 1: Global distribution of ground-based radar stations (blue dots) and their available coverage (white transparent circles) . Some regions have excellent coverage (USA, Europe, East-Asia), while other densely populated regions have little to no precipitation observations available (Africa, South America, Central Asia). RainAtlas consists of the three multi-source precipitation products highlighted in the figure, covering three Northern Hemisphere regions.

Variable	Description	Unit	Level
tp	Total precipitation	mm	surface
cp	Convective precipitation	mm	surface
cape	Convective potential energy	J $\cdot$ kg -1	surface
sp	Surface pressure	Pa	surface
tisr	Top-of-the-atmosphere incident solar radiation	J $\cdot$ m -2	surface
tcw	Total column water	kg $\cdot$ m -2	total column

Table 1: ERA5 and static input data variables.

Figure 2: Diffusion-based models outperform other baselines. In-domain performance evaluation on importance-subsampled test sets. MAE/CRPS, and bias are computed without artificial coarsening. MAE/CRPS, and variogram rows have region-specific

y

-axis ranges to facilitate the visualization. Arrows indicate the desired direction for each metric. Numerical results are detailed in Table 4 .

Figure 3: Models trained on MRMS show more robust geographical generalization. Relative MAE, CRPS and VARIOGRAM degradation ( $\%$ ) with respect to in-domain training. Additional results are detailed in Figures 15 and 16 .

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Comparison of MRMS and ERA5 precipitation samples. The high-resolution MRMS crops (top) are compared with the low-resolution ERA5 crops (bottom) for two different locations. In Location A , the MRMS data appears to be upsampled from a coarser grid. In contrast, Location B shows the typical fine-scale details expected from MRMS.

Figure 5: (left) Raw maximum hourly precipitation per grid-cell for EURADCLIM (2013 to 2023), and (right) same using a 125 mm threshold for easier identification of problematic grid-cells.

	MRMS		EURADCLIM		EA
	Random	Importance	Random	Importance	Random	Importance
$\leq 0$ mm (%)	95.80	72.76	82.81	55.77	96.68	80.33
$0-1$ mm (%)	2.38	12.63	15.34	37.18	1.01	5.05
$1-4$ mm (%)	1.47	11.72	1.65	7.13	1.77	10.58
$4-5$ mm (%)	0.10	0.91	0.08	0.36	0.15	1.03
$5-8$ mm (%)	0.13	1.14	0.08	0.39	0.19	1.37

Appendix

Table 2: Statistics of RainAtlas ’s training datasets with

150

k samples, generated with random or importance subsampling. The first section contains grid cell occurrences of multiple precipitation intensities, while the second section gives information on crops: average precipitation, average maximum intensity, and proportion of crops with more than

10\%,25\%

and

50\%

non-zero precipitation grid cells.

Figure 6: Train, validation, and test temporal splits. Because MRMS only partially overlap with EURADLCIM and EASTASIA, we reserve 2021 and 2022 for subsampling validation and test datasets. Since atmospheric predictability is limited to approximately 14 days, training on data from later periods does not induce data leakage.

	UNet	Diffusion	Corrective Diffusion	SerpentFlow	Consistency
Loss	$L_{2}$	EDM denoising	EDM denoising	$L_{2}$	Pseudo-Huber
Optimizer	AdamW	AdamW	AdamW	AdamW	AdamW
Learning rate	$2{\times}10^{-4}$	$10^{-4}$	$10^{-4}$	$10^{-4}$	$10^{-4}$
Weight decay	$10^{-4}$	$10^{-2}$	$10^{-2}$	$10^{-2}$	$10^{-2}$
Residual	✓	—	✓(UNet)	—	—
Dropout	0.1	0.1	0.1	0.1	0.1

Appendix

Table 3: Training and architecture hyperparemeters for every model. Additional hyperparameters are specified under each model specification when they are not shared with others. We note that most of these hyperparameters are certainly sub-optimals as we didn’t perform any extensive hyperparameters seach. Most values are taken from previous work ( Watt and Mansfield, 2024 ; Karras et al., 2022 ; Song and Dhariwal, 2024 ; Keisler et al., 2026 ) .

		MAE / CRPS $(\downarrow)$		BIAS $(\rightarrow 0)$		LHD $(\downarrow)$		SSR $(\rightarrow 1)$		VARIOGRAM $(\downarrow)$
MRMS	Bilinear	0.7288		-0.0490		19.9497		-		-
	UNet	0.5750	±0.0026	-0.4033	±0.0069	21.0799	±0.1626	-		-
	Diffusion	0.4593	±0.0014	-0.0472	±0.0413	4.4261	±0.6763	0.9437	±0.0103	0.1671	±0.0004
	Corrective Diffusion	0.4639	±0.0031	-0.0341	±0.0358	3.3621	±1.0940	0.8373	±0.0436	0.1662	±0.0007
	SerpentFlow	0.5031	±0.0005	0.0353	±0.0271	8.5822	±0.9519	1.0268	±0.0616	0.1798	±0.0012
	Consistency	0.4867	±0.0107	0.2825	±0.0623	7.0539	±0.2034	1.2294	±0.0225	0.1825	±0.0027

Appendix

Table 4: In-domain performance evaluation on importance-subsampled test datasets. MAE, CRPS, and BIAS are computed on the original resolution without artificial coarsening. Probabilistic metrics (SSR and VARIOGRAM) are evaluated exclusively for generative models, omitting the deterministic Bilinear (ERA5) and UNet baselines. Arrows in the header indicate the optimal direction for each metric. The best and second-best results are bolded and underlined , respectively. Subscripts denote standard deviations.

Figure 7: Effect of spatial coarsening on pointwise error and bias (importance-subsampled test sets). MAE / CRPS ( top ,

\downarrow

) and bias ( bottom ,

\rightarrow 0

) evaluated across four effective spatial resolutions (native

2

km,

8

km,

16

km, and

24

km) obtained by average pooling. Shaded bands denote

\pm

standard deviation across three random seeds. Pointwise errors systematically decrease with spatial aggregation due to reduced sensitivity to small-scale displacement errors (double-penalty effect), while bias remains scale-invariant. Relative performance rankings remain unchanged across resolutions.

Figure 8: In-domain intensity distribution and spectral fidelity (importance-subsampled test sets). (Top) Precipitation intensity histograms comparing model predictions against ground truth up to

300

mm h -1 on a logarithmic scale. (Bottom) Radially Averaged Power Spectral Density (RAPSD) normalized by observed power spectrum across radial wavenumbers

k

. The horizontal dashed line at

y=1.0

indicates perfect spectral agreement with observations. Shaded bands represent

\pm

standard deviation across evaluation crops and ensemble members.

Figure 9: Fractions Skill Score across precipitation thresholds (importance-subsampled test sets). Fractions Skill Score (FSS,

\uparrow 1

) evaluated across neighborhood scales (

4,8,16,32,64

km) for thresholds

q\in\{1,5,10,20\}

mm h -1 ( left to right ). Shaded bands denote

\pm

standard deviation across three random seeds. Generative baselines, especially diffusion and corrective Diffusion, retain high skill for extreme precipitation (

q\geq 5

mm h -1 ), whereas deterministic baselines degrade sharply.

Figure 10: Ensemble calibration via rank histograms (importance-subsampled test sets). Talagrand rank histograms for generative models evaluated using an

8

-member ensemble (

9

rank bins). The horizontal dashed line at

y\approx 0.111

marks the ideal uniform distribution corresponding to perfect ensemble calibration. Random tie-breaking is applied to values below

0.1

mm h -1 .

		MAE / CRPS $(\downarrow)$		BIAS $(\rightarrow 0)$		LHD $(\downarrow)$		SSR $(\rightarrow 1)$		VARIOGRAM $(\downarrow)$
MRMS	Bilinear	0.1115		0.0144		18.7129		-		-
	UNet	0.0719	±0.0001	-0.0513	±0.0011	19.8704	±0.1937	-		-
	Diffusion	0.0609	±0.0003	0.0323	±0.0063	5.0397	±0.6527	1.2478	±0.0158	0.0314	±0.0004
	Corrective Diffusion	0.0616	±0.0008	0.0224	±0.0075	3.6442	±1.6567	1.0549	±0.0788	0.0304	±0.0006
	SerpentFlow	0.0742	±0.0002	0.0349	±0.0028	9.2778	±0.5459	1.2290	±0.0520	0.0340	±0.0005
	Consistency	0.0659	±0.0017	0.0716	±0.0109	7.3231	±0.2379	1.4877	±0.0318	0.0347	±0.0008

Appendix

Table 5: In-domain performance evaluation on randomly-subsampled test datasets. Details regarding metrics and formatting are identical to Table 4 . Probabilistic metrics are evaluated exclusively for generative models. The best and second-best results are bolded and underlined , respectively. Subscripts denote standard deviations.

Figure 11: Effect of spatial coarsening on pointwise error and bias (randomly-subsampled test sets). Settings and results are similar to Figure 7 .

Figure 12: In-domain intensity distribution and spectral fidelity (randomly-subsampled test sets). Settings are similar to Figure 8 . Generative baselines show much larger RAPSD, with high variance on EA, compared to evaluation on importance-subsampled test datasets.

Figure 13: Fractions Skill Score across precipitation thresholds (randomly-subsampled test sets). Settings are similar to Figure 9 . Generative models maintain notable skill for higher intensity thresholds even under climatologically unweighted random sampling.

Figure 14: Ensemble calibration via rank histograms (randomly-subsampled test sets). Settings are similar to Figure 10 . Ensemble calibration patterns remain consistent with the importance-subsampled results.

Figure 15: Out-of-domain generalization relative degradation on importance-subsampled test sets. Relative degradation ( $\%$ ) with respect to in-domain training across five evaluation metrics (MAE/CRPS, BIAS, LHD, SSR, and VARIOGRAM). Best-performing out-of-domain training domains per column are boxed in black. Scores are averaged over three random seeds, except for configurations with runs still in progress at submission, which are averaged over the seeds available. Full results will be updated during the rebuttal period.

Figure 16: Out-of-domain generalization relative degradation on randomly-subsampled test sets. Relative degradation ( $\%$ ) with respect to in-domain training across five evaluation metrics (MAE/CRPS, BIAS, LHD, SSR, and VARIOGRAM). Scores are averaged over three random seeds, except for configurations with runs still in progress at submission, which are averaged over the seeds available. Full results will be updated during the rebuttal period.

Figure 17: Intensity histograms and relative RAPSD under geographical generalization (importance-subsampled test sets). Line colors distinguish the different training domains (MRMS, EURADCLIM, EASTASIA, EUR+EA, MRMS+EA, MRMS+EUR) against ground truth (black).

Explore similar work

Evaluating Cross-region Generalization for Wavelet-Diffusion Precipitation Downscaling
Sep 23, 2026Weikang Qian, Yixin Wen, Chugang Yi +3Dynamical Downscaling Robust Cross-Domain Generalization
University of Florida, Department of Geography, Gainesville, FL, USA · School of Marine and Atmospheric Sciences, Stony Brook University, Stony Brook, NY, USA · Institute for Advanced Computational Science, Stony Brook University, NY, USA +5
Longwang: Zero-Shot Global Spatiotemporal Precipitation Downscaling with a Latent Generative Prior
May 17, 2026Yue Wang, Daniele VisioniDynamical Downscaling Precipitation Nowcasting
1*Department of Earth and Atmospheric Sciences, Cornell University, 112 Hollister Drive, Ithaca, 14853, NY, USA.
Lightweight Probabilistic Downscaling from a Deterministic Base Model
Sep 24, 2026Joseph McLean, Tiffany Vlaar, Sigrid Passano Hellan +1Dynamical Downscaling Probabilistic Forecasting
University of Glasgow · Climate System and Climate Services, NORCE Research AS and Bjerknes Centre for Climate Research, Bergen, Norway

Sep 23, 2026, cs.LGJ/K move · Enter open · S save

Evaluating Cross-region Generalization for Wavelet-Diffusion Precipitation Downscaling

Weikang Qian, Yixin Wen, Chugang Yi, Zhi Li+2

University of Florida, Department of Geography, Gainesville, FL, USA · School of Marine and Atmospheric Sciences, Stony Brook University, Stony Brook, NY, USA · Institute for Advanced Computational Science, Stony Brook University, NY, USA+5

Dynamical Downscaling Robust Cross-Domain Generalization Wavelets

Open paper AskPDF arXiv

Diffusion models have shown strong potential for kilometer-scale precipitation downscaling, but their performance in geographically unseen regions and event regimes remains insufficiently understood. Building on the wavelet diffusion model (WDM) framework, this study evaluates cross-region and cross-event generalization. Six 3 x 3 deg U.S. regions represent convective, winter, tropical, and atmospheric-river precipitation regimes. Low-resolution inputs are generated by block averaging NOAA Multi-Radar/Multi-Sensor (MRMS) composite reflectivity fields. A WDM trained only on Oklahoma (OK) samples and a WDM trained on all six regions are compared with nearest-neighbor and Bicubic interpolation. Model performance is evaluated using three metric families that measure image-domain reconstruction, spectral and distributional fidelity, and bin-wise precipitation detection. The OK-trained WDM remains competitive outside OK. Although the all-region WDM delivers the best and most consistent overall image-domain and detection performance, its gains are uneven across precipitation intensities. Bin-wise critical success index (CSI) over 5-dBZ reflectivity bins shows that WDM improvements concentrate in localized higher-reflectivity structures, which image-domain metrics partly obscure. In addition, the performance differences among samples are strongly associated with the spatial organization of the precipitation field, quantified by Moran's I as the spatial autocorrelation of each reflectivity bin. The sample-level Moran's I-CSI correlation stratified by sample intensity reaches 0.901 in all six regions, including regions unseen during training. Overall, these findings support future efforts to transfer downscaling models to regions with limited local training data and to generate globally consistent, high-resolution precipitation products.

Abstract

Explore similar work

Evaluating Cross-region Generalization for Wavelet-Diffusion Precipitation Downscaling

Longwang: Zero-Shot Global Spatiotemporal Precipitation Downscaling with a Latent Generative Prior

Lightweight Probabilistic Downscaling from a Deterministic Base Model