A Generative Deep Learning Approach to Stochastic Downscaling of Precipitation Forecasts

Lucy Harris; Andrew T. T. McRae; Matthew Chantry; Peter D. Dueben; Tim N. Palmer

Source-linked AI summary

A Generative Deep Learning Approach to Stochastic Downscaling of Precipitation Forecasts

Lucy Harris, Andrew T. T. McRae, Matthew Chantry, Peter D. Dueben, Tim N. Palmer

arXiv:2204.02028v2Apr 5, 2022physics.ao-phcs.AIcs.CVcs.LGstat.ML

TL;DR

Precipitation forecasts are limited by unresolved small-scale rainfall processes and forecast error. This paper uses conditional GAN and VAE-GAN models to generate calibrated, high-resolution forecast ensembles, achieving spatially coherent precipitation maps and performance comparable to leading downscaling methods.

Problem
Global weather models often poorly predict precipitation because their resolution cannot represent fine rainfall structures and extremes, which matter for impact assessment.
Method
Conditional GAN and VAE-GAN models learn mappings from low-resolution weather-model fields to high-resolution radar rainfall, generating multiple forecast samples with explicit noise.
Results
Both models increase forecast resolution by a factor of 10, add forecast skill, and match or slightly improve upon ecPoint while producing spatially coherent precipitation maps.
Takeaways & Limitations
Stochastic generative downscaling can improve precipitation forecasts while preserving calibrated uncertainty and fine-scale spatial structure across scenarios including heavy rainfall.
Takeaways & Limitations
Forecast error between the IFS and NIMROD datasets, especially misaligned fronts and precipitation events, limits model success; the authors propose post-processing forecast ensembles as future work.

Abstract

from arXiv · show

Despite continuous improvements, precipitation forecasts are still not as accurate and reliable as those of other meteorological variables. A major contributing factor to this is that several key processes affecting precipitation distribution and intensity occur below the resolved scale of global weather models. Generative adversarial networks (GANs) have been demonstrated by the computer vision community to be successful at super-resolution problems, i.e., learning to add fine-scale structure to coarse images. Leinonen et al. (2020) previously applied a GAN to produce ensembles of reconstructed high-resolution atmospheric fields, given coarsened input data. In this paper, we demonstrate this approach can be extended to the more challenging problem of increasing the accuracy and resolution of comparatively low-resolution input from a weather forecasting model, using high-resolution radar measurements as a "ground truth". The neural network must learn to add resolution and structure whilst accounting for non-negligible forecast error. We show that GANs and VAE-GANs can match the statistical properties of state-of-the-art pointwise post-processing methods whilst creating high-resolution, spatially coherent precipitation maps. Our model compares favourably to the best existing downscaling methods in both pixel-wise and pooled CRPS scores, power spectrum information and rank histograms (used to assess calibration). We test our models and show that they perform in a range of scenarios, including heavy rainfall.

Plain Language Summary

The paper develops a machine-learning method that improves rainfall forecasts by increasing their resolution tenfold, generating probabilistic rainfall patterns, and outperforming existing approaches, including for extreme rainfall.

The method learns from past weather forecasts and observed rainfall to improve forecast accuracy automatically.
It increases rainfall-forecast resolution by a factor of 10.
The model generates many possible rainfall patterns, representing forecast uncertainty.
The approach uses Generative Adversarial Networks, adapting image super-resolution methods to rainfall forecasting.
It outperforms a variety of existing approaches and produces good predictions for the dataset’s most extreme rainfall situations.

1 Introduction

Precipitation varies at scales much finer than global forecast models resolve, making direct forecasts inadequate for local and extreme rainfall. The paper extends stochastic downscaling from pure super-resolution to forecast post-processing that also accounts for forecast error.

Global weather models typically operate at 10–80 km grid spacings, limiting their representation of small-scale rainfall processes, topography, and land-sea effects.
Direct application of weather and climate model outputs is inadequate for extreme rainfall and areas with heterogeneous orography or coastlines.
Precipitation intensity can vary over spatial scales of 1 km or less, making high-resolution downscaling important for assessing extreme-rainfall impacts.
GAN-based super-resolution methods provide the conceptual basis for generating realistic fine-scale atmospheric structure from coarse inputs.
Unlike pure super-resolution methods, the proposed approach maps multiple low-resolution forecast fields to high-resolution radar truth while correcting forecast error probabilistically.

2 Data

The study maps hourly ECMWF IFS forecasts over mainland UK to hourly high-resolution NIMROD radar rainfall, using static geographic fields and carefully separated training, validation, and test data.

The model maps hourly Integrated Forecast System data to hourly accumulated rainfall from the NIMROD radar network.
The domain covers mainland UK between latitudes 49.5°–59° and longitudes -7.5°–2°.
Inputs use 7–17-hour IFS forecasts initialized at 00Z and 12Z, with later lead times expected to have poorer alignment with observations.
The target dataset is 1 km-resolution UK composite radar rainfall from NIMROD, aggregated to hourly values for 2016–2020.
High-resolution surface geopotential and land-sea masks provide static geographic information that may help reconstruct sub-input-scale rainfall structure.
Training uses 2016–2018, validation uses 2019, and all quantitative evaluation uses held-out 2020 data.

3 Methods

The paper uses conditional GAN and VAE-GAN models to generate calibrated, high-resolution precipitation ensembles from atmospheric forecasts and geographic conditioning fields. Their architectures combine convolutional residual networks, noise or latent variables, and adversarial discrimination against high-resolution rainfall truth.

Both models post-process lower-resolution atmospheric forecasts into well-calibrated ensembles of high-resolution precipitation forecasts.
The conditional GAN supplies forecast fields, orography, and land-sea masks to a noise-driven generator, while a discriminator distinguishes generated rainfall from high-resolution truth.
3.2 Model 2: VAE-GAN: The VAE-GAN replaces the GAN generator with a variational autoencoder and uses the discriminator as its content-loss function, producing sharper and better-calibrated results than the initial VAE.
3.3 Model Architecture: The networks are deep convolutional models built largely from residual blocks, with fully convolutional generators that can accept different image sizes.
The models use high-resolution geographic fields, low-resolution atmospheric conditioning, and stochastic noise or latent variables, with downscaling factor K = 10.
The GAN uses approximately 3.2 million generator parameters and 64 million discriminator parameters with the selected filter settings.
3.4 GAN architecture: In the VAE-GAN, additional residual blocks after latent sampling help sampled variables develop coherent larger-scale spatial variations before upsampling.

4 Training and Validation

The paper trains conditional WGAN and VAE-GAN precipitation downscalers with adversarial, content, and regularization losses, then evaluates them using probabilistic, spatial, spectral, and calibration metrics. Training is non-monotonic and unstable, so checkpoints and multiple metrics guide model selection.

Training objectives: Conditional GANs provide IFS and geographic data to both networks, while the generator additionally receives noise and the discriminator distinguishes truth from generated precipitation.
Training objectives: Wasserstein GAN training uses discriminator and generator losses based on the Wasserstein distance, with a gradient penalty and a content-loss term.The gradient penalty uses randomly weighted averages of real and generated samples; the content loss is the mean squared error between truth and an eight-member ensemble mean.
Training objectives: The VAE-GAN generator adds a Kullback–Leibler divergence term weighted by 10^-5, and results were not especially sensitive to that factor.
Training procedure: Training alternates five discriminator iterations with one generator iteration, uses Adam, and required small learning rates because larger rates produced unstable training.The GAN used a learning rate of 10^-5 and the VAE-GAN 5 × 10^-6; training used batch size 2 for 320,000 batches.
Evaluation metrics: Performance is assessed with CRPS, rank histograms, ensemble-mean RMSE, RALSD, and fractions skill scores to measure probabilistic accuracy, calibration, image quality, spectra, and spatial consistency.CRPS averages pixel-wise ensemble scores, rank histograms assess variability, and FSS evaluates thresholded precipitation at spatial scales.
Evaluation metrics: Basic FSS can be artificially inflated by small-scale noise at intermediate spatial scales, whereas ensemble FSS does not appear to have the same flaw.

5 Results

On unseen 2020 cases, the GAN and VAE-GAN generally matched or exceeded existing downscaling methods across probabilistic, spatial, visual, and extreme-rainfall evaluations, while retaining some calibration limitations.

Quantitative evaluation: The VAE-GAN achieved the best CRPS scores, marginally ahead of the GAN, while both compared favourably with ecPoint on pixel-wise CRPS.The GAN produced the best ensemble-mean RMSE, marginally better than the deterministic CNN.
Quantitative evaluation: Only the GAN and VAE-GAN produced realistic power spectra, indicating spatial structure absent from the alternative methods.The RALSD results assessed power-spectrum realism; ecPoint was not designed to produce coherent spatial forecasts.
Example predictions: Across four cases, GAN predictions were more detailed and visually realistic than competing methods, and more robust for intense rainfall.RainFARM added repetitive texture, ecPoint was effectively bias-corrected IFS, and the deterministic CNN produced blurry predictions unable to predict extremes.
Example predictions: GAN and VAE-GAN ensembles produced sharply varying spatial structures resembling the truth, while the GAN more often predicted intense rainfall than the VAE-GAN.Both models generated multiple realisations for each forecast situation, representing forecast uncertainty.
Extreme events: For Storm Ciara, both models captured peak intensities and fine-scale rainfall structure better than the IFS forecast.The example covered 09:00–10:00 UTC on 9 February 2020, one of the dataset’s most extreme events.
Calibration: Both neural ensembles were slightly underdispersive in rank statistics, while the GAN performed better than the VAE-GAN on extreme events but none was particularly reliable.For the top 0.01% of forecasted precipitation events, ecPoint was over-dispersive and the neural models were especially under-dispersive on the right tail.
Fractions skill: At 0.5 mm/hr, both neural models outperformed ecPoint in ensemble FSS, whereas at 5.0 mm/hr the GAN significantly outperformed the VAE-GAN and ecPoint at small and intermediate scales.The VAE-GAN struggled to produce the highest precipitation intensities.
Lead time: With increasing lead time, GAN and VAE-GAN CRPSS decreased, consistent with application beyond their training lead times, whereas ecPoint CRPSS increased.The investigation covered 00Z forecasts at 6–72-hour lead times using 100-member ensembles.

6 Discussion and conclusions

The GAN and VAE-GAN models increase forecast resolution by 10×, add forecast skill, and produce spatially coherent ensembles, while forecast-error misalignment remains the main limitation.

The models achieve scores similar to or slightly better than ecPoint while producing spatially coherent and visually realistic precipitation maps.ecPoint remains better calibrated according to rank histograms.
The models generate ensembles that provide uncertainty estimates for weather-forecasting applications.
In NIMROD-to-NIMROD mapping, CRPS decreases from 0.0856 mm/hr to 0.0230 mm/hr, indicating forecast-error differences limit the full problem.The authors attribute this limitation to misaligned fronts and precipitation events between IFS and NIMROD.
Future work includes post-processing forecast ensembles, improving temporal consistency, exploring architectures, and extending the model geographically.

7 Open Research

The paper identifies several directions for extending the approach, including downstream flood modelling, ensemble post-processing, temporal consistency, architectural improvements, and broader geographic application.

The model and VAE-GAN code are publicly available, and the experiments primarily used TensorFlow 2.7.0.

Appendix A Ablation studies

The appendix evaluates the final GAN against ablated versions, examining training-data weighting, loss terms, and geographic input fields.

Table A1 compares the final GAN with ablated versions of the model.
The ablations vary training-data distribution, content-loss inclusion, and high-resolution geographic inputs.

A1 Varying training data distribution

Training-data weighting materially affects rainfall predictions and evaluation metrics, with an intermediate weighting selected as the best overall balance.

41× natural weighting causes the network to under-predict rainfall, despite marginally better CRPS scores.Equal sampling instead causes rainfall over-prediction.
A weighting of k=4 offers the best overall results across CRPS, rank histograms, and RALSD.Lower k values tend to improve rank histograms and RALSD, whereas higher values improve CRPS.
Differences between training-weighting choices are subtle in the final models, but data balance remains important.
Figure A1 compares GAN outputs under different input-sample weightings with low-resolution IFS inputs and high-resolution NIMROD truth data.
Adding an ensemble-mean-MSE content loss improves CRPS, RALSD, and ensemble-mean RMSE.
Removing high-resolution geographic fields can noticeably worsen validation performance across a larger number of cases.

Appendix B Pure super-resolution problem

The pure super-resolution experiment maps coarsened NIMROD radar data to full-resolution radar data, isolating resolution enhancement from forecast error. It performs substantially better than the full downscaling problem, while reduced-precision inference preserves performance.

Experimental setup: The experiment replaces the nine IFS input fields with coarsened NIMROD data and predicts the original full-resolution NIMROD truth.High-resolution orography and land-sea masks remain auxiliary inputs.
Results: The model performs very well on pure super-resolution, whereas the full downscaling problem is more challenging because it includes forecast error.The pure super-resolution CRPS is 0.0230 mm/hr, compared with 0.0856 mm/hr for the full problem.
Results: The pure super-resolution model obtains a CRPS of 0.0230 mm/hr, less than one-third of the 0.0856 mm/hr CRPS for full downscaling.This comparison separates recovering lost resolution from accounting for forecast error.
Reduced numerical precision: TF-32-only inference produces practically identical metrics to 32-bit inference while reducing runtime by approximately one-third.Explicit 16-bit training instead led to overflow and was unsuccessful.

Corresponding author: L. J. Harris, Department of Atmospheric, Ocean and Planetary Physics,

The supporting analyses examine calibration, discrimination, spatial spectra, ablations, and additional scenarios for GAN, VAE-GAN, and ecPoint models. Results indicate slight degradation with lead time, under-dispersion in the neural networks, and consistently stronger spatial spectra than simple interpolation methods.

Lead-time calibration: All three methods show only small differences between 24h, 48h, and 72h lead times, with very slight worsening at increasing lead times.They were trained on 7–17h lead-time data.
Lead-time calibration: At the top 0.01% of IFS predictions, all methods perform noticeably worse at 48h and 72h than at 24h lead time.This indicates that the most extreme forecast cases are more sensitive to lead time.
Calibration: The neural-network approaches are under-dispersive, whereas ecPoint is over-dispersive.This comparison comes from rank-histogram calibration assessments.
Spatial spectra: GAN and VAE-GAN models clearly outperform Lanczos and RainFARM in radially averaged power-spectrum comparisons across example images.The supporting plots are described as generally consistent across the examples.
Additional scenarios and ablations: Additional examples cover four weather scenarios and are included to demonstrate the models’ wide-ranging capabilities.Ablation studies compare models without content loss or geographic fields against the original GAN.

2. Description of case studies used in the main paper

Case studies show that GAN and VAE-GAN predictions preserve broad precipitation structure while adding fine-scale, spatially coherent variation and correcting several intensity or coverage errors in IFS forecasts. Ensemble members vary locally while remaining consistent at larger scales, expressing forecast uncertainty.

Heavy rainfall: For intense localized rainfall, the GAN and VAE-GAN make bold local predictions while the IFS forecast fails to capture peak intensities.The models also show fine-scale variation within the rainfall band.
Case-study predictions: GAN and VAE-GAN predictions add finely detailed precipitation structure while maintaining broadly correct large-scale organization.In one scattered-rain case, they also remove the IFS over-prediction of light rain over Northern Ireland.
Ensemble uncertainty: The models produce significant variation between predictions, corresponding to uncertainty in the rainfall scenario.Fine-scale variation occurs while larger-scale structure remains consistent.
Heavy and structured rainfall: For a structured two-band rainfall case, GAN predictions capture peak intensity and avoid the IFS over-prediction of light rain over a broad area.VAE-GAN corrects intensity somewhat but not enough in that example.
Orographic structure: The models use orographic information to produce consistent rainfall over the Highlands, Brecon Beacons, and western Pennines.The case study reports rainfall repeatedly concentrated over these regions.
Case-study predictions: In a light-rain case, GAN and VAE-GAN outputs mirror the radar pattern more realistically than the blurrier IFS forecast.The south-UK rain patch has location and intensity closely matching NIMROD for all three GAN predictions, while VAE-GAN slightly under-predicts it.

comparing our VAE-GAN model to both Lanczos interpolation and the RainFARM method

The supporting comparisons assess VAE-GAN and GAN spatial and threshold-based behavior against interpolation and RainFARM baselines. The available text emphasizes repeated evaluation across scenarios, models, thresholds, and ensemble predictions.

Ablation comparisons: Calibration plots compare ablated GAN variants, including versions without content loss or geographic fields, with the original GAN.The comparisons include all events and the top 0.01% of forecast predictions.
Calibration evaluation: The supporting analyses use rank histograms and CDFs to assess per-pixel normalized ranks for the trained models on the 2020 hold-out dataset.Thresholded versions separately assess the top 0.01% of predictions.
Scenario coverage: The supporting examples evaluate model predictions over four additional weather scenarios.The examples are intended to demonstrate the models’ wide-ranging capabilities.