[Paper Review] Increasing the accuracy and resolution of precipitation forecasts using deep generative models

Ilan Price, Stephan Rasp|arXiv (Cornell University)|Mar 23, 2022

Meteorological Phenomena and Simulations39 citations

TL;DR

The paper introduces CorrectorGAN, a conditional GAN that corrects and downscales global coarse precipitation forecasts to high-resolution, probabilistic forecasts, achieving near-regional-model performance at a fraction of the cost.

ABSTRACT

Accurately forecasting extreme rainfall is notoriously difficult, but is also ever more crucial for society as climate change increases the frequency of such extremes. Global numerical weather prediction models often fail to capture extremes, and are produced at too low a resolution to be actionable, while regional, high-resolution models are hugely expensive both in computation and labour. In this paper we explore the use of deep generative models to simultaneously correct and downscale (super-resolve) global ensemble forecasts over the Continental US. Specifically, using fine-grained radar observations as our ground truth, we train a conditional Generative Adversarial Network -- coined CorrectorGAN -- via a custom training procedure and augmented loss function, to produce ensembles of high-resolution, bias-corrected forecasts based on coarse, global precipitation forecasts in addition to other relevant meteorological fields. Our model outperforms an interpolation baseline, as well as super-resolution-only and CNN-based univariate methods, and approaches the performance of an operational regional high-resolution model across an array of established probabilistic metrics. Crucially, CorrectorGAN, once trained, produces predictions in seconds on a single machine. These results raise exciting questions about the necessity of regional models, and whether data-driven downscaling and correction methods can be transferred to data-poor regions that so far have had no access to high-resolution forecasts.

Motivation & Objective

• Address the need for accurate, high-resolution precipitation forecasts amid increasing extremes due to climate change.
• Develop a data-driven approach that corrects biases in global low-resolution forecasts and downscales them to 4 km resolution.
• Produce ensembles of high-resolution forecast fields with realistic spatial dependencies and calibrated probabilities.
• Evaluate against baselines including interpolation, pure super-resolution, CNN-based methods, and a regional high-resolution model (HREF).
• Demonstrate computational efficiency enabling rapid inference on a single machine.

Proposed method

• Develop a two-stage generator: a Corrector that produces corrected low-resolution representations and a Super-resolver that upscales to high resolution.
• Train using a three-stage regime: Stage 1 (low-res correction) with a log-transform preprocessing and FSS-based loss; Stage 2 (high-res pre-training) adding an L1 loss on the high-res output; Stage 3 (GAN training) with Wasserstein GAN loss and auxiliary L_LR and L_HR terms.
• Discriminator processes high-res observation and low-res forecasts, fusing their representations to judge plausibility and ground-truth consistency.
• Use a 4-channel input (16x16 TIGGE ensemble members, plus contextual variables) and a 16x16 to 128x128 super-resolver.
• Optimize with a Wasserstein loss with gradient penalty for the discriminator and a multi-term generator loss that encourages short-term and long-term consistency between corrected, high-resolution outputs and ground truth.

Experimental results

Research questions

• RQ1 Can a data-driven generative model correct biases in coarse global forecasts while also downscaling to high resolution?
• RQ2 Do corrected high-resolution forecasts produced by CorrectorGAN provide reliable probabilistic forecasts compared to baselines and regional models?
• RQ3 How close can a downscaled, corrected global-model approach get to a state-of-the-art regional high-resolution forecast in terms of probabilistic metrics?
• RQ4 What is the impact of including additional meteorological context and ensemble inputs on downscaling and bias correction performance?

Key findings

• CorrectorGAN outperforms interpolation, Pure-SR GAN, and BG-CNN baselines on CRPS and Brier Scores across 1, 5, 10, and 30 mm thresholds.
• CorrectorGAN’s CRPS is 0.574, close to HREF’s 0.562, and significantly better than TIGGE interpolation (0.605) and Pure-SR GAN (0.61).
• For 1mm, 5mm, 10mm, and 30mm thresholds, CorrectorGAN achieves Brier Scores of 0.06, 0.034, 0.02, and 0.0024 respectively, approaching HREF at extreme thresholds.
• HREF generally performs best among all models, with CorrectorGAN closely matching its reliability, especially for extreme precipitation (30 mm).
• The model is fast at inference, delivering high-resolution forecasts on a single machine, enabling scalable application.