Skip to content

NowcastNet (DGMR): Skilful precipitation nowcasting using deep generative models of radar

Status: completed

Authors: Suman Ravuri, Karel Lenc, Matthew Willson, et al.

Venue / Year: Nature, 2021

Affiliations: DeepMind, Met Office (UK), University of Exeter, University of Reading

Links: Nature | GitHub

Tags: [[precipitation-nowcasting]], [[generative-models]], [[GAN]], [[radar]], [[spatiotemporal-forecasting]], [[operational-meteorology]]

One-Sentence Summary

A Deep Generative Model (DGMR) using dual-discriminator GAN framework produces probabilistic precipitation nowcasts that maintain spatiotemporal consistency and spectral characteristics, outperforming PySTEPS and UNet baselines in expert evaluation (89% preference rate).


Problem

Core Challenge

Existing deep learning nowcasting models (predominantly deterministic UNet architectures) suffer from spectral blurring at increasing lead times as they optimize for per-grid-cell MSE, failing to capture the stochastic nature of rainfall evolution. Meanwhile, operational methods like PySTEPS rely on advection equations that struggle with non-linear convective initiation.

Motivation

Accurate 0-2 hour nowcasts are critical for emergency services, flood early-warning systems, and aviation—domains where current NWP systems struggle due to spin-up lag. The socioeconomic impact of improved high-resolution decision-making justifies this research direction.


Method

Task Formulation

Conditional generative model predicting future radar fields (90 min) given past context fields (20 min):

Architecture

ComponentDescription
Conditioning StackCNN processing four 256×256 radar fields at multiple resolutions
Latent Conditioning StackSpatial attention block producing latent representations (1/32 of input size) from samples
Sampler (convGRUs)Convolutional GRUs with spectrally normalized convolutions for training stability
Spatial DiscriminatorEvaluates 8 random frames with hinge loss for spatial consistency
Temporal Discriminator3D CNN evaluating 128×128 crops across sequence to penalize non-physical transitions

Loss Function

Grid-cell regularization:

where clips at 24 mm/h to prioritize heavy precipitation accuracy.


Evidence

Key Results

MetricDGMRPySTEPSUNetAxial Attention
CSI (1 mm/h)0.680.620.580.55
CSI (4 mm/h)0.520.450.380.35
CSI (8 mm/h)0.380.310.220.18
PSD Resolution (T+90)~1 km~8 km32 km16 km

Expert Evaluation (n=56 meteorologists)

  • 89% ranked DGMR 1st for accuracy and value
  • PySTEPS criticized for "positional errors"
  • UNet deemed "too blurry"
  • Axial Attention labeled "too bland/blocky"

Computational Efficiency

  • Inference time: ~1.3 seconds per realization (NVIDIA V100)
  • Enables 20-member ensembles within operational windows

Limitations

  1. Intensity Decay: Underestimation at T+90 for heavy rainfall (>10 mm/h)
  2. Initiation Challenge: Struggles with small-scale isolated showers lacking contextual precursors
  3. Data Bias: Training on UK/US data may not generalize to tropical convective regimes
  4. Advection vs. Initiation: Handles advection of existing features well; initiation physics need refinement

Critical Assessment

Strengths

  • Resolution Maintenance: First DL model to prevent effective resolution collapse (maintains ~1km at T+90)
  • Operational Validation: Systematic expert endorsement beyond standard metrics
  • Training Stability: Successful spectral normalization in high-stakes physical domain

Risks

  • Ensemble mean remains blurry despite individual realizations maintaining detail
  • Intensity decay at longer lead times for extreme events
  • Importance-sampling scheme may create distribution shift issues

Innovation Score: 9/10

Paradigm shift: proves GANs can satisfy rigorous atmospheric science demands while redefining "good" forecast criteria through expert utility.


Reusable Ideas

  1. Dual-Discriminator Design: Spatial + Temporal discriminators for spatiotemporal consistency
  2. Spectral Normalization: Critical for stabilizing GAN training in physical domains
  3. Grid-cell Regularization: Weighted MAE with intensity-aware clipping for rare event focus
  4. Expert-in-the-Loop Evaluation: 56 expert meteorologists provide ground-truth utility assessment

Follow-up Directions

  • Time-dependent regularization weights to target T+90 intensity decay
  • Hybrid loss functions incorporating mass conservation constraints
  • Validation on tropical convective regimes

ModelApproachEffective Resolution
DGMRGAN-based generative~1 km
PySTEPSOptical flow + stochastic perturbations~8 km
UNetDeterministic CNN32 km
Axial AttentionAttention-based16 km
MetNetAxial attention + CNN2 km (deterministic)

Personal Notes

Why this matters: Demonstrates that generative models can bridge the gap between pixel-wise metrics and operational utility—critical insight for any spatiotemporal forecasting task where "realism" matters beyond MSE.

Questions for future work:

  • Can intensity decay be addressed with time-dependent loss weights?
  • How does this extend to multi-modal inputs (satellite, lightning)?
  • What is the minimum viable ensemble size for operational deployment?

Static research notes built with VitePress and KaTeX.