Source-linked AI summary
MetNet: A Neural Weather Model for Precipitation Forecasting
Casper Kaae Sønderby, Lasse Espeholt, Jonathan Heek, Mostafa Dehghani, Avital Oliver, Tim Salimans, Shreya Agrawal, Jason Hickey, Nal Kalchbrenner
TL;DR
Weather forecasting is important but remains constrained by the computational demands and physical approximations of Numerical Weather Prediction. MetNet uses a probabilistic deep neural network with radar and satellite inputs, lead-time conditioning, and large-context axial self-attention. It improves upon HRRR for up to 8 hours, while extending beyond 8 hours requires larger contexts, engineering, and deeper networks.
Problem
Weather forecasting has major scientific, social, and economic importance, while NWP faces resolution–computation trade-offs and potentially imprecise or unresolved physical formulations.
Method
MetNet predicts probabilistic precipitation maps from mosaicked radar and satellite imagery plus forecast lead time, using axial self-attention to aggregate large spatial context.
Results
MetNet improves upon HRRR for up to 8 hours of precipitation forecasting and is reported to outperform NWP at this range and resolution.
Takeaways & Limitations
MetNet demonstrates that a neural weather model can outperform operational NWP on structured precipitation forecasting at continental scale and high spatiotemporal resolution.
Takeaways & Limitations
Forecasting beyond 8 hours requires ever larger input contexts, rigorous engineering, and deeper neural networks.
Abstract
from arXiv · showhide
Weather forecasting is a long standing scientific challenge with direct social and economic impact. The task is suitable for deep neural networks due to vast amounts of continuously collected data and a rich spatial and temporal structure that presents long range dependencies. We introduce MetNet, a neural network that forecasts precipitation up to 8 hours into the future at the high spatial resolution of 1 km$^2$ and at the temporal resolution of 2 minutes with a latency in the order of seconds. MetNet takes as input radar and satellite data and forecast lead time and produces a probabilistic precipitation map. The architecture uses axial self-attention to aggregate the global context from a large input patch corresponding to a million square kilometers. We evaluate the performance of MetNet at various precipitation thresholds and find that MetNet outperforms Numerical Weather Prediction at forecasts of up to 7 to 8 hours on the scale of the continental United States.
1. Introduction
Weather forecasting is socially and economically important, but Numerical Weather Prediction faces resolution, computational, and physical-modeling challenges. MetNet applies deep neural networks to precipitation forecasting and outperforms HRRR for up to 7 to 8 hours.
- Accurate weather forecasts support science, agriculture, energy, transportation, and the prevention of storm- and flood-related losses.
- NWP faces a trade-off between forecast resolution and computation time, while its physical formulation may be imprecise or unresolved at model resolution.
- Deep neural networks can exploit large spatial and temporal contexts while producing probabilistic outputs that represent predictive uncertainty.
- MetNet forecasts precipitation over the continental United States for up to 8 hours at 1 km^2 spatial and 2-minute temporal resolution using radar, satellite imagery, lead time, and axial self-attention.Its input context covers 1024×1024 km, and predictions take seconds independently of lead time and can run in parallel.
- MetNet outperforms HRRR for up to 7 to 8 hours, while ablations indicate that large spatial context supports accurate predictions at longer lead times.
- MetNet is reported as the first machine learning model to outperform NWP on a structured prediction task at this range and resolution.Visualizations suggest it captures advection and the formation of new precipitation regions.
2. Precipitation Forecasting
Precipitation forecasting is framed as a structured prediction benchmark using densely measured radar-derived targets. Each tensor value represents precipitation rate at a particular time and location.
- Precipitation provides a highly varying, densely measured target for structured prediction.
- The target is a three-dimensional tensor whose values represent precipitation rates in mm/h for specific times and locations.Target rates are estimated from Multi Radar Multi Sensor ground-based radar echoes.
3. Neural Weather Models
Neural weather models predict probability distributions over future weather conditions from observed inputs rather than single deterministic outputs. They use learned architectural constraints and discretized target intervals to model structured weather variables.
- Neural weather models predict probabilities of target weather conditions y at time T_y from input conditions x available up to T_x.
- The probabilistic formulation represents uncertainty by assigning a distribution over possible outcomes instead of producing one deterministic forecast.
- MetNet combines a lead-time-conditioned input patch with spatial downsampling, temporal encoding, and axial self-attention to aggregate global context.The figure describes shared convolutional processing, a Convolutional LSTM, and 8 axial self-attention blocks.
- These models encode generic spatial and temporal relations through architectural constraints and learn their parameters by minimizing forecast error.
- Continuous weather variables are discretized into intervals, and the model predicts a categorical distribution over those intervals.The discrete distribution is described as flexible and as stabilizing training of the underlying deep neural network.
4. MetNet Architecture
MetNet combines a large multi-source input patch, lead-time conditioning, temporal recurrence, and axial self-attention to produce probabilistic precipitation forecasts. Its design preserves broad spatial context while controlling computation and supports predictions from 2 to 480 minutes ahead.
- Input Patch: MetNet receives radar, satellite, geographic, temporal, and lead-time features from a 1024 × 1024 km input patch.The input spans 90 minutes sampled every 15 minutes, while the target covers 64 × 64 km.
- Lead-Time Conditioning: A forward pass conditions the computation on a requested lead time represented for forecasts from 2 to 480 minutes.The lead-time encoding is concatenated with descriptive input features at the outset.
- Output Layer: The output is a 512-way categorical distribution over precipitation-rate intervals from 0 to 102.4 mm/h.Threshold probabilities are obtained by summing the relevant categorical bins.
- Temporal Encoding: MetNet uses convolutional and pooling layers to contract spatial inputs before a recurrent network encodes temporal dynamics.The recurrent component is a convolutional LSTM processing time slices in sequence.
- Spatial Aggregator: Eight axial self-attention blocks provide a full spatial receptive field while avoiding vanilla self-attention’s quadratic computational factor.Four blocks operate along width and four along height; global context requires two axial blocks instead of 32 standard 3×3 convolutional blocks.
- Evaluation: MetNet outperforms HRRR up to 400 to 480 minutes and exceeds optical flow and persistence across the full 480-minute evaluation range.The comparison uses F1 scores at 0.2, 1.0, and 2.0 mm/h precipitation thresholds.
5. Discussion of NWMs and NWP
Neural Weather Models estimate future-condition distributions directly from observations, whereas NWP simulates atmospheric dynamics using physical equations and ensembles. Their computational behavior, uncertainty representation, and dependence on domain data differ accordingly.
- NWP and NWM Properties: NWP performs deterministic physical simulations from initial conditions and estimates uncertainty using ensembles with perturbed initial conditions.NWMs instead directly estimate p(y|x) from current observations.
- Model Structure: MetNet uses general convolutions, recurrence, and attention modules, while NWP relies on explicit phenomena-dependent physical equations.MetNet’s performance therefore correlates with the availability of task data.
- Computational Properties: MetNet’s latency is independent of the target lead time.
6. Experiments
Experiments compare MetNet with operational, optical-flow, and persistence baselines, then probe spatial and temporal context, data sources, and forecast visualizations. Results show broad long-horizon advantages and identify spatial context as especially important.
- Eight Hour Forecasts: The benchmark compares MetNet with HRRR, optical flow, and persistence using F1 scores at 0.2, 1.0, and 2.0 mm/h thresholds.The evaluation covers precipitation forecasts over the continental United States.
- Eight Hour Forecasts: MetNet outperforms HRRR through 7 to 8 hours and exceeds optical flow and persistence at all evaluated lead times.The paper identifies this as the first such machine-learning result on this structured weather benchmark at the stated scale and range.
- MetNet Ablation Experiments: Reducing the input patch to 512 km leaves performance similar through 150 minutes but causes progressively worse performance afterward.This ablation indicates the importance of large spatial context for longer lead times.
- MetNet Ablation Experiments: Reducing temporal context from 90 to 30 minutes does not significantly affect performance, while GOES-only performance approaches full MetNet with increasing lead time.The results suggest that MRMS data becomes less necessary as lead time increases.
- Visualization: MetNet predictions capture precipitation increases, decreases, and the formation of new high-precipitation regions across forecast times.Predictions become increasingly blurry with lead time, reflecting greater uncertainty in event timing and location.
7. Conclusion
MetNet improves upon HRRR for precipitation forecasting up to 8 hours, while extending beyond 8 hours requires larger input contexts, engineering, and deeper networks.
- MetNet improves upon the operational NWP system HRRR for up to 8 hours of lead time.
- Table 2 compares ground-truth MRMS observations with MetNet and HRRR predictions at a 1 mm/h precipitation threshold.
- Extending MetNet beyond 8 hours requires ever larger input contexts, rigorous engineering, and deeper neural networks.
Appendix A. Data
The appendix describes MetNet’s input construction, feature transformations, and masking strategy for precipitation targets.
- Input data are selected relative to an anchor time Tx using MRMS and GOES observations immediately preceding Tx.The preceding times are [−90 min, −75 min, ..., 0 min] relative to Tx.
- Figure 7 details MetNet input-data downsampling and the space-to-depth transformation, including resulting feature dimensions.
- Figure 8 shows NOAA quality masks used so loss is calculated only for targets covered by the mask.The masking strategy is intended to minimize issues with wrongly labelled targets.
Appendix B. Numerical Weather Prediction
Numerical Weather Prediction uses atmospheric equations and assimilated observations to forecast weather, but its accuracy depends on formulation and initial conditions.
- Numerical Weather Prediction uses PDEs and other equations to represent atmospheric dynamics.
- NWP model variables include surface-, airborne-, and satellite-based measurements such as temperature, moisture, precipitation, wind fields, and pressure.
- Forecast accuracy depends on the mathematical formulation and on initial conditions estimated through complex data assimilation.Initial-condition errors remain and negatively impact model performance during spin-up.
Appendix C. HRRR Baseline
The study compares MetNet with NOAA’s operational HRRR baseline after aligning HRRR forecasts and MRMS measurements on a common grid and time basis.
- HRRR produces hourly forecasts 1–18 hours into the future at 3-by-3 kilometer resolution.
- HRRR predictions are resampled to the NOAA CONUS grid used for comparison.
- Samples are discarded when no MRMS target is found within 5 minutes of the HRRR lead time.
Appendix D. Optical Flow Baseline
The appendix benchmarks optical-flow precipitation-forecasting baselines using MRMS data and compares three methods under a 0.1 mm/h cutoff. It also examines how spatial context affects the Dense Inverse Search baseline.
- The benchmark evaluates optical-flow precipitation-rate forecasting algorithms from the RainyMotion and PySteps libraries using MRMS data for tracking.The remaining data are otherwise left unchanged.
- Figure 9 compares Dense Inverse Search with constant-vector extrapolation, Lucas-Kanade with semi-Lagrangian extrapolation, and STEPS using 512 pixels of spatial context.
- The figure uses 0.1 mm/h as the precipitation cutoff level for comparing optical-flow baselines and context size.
- The Dense Inverse Search baseline is evaluated for the effect of spatial context size.The figure’s right panel focuses on this context-size comparison.