[Paper Review] MetNet: A Neural Weather Model for Precipitation Forecasting

Casper Kaae Sønderby, Lasse Espeholt+7 more|arXiv (Cornell University)|Mar 24, 2020

Meteorological Phenomena and Simulations26 references306 citations

TL;DR

MetNet forecasts precipitation up to 8 hours ahead at 1 km2 spatial resolution and 2-minute temporal resolution using radar and satellite inputs with axial self-attention, outperforming NOAA HRRR on continental US scales.

ABSTRACT

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.

Motivation & Objective

• Demonstrate that a neural network can forecast precipitation with high spatial/temporal resolution over large regions.
• Show that direct probabilistic forecasts can capture uncertainty without physical PDE-based modelling.
• Investigate the benefits of large spatial context and fast, parallel inference for weather prediction.
• Compare MetNet to traditional NWP systems and baselines across lead times up to 8 hours.

Proposed method

• Use a four-dimensional input patch capturing MRMS radar, GOES-16 satellite data, and static geolocation features over a 1024x1024 km region.
• Condition MetNet on target lead time by encoding lead time information as an input feature.
• Output a 512-bin categorical distribution representing precipitation rates from 0 to 102.4 mm/h in 0.2 mm/h intervals.
• Process each time slice with a Spatial Downsampler, then a Temporal Encoder (ConvLSTM) to capture temporal dynamics.
• Apply an 8-layer axial self-attention Spatial Aggregator to achieve a global receptive field efficiently.
• Train with a discretized target distribution to stabilize learning and capture uncertainty.

Experimental results

Research questions

• RQ1 Can MetNet outperform operational NWP systems (e.g., HRRR) for 2–8 hour precipitation forecasts over the continental US?
• RQ2 How does increasing spatial and temporal context affect forecast accuracy for various precipitation thresholds?
• RQ3 What is the contribution of MRMS versus GOES-16 data to MetNet’s performance over different lead times?
• RQ4 Can a neural weather model provide calibrated probabilistic forecasts with practical latency suitable for real-time use?
• RQ5 What are the trade-offs between spatial context, temporal context, and model architecture in NWMs?

Key findings

• MetNet substantially outperforms HRRR for forecasts up to 8 hours over the continental US at multiple thresholds.
• MetNet’s performance improves with larger spatial context; reduced-context ablations show degradation beyond about 150 minutes lead time.
• Including MRMS data is more critical at shorter lead times, while GOES-16 data becomes relatively more useful at longer lead times.
• MetNet achieves near-seconds latency per lead time, with parallel computation irrespective of lead time.
• MetNet uses a 1024x1024 km input patch and 64x64 km target patch to balance context and resolution.
• Ablation studies indicate the importance of capturing spatial context beyond 512 km and temporal context up to 90 minutes.