MetNet: A Neural Weather Model for Precipitation Forecasting

TL;DR

MetNet tackles high-resolution, long-lead precipitation forecasting by learning a direct probabilistic map from multimodal radar and satellite inputs to a large-scale spatial-temporal distribution. The model architecture combines a spatial downsampling path, a ConvLSTM temporal encoder, and a global-context axial self-attention-based spatial aggregator to produce a 512-bin precipitation distribution for up to 8 hours ahead at 1 km^2 resolution with seconds-scale latency. Across continental US forecasts, MetNet outperforms the operational HRRR baseline and strong baselines like optical flow for extended lead times, demonstrating the viability of large-context neural weather models. The work highlights the value of continuous probabilistic outputs and scalable context-rich architectures for practical, fast forecasting in real-world weather prediction.

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.

Figures (9)

• Figure 1: Data sources. Left: GOES-16 visual bands. Right: MRMS precipitation rates. (Preliminary)
• Figure 2: Comparison of continuous and discrete predictions. The discrete prediction is arbitrarily flexible and better able to approximate a complex target distribution.
• Figure 3: Diagram of the input patch and the three parts of the MetNet architecture. The target lead time represented by the integer ${\bm{y}}$ is concatenated at the input to indicate to MetNet the desired lead time for the prediction. The Spatial Downsampler encodes the time slices sampled every 15 minutes using a shared convolutional neural network. The Temporal Encoder processes the downsampled time slices in the direction of time using a Convolutional LSTM. The Spatial Aggregator uses 8 axial self-attention blocks (only 2 shown) to cover a global receptive field over the input patch.
• Figure 4: Properties of NWP and NWMs. Left: NWP performs a deterministic physical simulation starting from the initial conditions. The predictive uncertainty is estimated from an ensemble of predictions each run with slightly different initial conditions. Right: The NWM treats the current observations as direct inputs to a DNN, directly estimating the distribution over future conditions $p(\bm{y}|\bm{x})$.
• Figure 5: F1 scores for MetNet, HRRR, Optical Flow and Persistence for 2 to 480 minutes of lead time evaluated at $0.2$ mm/h (top), $1.0$ mm/h (middle) and $2.0$ mm/h (bottom) precipitation rates. MetNet outperforms HRRR up to 400 to 480 minutes and outperforms a strong optical flow method and the persistence baseline throughout the 480 minute range.