Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
125 tokens/sec
GPT-4o
53 tokens/sec
Gemini 2.5 Pro Pro
42 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

MetNet: A Neural Weather Model for Precipitation Forecasting (2003.12140v2)

Published 24 Mar 2020 in cs.LG, physics.ao-ph, and stat.ML

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.

Citations (251)

Summary

  • The paper introduces MetNet, a neural network model that forecasts precipitation up to 8 hours ahead, outperforming traditional NWP methods with lower latency.
  • It employs a novel architecture combining spatial downsamplers, ConvLSTM-based temporal encoders, and axial self-attention for capturing expansive geographic context.
  • Empirical evaluations reveal robust F1 scores across precipitation thresholds, underscoring its potential for fast and accurate short-term weather predictions.

MetNet: A Neural Weather Model for Precipitation Forecasting

The paper introduces MetNet, a neural network-based model designed to forecast precipitation with a lead time of up to 8 hours, spatial resolution of 1 km², and temporal resolution of 2 minutes. This model exhibits significant advancements in the field of short-term weather prediction by leveraging deep learning techniques, thus challenging traditional Numerical Weather Prediction (NWP) methods, particularly at smaller temporal scales.

Core Attributes and Methodology

MetNet capitalizes on the expansive, continuously-collected weather data, utilizing both radar and satellite sources. The innovative model architecture, predominantly featuring axial self-attention, enables the effective aggregation of spatial context from substantial geographic areas—up to 1024 x 1024 km—facilitating the capture of long-range dependencies crucial for enhanced predictive capabilities.

MetNet processes a four-dimensional data tensor comprised of time, geography, and satellite channels. The model's underlying structure consists of a Spatial Downsampler to contract spatial dimensions, a Temporal Encoder using ConvLSTM to perceive temporal dynamics, and a Spatial Aggregator employing axial self-attention layers to ensure a comprehensive receptive field. The integration of these components results in a model that can predict precipitation probability distributions over a high-dimensional space, addressing the structured prediction constraints in weather forecasting.

Results and Comparison

The empirical evaluation presented in the paper demonstrates that MetNet consistently outperforms the high-resolution HRRR NWP model up to 8 hours, evidencing its superiority in short-term precipitation prediction. Intriguingly, the model maintains a reduced computation latency—achieving results within seconds—contrasting with the extensive computations and longer latencies typical of NWP. The F1 scores across different precipitation thresholds (0.2, 1.0, and 2.0 mm/h) indicate that MetNet's predictive performance is statistically robust against both optical flow methods and persistent baseline approaches.

The ablation studies further corroborate the importance of capturing comprehensive spatial context, with MetNet's performance markedly enhancing when larger input patches are utilized. This high spatial awareness, coupled with temporal modeling, allows MetNet to model convective and non-convective precipitation patterns more effectively than traditional and even some contemporary optical flow models.

Implications and Future Directions

The ability of MetNet to leverage large volumes of data with computational efficiency has profound implications for operational weather forecasting, especially in regions with dense observational networks but where fast, real-time prediction capabilities are paramount. MetNet's prowess in processing and making contextual use of both historical and geographical data heralds a shift from deterministic physical simulations towards probabilistic, data-driven modeling in meteorology.

Future work may focus on extending MetNet's architecture to cover longer lead times by incorporating larger datasets and even more sophisticated attention mechanisms. Additionally, the exploration of incorporating other atmospheric variables beyond precipitation could potentially transform this architecture into a comprehensive neural weather modeling system.

In conclusion, MetNet represents a significant stride in the field of neural weather modeling, offering a viable alternative to classical NWP methods for short-term forecasts. This paper underscores the potential of deep learning models in enhancing predictive accuracy and efficiency, paving the way for future developments in AI-driven weather prediction methodologies.

X Twitter Logo Streamline Icon: https://streamlinehq.com
Youtube Logo Streamline Icon: https://streamlinehq.com