Key Points

1. The paper introduces MetNet-3, a deep learning model for weather forecasting that extends lead time range and predicts multiple variables including precipitation, wind, temperature, and dew point up to 24 hours ahead.

2. MetNet-3 utilizes densification techniques, which produce spatially dense forecasts and capture data assimilation, overcoming the challenge of learning from sparse observations.

3. The model outperforms state-of-the-art probabilistic Numerical Weather Prediction models over the CONUS region for up to 24 hours lead time, setting a new performance milestone.

4. MetNet-3 focuses on operational latency, with forecasts available in Google Search, and has high temporal and spatial resolution capabilities.

5. The paper emphasizes the challenges of data assimilation, incorporating observational data, and capturing atmospheric information from diverse sources, such as weather stations, geospatial images, and radar signals.

6. MetNet-3's neural network architecture uses topographical embeddings, U-Net backbone, and MaxVit transformer to capture long-range interactions.

7. The evaluation metrics include Continuous Ranked Probability Score (CRPS), Critical Success Index (CSI), and Mean Absolute Error (MAE) to assess forecast quality for precipitation and other variables.

8. The paper includes detailed technical aspects such as input data processing, network architecture, lead time conditioning, topographical embeddings, and regularization techniques.

9. Additional results, ablations, comparisons, and visualizations are provided, showcasing the model's forecasting performance and improvements over its predecessors.

Summary

The paper presents MetNet-3, a high-resolution neural weather forecasting model that extends the lead time range to 24 hours and predicts multiple variables including precipitation, wind, temperature, and dew point. MetNet-3 achieves a temporal resolution of up to 2 minutes and a spatial resolution of 1 km with low operational latency. The model outperforms traditional Numerical Weather Prediction (NWP) models for up to 24 hours of lead time. MetNet-3 introduces a novel densification technique that allows it to produce spatially dense forecasts despite training on extremely sparse targets. It is operational and its forecasts are served in Google Search in conjunction with other models.

Sparsity Challenges and Novel Densification Process
The paper addresses the challenges in incorporating observational data with varying degrees of sparsity, including individual weather stations and dense geospatial images from ground-based radars and orbiting satellites. The novel densification process used by MetNet-3 aims to capture data assimilation and produce dense forecasts from sparse input data. The paper emphasizes that neural models offer an alternative paradigm for atmospheric modeling, with advantages such as high temporal and spatial resolution, minimal latency, and the ability to learn directly from atmospheric observations.

Evaluation Metrics and Probabilistic Outputs
Specific evaluation metrics including Continuous Ranked Probability Score (CRPS), Critical Skill Index (CSI), and Mean Absolute Error (MAE) are used to assess the quality of forecasts. MetNet-3's probabilistic outputs are compared with those of ensemble NWP models, showing its superior performance in forecasting precipitation rates, especially for longer lead times.

Technical Details and Results
Additionally, the paper provides technical details about the network architecture, inputs, normalization, lead time conditioning, and topographical embeddings used in MetNet-3. It also presents results for surface wind components, ablations with topographical embeddings, and a comparison between MetNet-2 and MetNet-3.

Overall, the paper highlights the advancements and challenges involved in implementing high-resolution neural weather forecasting models, particularly the unique capabilities and performance of MetNet-3 in predicting various weather variables with extended lead times and high spatiotemporal resolution.

Reference: https://arxiv.org/abs/2306.06079