Introduction
Meteorology is drowning in data but starving for accessible, high-quality training sets. While systems like GOES (satellites) and NEXRAD (radar) generate petabytes of information, using this data for deep learning is notoriously difficult. Researchers face three main hurdles: the sheer volume and complexity of the archives, alignment issues between different sensors operating on varying spatial grids and time steps, and the "rare event" problem where randomly sampling weather data often yields uninformative "clear sky" images.
The SEVIR (Storm EVent ImagRy) dataset bridges the gap between petabytes of raw sensor data and "machine-learning ready" formats. By providing over 10,000 curated, spatio-temporally aligned weather events, SEVIR accelerates deep learning in meteorology, enabling high-fidelity precipitation nowcasting and the generation of synthetic radar imagery from satellite data.
The SEVIR Dataset
The core contribution of this work is the SEVIR dataset itself, an approximately 1TB collection hosted on the Amazon Open Data Registry. It consists of over 10,000 weather events, each represented as a 4-hour sequence of 384 km x 384 km patches sampled at 5-minute intervals.
Crucially, SEVIR standardizes and aligns five distinct data modalities:
- Three GOES-16 satellite channels: Visible (VIS), Water Vapor (IR), and Longwave IR (Brightness Temperature).
- NEXRAD Vertically Integrated Liquid (VIL) mosaics (radar-derived).
- GOES-16 Geostationary Lightning Mapper (GLM) flashes.
Many of these events are cross-referenced with the NOAA Storm Events database, allowing researchers to link imagery to real-world impacts like hail, tornadoes, or flooding.
Methods and Modeling
Beyond providing the data, the authors establish benchmarks using U-Net architectures, a staple in image-to-image translation tasks. Their methodological exploration focuses heavily on how different loss functions affect meteorological outputs:
- Reconstruction Losses (MSE/MAE): These are standard but often result in "blurry" forecasts because they average out uncertainty.
- Perceptual Losses (VGG16 Content/Style): Borrowed from computer vision, these help models capture the "texture" and structure of storms.
- Conditional GANs (cGAN): Used to generate highly realistic synthetic radar imagery that mimics the statistical properties of real radar, even if pixel-wise accuracy is slightly lower.
Evaluation relies on domain-specific metrics like the Critical Success Index (CSI) and Success Ratio (SUCR), alongside the LPIPS metric for perceptual similarity.
Key Insights and Results
The creation and application of SEVIR yielded several important insights for the intersection of AI and meteorology:
- Alignment as a Service: The bottleneck in meteorological AI is often data engineering, not model architecture. By doing the heavy lifting of spatio-temporal alignment, SEVIR significantly lowers the barrier to entry for the broader AI community.
- Texture vs. Accuracy: In weather forecasting, a "blurry" but statistically accurate prediction (MSE) is often less useful to a human forecaster than a sharp, realistic-looking prediction (GAN) that might be slightly off in exact placement. SEVIR allows researchers to explicitly explore this trade-off.
- Synthetic Radar is Viable: The paper demonstrates that satellite and lightning data can be used to "hallucinate" radar imagery. Models trained with cGAN + MAE produced imagery that was qualitatively indistinguishable from real radar, capturing the fine-scale structures of intense storm cells. This is a game-changer for regions lacking expensive ground-based radar infrastructure.
- Nowcasting Performance: The U-Net models significantly outperformed traditional baselines like "Persistence" (assuming the weather won't change) and "Optical Flow" (moving pixels based on recent motion) for lead times up to 60 minutes.
Limitations and Future Work
While SEVIR is a major step forward, the authors acknowledge several limitations. Weather datasets are inherently susceptible to sensor noise and calibration discrepancies, which can affect model robustness. Furthermore, like many deep learning applications, these predictive models act as "black boxes." Future work must focus on explainability to help human forecasters trust AI-generated outputs.
Looking ahead, expanding this methodology from the US (CONUS) to global satellite constellations (like Meteosat or Himawari) is a logical next step. The SEVIR dataset also lays the groundwork for public challenges, driving further innovation in AI-driven meteorology.
Summary updated