English | 简体中文
The repository for this project is the code implementation of the paper Transforming Weather Data from Pixel to Latent Space.
If you find this project helpful, please give us a star ⭐️.
Main Contribution
-
We are the first to propose the novel idea of transforming weather data from pixel space to latent space for weather tasks. By transforming data into latent space, we decouple weather reconstruction from the downstream tasks, enabling the model to generate sharp and accurate results. The unified representation of pressure levels and variables allows task models to handle multiple pressure-variable subsets, while the latent space framework significantly reduces data storage and computational costs.
-
We introduce the Weather Latent Autoencoder for the pixel-to-latent transformation of weather data. WLA can effectively transform any pressure-variable subset from pixel space to a unified latent space, providing excellent compression and reconstruction performance. This allows weather task models to operate in latent space, achieving high accuracy with low data cost across multiple pressurevariable subsets.
-
We have constructed the ERA5-latent dataset, which provides large-scale ERA5 weather data with multiple pressure-variable subsets in a smaller data storage footprint and unified latent space. This transformation reduces the data costs of original ERA5 data, offering a robust data foundation for broader meteorological research
- Open source the model code
- [] Open source the training code
- [] Open source the pretrained model
- [] Open source the Era5-latent dataset
This project is licensed under the Apache 2.0 License。