This project combines the historical record of real cyclones with data-assimilated global weather model output (reanalysis) to train a deep convolutional network capable of detecting cyclones in simulated climates. See the testing notebook for more explanation and demonstration code.
I would be happy to share the full dataset, which is standardized and stored convieniently in zarr format (inputs and targets) on Google Cloud Storage.
Final model training hasn't happened yet because I ran out of GCP free credits doing other stuff and I'd like to avoid spending a bunch of money for time on the bigger GPUs.
- ERA5 reanalysis
- vorticity and temperature are downloaded on pressure levels
- surface pressure is downloaded on a single level
- the climate data store API (cdsapi) was used to download files directly to Google Cloud Storage
- Storm Tracks
- provided by NOAA NCEI under the description: International Best Track Archive for Climate Stewardship (IBTrACS)
- technical details and column documentation are included in this repo for reference