- ShearWater (the bird)
- ShearWater is a bird to produce AI-enhanced tropical cyclone (TC) activity forecasts.
This WPS is designed to produce AI-enhanced tropical cyclone (TC) activity forecasts. Given a user-defined initialisation date, lead time and region, the WPS generates a probabilistic forecast of TC activity by evaluating the environmental conditions using an AI model. We here define TC activity as the passage of at least one TC within a distance of 300 km and a time window of 48 h, following a definition already used at ECMWF for a medium-range TC product based on IFS ensemble forecasts (Vitart et al. 2012).
Following this definition of TC activity, the target variable used to train the AI model was derived from the International Best Track Archive for Climate Stewardship (IBTrACS; Knapp et al. 2010, Gahtan et al. 2024). Only those parts of the observed TC tracks were considered, where the intensity exceeded tropical storm strength (>= 17 m/s). Meteorologically relevant features were extracted from the ERA5 atmospheric reanalysis dataset (Hersbach et al., 2020), produced by the Copernicus Climate Change Service at ECMWF. The problem under consideration is of a classificatory nature with binary targets. The default model is a purely data-driven model, based on a Unet implementation that takes as input the set of meteorological features, treating them as channels of an image. The model is designed to make a probabilistic prediction in the form of a grey-scale image.
The years 1980 to 2010 (11323 daily values) were selected for training the AI model in order to base it on more reliable data from the satellite era. The following years, 2011 to 2015 (1826 daily values), were used to fine-tune the AI model for better performance. The developed AI model was independently tested on unseen data and evaluated against a set of benchmarks for the period 1 June 2016 to 31 December 2022 (2405 daily values). The results of this thorough evaluation study will be published shortly. Thanks to a live connection to the DKRZ data pool, the AI model can be used to generate forecasts for any initialisation date in the period 1 January 1940 to 31 December 2024. Note, however, that input data prior to 1 January 1980 are less reliable due to a poorer observation coverage and quality, while dates between 1 January 1980 and 31 December 2015 have been considered during model training and fine-tuning already and are therefore not independent.
To generate a forecast, the WPS requires three inputs from the user. First, an initialisation date must be specified, which defines the reference at which the input features are extracted and from which the (possibly retrospective) forecast is issued. Second, the lead time can be selected from a predefined list, where the starting point of the 48-hour target window can vary from immediately after initialisation (‘0-48 h’) up to 13 days (‘312-360 h’) into the future. Third, the region of interest must be selected from the following seven basins, which have been used to train the AI model: Australia, East Pacific, North Atlantic, Northern Indian, Northwest Pacific, South Pacific, Southern Indian.
Once all the necessary information has been provided, the user can click on the ‘Submit’ button, which will run the process and present the output in two formats: 1) a CSV table containing the predicted values at each geographic grid point in the region of interest, and 2) a PNG image displaying the predicted TC activity map using the same colour coding as for the equivalent IFS-based product.
Future plans are to include additional AI models and benchmarks, such as ECMWF’s IFS ensemble probability, the climatological probability and the probability predicted by the more sophisticated hybrid approach. The latter takes the input features from the IFS ensemble and applies the trained AI model to combine the benefits of both approaches. Another idea is to let the user select the region of interest.
References
Gahtan, J., K. R. Knapp, C. J. Schreck, H. J. Diamond, J. P. Kossin, and M. C. Kruk, 2024: International Best Track Archive for Climate Stewardship (IBTrACS) project, version 4r01. NOAA National Centers for Environmental Information, accessed 1 July 2022, https://doi.org/10.25921/82ty-9e16.
Hersbach, H., B. Bell, P. Berrisford, S. Hirahara, A. Horányi, J. Muñoz‐Sabater, J. Nicolas, C. Peubey, R. Radu, D. Schepers, an A. Simmons, 2020: The ERA5 global reanalysis. Quart. J. Roy. Met. Soc., 146, 1999-2049, https://doi.org/10.1002/qj.3803.
Knapp, K. R., M. C. Kruk, D. H. Levinson, H. J. Diamond, and C. J. Neumann, 2010: The International Best Track Archive for Climate Stewardship (IBTrACS): Unifying tropical cyclone best track data. Bull. Amer. Meteor. Soc., 91, 363–376, https://doi.org/10.1175/2009BAMS2755.1.
Vitart, F., F. Prates, A. Bonet, and C. Sahin, 2012: New tropical cyclone products on the web. Tech. Rep. 130, ECMWF Newsletter, 17–23 pp. https://doi.org/10.21957/ti1191e2.
Learn more about ShearWater in its official documentation at https://shearwater.readthedocs.io.
Submit bug reports, questions and feature requests at https://github.com/cehbrecht/shearwater/issues
You can find information about contributing in our Developer Guide.
Please use bumpversion to release a new version.
- Free software: Apache Software License 2.0
- Documentation: https://shearwater.readthedocs.io.
This package was created with Cookiecutter and the bird-house/cookiecutter-birdhouse project template.