NOAA GSL's Machine Learning Tropical Cyclone Detection System

NOAA Global Systems Laboratory (GSL) scientists, in collaboration with scientists from the European Centre for Medium-Range Weather Forecasts (ECMWF) and NVIDIA, investigated the use of machine learning (ML) and artificial intelligence (AI) to extract information for tropical cyclone (TC) forecasting. The goal of the collaboration was to develop the infrastructure for ML workflows into time-critical operations at ECMWF and to assess the feasibility of an operational product that uses ML to detect tropical cyclones.

Tropical cyclones are rapidly rotating storm systems with a low-pressure center that bring strong winds and heavy rainfall. Extracting meteorological data is a crucial component in cyclone forecasting, yet the process is very time-intensive. Great research has been devoted to easing this process to develop faster and more accurate convection prediction models, one such method is using machine learning.

Data is used to "train" a machine-learning model, teaching the model how to predict outputs. Operationally, forecast or satellite convection fields are entered in the model. The ML algorithm processes the data and produces an output prediction field known as a Region of Interest (ROI). Machine learning in the context of tropical cyclones applies a statistical technique known as clustering, to group satellite pixels that are similar to each other in order to produce an ROI.

"ML techniques can improve how fast we identify regions of interest in large data, such as tropical cyclones, that might otherwise be too time-consuming using more traditional methods. The ML methods might also identify more regions missed by these existing methods which could improve forecasts and warnings,” said Christina Kumler, a researcher on the project.

The ML algorithm under development is known as TCycl, and it can be used to detect the presence and location of TCs from certain meteorological fields. The algorithm is currently trained on precipitable water content to mimic satellite images. But, the variables in the training data set can be extended to include surface pressure or wind speeds.

Researchers compared the ECMWF non-ML TC-Tracker with the TCycl. The TC-Tracker was more sensitive, and TCycl tended to miss weaker TC’s, but detected hits for longer. The next steps will be to investigate when and where TCycl and TC-Tracker differ and prepare for the pre-operational stage.
Faster and more accurate convection forecasting enabled by machine learning can significantly improve public safety. The ability of the models to detect more ambiguous storm formations can allow for early detection and more time to prepare, leading to increased resiliency in the face of more storms.