Machine Learning-Driven Enhancement of Predictors and Prediction for Tropical Cyclone Activity

The H2020-CLINT project aims to improve the detection, causation and attribution of extreme climate events through artificial intelligence for the provision of enhanced climate services. Part of this project is dedicated to improving predictions of tropical cyclone activity on the medium range by training various types of machine learning (ML) models, ranging from classical baseline models (e.g., logistic regression, randomized trees) to more complex convolutional neural network (CNN) architectures (e.g., LSTM, U-Net). An extensive pool of predictor variables combines local and remote, oceanic and atmospheric, as well as tropical and extratropical predictors, represented by either 2-dimensional fields or index values. For the best-performing ML models, we also test a hybrid approach, in which dependencies learned for shorter lead times are applied to predictor fields taken from dynamical model output, to extend the predictive skill to longer lead times.

Since operational forecasts are largely based on dynamical models, the ECMWF ensemble is used as a skilful representative and serves as benchmark for evaluation along with a model predicting climatological probability. Results show that the best-performing models are CNN-based on the short-range, which can exploit spatial correlations in the input fields. For longer lead times, plain neural networks are more useful as they still allow for modelling nonlinearities but without identifying spatiotemporal patterns, before the predictive signal becomes too weak, and climatological forecasts take over. Even though a variety of predictors was tested, considerable improvements were mainly found when including previous predictions or near real-time observations. Insight into learned dependencies is gained through visualization methods for ML model interpretation.