Robust and interpretable statistical models for predicting the intensification of tropical cyclones

Hurricane intensity prediction has proven to be a challenging task for the tropical cyclone forecasting community. One way this problem has been approached is through the use of statistical models. In this paper, three different aspects of intensity forecasting through statistical inference will be explored. To date, the prediction of intensity change for up to five days has been sufficiently addressed using multiple linear regression (MLR) in practice, while efficient non-linear regression methods, like neural networks (NNs), can be found in the literature. On the other hand, the procedures for reporting the prediction performance of the models and the feature selection techniques have not been investigated to a great extent, in the sense that (1) they widely vary and (2) the derived models have an inherent bias that prohibits good generalization behavior. Machine learning theory and a wide range of experiments in both artificial and real life domains have provided us with a large repository of algorithms and heuristics that help improve upon the methodologies used so far, while simultaneously building more robust models. Based on that, a large range of prediction performance and feature selection methods are investigated and the results are presented against the reference model, the Statistical Hurricane Intensity Prediction Scheme (SHIPS). As a final contribution, recently developed rule based regression techniques are applied to the dataset in order to identify more elaborate structure behind the intensity predictions. MLR and NNs fail to provide the human expert with interpretable results regarding possible multiple interdependencies of the inputs and the output. In contrast, rule based methods are not only competitive with respect to prediction performance, denoting that the rules are of good quality, but also support the capability of identifying multiple correlations in the dataset in an easy to read and validate manner.