Using Uncertainty-enhanced Deep Learning to Predict the Intensification of Tropical Depressions to Tropical Storms

Predicting whether a tropical depression (TD) will intensify to tropical storm (TS) strength is a problem of great societal importance.For each TD around the globe, our task is to predict whether it will intensify to TS strength in the next N hours -- N being a lead time in the set {6, 12, ..., 168}. To create labels, we use the IBTrACS dataset, which merges multiple data sources to create a best-guess archive of intensity/position for every TC globally. We use two types of predictors: TC-centered satellite images and environmental indices. Satellite images contain brightness temperature from the ~10.8-micron channel at 15-minute time intervals and 4-km grid spacing. Environmental indices come from the Statistical Hurricane Intensity Prediction Scheme (SHIPS) developmental dataset and include wind shear, air temperature, sea-surface temperature, maximum potential intensity, etc.Our specific deep-learning algorithm is the convolutional neural network (CNN). Because uncertainty quantification (UQ) is important for high-stakes decision-making, we combine the CNN with three UQ methods: Monte Carlo dropout, quantile regression, and continuous ranked probability score as loss function (CRPS-LF). We use a global dataset for training and evaluation, including the southern hemisphere. The CNN performs well on the testing set -- aggregating over all 28 lead times, the Peirce score (sensitivity minus specificity) is 0.76, and the critical success index (CSI) is 0.79. On the same data, the widely used SHIPS model achieves a Peirce score of 0.53 and CSI of 0.64. When combined with the CRPS-LF method, the CNN also produces good estimates of its own uncertainty, yielding a spread-skill ratio near 1 across all spread values. MC dropout and quantile regression both lead to overconfident (underspread) models. We will show these results in detail along with case studies.