Regressing the Cyclones with CNN via Continuous Satellite Image Classification

In this work, tropical cyclone intensity estimation is approached as a continuous image classification problem, with gradual changes in the intensity values. The key objective of existing methods that approach cyclone intensity estimation as an image classification task is to generate accurate and differentiating representations of the cyclone images from different categories. These techniques do not, however, address the long-tailed distribution of the cyclone image datasets, which results in the under-representation of some cyclone categories. Therefore, we propose an approach that deals with the long-tailed distribution of cyclone image datasets. In this study, a multi-class classifier that serves as a feature extractor is trained using class-balancing strategies. The classification of test features using previously learned categorical feature representations by feature extractor is followed by intensity estimation. A cascade model approach, where sub-models are trained on sub-categories of the dataset and then cascaded to form a model that tackles heavy class imbalance is also explored. Additionally, a probabilistic approach to comprehend the model’s performance in scenarios including continuous image frames with reduced class diversity at the decision boundary is presented in the discussion section. In all the experiments, the benchmark dataset—the TCIR dataset—for tropical cyclone intensity estimation is utilised. Images in this dataset belong to various categories of tropical cyclones, ranging from tropical depressions to category 5 cyclones. This categorization is made based on the maximum rotational velocity of the cyclone measured in knots (kt). Extensive experiments indicate that the current proposed method outperforms the baseline and has achieved an RMSE of 11.15 kt (knots).