This study develops a deep-learning framework for detecting tropical cyclone (TC) life-cycle stages from infrared satellite imagery across six global ocean basins. You Only Look Once (YOLO) v8 models were trained separately to detect genesis, intensification, and decay stages within individual basins as well as in generalized settings. Detection accuracy was consistently highest during intensification, with F1 scores exceeding 0.80 in several basins, while genesis and decay remained more challenging due to weaker or more disorganized storm structures. Cross-basin experiments revealed substantial performance degradation when Northern Hemisphere models were applied to Southern Hemisphere storms, highlighting the influence of hemispheric rotation and storm asymmetry. Rotation-based augmentation (180° rotation via image flipping) improved detection in the South Pacific and South Indian basins, particularly for genesis and intensification. Nonetheless, cross-basin transferability was constrained even within the same hemisphere, with notable weaknesses in the North Atlantic and variability across other basins. To interpret these disparities, three structural indices were evaluated; symmetry, texture, and shape complexity and Detected versus Non-detected distributions were statistically tested. Detected cases were generally associated with stronger texture and more symmetric structure, whereas shape-complexity contrasts were more basin- and stage-dependent. Spatial localization was evaluated using a YOLO box-midpoint center proxy, which can be biased for asymmetric systems. Overall, the results indicate that detection performance depends on life cycle stage, hemispheric orientation, and storm structural organization, underscoring the need for regionally adapted, structure-aware-specific generalization strategies for robust global TC detection.
Gunawardena et al. (Wed,) studied this question.