Abstract
Tropical cyclone (TC) intensity estimation remains a critical yet challenging task, especially for intense storms that are often underestimated by both conventional and deep learning satellite‐based methods. Among various structural features, eye formation is the most distinct transformation during TC development and is closely linked to rapid intensification and peak intensity. Despite its significance, TC eye occurrence information remains underutilized in existing deep learning‐based intensity estimation models. In this study, we improve TC intensity estimation by incorporating TC eye occurrence information as a physical auxiliary input to deep learning models. We first develop a deep learning‐based classifier (DeepTCEye) to detect TC eyes from infrared imagery, achieving classification accuracies of 96.54% for eye scenes and 95.93% for non‐eye scenes. Using the output probabilities from DeepTCEye, we construct eye persistence index (EPI) of varying time windows and use them as auxiliary inputs to the intensity estimation model. Results show that incorporating EPI leads to substantial improvements on TC intensity estimation, especially for intense TCs (intensity >96 kt), reducing the root mean square error by up to 15.65%. Moreover, EPI remains effective when combined with other known physical auxiliary input such as TC fullness, a structural indicator. Our results demonstrate the value of bringing the TC eye occurrence information into focus, showing that temporal evolution of TC eye provides deep learning models with a simple but physically grounded signal that significantly improves intensity estimation, particularly for intense storms. Plain Language Summary Tropical cyclones (TCs) are powerful storms that can cause severe damage. Accurately estimating their intensity is crucial but remains challenging, especially for intense storms. Although the eye of a TC is a key indicator of storm intensity, existing deep learning models have not fully used this information. This study develops a deep learning method to detect the presence of the eye using satellite infrared images. Based on these detections, we construct eye persistence index that captures how the eye comes and goes over time. Incorporating these sequences significantly improves deep‐learning based intensity estimates, particularly for strong TCs. Moreover, the longer EPI time series lead to better performance. Importantly, EPI can work not only on its own but also together with other physical features to support more accurate intensity estimation. Our findings show that including detailed eye evolution data helps deep learning models better understand storm intensity, paving the way for more accurate forecasts of high‐impact TCs.
Showing the abstract — retrieve the full paper via the Exa API.