Hierarchical attention-enhanced multihead CNN and level sets segmentation: A proposed approach to enhance the cyclone intensity estimation

• The method studies the sway in a cyclone that finds to extract contours, patterns. • The method establishes as a robust tropical cyclone intensity estimation technique. • The method has a superior performance in comparison with AlexNet, CycloneNet, ResNet Tropical cyclones are considered to be one the most devastating natural disasters, that pose severe threats if not accounted for before its arrival. Predicting the intensity of such tropical cyclones becomes critically important for effective disaster management and response. However, existing techniques of Dvorak’s estimation do not take into account the physical attributes of a cyclone contributing to its intensity and assume a steady state for the same failing to incorporate the temporal aspect to them. This study proposes the use of a Nadam-optimizer based Hierarchical Attention-Enhanced Multihead for real time estimation and Bidirectional Long Short-Term Memory Networks for time series forecasting, with additional pre-processing of difference of Gaussians and level sets segmentation applied. Unlike the aforementioned existing methods, this approach accounts for all physical factors of influence in a cyclone that level sets segmentation leverages to extract contours and cyclonic patterns. Subsequently, the Hierarchical Attention-Enhanced Multihead CNN focuses on these highlighted features to extract intensity aided by the Nadam optimizer, which adapts to any further noise during the gradient estimates in the real time estimation. Likewise, for time series based forecasting, the bidirectional LSTM uses the temporal cyclonic features to predict for next successive time units. Additionally, for further focus on the Region of Interest, being the Indian Ocean as per the used dataset of (IR) satellite images from INSAT-3D Imagery and cyclone maximum sustained wind speed (MSW) details from IBTrACS, coordinate based cropping was applied for the cyclonic extraction. Therefore, the proposed approach establishes itself as a robust tropical cyclone intensity estimation technique, whilst providing superior performance in comparison with other potential deep learning methods like those of AlexNet, CycloneNet, ResNet, LeNet for real time and BiGRU, GRU, Stacked LSTM and such for time series.