Estimation of the rapid intensification of tropical cyclones over the North Indian Ocean using attention-based deep learning models

The estimation of rapid intensifications (RI) of cyclones is crucial in tropical cyclone (TC) forecasting. This challenge is deepend over the north Indian Ocean (NIO) basins due to the non-availability of real-time aircraft atmospheric observations and scarce ocean observations with highly vulnerable dense coastal populations. Deep learning (DL) encompasses the potential to enhance RI estimation even more if deep neural networks can effectively incorporate satellite images of TC convection. In this study, we used the novel Mach Band Attention Model (MBAM) that uses attention weights, which are calculated as the percentage of response normalization-induced Mach band overshoot or undershoot, to either increase or decrease the prominence of feature locations inside the convolutional feature space. The dataset comprises 6115 infrared images from the INSAT-3D satellite over the NIO, of which 5456 were used for training the model and 659 for testing purposes. This is the first time the model has been tested on RI cyclones (i.e., Bulbul, Maha, and Nilofar) apart from random testing as in previous studies. Results demonstrate that the MBAM showed the least (2.87 kts) root mean square error (RMSE) for all testing cyclones. Furthermore, overall, the model captures the RI phase of the TC efficiently with an RMSE of 2.63 kts. Additionally, the MBAM proficiently estimates the peak intensity for very severe cyclonic storms (VSCS) and extremely severe cyclonic storms (ESCS) during the rapid intensification (RI) phase of tropical cyclones (TCs) with respective accuracies of 76.27% and 100%.