SSN: A Novel CNN-Based Architecture for Classification of Tropical Cyclone Images From INSAT-3D

Tropical cyclone (TC) is a natural phenomenon that adversely affects human livelihoods. Analyzing TC intensity remains challenging despite recent advances in traditional numerical weather prediction models, which offer weather forecasts with reasonable accuracy. In this study, a novel, lightweight convolution neural network (CNN)-based architecture, inspired from residual networks, called Small Skip Net (SSN), is proposed to classify the satellite images of TC into six predefined classes. The SSN model is trained and tested on a dataset prepared from Indian National Satellite (INSAT-3D) images of 68 cyclonic disturbances during 2013–2023 over Bay of Bengal (BoB) obtained from Meteorological and Oceanographic Satellite Data Archival Centre (MOSDAC) of Indian Space Research Organization (ISRO), Government of India. An overall classification accuracy of 92.35% is obtained on the test set, comprising TC that occurred in 2023, with the proposed architecture. The number of parameters to train is ~5 million, which is far less than comparable state-of-the-art works. This makes SSN a lightweight and energy-efficient model, suitable for obtaining quick classification results required for achieving TC intensity analysis capabilities.