A Comparative Analysis of Deep Neural Networks (DNN) for Recognition of Tropical Cyclones in Real-Time Satellite Images over the Indian Sub-Continent

Supervised learning is typically required to train a Deep Neural Network (DNN) to identify satellite cyclone images with noise and blur in the visible and infrared spectrum. This requires input-target pairs of noisy images and corresponding blurry photos. In this research, we propose a self-supervised learning method to train a Deep Neural Network (DNN) employing only real-time images from the visible and infrared spectrums. The suggested technique, which serves as a self-supervision tool, can identify convective activity, the eye of the storm, and wall clouds in the tropical cyclone cloud distribution. Our approach involves two stages: Offline pre-training on Cyclonic Storm (CS) images over the Indian sub-continent, North Indian Ocean, Arabian Sea, and Bay of Bengal was followed by real-time testing of the localization on INSAT-3D satellite images. This allows for efficient testing of the model. Satellite cyclone images of recent tropical cyclones from 2018 to 2023 are used to assess the algorithm's efficacy thoroughly. An analysis of performance metrics is attempted with graphical plots and a precision and recall matrix. Furthermore, according to the experimental results, our suggested algorithm outperforms the state-of-the-art models in terms of both classification accuracy and localization learning models' test-time performance.