Detecting Tropical Cyclones in INSAT-3D Satellite Images Using CNN-Based Model

Abstract

The devastation caused by a natural disaster like tropical cyclone (TC) is beyond comprehension. Livelihoods are damaged and take years on end to fix. An automated process to detect its presence goes a long way in mitigating a humanitarian crisis that is waiting to occur. With the advent of deep learning techniques, Convolutional Neural Network (CNN) has had significant success in solving image-related challenges. The current study proposes a CNN based deep network to classify the presence or absence of TC in satellite images. The model is trained and tested with multi-spectral images from INSAT-3D satellite obtained from Meteorological & Oceanographic Satellite Data Archival Centre (MOSDAC) of Indian Space Research Organization (ISRO), Government of India, and an average accuracy of 0.99 is obtained with the proposed architecture. The number of parameters trained are only 1.9 million, which is far less than earlier studies. The detection process described in this study can serve as the first step in better predicting TC tracks and intensity for disaster management and to minimize the impacts on human lives and economy of the country.