Tropical Cyclone Detection Using Optical Flow Estimation Deep Learning

Abstract

AbstractTropical cyclones (TC) are amongst the deadliest natural disasters that cause massive damage to property and lives. Meteorologists track these natural phenomena using satellite imagery. The spiral rain bands appear in a cyclic pattern with an eye as a centre in the satellite image. Automatic identification of the cyclic pattern is a challenging task due to the clouds present around the structure. Conventional approaches use only image data to detect the cyclic structure using deep learning algorithms. The training and testing data consist of positive and negative samples of TC. But the cyclic structure’s texture pattern makes it difficult for the deep learning algorithms to extract useful features. This paper presents an automatic TC detection algorithm using optical flow estimation and deep learning algorithms to overcome this drawback. The optical flow vectors are estimated using the Horn-Schunck estimator, the Liu-Shen estimator, and the Lagrange multiplier. The deep learning algorithms take the optical flow vectors as input during the training stage and extract the features to identify the cyclone’s circular pattern. The proposed method increases the accuracy of detecting the cyclone pattern through optical flow vectors compared to pixel intensity values.KeywordsTropical cycloneConventional approachesDeep learning algorithmHorn-Schunck estimatorLiu-Shen estimatorLagrange multiplier