A study and analysis of prediction of intensity and track of tropical cyclone using deep learning techniques on satellite imagery

Abstract

The capability of forecasting the intensity as well as the trajectory of newlinea Tropical Cyclone (TC) seems to be a crucial task without satellite imagery. newlineWith contemporary progress in satellite imaging, the conventional Dvorak newlinetechnique is not sufficient and accurate for the prediction of cyclone intensity newlinetrack. Hence, it is a challenging task to develop a TC Intensity prediction model newlinewith high accuracy and reduced intensity forecast error that provides actionable newlineinformation to mitigate the disasters caused by cyclones and their associated newlinestorm surges reflecting in the prodigious amount of economic losses. Therefore, newlinethe proposed research work is focused on using variations of Recurrent Neural newlineNetwork integrated with several Metaheuristic Optimization algorithms on newlinedynamic cloud patterns of satellite imagery resulting in more reliable intensity newlineand track estimates of Tropical Cyclone with fast convergence rate and global newlineoptimum solution. newlineA single-dimensional Convolution neural network Autoencoder along newlinewith the Independent Recurrent Neural Network based student psychology newlineoptimization algorithm (single-dimensional CAE-IRNN based SPOA) is utilized newlinefor the prediction of a Tropical Cyclone track with the image to Intensity newlineRegression dataset (TCIR dataset) resulting in reduced error rate and more newlineaccuracy. TCIR dataset with passive microwave channel, water vapour channel, newlinevisible channel and infrared channel provides TC information such as TC center newlinelocation, minimum sea-level pressure, the maximum sustained wind in knots, the newlinemean of radii of 35-knot wind in the four quadrants in the nautical mile. Radial newlinefeatures and angular features are used to find the image similarities in the newlineestimation of tropical cyclone tracks. The dataset features climatological newlinepredictors, oceanic predictors, environmental predictors, time-dependent newlinepredictors etc. newline