CycloCast – Predicting Cloud Formations in Cyclonic Conditions Using Indian Satellite Imagery

Abstract

CycloCast is a deep learning-based forecasting model which predicts the future development of cyclonic weather cloud patterns from the satellite imagery taken from ISRO INSAT Series of satellites. Using a ConvLSTM network, it predicts future developments of cloud patterns at every 30 minutes based on a single input image sequence. The potential of predicting cloud coverage patterns makes CycloCast a valuable tool for meteorologists and disaster management authorities. In this paper we detail the data collection methods, data processing approaches such as adaptive thresholding and cyclone-based cropping, and the training approaches to achieve stable forecasting performance. CycloCast achieves a very high level of performance in forecasting cloud coverage patterns with a Cloud Coverage % of 85.6 and a Structural Similarity Index (SSIM) of 0.7765.