Cyclone Frame Prediction by Gaussian Mixture Modeling of the Three Penultimate Time-Lapse Frames

A fast cyclone frame prediction is proposed in this paper that fits a Gaussian Mixture model on the spatio-temporal data extracted from the three penultimate time-lapse frames, prior to fuzzy regression. Unlike the previous work in Verma and Pal (In: Applied Imagery Pattern Recognition Workshop (AIPR), pp. 1–8, 2010) that models the entire history on a per pixel basis, a single Gaussian mixture is used for fitting the spatio-temporal data within the time-span of the last three frames, making the process faster and more accurate. The increase in accuracy is attributed to the fact that cyclones evolve over time and thus the recent frames give a more meaningful insight into the predictions for the next frame. The number of components in the Gaussian mixture is determined from the occurrence of equally likely modes that correspond to high entropy peaks. Our results on satellite videos of recent cyclones that hit the Indian seas show a high accuracy of frame prediction.