Research Projects

As we are aware that many intense storms such as cyclone, typhoon and hurricane occur every year around the world, which results in high numbers of casualties and serious threat to property safety and forest vegetation. Storms are non-rigid in nature and known as fluid objects. According to WHO, storms, including heavy rain have affected the lives of more than 726 million people worldwide between1998 and 2017. Therefore, it becomes particularly important to predict the near future weather conditions as soon as possible to prevent the fatalities. Previous approaches such as thunderstorm identification, tracking and analysis, and tracking radar echo by correlation, used satellite images are either based on linear extrapolation or utilized storms optical flow as a tool for all sky-conditions. Thus, no deep learning and physics-based model is given for the motion of fluids in these techniques. Hence, these methods fail to handle the high degree of storm motion, robustness against outliers and discontinuities in the motion field. In the proposed project, a level set segmentation based fractional order Tikhonov regularization variational model is presented for the motion estimation of storm using the continuum equation of fluids. The storm motion or velocity is estimated in terms of optical flow from a sequence of satellite images taken at some time intervals. Optical flow helps to localize the high alert region. The Tikhonov regularization model is formulated with the help of non-quadratic Charbonnier norm and Marchaud fractional derivative. This non-quadratic penalty provides an effective robustness against outliers, whereas the fractional derivative possesses a non-local character, and therefore is capable to deal with the discontinuities of dynamic textures and edges. Moreover, level set segmentation is based on active contours, which helps in detecting the topological changes taking place on the boundaries. Now, these nicely segmented high alert regions are also assisted with the probability score of storms. Further, 4D feature vectors of the segmented optical flow fields are obtained for the classification based on the properties of fluids such as transport energy, flow magnitude, directional variance and matching ratio. Finally, the storms detection is carried out by implementing a mixed data deep learning models. The mixed data employed in the proposal is composed of satellite images and the corresponding 4D feature vector sequence. Thus, the proposed algorithm used the dynamic as well as static features of storms. One branch of the algorithm works with static features of the images, while other branch will be applied on 4D dynamic feature vector. At last, prediction of near future weather conditions in satellite images is performed based on the advection and anisotropic diffusion equation of fluid dynamics. Satellite datasets comprising of different weather conditions and regions of worldwide will be collected for broad analysis of the model.