Rapid Detection of Tropical Cyclones from Climate Simulations

Climate simulations provide valuable information to represent the situations of the atmosphere, ocean and land. Increasingly advanced computational technologies and Earth observation capabilities have enabled the climate models to have higher spatial and temporal resolution, providing an ever realistic coverage of the Earth. The high spatiotemporal resolution also provides us the opportunity to more precisely pinpoint and identify/segment the occurrence of extreme weather events, such as tropical cyclones, which can have dramatic impacts on populations and economies. We propose to utilize deep learning techniques on the rapid detection of tropical cyclones. Deep learning models trained on past climate simulations will inform the effectiveness of the approach on future simulations. Deep learning methods performing real-time segmentation of relevant features for extreme weather events can generate such list or database storing these features, and detailed information can be obtained by rerunning the simulation with high spatiotemporal data when needed. The objectives of this research are to 1) identify atmospheric features and segment relevant features from climate simulation results, e.g. rain bands and hot towers associated with cyclones; and 2) build a deep learning model to train and test the atmospheric features for the automatic classification, identification, and relevant feature segmentation.