Detection and Prediction of Natural Hazards Using Large-Scale Environmental Data

Recent developments in remote sensing have made it possible to instrument and sense the physical world with high resolution and fidelity. Consequently, very large spatio-temporal environmental data sets, have become available to the research community. Such data consists of time-series, starting as early as 1973, monitoring up to thousands of environmental parameters, for each spatial region of a resolution as low as \(0.5'\times 0.5'\). To make this flood of data actionable, in this work, we employ a data driven approach to detect and predict natural hazards. Our supervised learning approach learns from labeled historic events. We describe each event by a three-mode tensor, covering space, time and environmental parameters. Due to the very large number of environmental parameters, and the possibility of latent features hidden within these parameters, we employ a tensor factorization approach to learn latent factors. As the corresponding tensors can grow very large, we propose to employ an outlier-score for sparsification, thus explicitly modeling interesting (location, time, parameter) triples only. In our experimental evaluation, we apply our data-driven learning approach to the use-case of predicting the rapid-intensification of tropical storms. Learning from past tropical storms, we show that our approach is able to predict the future rapid-intesification of tropical storms with high accuracy, matching the accuracy of domain specific solutions, yet without using any domain knowledge.