Monday, 7 January 2019: 2:30 PM
North 125AB (Phoenix Convention Center - West and North Buildings)
Atmospheric data used for data-driven prediction is increasing in size and variety, posing a challenge and an opportunity to create informative models that can be used to predict extreme weather events. Predicting such events pertains to inferring events such as hurricanes, cyclones, weather fronts, and blocking events from a combination of different atmospheric variables and their associated spatio-temporal patterns. Some of the challenges associated with predicting extreme weather events is the small frequency of the events, difficulty in mining spatio-temporal patterns in atmospheric variables to form a regression model and uncertainty pertaining to atmospheric predictors. Defining an extreme event can be posed as a classification problem within a CNN framework, however this formulation does not address uncertainty in the predictors. In this work, a deep Bayesian convolutional neural network is presented to predict extreme weather events. Proposed work addresses the uncertainty in atmospheric data sources and incorporates prior knowledge pertaining to the system being modeled. Conceptual ideas behind Bayesian deep learning related to atmospheric sciences is elaborated and a case study is built using the "ExtremeWeather" dataset as training. The importance of addressing uncertainty in extreme event prediction is demonstrated qualitatively by comparing Bayesian deep learning predictions to a commonly-used, deterministic deep learning methodology.
![](images/flag.png)
![](images/blue_ribbon.png)