Computational Models for Detection and Analysis of Synoptic-Scale Ice Storm Patterns

Ice storms are responsible for billions of dollars of loss in the continental United States and Canada. Despite their potential for damaging impacts, our ability to accurately predict ice storms and their behaviour, and mitigate their impacts, is challenged by the facts that: (a) the atmospheric physical processes involved in causing ice storms are complex and difficult to model or identify; (b) very high resolution temporal and spatial data is required in run time for good forecast accuracy; (c) ice storm intensity information is typically expressed in terms of physical damage to life, property and natural resources, thus being more a function of population density than actual storm intensity; and (d) the prevalence and/or intensity of future ice storms could be altered as a result of the influence of human-induced climate change on atmospheric circulation.

We address these challenges by transforming detection and analysis of ice storms into a pattern recognition problem where we identify complex multidimensional atmospheric patterns at synoptic scales using machine learning algorithms. We identify surface and upper atmospheric data for relevant atmospheric variables such as temperature, geopotential height and specific humidity at synoptic scales and frame ice storm identification as a binary classification problem to distinguish between ice storm and non-ice storm conditions in winter months (October-April). The complex geo-spatial data patterns were well suited for a Support Vector Machine (SVM) based classifier using an RBF kernel with which we were able to achieve close to 80% accuracy in identifying patterns in independent reanalysis data. We then apply the learned models to experimental high-resolution simulations by the HIRAM GFDL model under both historical and future climate scenarios. Although the learned algorithm is able to identify ice storm events in the GCM simulations, there are essential differences along certain variables or dimensions between the reanalysis and GCM simulations necessitating bias corrective measures.

To characterize ice storm intensity, we further experiment with grouping or clustering storm patterns and show that the storms can be further categorized based on physical properties seen as features or feature interactions that can be used in conjunction with past storm data for assigning intensity levels. Results from clustering experiments show that ice storm patterns can indeed be grouped into more than one category. Our next step in this direction is to identify signatures for such clusters and map them to the corresponding physical properties of the storms as understood by atmospheric scientists.

Our experiments so far have been restricted to the continental north eastern United States, but our methods can be readily applied globally to other regions, with the goal of continuing to develop our ability to apply machine learning and computational intelligence to identifying and understanding ice storm conditions and their impacts on human systems and the natural environment.