Prediction of Extreme Weather Event Devastation

Team Windows is comprised of Kate Brown, Cassie Chou, Grace Brooks, and Dhruthi Mandavilli.

Research Question

Can we model extreme wind devastation (aggregate statistic including direct and indirect injury/deaths, damage to property, and damage to crops) based on factors like location, magnitude, time of year, etc.?

Motivation

We seek to protect vulnerable communities by predicting storm devastation to allocate more resources and save lives through wind event preparation. Ultimately, we hope to preemptively reduce devastation and speed up recovery after the event.

Goals

Data Source

Our data is sourced from the National Oceanic and Atmospheric Administration (NOAA) Storm Event database. We used storm event data from the years 2008, 2013, and 2018.

Main Takeaways and Future Directions

We successfully created an aggregate statistic that represents wind event devastation. In the end, our model with the best hyperparameters had an accuracy of 0.6898446 and a kappa value – which measures how much better our model is than random chance (on a scale from 0 to 1) – of 0.4593361. The most important predictor variables were wind magnitude and the state the event took place in. These results indicate that our statistic for wind categorization can be predicted from our predictors, and that future work can be done to improve the model. This may include using different machine learning models and testing our model on different wind datasets to improve and test accuracy. One might also try different strategies for handling missing data – we chose to list everything missing in the devastation categorization as zero, but perhaps there are better strategies for this.