Predicting property damage from tornadoes with zero-inflated neural networks

Authors: Jeremy Diaz and Maxwell Joseph, Ph.D
Affiliation: Earth Lab, Cooperative Institute of Research for Environmental Sciences, University of Colorado Boulder
Preprint: arXiv - stat.ML
Research data: figshare (this workflow pulls all the data, however Storm Events updates periodically)

Sample Figures

Summary of the Project

This project aims to predict dollar-amount, tornado-induced property damage using the following types of variables:

• Tornado-describing
• Land cover of area struck
• Census estimates of the area struck (socioeconomic and demographic)

Multivariable linear regression, zero-inflated log-normal, and various neural networks (some mimicking a zero-inflated log-normal) are all models utilized in this project, which uses a combination of python, R, and bash/shell.

Summary of the Repository

This repository serves as the workflow-documentation to the paper "Predicting property damage from tornadoes with zero-inflated neural networks" by Jeremy Diaz and Maxwell Joseph. R scripts, R Markdown files, and Jupyter notebooks are spread among 3 main directories:

• Revisions
• Complete Workflow
• Old Code

Revisions contains two subdirectories (1) Explorations and (2) Complete Workflow. Revisions/Explorations contains prototype code from the revision stage of our paper, while Revisions/Complete_Workflow contains the final form of this project, which varies drastically from Complete_Workflow, the first draft form of this project. Lastly, Old Code contains early prototype code from the first draft.

Both Complete_Workflow directories are organized such that the numbered files should be ran in order to fully reproduce the entire project (data gathering, data integration, preprocessing, model fitting, and evaluation).

Interactive Dashboard

Click here for some interactive maps produced using the best model