Abstract

This project details the development and implementation of a near real-time geospatial data pipeline of fatal and serious traffic accidents across the State of Minnesota. To accomplish this, data is scraped and preprocessed via a serverless, automated pipeline on the cloud, before various spatial and temporal analytics are run on the resulting data to provide insights into crash hotspots and clusters across both space and time. The novelty of this pipeline lies in the creation of an open, near-real-time, geospatial web application to view serious traffic crashes in Minnesota, as well as investigations made into the integration of generative artificial intelligence (AI) into the geocoding process and the use of simple spatial analyses to study the stability of point clustering footprints over time. This project shows several insights into the historical patterns of car crashes across the State, specifically concerning: 1) the need for advances in transportation engineering and increased enforcement and emergency services in areas of high concern; and 2) the association between the presence of road construction projects and changes in accident trends.

Methods

The focus of this project was to conduct an analysis of competing methods/algorithms, to understand if the different methods told consistent stories about the underlying dynamics of crashes across the State. First, the Local Indicators of Spatial Association (LISA) method was used to identify spatial clusters and outliers of accidents, aggregated by the various Cities, Townships, and Unorganized Territories that make up Minnesota (Fig. 2). Second, the Approximate DBSCAN (A-DBSCAN) method was used to identify robust spatial clusters directly from the point patterns (Fig. 3). Lastly, to understand the temporal robustness and stability of the clusters created by A-DBSCAN, the algorithm was run on yearly subsets of the data, before the resulting cluster footprints for each run were overlaid to understand which areas are consistently included within clusters across different years (Fig. 4). Lastly, ad-hoc analyses were conducted on the global time series, as well as the individual time series of different A-DBSCAN clusters to understand trends in the various signals (Fig. 5). Namely, offline change point detection algorithms were used to identify discontinuity within the time series signals.

Results

The various methods used all share different perspectives of the underlying data, but all tell a consistent story of where accidents are occurring. There are around 7 to 15 different accident corridors around the State, based on the outcomes of the analyses conducted in this project. In addition to the methods all sharing a similar story to one another, qualitative evidence in the form of news articles also seems to back up these stories as well. Lastly, based on the qualitative evidence from various news articles as well as time series analysis, the results suggest that road construction may be a driver in accident frequency, however, more thorough work is needed to investigate these findings.

Next Steps

With a large amount of data and analytical outputs created through the data pipeline, an API and web application were created to enable both technical and non-technical users to use and interact with the data for various purposes. One future goal is to expand the capabilities of these applications. In addition to this, with an easy-to-use data source established, the ability to make use of different analytical methods is more easily attainable. Opportunities such as network-based analysis, spatiotemporal prediction, and agent-based modeling all could be potential avenues for further exploration. 

Project Advising

Dr. Bryan Runck, Associate Director, GeoCommons

Award

This project was awarded first place at the graduate student competition held by the Minnesota GIS/LIS Consortium at their 2023 annual conference.