Traffic safety is rarely shaped by a single factor. Road layout, intersections, traffic conditions, neighborhood design, and community characteristics can interact in ways that are difficult to capture with conventional analysis.

A 2026 study by Jiaqing Lu, Ziqi Li, Lei Han, and Qianwen Guo examines this challenge using a geospatial explainable artificial intelligence approach known as GeoXAI. The research looks at crash density across Florida and asks a practical question: why can the same type of road or neighborhood characteristic have different relationships with crashes depending on where it occurs?

Why Location Matters

Traffic crash patterns are not evenly distributed across a large geographic area. Urban neighborhoods can have many roads, intersections, and travel activities packed into relatively small spaces, while rural areas often have a very different physical environment.

The study focuses on Florida census tracts, allowing the researchers to examine crash patterns at a relatively local level rather than treating the entire state as one uniform area. The analysis covered crashes recorded between January 2023 and December 2024 and ultimately included 4,699 census tracts. Crash density was calculated according to the number of crashes occurring within each tract relative to its land area.

This local approach is important because an overall statewide relationship can hide differences between neighborhoods. A roadway characteristic associated with higher crash density in one setting may have a different relationship in another.

What GeoXAI Adds

The research combines machine learning with a method called GeoShapley. In simple terms, the framework is designed to explain not only which factors are associated with predicted crash density, but also how those relationships change geographically.

The researchers first used an automated machine-learning process to compare several predictive models. XGBoost was selected as the strongest-performing model within that process. GeoShapley was then used to interpret the model's predictions while considering both location and individual factors.

This differs from conventional approaches that may assume relationships remain broadly consistent across an entire study area. The researchers also compared GeoXAI results with SHAP, a widely used explainability method, and Multiscale Geographically Weighted Regression, or MGWR, which is designed to examine spatially varying relationships.

Road Density Is Not the Whole Story

Several roadway and neighborhood characteristics showed complex relationships with crash density. Among the factors highlighted in the research were road density, intersection density, neighborhood compactness, and educational attainment. Their relationships with crash density were not simply straight lines.

For example, the study found that very compact urban environments could experience a sharp increase in predicted crash density after certain levels of neighborhood compactness were reached. This suggests that adding more roadway connections or concentrating activity does not necessarily produce a predictable change in crash patterns at every location.

Different Areas Show Different Patterns

The geographic results also showed substantial differences across Florida. Highly urbanized locations, including Miami, displayed much higher intrinsic crash contributions than many rural areas. Orlando, Tampa, and Jacksonville were also identified as metropolitan areas with elevated baseline contributions in the model.

The researchers emphasize that these differences should not be interpreted as meaning that one factor automatically causes crashes everywhere. Instead, the findings demonstrate that the relationship between roadway, socioeconomic, and built-environment characteristics can depend strongly on local circumstances.

The study also reported a nonlinear relationship involving educational attainment. At lower-to-moderate levels, a higher percentage of residents with a bachelor's degree was associated with lower predicted crash contributions, while the relationship changed direction at higher levels. The reported threshold was approximately 35%, illustrating why assuming a simple linear relationship can overlook important patterns.

How the Findings Could Help

The value of this approach extends beyond producing another prediction model. By identifying where particular factors have stronger relationships with crash density, transportation planners can potentially examine safety conditions at a more targeted geographic scale.

The researchers discuss strategies such as traffic-calming measures in compact neighborhoods, changes to intersection design, and speed-management approaches on busy corridors. The framework could also help identify areas where additional investigation is warranted rather than applying identical safety measures across every neighborhood.

The study presents GeoXAI as a way to make traffic-safety analysis more geographically detailed. By bringing machine learning, spatial information, and model interpretation together, GeoXAI offers a more detailed way to examine those differences and identify areas that may deserve closer safety attention.

The study presents GeoXAI as a way to make traffic-safety analysis more geographically detailed. By bringing machine learning, spatial information, and model interpretation together, GeoXAI offers a more detailed way to examine those differences and identify areas that may deserve closer safety attention.