Presenting Spatial Data Mining Based on a Stage-Interpretable Multilayer Deep Learning Approach to Analyze the Factors Affecting the Crash Frequency of Light Vehicles in Rural Freeways

Abstract

Some studies show that most light vehicle crashes occur on rural freeways. Crash frequency analysis presents significant challenges due to the complexity of large-scale datasets, hidden intervariable relationships, limited interpretability, and the influence of segmentation methods used. The present study aims to provide an innovative stage-interpretable multilayer perceptron (SI-MLP) method based on deep learning for quantitative and spatial analysis of the effect of geometric factors, weather conditions, and traffic parameters on the crash frequency by presenting a combined dynamic–static segmentation approach in rural freeways. The stage-interpretable method makes the interpretability of deep learning algorithm results possible by stepwise feature omission and assessing its impact on model accuracy. To this aim, significant variables were selected using a decision tree algorithm in conjunction with the Pearson correlation test. The Poisson regression model was utilized to compare the results with those of the proposed SI-MLP method due to the discrete nature of the data and the ability of such an algorithm to model the crash frequency. TabNet was also considered as a benchmark to assess the proposed SI-MLP, since it supports both crash frequency classification (enabling direct accuracy comparison with the MLP-based formulation) and regression-based crash frequency estimation as a numerical outcome. The results revealed that the proposed SI-MLP outperforms the Poisson regression model and the TabNet benchmark in identifying influential crash factors, owing to its ability to capture complex patterns and relationships and to extract high-level features from crash data enriched with geometric and weather information. In the three-class setting (low, medium, and high crash frequency levels), the SI-MLP achieved an overall accuracy of 0.81 and a Macro-F1 of 0.81. Based on the results, the vehicle type and color, license type, temperature, driver’s age, lighting conditions, and front collision with a fixed object were identified as variables affecting the crash frequency. In the next step, the effective variables identified by the proposed method were analyzed using thematic kernel density maps to model the clustering patterns and spatial autocorrelation of the crash frequency, with an emphasis on the most effective factors. Identifying key contributing factors within high-crash segments using the proposed approach allows for the formulation of targeted safety priorities and supports the deployment of proactive countermeasures in segments with the highest risk.