Beta Distribution Learning for Reliable Roadway Crash Risk Assessment

TL;DR

This paper addresses the challenge of estimating roadway crash risk with meaningful uncertainty, proposing a probabilistic framework that maps multi-scale satellite imagery to a Beta distribution per location. By modeling $P\sim\mathrm{Beta}(\alpha,\beta)$ and using $R=\mathbb{E}[P]=\frac{\alpha}{\alpha+\beta}$ as the risk score, the method delivers uncertainty-aware predictions and improved calibration. A joint training objective combining a Wasserstein-2 surrogate for distribution alignment and a binary classifier loss, along with a novel target Beta labeling mechanism, enables reliable risk estimation at both highways and local roads. Empirical results on a Texas MSCM dataset show notable recall gains (up to ~17-23% relative) and better interpretability through per-prediction Beta distributions, demonstrating practical benefits for safer routing, autonomous navigation, and urban safety planning while highlighting ethical deployment considerations and limitations like static risk framing and regional generalizability.

Abstract

Roadway traffic accidents represent a global health crisis, responsible for over a million deaths annually and costing many countries up to 3% of their GDP. Traditional traffic safety studies often examine risk factors in isolation, overlooking the spatial complexity and contextual interactions inherent in the built environment. Furthermore, conventional Neural Network-based risk estimators typically generate point estimates without conveying model uncertainty, limiting their utility in critical decision-making. To address these shortcomings, we introduce a novel geospatial deep learning framework that leverages satellite imagery as a comprehensive spatial input. This approach enables the model to capture the nuanced spatial patterns and embedded environmental risk factors that contribute to fatal crash risks. Rather than producing a single deterministic output, our model estimates a full Beta probability distribution over fatal crash risk, yielding accurate and uncertainty-aware predictions--a critical feature for trustworthy AI in safety-critical applications. Our model outperforms baselines by achieving a 17-23% improvement in recall, a key metric for flagging potential dangers, while delivering superior calibration. By providing reliable and interpretable risk assessments from satellite imagery alone, our method enables safer autonomous navigation and offers a highly scalable tool for urban planners and policymakers to enhance roadway safety equitably and cost-effectively.

Figures (11)

• Figure 1: Training Architecture with Joint Optimization
• Figure 2: Streamlined Architecture for Inference
• Figure 3: Uncertainty Quantification and Interpretability
• Figure 4: Analysis of Predicted Probability Distributions
• Figure 5: Qualitative Results for Unambiguous ("Easy") Cases (a: True Negatives, b: True Positives)