Ahmad Elallaf1, Nathan Jacobs2, Xinyue Ye3, Mei Chen4, Gongbo Liang1
1Texas A&M University-San Antonio, 2Washington University in St. Louis, 3University of Alabama, 4University of Kentucky
AAAI Conference on Artificial Intelligence, 2026
Figure 1: A case study for the San Antonio River Walk. Our model (right) successfully identifies more crash sites with elevated risk scores (yellow, orange) and produces a more spatially coherent map compared to the baseline model (middle).
Roadway traffic accidents represent a global health crisis, yet conventional AI-based risk estimators typically generate point estimates without conveying model uncertainty, limiting their utility in critical decision-making. To address this, we introduce a novel geospatial deep learning framework, BetaRisk, that reframes risk estimation as a probabilistic learning problem. Rather than producing a single deterministic output, our model estimates a full Beta probability distribution over fatal crash risk from satellite imagery, yielding accurate and uncertainty-aware predictions.
A key innovation of our framework is the procedural generation of supervisory signals. Instead of using static labels, we dynamically create a target Beta distribution for each training sample based on the properties of a random crop augmentation. For positive samples (crash locations), the target distribution reflects the quality of the visual evidence in the crop, quantified by an "influence score" based on its size and centrality. This technique transforms data augmentation from a simple regularizer into a rich source of continuous supervision for learning both risk and uncertainty. Our model significantly outperforms baselines, achieving a 17-23% improvement in recall while delivering superior calibration.
Figure 2: The training architecture. Multi-scale images are passed through a shared feature extractor. The concatenated features are fed into two parallel heads: a Distribution Learning Module and a Classification Module.
A key advantage of BetaRisk is its ability to provide richer, more interpretable predictions. The figures and table below demonstrate how the model visualizes its own uncertainty and resolves the ambiguity inherent in standard, deterministic models.
Figure 3: Qualitative prediction examples. The model demonstrates well-calibrated confidence, correctly assigning high risk with high confidence to a complex coastal road (far left) and low risk with high confidence to a simple suburban street (second from right), while expressing appropriate uncertainty for more challenging scenes.
Figure 4: Our model's ability to quantify uncertainty. The plots show that uncertainty is highest for ambiguous predictions (left) and that every prediction is accompanied by a 95% confidence interval (center), providing a level of interpretability baseline models cannot match.
The table below expands on the concept shown in Figure 4, demonstrating how a single, ambiguous score from a baseline model can correspond to multiple, distinct scenarios in BetaRisk. For any given risk score, our model reveals the underlying confidence level, providing crucial context that is unavailable in standard deterministic models.
To train and evaluate our model, we use the MTSL-RoadRisk dataset, a large-scale collection of satellite imagery from Texas, USA, with over 16,000 locations labeled with historical fatal crashes. The dataset is publicly available for research purposes.
If you find this work useful in your research, please consider citing our paper:
