Novel Deep Learning Architecture for Modeling Crash Severity Using Satellite Imagery Considering Land Use and Geometry

Crash severity prediction has been significantly enhanced by deep learning (DL) techniques, which overcome the limitations of traditional statistical and machine learning models that rely on manually engineered features and fixed relationships. These conventional approaches often fail to account for the complex spatial and structural factors influencing crash severity. In contrast, convolutional neural networks (CNNs) can automatically extract hierarchical features from high-dimensional data, such as satellite imagery, allowing for more scalable and accurate analysis. However, most existing deep-learning-based models use a single-input architecture, which restricts the integration of multiple interacting factors. To address this limitation, this study introduces ArroNet42, a novel dual-branch CNN designed to classify crash severity into four categories—no crash, simple collision, injury crash, and fatal crash—by concurrently processing land use and road geometry imagery. The model demonstrates superior performance, achieving a validation accuracy of 80.8%, outperforming established architectures such as GoogleNet (73.3%), ResNet-50 (72.9%), EfficientNet-B0 (72.8%), and DarkNet-53 (72.2%), which were limited to single-input land use imagery. An extended model, ArroNet43, which integrates roadway curvature as an auxiliary numerical input, achieved a slightly lower validation accuracy (79.3%), suggesting that curvature-related spatial features are already implicitly learned from the imagery inputs. These results underscore the advantage of multi-branch DL architectures in capturing complex crash-related factors. ArroNet42 provides a robust framework for proactive safety analysis and supports data-driven interventions by reducing dependence on historical crash records.