Spatio-Temporal Feature Fusion via Transformer-Based Architectures for Autonomous Classification of Cryospheric Road Surface Pathologies in Sub-Arctic Urban Environments

City conditions in urban sub-arctic areas, such as Saskatoon, Saskatchewan, Canada, areincredible impediments to autonomous vehicle perception, with cryospheric pathologies of theroad surface (including black ice, compacted snow, slush formation, and wet frost formation)exemplified in Saskatchewan, Canada. All these conditions lead to deficiency in the detectionsystems based on cameras and also contribute significantly to the statistics of traffic accidentsduring winter time. This paper introduces the Saskatoon Cryospheric Road Safety System(SCRSS), which is a new architecture of combining the object detection in YOLOv8 basedarchitecture and spatio-temporal transformer feature fusion in the framework of real-time roadsurface conditions classification. The system is trained on the DAWN (Detection in AdverseWeather Nature) benchmark dataset that is augmented with synthetically generated sub-arcticwinter imagery and deployed as the interactive Streamlit web application available athttps://saskatoon-cryospheric-road-ml-cj68z2ta6kwlseyyyyaytr.streamlit.app/. Experimentalanalyses show that the mean Average Precision (mAP@50) is 0.847 across five hazardcategories with inference latency per frame of 34 ms with standard CPU hardware. The proposedarchitecture incorporates a spatial attention unit and temporal context aggregator into theYOLOv8n backbone, and achieves an improvement of 12.3% in the accuracy of black icedetection when compared to baseline YOLOv8n backbone. The system also offers geospatialincident mapping through Folium and the generation of downloadable reports on safety to themunicipal traffic management. This work lays the foundation of a deployable, open-sourcebaseline of cryospheric road intelligence in high-latitude metropolitan regions