Exploring weather and traffic conditions in traffic accidents using one-class learning

Ensuring road safety is a critical responsibility for public organizations such as road network operators, emergency services, and national meteorological services (NMS). Traffic accidents arise from a complex interplay of environmental and human factors, making proactive risk management essential for road network operations. In practice, emergency services and road operators predominantly collect high-precision records of accident locations, resulting in presence-only datasets that lack explicit non-accident observations. 

Unlike traditional accident modeling approaches that rely on labeled non-accident data or synthetically constructed negative classes, this study investigates one-class learning as a natural and operationally realistic framework for traffic accident analysis. Researchers at the Royal Netherlands Meteorological Institute (KNMI) explore the use of AI/ML methods to model high-resolution presence-only accident data using five years of traffic accident locations (2018–2022) provided by the Dutch road authority. Each accident is characterized by a set of weather and traffic intensity features describing the conditions under which it occurred. 

Traffic accidents are modeled using neural one-class classification to obtain a high-dimensional embedding of accident conditions, which is subsequently analyzed using dimensionality reduction techniques to identify clusters of accidents with similar environmental signatures. By learning directly from observed accident occurrences, the approach enables the identification and comparison of recurring accident patterns associated with specific weather and traffic conditions, providing a structured basis for further analysis of weather-related traffic risk.