Toward explainable automatic incident detection: a spatiotemporal self-supervised framework with multi-probing classifiers
Xuehao Zhai, Fangce Guo, Aruna Sivakumar · Journal of Intelligent Transportation Systems · 2025
Automatic Incident Detection (AID) in the case of no pre-labeled anomalies is a challenging problem in Intelligent Transport Systems. A considerable number of emerging self-supervised machine learning models have been deployed to detect traffic anomalies in recent years. However, it is difficult to verify their performance since ground truth labels are inaccessible in many real-world applications. Another limitation is that these state-of-the-art models have low explainability due to their complex structure and nontransparent working process. In this paper, we propose an explainable AID framework that integrates spatiotemporal deep learning with a proxy-based probing classifier to address these limitations. Given that the common traffic patterns being learned are expected to vary across regions and be dynamic temporally, we propose a sort of model with different combinations of spatiotemporal deep learning structures to generate the common patterns and identify anomalies. We then introduce two types of probing classifiers—land-use-related and weather-related—to assess how effectively the model captures recurrent and non-recurrent information in its reconstruction layer and output error layer. Among the test models, the combination of Graph Attention Network (GAT) and Long Short-Term Memory (LSTM) performs the best ability to capture the recurrent anomaly and non-recurrent anomaly information. Furthermore, our analysis confirms that adding explicit spatial modules helps the system learn regional heterogeneity in its reconstruction layer, while random disruptions from external factors are naturally highlighted in the reconstruction errors. By focusing on proxy-based explainability, our framework provides a transparent and scalable solution for AID in scenarios where labeled data is unavailable.