Urban road traffic state classification is essential for identifying early-stage
deterioration and enabling proactive traffic management. This study presents a
novel method to accurately assess the traffic state of urban roads while
addressing the limitations of existing methods in spatial generalization
performance. The approach consists of three key components. First, several
indicators are designed to capture the spatial-temporal evolution mechanisms of
traffic state, speed freedom, flow saturation, and their variations over time
and space. Then, a feature learning module based on an AutoEncoder network is
introduced to reduce the dimensionality of the constructed feature set. This
enhances feature distinction while mitigating noise effects on classification
results. Third, k-means clustering is applied to analyze significant features
extracted from the AutoEncoder latent space, categorizing road traffic states
into fluent, basic fluent, moderate congested and severe congested. Finally, a
road network in Xuancheng, a city in Anhui Province, China, is selected as the
test area. The results of road state categorization for both the entire network
and single roads are visualized and analyzed, demonstrating the interpretability
and practical utility of the approach. The proposed method is also compared with
classical k-means clustering, the threshold-based classification, and FCM. To
quantify performance, we introduce a traffic state fluctuation rate index,
defined as the ratio of state transitions between adjacent time windows. The
results show that during the daytime (06:00-20:00), the fluctuation index of the
proposed method increases by 13.1%, 22.7%, and 29.4% compared to the classical
k-means, threshold-based method, and FCM, respectively. Meanwhile, during the
nighttime (20:00-24:00 and 00:00-06:00), the fluctuation index decreases by
12.7%, 22.5%, and 9.0%, aligning more closely with the real changing patterns of
daytime and nighttime traffic conditions.