Sota TOZAWA

Construction of a Road Surface Condition Estimation Model for Winter Expressways

Kazushi SANO

In recent years, short-term heavy snowfall events have increased mainly along the Sea of Japan side of Japan, causing large-scale traffic congestion on expressways triggered by stalled heavy vehicles. In winter traffic management, wide-area monitoring of road surface conditions is essential for improving road closure decisions and snow removal planning. However, visual judgment using camera images and on-site observations imposes a heavy burden on personnel, making continuous monitoring difficult.
This study develops a framework for estimating road surface conditions using operational meteorological, traffic, and snow removal data. Road surface labels obtained from fixed-point camera observations on the Kan-Etsu Expressway were used as training data, and two models were evaluated: a machine learning model (LightGBM) and a state transition model considering Markov properties.
The analysis targeted four locations—Yuzawa, Shiozawa-Ishiuchi, Muikamachi, and Ojiya. Road surface data labeled at 30-minute intervals from December 2023 to March 2024 were used. Five road surface categories were defined: no snow, wet snow, frozen surface, dry snow (1–3 cm), and dry snow (over 3 cm). Explanatory variables included snowfall, cumulative snowfall, air temperature, traffic volume, average speed, heavy vehicle ratio, occupancy rate, and elapsed time since the last snowplow pass.
First, a multiclass classification model using LightGBM was developed and evaluated through K-fold cross-validation, achieving an overall accuracy of approximately 77%. Next, a Markov-based road surface estimation model was constructed to represent the persistence of road surface states by expressing transition probabilities as functions of explanatory variables. The regularized Markov model showed higher accuracy than a comparison logit model, confirming its effectiveness.
Furthermore, road surface categories were extended by considering wheel ruts, distinguishing wet and dry snow with ruts as black slush and white slush. The results indicate that the proposed approach can capture operationally important deterioration phases. Future work includes real-time implementation and improving estimation performance for rare and ambiguous road surface conditions.