In recent years, the aging of snow-removal personnel and increasingly severe labor shortages have made it critical to establish efficient winter road management methods that utilize automation and labor-saving technologies. These methods are required to rapidly identify the risks leading to traffic disruptions arising from the deterioration of road surface conditions. In this study, we integrate various time-series data, including meteorological data, traffic volume, and snow removal records, to construct a prediction model for winter road surface conditions six hours ahead. Instead of relying on image recognition, which is prone to misclassification during nighttime and snowfall, we employ a deep learning model combining Long Short-Term Memory (LSTM) networks and Multi-Head Attention as methods that are capable of stable prediction under such conditions. They can help estimate road surface conditions (dry/wet, slush, and compacted snow) six hours ahead. Using approximately one month of observational data collected at a mountain pass in Nagaoka City, Niigata Prefecture, Japan, we trained and validated the model, achieving an accuracy of approximately 72 % on the test data with a particularly high recall for compacted-snow conditions. Feature-importance analysis indicates that snowfall amount, traffic volume, air temperature, and elapsed time since snow removal all contribute to improved prediction accuracy, whereas ablation studies more clearly demonstrate the necessity of including the elapsed time since snow removal. The proposed method can reduce the labor burden compared with conventional visual monitoring and enable planned snow-removal operations that proactively anticipate deterioration in road surface conditions, thereby contributing to more efficient winter road management.

This study aims to clarify the impact of global warming on the annual maximum snow depth in the Shinetsu region. To achieve this, statistical analysis was performed on the following data: (i) annual maximum snow depth, (ii) winter mean air temperature, and (iii) winter precipitation, over a 100-year period from 1926 to 2025. The data were collected from seven meteorological observation stations in the Nagano and Niigata prefectures. The results showed that the annual maximum snow depth had significantly decreased at a significance level of ≤1 % in Aikawa and Takada, ≤5 % in Matsumoto and Iida, and ≤10 % in Niigata. The winter mean air temperature exhibited an increase at a ≤1 % significance level at all observation stations. The winter precipitation exhibited a decrease at a significance level of ≤5 % in Takada and ≤10 % in Nagano. There were statistically significant negative correlations found between the annual maximum snow depth and the winter mean air temperature at a ≤5 % significance level at all observation stations. Additionally, statistically significant positive correlations were found between the annual maximum snow depth and winter precipitation at a ≤1 % significance level at all observation stations except Iida. Based on these results, it was concluded that global warming had increased the winter mean air temperature, making it likely that winter precipitation would fall as rain rather than snow. Consequently, this would decrease the annual maximum snow depth, particularly in relatively warm, low-altitude regions. At Takada, a decrease in winter precipitation may also have contributed to the decrease in annual maximum snow depth.