Evaluation of ConvLSTM-Based Road Surface Temperature Prediction Model Performance using Mobile Spatio-Temporal Data

With increasing traffic safety risks from winter road icing, predicting road surface temperature (RST) is vital for analyzing icing mechanisms and ensuring road safety. This study built RST prediction models using mobile observational data and compared machine learning methods (Random Forest, XGBoost, and Light GBM) with a deep learning model, the Convolutional Long Short-Term Memory (ConvLSTM) network. Mobile data were collected from Automatic Weather Station (AWS) and non-contact sensors along a 330-km section of the Seohaean Expressway during November-December in 2024. ConvLSTM achieved the lowest RMSE and MAE across all horizons, showing high short-term accuracy (≤ 600 s). In long-term forecasts (≥ 1,800 s), errors increased slightly but overall performance remained superior. Physical and geographical factors, including Solar Zenith Angle (SZA), Sky View Factor (SVF), and bridge/tunnel indicators, improved short-term accuracy but had limited long-term effects. These results highlight ConvLSTM’s potential for road icing prediction and traffic safety management.