Two-Stage Prediction of Snowplow Dozer Operation Counts from GPS Data: A Case Study of Akita City, Japan

For effective winter road management in snow-prone regions, timely snow removal that reflects weather and traffic conditions is required. In Akita City, Japan, city hall staff measure snow depth and dispatch contracted snow removal crews only when a predefined threshold is exceeded. Consequently, dispatch decisions depend heavily on staff experience. This study demonstrates objective, experience-independent dispatching based on predicting the number of snowplow dozers in operation, thereby reducing the municipal decision burden and improving contractor efficiency. The target variable is highly imbalanced, with long non-operational periods and wide variations in the number of deployed units during snowfall events. When trained directly on such data, models tend to regress toward near-median values and face difficulty capturing operational dynamics. To address this issue, we propose a two-stage framework: firstly, a classifier predicts whether snow removal operations will occur; secondly, a regressor estimates the number of operating dozers based on the operation. We further integrate multi-year datasets to enhance generalization across diverse snow conditions. Experiments showed that the proposed approach achieved an AUPRC of 0.84 for operation occurrence and an RMSE of 1.85 for dozer-count estimation, outperforming models trained on a single year.