This study was conducted to develop a machine learning model that classifies hazardous winter road surface conditions, thereby supporting snow removal decision-making in place of the visual inspection and operator experience currently relied upon in practice. Road surface imagery and surface temperature were collected at 10-minute intervals from four field sites in Seoul during the 2023-2024 winter season, and were merged with meteorological data obtained from the nearest weather stations. Road surface conditions were labeled into two classes according to accident risk: "Ice or Snow" and "Normal or Wet." The Synthetic Minority Over-sampling Technique (SMOTE) was then applied to the training set only, leaving the test set unaltered. A random forest classifier was trained, and twelve input variable cases were compared to determine whether excluding highly correlated variables improves predictive performance. Feature importance was assessed using both Gini importance and permutation importance to verify the robustness of the results. Excluding highly correlated variables generally degraded predictive performance rather than improving it, indicating that air temperature, surface temperature, dewpoint, and relative humidity each retain unique information that cannot be fully explained by the others. The best-performing input set excluded only wind speed, achieving a recall of 0.96, an F1-score of 0.90, and a precision of 0.86 for the hazardous class. Wind speed exhibits strong local variability, and values measured at weather stations were therefore considered inadequate for representing field conditions. The two feature importance measures produced largely consistent rankings, with surface temperature, relative humidity, and time elapsed since the end of precipitation ranking highest. Notably, surface temperature ranked first in permutation importance despite its high correlation with air temperature and dewpoint, indicating that it carries information that cannot be substituted by other temperature-related variables. A random forest model was developed to classify hazardous winter road surface conditions from meteorological data and surface temperature. Surface temperature and relative humidity were identified as the dominant variables. Since the model relies on variables observed at a single point in time rather than timeseries structures, it can be extended to predict future road surface conditions.

ABSTRACT
1. 서론
    1.1. 연구배경 및 필요성
    1.2. 연구목적 및 연구방법, 연구범위
2. 데이터 수집 및 전처리
    2.1. 데이터 수집 및 기초통계
    2.2. 데이터 전처리
3. 노면상태 예측모델 개발
    3.1. 학습데이터 구성 및 상관성 분석
    3.2. 모델 입력데이터 케이스 분류
    3.3. 모델 구조 및 하이퍼파라미터 설계
4. 학습결과
    4.1. 모델성능 평가지표
    4.2. 성능평가 결과 분석 및 입력변수 선정
    4.3. 변수중요도
5. 고찰
    5.1. 데이터의 공간적 특성과 모델 적용 범위
    5.2. 모델의 공간적 확장 가능성
    5.3. 모델의 시간적 확장 가능성
6. 결론
REFERENCES

• 마경훈(서울연구원 인프라기술연구실 연구원) | Ma Gyeonghoon
• 오한진(서울연구원 인프라기술연구실 연구위원) | Oh Han Jin Corresponding author
• 이진욱(서울연구원 친환경도로기술연구센터 연구위원) | Lee Jin Wook