Automated Estimation of Winter Driving Conditions

Baldwin, Michael E; Burris, Kevin D; Elmore, Kimberly L · 2018 · ROSA P / Purdue University. Joint Transportation Research Program

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This study addresses the need for automated, real-time estimation of winter driving conditions to support transportation agency operations and traveler safety. The Indiana Department of Transportation (INDOT) currently relies on the Condition Acquisition Reporting System (CARS), where staff manually report road conditions as "good," "fair," "difficult," or "hazardous." This manual process is limited during intense storms when maintenance staff are focused on plowing rather than reporting. The authors developed a machine learning model to automatically estimate these conditions using weather data, aiming to provide rapid updates and reduce the burden on personnel. The researchers integrated high-fidelity weather variables from the North American Mesoscale (NAM) model and NCEP Stage IV precipitation analysis with historical CARS reports from 2014–2016. Weather data were interpolated to a 1/8-degree grid, and CARS reports were spatially matched to these grids to create a training dataset of 184,884 labeled examples. The team tested 16 classification models using the Scikit-learn library. Decision tree-based methods, particularly the random forest classifier, were selected for their high accuracy and ability to handle complex input variables. The model was trained on the 2014–2016 data and then deployed as an experimental system during the 2017–2018 winter season to evaluate its performance on real-time data. Additionally, the study merged crowdsourced precipitation observations with standard airport data to generate seasonal analyses of winter precipitation frequency. The random forest model achieved approximately 90% accuracy on the held-out test portion of the training data. However, when applied to the 2017–2018 season, performance dropped to roughly 70% accuracy. The model significantly underestimated "difficult" and "hazardous" conditions, likely due to overfitting to the training data and changes in the National Weather Service’s short-term forecast systems. Feature importance analysis indicated that visibility and snow depth were critical predictors, with lower visibility and deeper snow correlating with worse driving conditions. The seasonal precipitation analysis revealed significant differences from previous studies, incorporating a larger volume of observations. The findings demonstrate the potential for automated systems to estimate winter driving conditions but highlight the need for further refinement to address overfitting and adapt to changing weather forecast inputs. The authors propose implementing a notification system to alert staff of discrepancies between automated estimates and manual reports, serving as a transitional step toward using automated estimates as initial guesses in the CARS system. This approach aims to improve the timeliness and coverage of road condition information, enhancing both maintenance efficiency and traveler awareness.

Key finding

The random forest model correctly classified winter driving conditions at approximately 90% accuracy on training data but dropped to roughly 70% accuracy during the 2017-18 experimental season.

Methodology

dataset

Sample size: 184884

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• weather rain fog snow