Optimal RWIS Sensor Density and Location – Phase III: Continuous Mapping of Winter Road Surface Conditions via Big Data and Deep Learning Project TPF-5(290)

(2021) Optimal RWIS Sensor Density and Location – Phase III: Continuous Mapping of Winter Road Surface Conditions via Big Data and Deep Learning Project TPF-5(290). Transportation, Department of

Preview	PDF optimal_RWIS_sensor_density_and_location_phase-III_w_cvr.pdf File Size:2MB
Preview	PDF optimal_RWIS_sensor_density_and_location_phase-III_t2.pdf File Size:174kB

Abstract

Road weather information systems (RWIS) have long been regarded as one of the most advanced technologies for monitoring road surface conditions (RSCs) during the winter season. While RWIS provide information essential for winter road maintenance (WRM) services, they can only be implemented at select areas largely due to budgetary constraints. It is therefore indispensable to fill large spatial gaps that exist between RWIS stations to promote safer driving conditions and lower the cost of WRM activities. Furthermore, most RWIS stations nowadays are equipped with cameras that provide users with a direct view of the road conditions being covered; however, checking RSCs via these cameras is still being done manually, which hinders the full utilization of the rich image-based road condition data. To help tackle these challenges, this project aimed to develop a systematic, yet transferrable, method for estimating key RSC variables (i.e., road surface temperature and slipperiness) between RWIS stations using large-scale data and two advanced modeling techniques—kriging and deep learning (DL). Road surface temperatures, dash camera images, and remotely sensed data collected along selected Iowa interstate highway segments between October 2018 and April 2019 were used to develop the models for estimating RSCs. A total of 262 hourly events and more than 10,000 images were processed and utilized for model development. The findings suggest that the proposed kriging method is able to capture the general RSC pattern along the highway stretch with as few as one RWIS input. In addition, the DL model developed in this study showed promising performance in automatically classifying dash camera images. The road condition images labeled by the DL model were later used for road slipperiness estimations between existing RWIS stations. Although additional data sets would be required to further confirm the validity of the developed models and the conclusiveness of the results documented herein, the proposed method will undoubtedly provide decision makers with a tool that helps to implement WRM activities more quickly, efficiently, and cost effectively.