GIS-Based,Machine Learning Approach for Model Inventory Roadway Element Surface Type Classification from Aerial Imagery

Abstract

Roadway surface type classification (e.g., unpaved, asphalt, concrete) is an essential data element for road safety analysis and required as part of the Model Inventory of Roadway Elements (MIRE) Fundamental Data Elements (FDEs). Supplementary guidance issued by the Federal Highway Administration (FHWA) requires detailed surface type classification for the National Highway System (NHS) and roads functionally classified as interstates; however, less detailed surface type attribute group values (i.e., unpaved surface, asphalt pavement, concrete pavement, and other paved surface) are required for all remaining road segments that are not a part of the NHS. As part of a FHWA data analysis technical assistance (DATA) team, the research team worked with the Kansas Department of Transportation (KDOT) to develop a machine learning model capable of initially categorizing roads to reduce the level of effort required to satisfy the full MIRE FDE requirements. The research team conducted a pilot image classification analysis and determined a potential workflow for a broader statewide study area. This paper outlines the pilot methodologies and results. The team developed a support vector machine (SVM) learning model combined with Bayesian statistics that classified roads as paved or unpaved from aerial imagery with an overall model accuracy of 95%.