GIS-Based Multivariate Spatial Clustering for Traffic Pattern Recognition using Continuous Counting Data

Abstract

Traffic count stations play a key role in measuring roadway characteristics and traffic performance by collecting and monitoring travel behavior and vehicle data. Continuous counting stations (CCSs), which count traffic volumes continuously throughout the year, are used to develop seasonal adjustment factors to convert short-term traffic counts (average daily traffic) to annual average daily traffic (AADT). As data collection is conducted at limited locations, many state Departments of Transportation (DOTs) group the CCSs based on different traffic patterns and estimate the AADT at specific locations by considering seasonal adjustment factors. Computer-based clustering approaches are widely used in grouping continuous traffic data for their accuracy in traffic pattern recognition. However, most of the clustering techniques do not consider the spatial constraints and therefore overlooked the locational proximity and inference from nearby traffic data. In this study, a GIS-based multivariate spatial clustering approach was developed to recognize statewide traffic patterns based on temporal and spatial variables. A total of 12 optimal clusters were automatically computed and labeled based on the proposed clustering algorithm. The proposed clustering approach was compared and validated based on machine learning classifiers. The results showed that it outperformed the traditional Michigan DOT clustering approach and was consistent in nature across different years. The model was applied to estimate the AADT, and good accuracy was detected relative to other approaches. The proposed clustering method offers a new approach to group traffic patterns by simultaneously incorporating proximity and similarity of traffic data.