Trip Purpose Estimation for Urban Travel in the U.S.: Model Development, NHTS Add-on Data Analysis, and Model Transferability Across Different States

It's becoming a trend that the traditional travel survey will be supplemented or replaced by advanced survey based on GPS technology. However, the pivotal key to successfully establish the dominance of the GPS-based travel survey in future is the efficient post-processing methods that can generate the essential components such as travel time, trip purpose, travel mode, and trip length as accurately as possible. This paper therefore concentrates on part of the geospatial data post-processing: trip purpose derivation. Two released 2009 NHTS add-on data sets (Georgia and Arizona) containing geospatial location data provide the possibility of imputing trip purpose, validating the trip purpose model and further evaluating the model transferability. Multiple classifiers are explored employing machine learning methods with 2009 NHTS add-on data sets and land use data at both parcel level and point level. Different validation methods including 10-fold cross validation, within-sample and cross-sample validation are used to evaluate, test and validate the developed models. Results indicate that the trip purpose models perform well for Home, Work, School/Daycare and Shopping/Errands trips with accuracy above 80%, but present unsatisfactory results for transport someone, meals, social/recreation, family personal business/obligations and other trips. In addition, aggregating discretionary trips into one trip purpose category usually improves the trip purpose imputation accuracy. Furthermore, when applying the models to a different geographic place to predict the trip purpose without any model calibration, the models present non-striking model transferability.