RESUMO
Road grade or slope is a key factor for walking and cycling behavior and outcomes (influencing route, speed, energy, etc.). For this reason, the scarcity of precise road grade data presents a challenge for travel information and analysis. This paper examines the accuracy of using crowd-sourced GPS data from a fitness application to estimate roadway grade profiles, which can then be used to develop network-wide road grade datasets. We externally validate an elevation estimation method described by McKenzie and Janowicz using field surveying data, and then propose and evaluate modifications for estimation of road grade (which is more directly relevant than elevation for walking and cycling analysis). We find that a modest amount of crowd-sourced GPS data can be used to generate relatively accurate road grade estimates: better than commonly-used low-resolution elevation models, but not as accurate as high-resolution data derived from LiDAR (Light Detection and Ranging). We also find that the grade estimates are more reliable than the elevation estimates, relative to alternative data sources. The most accurate method to aggregate crowd-sourced GPS data builds a composite roadway grade profile using partition-around-medoid clustering of individual grade sequences, first smoothed with a Savitzky-Golay filter and cleaned with Density-Based Spatial Clustering of Applications with Noise (DBSCAN). Implementing this method with an average of 150 GPS traces per location yields a root mean square error (RMSE) of 1% road grade. The findings in this paper can be used to incorporate precise road grade information into street network datasets over a wide spatial scale, which is necessary for walking and cycling analysis that fully considers the physiological aspects of active transportation.