Global positioning systems (GPS) tracking data have been widely used to collect mobility data to investigate travel behaviors and identify travel patterns. Some critical travel information, such as frequently visited locations, speeds, temporal changes, can be easily extracted from the raw GPS data. However, travel information like transport modes that have been used are difficult to acquire, and more complex analytical processes are required. Previous studies have mostly adopted context-specific methods or stand-alone methods in detecting transport modes from GPS data. Most of these context-specific methods are based on a limited number of datasets since the required data labeling process is time-consuming. This paper proposes a generic stepwise methodology by integrating unsupervised learning algorithms, GIS multi-criteria process, and supervised learning algorithms for data labeling and transport mode detection. The performances of five commonly used supervised algorithms are evaluated by applying them to a large-scale GPS tracking dataset. The results indicate that the proposed stepwise methodology can reduce data labeling time while providing high precision in detecting transport modes. The evaluation shows that the Random Forest algorithm is the most preferable, with only 10% labeled data needed and it can achieve a precision of 99%.