Convolutional Neural Network-Bidirectional Long Short-Term Memory-Based Framework for Joint Transportation Mode and Transition Point Detection From GPS Trajectories

With the increasing availability of GPS-enabled devices, studying human mobility through recorded trajectories is becoming increasingly vital for shaping intelligent transport systems and supporting urban development. However, accurately classifying transportation modes and detecting transition points within multimodal trips remains a complex challenge because of spatial-temporal variability and noisy data. This study proposes a deep learning-based framework that jointly estimates transportation modes and transition points using a multi-output model built on a convolutional neural network and bidirectional long short-term memory architecture. The proposed framework incorporates an effective STAY/MOVE segmentation method that separates stationary and moving intervals, thereby improving both transportation mode recognition and transition point detection. The model leverages rich kinematic features, including capturing speed, acceleration, direction changes, and turning behavior, to learn spatial and temporal patterns in mobility data. To assess the generalizability of the model, cross-validation is conducted across multiple provinces in China using GPS data collected from GPS loggers and GPS-enabled phones. The best-performing configuration under random split achieves a 96% F1-score and 97% recall for transportation mode classification, as well as a 95% F1-score and 89% recall for transition point detection, successfully identifying the transport modes (designated as walk, bike, car, bus, and train). These results demonstrate the robustness of the proposed approach across diverse urban contexts and reinforce the role of deep learning in large-scale mobility analytics.