Travel mode identification is among the key problems in transportation research. With the gradual and rapid adoption of GPS-enabled smart devices in modern society, this task greatly benefits from the massive volume of GPS trajectories generated. However, existing identification approaches heavily rely on manual annotation of these trajectories with their accurate travel mode information, which is both economically inefficient and error-prone. In this work, we propose a novel semi-supervised deep ensemble learning approach for travel mode identification to use a minimal number of annotated data for the task. The proposed approach accepts GPS trajectories of arbitrary lengths and extracts their latent information with a tailor-made feature engineering process. We devise a new deep neural network architecture to establish the mapping from this latent information domain to the final travel mode domain. An ensemble is accordingly constructed to develop proxy labels for unannotated data based on the rare annotated ones so that both types of data contribute to the learning process. Comprehensive case studies are conducted to assess the performance of the proposed approach, which notably outperforms existing ones with partially-labeled training data. Furthermore, we investigate its robustness to noisy data and the effectiveness of its constituting components.