Deep clustering of traffic signals using a single seismic station

Abstract

Vehicle traffic generates vibrations propagating in the subsurface. Identification and clustering of these seismic sources are crucial for traffic monitoring and subsurface imaging. We propose a novel method which uses a single seismic station and deep clustering to categorize the traffic signals. We utilize a deep embedded clustering (DEC) to extract features from frequency-time spectrograms of the recorded seismic signals. The similar traffic signals are grouped according to their key features and further used to infer the type of the vehicles. This deep clustering framework is unsupervised without manual labeling. Synthetic tests achieve a clustering accuracy of more than 99 %. We apply the method to field seismic recordings at three sites nearby the roadside with traffic videos for label validation. Results show an average accuracy of approximately 83 % and 91 % for vehicle type classifications at the intersection sites (Sites 1 and 2), respectively, where there are speed bumps in the roads. The vehicles moving in the near and opposite lanes are also distinguished from each other, with an accuracy of 73.3 % and 90.2 % at Site 1, and 88.4 % and 86.3 % accuracy at Site 2, respectively. At Site 3 along a straight road, the deep clustering model maintains 82 % accuracy for identifying heavy vehicles (buses and trucks), although the classification of small vehicles (cars and bikes) is limited to 58 % due to the relatively weak seismic signals generated by the light vehicles. The results confirm the framework's ability to cluster traffic seismic signals. By addressing the lack of single-station methods for traffic signal classification with unsupervised deep clustering, the proposed method offers a low-cost and scalable alternative to traditional camera-based traffic sensing systems, providing an effective tool for traffic seismic monitoring at the city scale.