Iterative self-training segmentation for scalable traffic monitoring using distributed acoustic sensing

Distributed Acoustic Sensing (DAS) enables long-distance and high-resolution monitoring using existing fiber-optic infrastructure, making it a promising sensing technology for transportation monitoring. However, DAS measurements recorded in real deployments are often affected by environmental disturbances, installation conditions, and instrument noise, which significantly complicates signal interpretation. Most existing DAS-based traffic monitoring systems rely on handcrafted signal preprocessing or supervised learning pipelines that require extensive manual annotations. In particular, segmentation of DAS traffic measurements without manually annotated training masks remains insufficiently explored, especially when combined with pseudo-label refinement. To address these limitations, this study proposes a unified framework that integrates a physics-aware learnable filter (PALF) with an iterative self-training semantic segmentation scheme. Vehicle-related regions are automatically identified and progressively refined without requiring manually annotated segmentation masks. The segmented vehicle regions are then used for downstream traffic analysis, including speed estimation and unsupervised vehicle grouping. Experimental evaluation on real-world DAS traffic data demonstrates that the proposed framework enables stable vehicle localization under noisy conditions. Among the evaluated architectures, the best-performing configuration achieved reliable speed-estimation performance, with an MAE of 5.08 km/h and an RMSE of 6.18 km/h, while also supporting meaningful unsupervised vehicle grouping. These results highlight the potential of the proposed approach for scalable, label-free DAS-based traffic monitoring in real-world deployment scenarios.