Monitoring Urban Traffic Dynamics at Fine Spatiotemporal Resolution Using Distributed Acoustic Sensing and Deep Learning

Motivation

• Existing traffic sensing approaches can offer fine resolution but face privacy, sparse-coverage, and operational vulnerabilities, including disruption during emergencies.These constraints motivate complementary sensing for timely situational awareness.Source: §1 · p. 2

DAS as an urban sensing modality

• DAS repurposes roadway fiber-optic cables as dense, continuous vibration-sensor arrays, enabling meter-scale monitoring without collecting personally identifiable information.The sensing approach also offers broader coverage and lower investment and maintenance costs than traditional node-based systems.Source: §1 · p. 2

Research gap and contribution

• Prior DAS traffic studies largely emphasize vehicle detection and trajectories, while system-level traffic-state inference and event-scale analysis remain underexplored.Limited labels also constrain robustness under noisy and congested conditions, motivating the study’s hybrid training strategy and system-level analysis.Source: §1 · pp. 3–4

Study objectives

• The study develops a deep learning-enabled DAS framework combining hybrid synthetic-real training with traffic-state inference, vehicle detection, and aggregation into system-level traffic indicators.Its questions address robust, computationally efficient inference and how disruptive events reshape traffic over time and across roadway segments.Source: §1 · pp. 3–4

2.1 Distributed acoustic sensing (DAS) as an emerging urban sensing infrastructure

• DAS measures dynamic strain along optical fiber using Rayleigh backscatter, turning existing cables into dense, continuous sensing arrays with meter-scale spatial and millisecond-scale temporal resolution.Urban and geoscience deployments illustrate its use beyond conventional sensing applications.Source: §2.1 · pp. 4–5

2.2 DAS-based traffic monitoring using signal-processing methods

• Signal-processing workflows extract vehicle counts, speeds, and traffic responses, but their computational cost can become prohibitive and performance degrades when traffic wavefields overlap.Prior work includes wavelet-denoising, template matching, and 4th-root slant stacking.Source: §2.2 · p. 5

2.3 DAS-based traffic monitoring using deep learning methods

• Deep learning on DAS spatiotemporal images supports vehicle detection and trajectory tracking, but limited annotations and emphasis on vehicle-level tasks leave scalable traffic-state inference underexplored.Prior approaches include YOLO-based detection, transformer architectures, and specialized trajectory-extraction methods.Source: §2.3 · pp. 5–6

2.4 Event-driven traffic dynamics and urban resilience

• Event-driven traffic studies document spatially uneven roadway responses, while DAS’s potential for fine-resolution event analysis has not been systematically explored.Planned events can raise traffic volume by 10% to 50%, and conventional loop, GPS, and camera sensing faces coverage, privacy, and resilience limitations.Source: §2.4 · pp. 6–7

3.1 Data collection

• The 23-km campus fiber recorded continuously at 4-m channel spacing and 100 Hz during October and November 2023.The deployment generated approximately 6.2 TB of raw strain-rate measurements.Source: §3.1 · pp. 7–8

3.2 Data preprocessing and spatiotemporal representation

• DAS recordings were filtered and denoised, then segmented into 180-s by 50-channel images, with synthetic trajectories embedded in authentic background noise.The 50-channel window spans approximately 200 m and balances spatial specificity with stable aggregation; the synthetic dataset contains 4,000 images.Source: §3.2 · pp. 8–10

3.3 Traffic detection using deep learning models

• Detection training progressed from 4,000 synthetic images, to hybrid synthetic-real data, to real-only fine-tuning, with a real-only model serving as baseline.The 1,200 labeled real images were split into 900 training and 300 validation images; 500 separate images were held out for final evaluation.Source: §3.3 · pp. 11–12
• YOLOv8, YOLOv11, and RT-DETR were evaluated using precision, recall, mAP@0.5, and mAP@0.5–0.95 on an independent hold-out dataset.The stricter mAP@0.5–0.95 averages performance across multiple IoU thresholds.Source: §3.3 · pp. 11–12

3.4 Traffic activity inference and event-based comparison

• Vehicle speed was estimated from the local PCA trajectory slope, while trajectory counts and individual speeds were aggregated into traffic volume and average speed.PCA-based estimates are more stable than endpoint-based slope estimates under noise and partial trajectory detection; channel spacing is 4 m.Source: §3.4 · pp. 12–13
• The line-based congestion intensity indicator combines vehicle accumulation and mobility degradation, and event periods are compared with matched non-event baselines.The indicator is not physical traffic density; the comparisons characterize event-associated changes rather than establish event causation.Source: §3.4 · p. 14

4.1 Validation of the proposed DAS traffic-state inference framework

• Hybrid training raised mAP@0.5–0.95 to approximately 0.75–0.77; YOLOv8 and YOLOv11 each reached 0.768 on the hold-out dataset.Synthetic-only models scored below 0.35, while additional real-data-only fine-tuning yielded only marginal improvement.Source: §4.1 · pp. 14–15, Figure 3 · p. 15
• Vehicle signatures can overlap under heavy congestion or intersect when vehicles move in opposite directions, increasing detection uncertainty or causing missed detections.These are qualitative limitations despite the framework’s high quantitative detection performance.Source: §4.1 · p. 15
• Deep learning estimates matched manual traffic-volume annotations more closely than slant stacking (r = 0.931, R² = 0.866, MAE = 0.608 versus r = 0.882, R² = 0.779, MAE = 0.798).The comparison also found the methods’ apparent-speed estimates correlated but not identical (ρ = 0.278, p = 7.46 × 10⁻⁹), with slant stacking tending higher under congestion.Source: §4.1 · p. 16, Figure 4 · p. 17
• 23 s versus 1201 s: deep learning inference took approximately 50 times less time than slant stacking.This comparison excludes one-time annotation and model-training costs.Source: §4.1 · pp. 16–17, Figure 4 · p. 17

4.2 Event-scale traffic state dynamics and urban activity inference

• Football gamedays produced a 34.2% increase in average congestion intensity, while Thanksgiving produced a 49.9% decrease and flattened the diurnal profile.Gameday congestion featured pre-event peaks and a near-doubling around 20:00; Thanksgiving suppression was sustained rather than concentrated in peaks.Source: §4.2 · pp. 18–19, Figure 5 · p. 19

4.3 Spatial heterogeneity of event-driven traffic responses

• Football gameday congestion increased across roadway segments by roughly 10% to 65%, with the F&B Road arterial exceeding 50% and some secondary segments decreasing.High-impact segments showed multiple surges, whereas low-impact segments lacked the pre-event buildup and could fall below baseline.Source: §4.3 · pp. 20–21, Figure 6 · p. 21
• Thanksgiving congestion reductions ranged from approximately 20% to over 50%, strongest on high-volume F&B Road arterials; secondary segments mainly declined during daytime.The arterial morning peak fell by approximately 30–40%, while secondary-segment traffic before 06:00 and after 18:00 remained broadly comparable to baseline.Source: §4.3 · pp. 21–22, Figure 7 · p. 22

5.1 Deep learning-enabled DAS for urban traffic monitoring

• Under high traffic, deep-learning trajectory slopes yield lower speeds than slant stacking, whose energy maxima can favor faster or more energetic vehicles in overlaps.The authors characterize trajectory-level estimates as more conservative and potentially more representative of a broader range of vehicle behavior.Source: §5.1 · pp. 23–24
• Approximately 50-fold speedup over traditional slant-stacking workflows accompanies a deep-learning detector that recognizes weak-energy trajectories using shape, continuity, contrast, and context.The authors describe slant stacking as effective for coherent, separable wavefields but less effective when low-energy trajectories are obscured.Source: §5.1 · pp. 23–24
• Deep learning does not replace conventional signal processing: the former depends on labels, transferability, and calibration, while the latter retains physical interpretability and can use little or no labeled data.The discussion frames the approaches as complementary, with deep learning advantageous in noisy, high-congestion, and large-scale monitoring scenarios.Source: §5.1 · p. 24

5.2 Mobility reorganization under disruptive events

• Football gamedays amplify and temporally structure mobility, whereas Thanksgiving suppresses congestion and flattens the diurnal profile through different reorganizations of travel activity.The discussion relates event-day patterns to temporally concentrated arrivals and departures, versus disrupted routine activities and weaker commuting synchronization during holidays.Source: §5.2 · p. 25

5.3 DAS as a continuous urban traffic observatory

• DAS tracks traffic-state evolution continuously along corridors at meter-scale spatial and second-level temporal sensing resolution, but reported indicators aggregate corridor windows and do not resolve within-window variation.The discussion contrasts this coverage with conventional point-based or spatially discontinuous observations.Source: §5.3 · p. 26
• DAS is positioned as a targeted complement, not a universal replacement, because deployment depends on suitable fiber infrastructure and interrogator costs.The authors identify critical corridors and areas vulnerable to traffic disruptions as strategic deployment settings.Source: §5.3 · pp. 26–27

5.4 Limitations and future work

• A single deployment limits generalizability; apparent speed lacks independent validation, congestion intensity is relative rather than physical density, and event associations are not causal estimates.The authors call for validation across deployments and against independent measurements, and for contextual variables in statistical analysis of event effects.Source: §5.4 · p. 27
• Future work could combine DAS with loop-detector or GPS indicators and use semi-supervised, self-supervised, or adaptive learning to reduce manual labeling and improve transferability across deployments.The authors present these as directions for further development rather than demonstrated results.Source: §5.4 · pp. 27–28

6. Conclusion

• Deep learning with DAS improves vehicle-detection robustness under noisy and congested conditions while reducing computational cost and supporting continuous traffic-state inference from aggregated observations.The framework extends beyond vehicle-level detection to monitoring urban traffic states.Source: §6 · p. 28
• Football gamedays and Thanksgiving reorganize urban mobility in distinct ways, producing characteristic temporal patterns and spatially heterogeneous roadway responses.The conclusion says these event-driven dynamics can be difficult to observe using conventional point-based sensing systems.Source: §6 · p. 28
• The authors identify model transferability and evaluation across a broader range of roadway and disruptive-event scenarios as future work.Source: §6 · p. 28

• The authors report no competing interest in the manuscript's publication or creation.Source: Declaration of competing interest · p. 28