Distributed Acoustic Sensing for Urban Traffic Monitoring: Spatio-Temporal Attention in Recurrent Neural Networks

Izhan Fakhruzi; Manuel Titos; Carmen Benítez; Luz García

cs.LGMar 14, 2026

Distributed Acoustic Sensing for Urban Traffic Monitoring: Spatio-Temporal Attention in Recurrent Neural Networks

Authors: Izhan Fakhruzi, Manuel Titos, Carmen Benítez, Luz García

Organizations: Department of Signal Theory, Telematics and Communications, and the Center for Information and Communication Technologies Research (CITIC), University of Granada, 18014 Granada, Spain

Abstract

Effective urban traffic monitoring is essential for improving mobility, enhancing safety, and supporting sustainable cities. Distributed Acoustic Sensing (DAS) enables large-scale traffic observation by transforming existing fiber-optic infrastructure into dense arrays of vibration sensors. However, modeling the high-resolution spatio-temporal structure of DAS data for reliable traffic event recognition remains challenging. This study presents a real-world DAS-based traffic monitoring experiment conducted in Granada, Spain, where vehicles cross a fiber deployed perpendicular to the roadway. Recurrent neural networks (RNNs) are employed to model intra- and inter-event temporal dependencies. Spatial and temporal attention mechanisms are systematically integrated within the RNN architecture to analyze their impact on recognition performance, parameter efficiency, and interpretability. Results show that an appropriate and complementary placement of attention modules improves the balance between accuracy and model complexity. Attention heatmaps provide physically meaningful interpretations of classification decisions by highlighting informative spatial locations and temporal segments. Furthermore, the proposed SA-bi-TA configuration demonstrates spatial transferability, successfully recognizing traffic events at sensing locations different from those used during training, with only moderate performance degradation. These findings support the development of scalable and interpretable DAS-based traffic monitoring systems capable of operating under heterogeneous urban sensing conditions.

Figures & tables

Fig. 1: General attention module (adapted from [ 53 ] ).

Fig. 2: 2.5 km of fiber, the red triangles represent the locations of the fiber where the registered data are collected and used in this study. (1) Palacio de Congresos ; (2) Acera del Darro crossroads. (Courtesy of Google Maps)

Fig. 3: (Left) Schematic overview of fiber laid perpendicular to the lanes. Identified SPs on the fiber is denoted by #1, #2, …, #n. (Right) DAS registers of a 10-minute long signals with y-axis corresponds to the SPs. The fiber layout is depicted for: (a) Palacio de Congresos with 2 lanes, 1 direction, and 3 identified SPs; (b) Acera del Darro with 4 lanes, 2 directions, and 7 identified SPs, devided into 3 Groups of SPs: A(#1–#3), B(#3–#5), C(#5–#7).

Event	Description	Palacio. (#W)	Acera. (#W)	Duration (s)
Noise	Pedestrian or no event	43,129	3,482	11.73 $\pm$ 12.70
Car	All types of cars	19,460	2,297	5.72 $\pm$ 3.05
Bus	Large-size vehicles (all types of buses & trucks)	9,411	1,421	15.86 $\pm$ 6.07
Total		72,000	7,200

TABLE I: Summary of window counts (#W) and event durations (Mean

\pm

Std, in seconds (s)) generated from 10- and 1-hour observations at Palacio de Congresos and Acero del Darro Crossroads.

Model	$X$					$X$ ,+ $\Delta$
	Acc(%)	F1(%)	#Param(M)	RI-Acc(%)	RPI(%)	Acc(%)	F1(%)	#Param(M)	RI-Acc(%)	RPI(%)
LSTM	86.62 ±0.02	86.59 ±0.02	0.08	-1.65	-79.55	88.17 ±0.01	88.18 ±0.01	0.92	0.11	124.35
bi-LSTM	88.08 ±0.01	88.10 ±0.01	0.41	0.00	0.00	88.38 ±0.01	88.41 ±0.01	0.69	0.35	68.09
bi-TA	88.47 ±0.01	88.49 ±0.01	0.25	0.45	-38.85	88.13 ±0.01	88.11 ±0.01	3.36	0.06	715.96
TA-bi	87.97 ±0.01	88.06 ±0.01	0.32	-0.12	-23.21	88.41 ±0.00	88.45 ±0.01	1.34	0.37	224.72
bi-SA	88.23 ±0.01	88.25 ±0.01	0.90	0.18	119.50	88.02 ±0.01	88.02 ±0.01	0.92	-0.06	123.26

TABLE II: Performance Comparison of the Models using feature Vectors with/without Temporal Derivatives (

X

and

X

\Delta

). Evaluation metrics include accuracy (Acc), F1-score (F1), number of trainable parameters (#Param), relative improvement in accuracy over the baseline (RI-Acc), and relative increase in the number of trainable parameters with respect to the baseline (RPI), with

\pm

indicating standard deviation across 5-fold cross-validation. The bi-LSTM model serves as the baseline for all comparisons highlighted in gray and #Param is reported in millions (M).

Fig. 4: Model comparison in accuracy, F1-score, and standard deviation (whisker) across 5-fold cross-validation. Detail for the left-side y-axis range between 50%–90%. Models are sorted from left to right based on increasing number of trainable parameters (secondary y-axis) and trained with and without temporal derivatives (+ $\Delta$ ). (a) Baseline models (blue), (b) Single attention models (green), and (c) Cascade spatio-temporal attention models (red).

Fig. 5: Two representative architectures from the ablation study: (a) SA-bi-TA, (b) bi-TA-SA.

Fig. 6: Confusion Matrices of the SA-bi-TA model: (a) Group A, (b) Group B, (c) Group C of Acera del Darro , and (d) Palacio de Congresos .

Model	Acc(%)	F1(%)	#Param (M)	RI-Acc (%)	RPI(%)
LSTM	86.62 ±0.02	86.59 ±0.02	0.08	0.00	0.00
bi-LSTM	88.08 ±0.01	88.10 ±0.01	0.41	1.68	389.00
LSTM+ $\Delta$	88.17 ±0.01	88.18 ±0.01	0.92	1.79	997.20
bi-LSTM+ $\Delta$	88.38 ±0.01	88.41 ±0.01	0.69	2.03	722.05

TABLE III: LSTM vs. bi-LSTM comparison using feature vectors with/without temporal derivatives (+

\Delta

). Metrics: Acc, F1, #Param, RI-Acc, and RPI. Baseline Model: LSTM (gray shading). #Param is reported in millions (M).

Model	Acc(%)	F1(%)	#Param (M)	RI-Acc (%)	RPI(%)
bi-SA	88.23 ±0.01	88.25 ±0.01	0.90	0.18	119.50
bi-TA	88.47 ±0.01	88.49 ±0.01	0.25	0.45	-38.85
bi-SA+ $\Delta$	88.02 ±0.01	88.02 ±0.01	0.92	-0.06	123.26
bi-TA+ $\Delta$	88.13 ±0.01	88.11 ±0.01	3.36	0.06	715.96
bi-LSTM	88.08 ±0.01	88.10 ±0.01	0.41	0.00	0.00

TABLE IV: SA or TA module After bi-LSTM on feature vectors with/without temporal derivatives (+

\Delta