Study Overview

This study was conducted in Lisbon, Portugal—a city with diverse urban morphology suitable for examining how environmental conditions influence pedestrians’ physiological and emotional responses. From August to December 2024, ninety adult residents of the Lisbon Metropolitan Area participated in a two-week monitoring protocol. Participants were recruited via a university mailing list and screened through an online eligibility questionnaire. Selected individuals received detailed written and verbal instructions explaining the use of both devices and daily procedures.

Each participant wore an Empatica E4 wristband continuously to record heart rate, electrodermal activity, and skin temperature, and used the X-ing mobility app to collect GPS trajectories and self-report emotional states after each walking trip using brief semantic differential scales. Surveys were completed directly in the app immediately after each trip. Data were synchronized automatically at the end of each day to a secure server linked to the participant’s account. All procedures were approved by the institutional ethics committee, and participants received a €75 voucher as compensation.

In total, 2,207 walking trips were validated and enriched with spatial and environmental indicators, including NDVI, noise, temperature, precipitation, street imagery, and infrastructure variables. Participant characteristics (age, gender, education, income, and residence) and the distribution of daily and total walking trips per participant are presented in Fig. 1and Supplementary Table S6. The methodological workflow (Supplementary Figure S14) builds on a naturalistic mobility protocol previously developed in Copenhagen [25]. The final dataset, including all validated trips and variable descriptions, is publicly available in the repository [26] (Data_per_route.csv, Metadata.csv).

Fig. 1The alternative text for this image may have been generated using AI.

Sociodemographic characteristics, spatial distribution, and walking behavior of participants

Data Preprocessing and Variable ConstructionGPS Trajectories and Route Construction

GPS data from the X-ing app were used to reconstruct pedestrian routes within Lisbon. The app has been previously applied and validated in mobility research, including pedestrian travel behavior and environmental exposure studies [27, 28]. Points were chronologically ordered by trip ID, and geodesic distances were calculated using the Haversine formula. Points more than 500 m apart were excluded to eliminate signal errors or modal shifts. A simplified stop detection algorithm segmented the data into valid walking trips, removing stationary periods and non-walking segments [29].

Each trip was converted into a LineString, and a 25-m buffer was created to capture micro-scale environmental exposure. This buffer size was chosen to approximate the immediate pedestrian microenvironment in Lisbon, where typical road widths in central neighborhoods range from 10 to 20 m. The 25-m distance, thus, encompasses sidewalks and adjacent elements influencing sensory walking experience, while avoiding dilution from more distant urban features. Comparables buffer sizes (20–30 m) have been employed in previous walkability and environmental exposure studies to represent fine-scale pedestrian environments [1, 30].

Environmental data were extracted from a hexagonal grid (25 m diameter), which ensures uniform cell spacing, minimizes edge effects, and enhances spatial continuity. This resolution effectively captures features like vegetation, street furniture, and infrastructure [31,32,33].

An overview of the GPS walking trajectories, buffer zones, and spatial coverage of trips across the study area is presented in the Supplementary Figures S1 and S2. These maps illustrate both the extent of the analyzed routes and the variation in trip frequency per cell.

Contextual indicators were derived from Mapillary and POIs from Open Street Map and subsequently aggregated along each route using mean values and frequency distribution across the hexagonal grid.

Subjective and Sociodemographic Data

Sociodemographic data were collected via an online pre-survey, including age, gender, education level, and self-reported monthly income [25].

Subjective data were gathered daily through the X-ing app, where participants retrospectively rated four dimensions: (i) Trip Feeling Environment (0–5); (ii) Unwell–Well (1–6); (iii) Agitated–Calm (1–6); and (iv) Tired–Awake (1–6). Participants confirmed the timing and relevance of their responses, increasing ecological validity.

These variables supported the interpretation of physiological responses by capturing emotional and perceptual states during walking trips.

Physiological Data Processing

Physiological signals were collected using Empatica E4 wristbands, worn continuously throughout the two-week monitoring period. These devices recorded cardiac activity and skin conductance, allowing the extraction of markers related to stress and autonomic nervous system regulation. Although E4 has been validated for ambulatory monitoring [34, 35] of HRV, electrodermal activity (EDA), and skin temperature, it remains less precise than laboratory-grade systems (e.g., Biopac, Powerlab). Moreover, as data were collected during walking, autonomic variability was naturally constrained by elevated sympathetic tone. To minimize the impact of these limitations, we focused on robust summary measures (e.g., mean RMSSD) and applied standardized preprocessing procedures to enhance signal quality.

From the interbeat interval (IBI) data, we computed heart rate variability (HRV) metrics: RMSSD (short-term parasympathetic activity), pNN50 (percentage of successive IBI differences > 50 ms), and mean heart rate (bpm) as a general cardiac marker. HRV indices were derived using 10-s windows by applying standard formulas to all IBIs within each interval and then summarizing per walking trip (mean and dispersion measures).

Electrodermal activity (EDA) was first smoothed with a Savitzky–Golay filter, after which tonic and phasic components were separated using a high-pass filter. As the phasic component reflects rapid changes in response to environmental stimuli, we derived features including skin conductance response amplitude, rise time (onset to peak), and recovery time (time to return to 63% of peak). For these EDA-derived metrics, mean values and standard deviations were computed at the trip level. To illustrate the signal quality and preprocessing steps, examples of raw and filtered EDA waveforms are provided in Supplementary Figures S13a–b.

To ensure signal validity, only sessions lasting ≥ 2 min, containing non-empty IBI arrays, and showing evidence of proper sensor contact (skin temperature ≥ 30 °C or raw EDA ≥ 0.05 µS) were retained for analysis. Trips with poor signal quality were excluded.

All physiological indicators were computed per walking trip and synchronized with GPS trajectories. Trips with poor signal quality were excluded.

Data processing followed the Emotional Cities experimental protocol [36, 37].

Climatic Data Acquisition and Processing

Daily temperature and precipitation values were obtained from the Portuguese Institute of the Sea and Atmosphere (IPMA) [38]. Maximum, mean, and minimum temperatures, as well as daily precipitation totals, were matched to each walking trip based on date, meaning that climatic variation reflects temporal variation between days, not spatial variation within the city. These variables were included as covariates to account for weather effects. Full methodological details are available in Supplementary Methods S2.4.

NDVI Extraction and Processing

Vegetation exposure was measured using NDVI composites from Sentinel-2 imagery processed in Google Earth Engine for August–December 2024. Monthly cloud-filtered mosaics (10 m) were clipped to Lisbon and summarized per walking route using mean, min, max, and SD values within the 25 m buffer. NDVI metrics were then used as continuous predictors in subsequent models (details in Supplementary Methods S2.5).

Noise Pollution Data Processing

Environmental noise exposure was derived from Lisbon’s official Lden noise map (dB[A]), provided by the municipality [39]. Values were extracted for each walking trip by intersecting noise rasters with 25 m buffers around GPS tracks, generating mean, minimum, maximum, and SD noise metrics per route. These indicators were included as continuous predictors of perceptual and physiological responses (full processing details in Supplementary Methods S2.6).

Streetscape Feature Extraction from Mapillary Imagery

Streetscape characteristics were derived from Mapillary street-level imagery using semantic segmentation applied to geolocated photographs intersecting a 25-m buffer around each route. Images were classified into visual categories (vegetation, sky, road, sidewalks, vehicles, poles, etc.) using a pretrained deep-learning model, and the mean pixel-proportion per class was computed to quantify visual exposure along each trip (full pipeline in Supplementary Methods S2.7) [40, 41].

Points of Interest (POI) from OpenStreetMap

Points of Interest (POIs) from OpenStreetMap were used as indicators of urban functional diversity, including services, commerce, tourism, public facilities, and greenery. POIs were extracted via the Overpass API and aggregated within a 25-m buffer around each walking route. Counts per category were summed to represent exposure intensity along each trip, and these indicators were later incorporated into statistical analyses [41].

Full data processing workflow, category definitions, and extraction scripts are documented in Supplementary Methods S2.8.

Statistical Analysis

The unit of analysis was the individual walking trip, with multiple trips nested within participants. Environmental and infrastructural predictors were continuously monitored along each route and aggregated at the trip level to create summary statistics (mean, maximum, and standard deviation) within a 25 m buffer around each GPS trajectory. These trip-level summaries were used as predictors in the models.

To account for repeated measures and the nested data structure, linear mixed-effects models with random intercepts for participant ID were fitted. Each model included multiple environmental predictors simultaneously, allowing estimation of their independent associations with perceptual and physiological outcomes while accounting for correlations among them. Pearson correlations were computed to explore bivariate relationships and guide model specification by minimizing multicollinearity (see Supplementary Tables S1– S3). Person-level characteristics (age, gender, education, income) were collected but not included as fixed effects, as the sample was relatively homogeneous and the study focused on within-individual variation across environmental contexts.

Hypothesis 1 (H1) posited that greener and more vibrant environments are associated with more positive perceptions. Separate mixed-effects models were estimated for four subjective outcomes (Trip Feeling Environment, Unwell–Well, Agitated–Calm, Tired–Awake), with standardized environmental indicators (NDVI, Mapillary-derived vegetation and sky visibility, sidewalk presence) and POI densities (leisure, tourism, arts, greenery, water bodies, slope) entered jointly as predictors. Missing POI values were replaced with zero. Models were implemented in Python’s MixedLM function (statsmodels), with α = 0.05. Full specifications and preprocessing steps are detailed in “6.Model_H1.py” (Zenodo repository [26]).

Hypothesis 2 (H2) examined whether environmental stressors (noise, heat, infrastructure) are associated with increased physiological activation. Using the same framework, predictors were grouped into five thematic blocks (Noise, Temperature, Technical Infrastructure, Functional POIs, Visual Elements), and each block was tested in separate models across 27 physiological indicators of HRV, EDA, and skin temperature. All predictors were standardized before analysis. Associations meeting both statistical significance (p < 0.05) and effect-size criteria (|β|> 0.1) were retained. Full details are available in “7.Model_H2.py” (Zenodo repository [26]).

Hypothesis 3 (H3) evaluated whether distinct environmental typologies elicit differentiated physiological and perceptual responses. K-means clustering was applied to standardized environmental variables (NDVI, sky visibility, vegetation, noise, and POI categories), with missing POIs imputed as zero. The optimal number of clusters (k = 4) was determined using the elbow method, producing typologies representing distinct urban environments (e.g., green/quiet vs. dense/noisy). Physiological and perceptual outcomes were compared across clusters using Kruskal–Wallis tests. Details of the clustering workflow and comparison methods are provided in “8.Model_H3.py” (Zenodo repository [26]).

All analyses were conducted in Python (Pandas, Statsmodels, Scikit-learn), with spatial preprocessing in ArcGIS Pro and data cleaning in MS Access. Statistical significance was set at α = 0.05, with adjustments for multiple comparisons when appropriate. This analytical framework enabled both explanatory and exploratory evaluation of how urban environmental typologies shape emotional and physiological responses during walking.