UrbanPulse: Spatiotemporal Urban Analytics

• UrbanPulse is a paradigm that represents city activity as temporally varying, structured signals derived from diverse data modalities.
• The framework employs topology-driven and multi-resolution methods to detect, rank, and compare urban hotspots and mobility patterns.
• UrbanPulse systems integrate heterogeneous sources—mobile data, social media, and street-view imagery—to inform urban planning and real-time analytics.

Searching arXiv for UrbanPulse-related papers and adjacent work on urban pulse, mobility rhythms, street-view change, and perception. UrbanPulse denotes a family of urban-computing frameworks that model the city through temporally varying signals and represent its “pulse” as structured spatiotemporal activity. In its original formulation, Urban Pulse was introduced as a topology-driven, multi-resolution method for detecting, characterizing, comparing, and visually exploring prominent activity locations from spatiotemporal point data. Subsequent UrbanPulse-style systems extended the same organizing idea to aggregated mobile network activity, anonymized call detail records, Twitter, mobility traces, street-view time series, app usage, subjective perception, real-time traffic analytics, and ultra-fine-grained OD prediction, thereby broadening the notion of urban pulse from hotspot discovery to city-scale monitoring of movement, land use, physical change, and human-centered perception (Miranda et al., 2016, Kondor et al., 2015, Yang et al., 23 Jul 2025).

1. Conceptual foundations and historical development

In the 2016 topology-driven framework, the city is represented as a spatial domain $D \subset \mathbb{R}^2$ discretized as a planar triangular mesh $K$, with time handled through discrete resolution families such as All, Month-of-Year, Day-of-Week, and Hour-of-Day. For each resolution $r$ and time step $k$, a piecewise-linear scalar function $f_{r,k}: K \to \mathbb{R}$ captures activity intensity. Activity is constructed by Gaussian kernel density estimation,

$$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$

with $\epsilon = 100 \text{ m}$ in the reported experiments, and maxima are ranked by topological persistence. A pulse is defined as a pair $P=(L,B)$, where $L$ is a prominent location and $B$ is a collection of beats across temporal resolutions; the framework uses a default persistence threshold $K$0 to isolate prominent maxima (Miranda et al., 2016).

This definition was rapidly generalized. The mobile-network study on New York, London, Hong Kong, and Los Angeles operationalized the city’s pulse as a canonical weekly signature $K$1 computed from ten months of aggregated mobile activity, and used residuals

$K$2

to identify anomalous events such as Wimbledon and the 2013 UEFA Champions League Final. In Buenos Aires, the pulse became a home-to-presence redistribution matrix derived from anonymized CDRs and census-scaled user location distributions. In Madrid, it became normalized unique Twitter-user presence by zone and time slot. UrbanRhythm reframed the pulse as temporal evolution among “city states” inferred from staying, leaving, and arriving attributes. Later work further expanded the scope to street-level physical change, app-mediated social vibrancy, and perceptual urban quality, so that “pulse” came to denote not only where activity concentrates, but also how built form changes and how places are experienced (Kondor et al., 2015, Sarraute et al., 2017, Garcia-Palomares et al., 2017, Song et al., 2019, Huang et al., 2024, Collins et al., 2024, Muller et al., 2022).

A persistent characteristic across these formulations is that UrbanPulse is not restricted to one modality or one task. It is instead a representation principle: urban phenomena are converted into temporally indexed signals, signatures, or state sequences; these are then compared across locations, time scales, or cities; and the result is exposed through maps, temporal profiles, or predictive models. This suggests that UrbanPulse is best understood as an organizing paradigm rather than a single software system.

2. Core representations of the urban pulse

The original framework formalized three beat types at each pulse location $K$3 and resolution $K$4. The significant beat $K$5 indicates whether the location is a high-persistence maximum at time step $K$6; the maxima beat $K$7 indicates whether it is any local maximum; and the function beat $K$8 stores the normalized activity magnitude,

$K$9

Pulse ranking is defined from a feature vector containing, for each resolution, the fraction of significant beats, the fraction of maxima beats, and the peak function-beat magnitude; the rank is the $r$0 norm of that vector. Similarity between pulses is computed by Euclidean distance across beat sequences shared by two locations (Miranda et al., 2016).

Several later systems replaced topological beats with temporal signatures. In the mobile-network tool, the primary feature is the typical-week signature $r$1 at 15-minute resolution, optionally compared to a citywide baseline through

$r$2

Neighborhoods are grouped into functional clusters “formed on the similarity of typical activity time series,” and cluster means reveal business, residential, leisure, park, tourist, or event-oriented signatures. In Madrid’s Twitter analysis, normalized counts

$r$3

were used to compare zones across four time slots, and time-varying OLS coefficients linked land-use area to the city’s daily pulse. In Buenos Aires, each user is summarized by a Location Distribution Matrix, and the expected population present in commune $r$4 at day group $r$5 and hour group $r$6 is

$r$7

with the City Pulse Matrix defined by $r$8 (Kondor et al., 2015, Garcia-Palomares et al., 2017, Sarraute et al., 2017).

Mobility-centered formulations introduced additional latent or state-based representations. UrbanRhythm defines three attributes per cell and time slot—staying $r$9, leaving $k$0, and arriving $k$1—and treats each time slice as a three-channel city image. Multi-stage Saak features are reduced and clustered into city states such as Sleep, Home, Rush, Work, and Relax, so that urban dynamics become a state sequence $k$2. The lifestyle study instead represents each person by a non-negative visitation vector and factorizes the population matrix as $k$3 via NMF, where rows of $k$4 are 12 latent interpretable activity behaviors and rows of $k$5 are user-specific mixtures. In that formulation, a lifestyle is explicitly not a hard cluster but a mixture,

$k$6

and per-user mixture diversity is quantified by entropy. These two directions—state sequences and latent mixtures—extend UrbanPulse from “where and when” toward “what type of activity regime” (Song et al., 2019, Yang et al., 2022).

Recent deep-learning variants use graph and token representations. The 2025 UrbanPulse OD framework models each city at time $k$7 as a directed graph $k$8 whose nodes are POIs and whose edge weights are OD transition counts at $k$9 minutes. The target is a sparse OD matrix $f_{r,k}: K \to \mathbb{R}$0, and dynamic adjacency is row-normalized after self-loop insertion. A temporal graph-convolutional encoder produces node embeddings, edge tokens are constructed as

$f_{r,k}: K \to \mathbb{R}$1

and a Transformer decoder predicts future edge weights over a 3-hour horizon. This shifts the pulse representation from signatures over zones to directed flow forecasts over individual POIs (Yang et al., 23 Jul 2025).

3. Data modalities and sensing infrastructures

UrbanPulse has been instantiated on heterogeneous sensing substrates. Social-media variants use geotagged Flickr photos or Twitter posts as point events. The original topology-driven study used Flickr for New York City and San Francisco, and language-filtered geotagged tweets to compare cultural-community activity in NYC. Madrid’s Twitter study used geotagged tweets from January 2012 to December 2013, restricted to typical workdays and aggregated into unique users per 15-minute interval to reduce prolific-user bias. These sources privilege publicly expressed, socially mediated presence and are therefore informative but not population-complete (Miranda et al., 2016, Garcia-Palomares et al., 2017).

Telecom-based variants use broader mobile-network or CDR coverage. The four-city mobile-network visualization platform used ten months of aggregated operator activity, with per-antenna counts of calls placed, text messages sent, data downloaded and uploaded, and data requests at 15-minute resolution, later aggregated onto a regular grid approximating neighborhood scale. The Buenos Aires study used more than 200 million CDRs from about 4.95 million distinct users over five months, inferred each user’s home commune from weekday-night calling, and applied commune-specific census scaling factors ranging from 17.26 to 93.29. The lifestyle study analyzed 67 billion anonymized GPS pings from approximately 1.2 million opted-in devices across 11 U.S. CBSAs, with visits mapped to about 1.1 million Foursquare places and reduced to 248 venue-category fractions plus five temporal fractions per user (Kondor et al., 2015, Sarraute et al., 2017, Yang et al., 2022).

Street-level imagery introduced a different observational model. CityPulse for urban change curated roughly 931 locations and 10,878 images across Seattle, San Francisco, Oakland, Los Angeles, and Boston, using Microsoft Building Footprints for target sampling and the Google Static Street View API for historical imagery. The London perception protocol sampled roads at 20 m intervals, queried Street View metadata, and built a deduplicated corpus of 633,419 unique images covering 67% of London roads, with a stratified subset of about 25K images used for pairwise walkability ratings. UrbanFeel collected more than 4,000 multi-temporal single-view and panoramic street-view images across 11 representative cities from 2007 to 2024 and constructed 14.3K visual questions spanning static scene perception, temporal change understanding, and subjective environmental perception (Huang et al., 2024, Muller et al., 2022, He et al., 26 Sep 2025).

App-usage and real-time streaming systems pushed the sensing stack toward operational analytics. The social-vibrancy study used Orange NetMob23 service-level traffic over 77 days in 18 French metropolitan regions, at 100 m $f_{r,k}: K \to \mathbb{R}$2 100 m tiles and original 15-minute cadence, aggregating 68 services into 30 app categories and then into 12 two-hour bins per day. The CityPulse traffic pipeline simulated 11,000,000 traffic-related records including vehicle telemetry, GPS coordinates, and weather patterns, ingested them through Dockerized Kafka and ZooKeeper, processed them in Spark Structured Streaming, buffered results to local CSV, stored refined outputs in a central warehouse, and served them via Flask and React. Urban perception with gaze added yet another sensing layer by pairing 2,248 PP2 street-view images with 10,223 valid image–gaze recordings from 96 participants using a Tobii Pro Spectrum at 600 Hz (Collins et al., 2024, Teledjieu et al., 15 May 2025, Che et al., 1 May 2026).

4. Analytical methods and computational architectures

UrbanPulse methods span topological analysis, clustering, probabilistic scaling, matrix factorization, computer vision, streaming systems, and deep spatiotemporal learning. The original topology-driven framework depends on superlevel-set persistence of PL scalar functions: maxima are creators, saddles are destroyers, and persistence $f_{r,k}: K \to \mathbb{R}$3 measures saliency. The system clusters prominent maxima within distance $f_{r,k}: K \to \mathbb{R}$4 and supports cross-city comparison through beat-distance $f_{r,k}: K \to \mathbb{R}$5. In the mobile-network study, temporal signatures rather than topology are primary; neighborhoods are classified by similarity of typical-week signatures, with cluster interpretation supported by cross-city comparisons such as business clusters in New York and Hong Kong being more similar to each other than residential clusters (Miranda et al., 2016, Kondor et al., 2015).

Mobility and land-use variants rely on aggregation and statistical modeling. Buenos Aires constructs user-level LDMs, infers home by weekday-night argmax, and rescales by commune census populations before forming $f_{r,k}: K \to \mathbb{R}$6 and $f_{r,k}: K \to \mathbb{R}$7. Madrid fits per-slot OLS models,

$f_{r,k}: K \to \mathbb{R}$8

with reported $f_{r,k}: K \to \mathbb{R}$9 values of 0.763 in the morning, 0.683 in the afternoon, 0.613 in the evening, and 0.609 at night. UrbanRhythm computes multi-stage Saak transforms over three-channel mobility images, reduces them to 128 dimensions, performs Ward-linkage agglomerative clustering, and then discovers motifs in the resulting state sequence via a collision matrix and DBSCAN-based motif grouping. The lifestyle framework applies KL-divergence NMF,

$$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$0

with balanced 10,000-user-per-city basis learning and cross-city projection (Sarraute et al., 2017, Garcia-Palomares et al., 2017, Song et al., 2019, Yang et al., 2022).

Street-view and perception systems employ deep visual models and human-in-the-loop labeling. The CityPulse urban-change model is a Siamese architecture with DINOv2 (ViT-B/14), fusion by concatenating both embeddings and their element-wise difference, and binary cross-entropy training. It reports Accuracy $$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$1, Precision $$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$2, Recall $$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$3, and $$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$4 for fine-tuned DINOv2. The London perception protocol uses Microsoft TrueSkill to convert pairwise preferences into continuous image scores, and then trains ResNet-101 models with MSE loss; for walkability, the best model achieved MSE 2.58 and PCC 0.16, with the strongest transfer seed coming from the Place Pulse “beauty” model. UrbanFeel is a benchmark rather than a prediction system, but its evaluation pipeline is itself methodologically important: exact-match accuracy for closed-form questions, auxiliary-language-model semantic verification for verbose outputs, and semantic similarity scoring for open-ended answers. The gaze-guided perception framework uses fixation-token Transformers, AOI fusion through Mask2Former semantic labels, and patch fusion through a frozen ViT, all optimized with cross-entropy (Huang et al., 2024, Muller et al., 2022, He et al., 26 Sep 2025, Che et al., 1 May 2026).

Real-time and predictive architectures emphasize deployment constraints. CityPulse for traffic analytics is explicitly containerized, with Kafka/ZooKeeper for ingestion, Spark Structured Streaming in micro-batch style, temporary local CSV buffering, a Random Forest classifier with $$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$5, and a Flask/React serving layer. Stress testing reports throughput $$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$6 records/minute and average end-to-end latency $$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$7 seconds per 100,000-record batch, with only an approximately 10% latency increase under full-load single-push ingestion. The 2025 UrbanPulse OD predictor combines temporal convolutions, per-time graph convolutions, edge-token construction, Transformer self-attention, supervised pretraining on Los Angeles, selective cold-start adaptation on San Francisco, and PPO-based reinforcement-learning fine-tuning of the output layer. Its RL state is a 1560-dimensional summary, actions are 33-dimensional block-wise multiplicative scalings, and reward emphasizes weighted MSE and MAE on high-flow edges (Teledjieu et al., 15 May 2025, Yang et al., 23 Jul 2025).

5. Empirical regularities and cross-city findings

Across modalities, UrbanPulse research repeatedly identifies strong temporal regularity plus systematic local deviation. In the mobile-network signatures, business areas show pronounced weekday daytime peaks and low evenings and weekends, while residential areas show elevated evening activity and lower weekday daytime activity. Texting rhythms differ across cities: text messages peak in the morning in Hong Kong, in the evening in New York, and around midday in London. London also shows an abrupt evening decrease in cellular data traffic compared to New York and Hong Kong’s evening peaks, which the authors speculate reflects costly cellular data and switching to home Wi‑Fi at night. Event residuals are highly localized: Wimbledon raises activity in Merton, and Wembley’s district spikes around major matches (Kondor et al., 2015).

Social-media and CDR studies show comparable center–periphery temporal migration. In Madrid, activity zones related to offices, education, health, transport, and culture intensify during daytime, while residential areas dominate at night; retail rises toward evening, and parks remain comparatively stable. The OLS coefficients make these gradients explicit: retail increases from 0.000791 in the morning to 0.001333 in the evening, while residence increases through the day and peaks at 0.000386 at night. Buenos Aires shows weekday centripetal movement toward Commune 1, with the lightest weekday-noon diagonal element in the City Pulse Matrix at 24%, indicating strong inflow, whereas weekends invert the pattern and Commune 14 (Palermo) becomes a dominant attractor. Friday night and Saturday night profiles further reveal sustained nightlife extending into the early morning (Garcia-Palomares et al., 2017, Sarraute et al., 2017).

Mobility-state and latent-behavior studies reveal recurrent urban regimes that are not reducible to static land use. UrbanRhythm identifies interpretable states such as Sleep 1/2/3, Home, Rush 1/2, Work 1/2, and Relax 1/2/3 in Beijing, and analogous states in Shanghai; it also reports 59 motifs, including a canonical sleeping motif and distinct weekday, weekend, and holiday motifs. The lifestyle study finds that city dwellers’ behavior can be decomposed into 12 latent activity behaviors, that average mixture entropy is approximately $$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$8, and that these behaviors are “equally present across cities.” It also reports added explanatory power beyond demographics for experienced income integration, exploration, commute burden, travel distance, physical activity, and obesity-related indicators (Song et al., 2019, Yang et al., 2022).

Street-level and multimodal systems expose a different layer of regularity: physical change and perception do not track purely objective urban functions. CityPulse’s Seattle deployment acquired about 795,919 street-view images and detected 11,838 change points, with tract-level change rates showing significant correlations with ACS 5-year estimates of median household income change ($$f_{r,k}(p) = \sum_{x_i \in N(p) \cap X_{r,k}} \exp(- d(p, x_i)^2 / \epsilon^2),$$9) and population change ($\epsilon = 100 \text{ m}$0), but negligible linear correlation with construction permits. UrbanFeel shows that many MLLMs perform strongly on static perception and Time-Consistent Recognition, that Gemini-2.5 Pro reaches 65.9 overall versus a human 67.4, and that Temporal Sorting is notably harder, with Gemini-2.5 Pro at 52.1 versus a human 70.0. It also reports that single-view inputs outperform panoramas by 11.7% accuracy on average. The gaze-guided perception study shows that gaze alone already carries predictive signal, and that Gaze + Patch improves Macro-F1 over Image-only ViT by +1.8 for Wealthy, +1.4 for Safe, and +0.7 for Boring, while AOI-level analysis associates vegetation with higher perceived Wealthy and Safe and wall, sky, truck, or fence with lower Wealthy and Safe (Huang et al., 2024, He et al., 26 Sep 2025, Che et al., 1 May 2026).

Predictive UrbanPulse systems show that the paradigm is no longer limited to retrospective analysis. The CityPulse streaming traffic system reports Macro-averaged precision/recall/F1-scores above 0.96 across Low, Medium, and High congestion classes, with Accuracy and Macro F1 above 0.95 across 20 sequential batches despite one temporary dip. The cross-city OD predictor reports state-of-the-art performance on Los Angeles, with UrbanPulse achieving MSE 1.79 and MAE 2.08, outperforming Graph WaveNet, EGAT, and DCRNN; for LA$\epsilon = 100 \text{ m}$1SF transfer, RL+cold-start yields MSE 0.64 and MAE 1.08, which is the best reported adaptation result (Teledjieu et al., 15 May 2025, Yang et al., 23 Jul 2025).

6. Validation, applications, limitations, and open directions

Validation in UrbanPulse research is modality-specific and often hybrid. The four-city mobile-network study points to comparison with census-based data in related work and validates event detection qualitatively through Wimbledon, the UEFA final, and Christmas. Buenos Aires validates commune-level presence estimates against the ENMODO 2010 OD survey, reporting an average difference of 5% and largest deviations of 20% in Commune 1 in the morning and 11% in Commune 6 at noon. Madrid validates residential night-time Twitter distributions against the Register of Inhabitants with $\epsilon = 100 \text{ m}$2. UrbanRhythm validates state semantics through Shanghai app usage, CityPulse change detection reports performance approaching human-level annotation performance, and UrbanFeel uses two independent groups of ten geography-trained participants as human baselines (Kondor et al., 2015, Sarraute et al., 2017, Garcia-Palomares et al., 2017, Song et al., 2019, Huang et al., 2024, He et al., 26 Sep 2025).

Applications follow directly from these validation regimes. Reported uses include land-use inference and updating of official maps, comparative urban analytics across global cities, event detection and monitoring, transit and network operations, city-specific perception mapping, fine-grained assessment of physical urban change, social-vibrancy monitoring from app usage, real-time congestion prediction, and ultra-fine-grained OD forecasting for urban planning, transportation management, and public health. The original Urban Pulse interface also emphasizes precedent finding: pulses with similar beats can be compared across cities even when their land-use labels differ, as in Rockefeller Center and Alcatraz or Bryant Park and Mission Dolores Park (Miranda et al., 2016, Collins et al., 2024, Teledjieu et al., 15 May 2025, Yang et al., 23 Jul 2025).

Several misconceptions are explicitly corrected by the literature. First, UrbanPulse outputs are usually proxies rather than direct ground truth: geotagged Twitter users are a nonrandom slice of the population, CDR-based OD in Buenos Aires is a home-to-current-location presence matrix rather than trip-by-trip trajectory reconstruction, and mobile-operator feeds reflect the operator’s user base and network geometry rather than a census universe. Second, the paradigm is not intrinsically real-time: some systems are batch historical analyses, while others, such as the traffic pipeline, are engineered for near-real-time streaming. Third, subjective-perception scores are not neutral urban facts: UrbanFeel shows city-identity label sensitivity, and the gaze-guided perception work reports substantial inter-rater variability, especially for Boring (Garcia-Palomares et al., 2017, Sarraute et al., 2017, Teledjieu et al., 15 May 2025, He et al., 26 Sep 2025, Che et al., 1 May 2026).

The principal limitations recur across modalities. Data biases include smartphone ownership and opt-in selection bias, platform bias in Twitter or Street View, market-share effects in telecom feeds, coverage gaps in imagery, and temporal mismatch between behavioral and POI datasets. Some studies leave algorithmic details unspecified, such as exact similarity metrics or clustering parameters in the mobile-network clustering tool, stream-processing semantics in the traffic pipeline, or cross-city transfer robustness in gaze-based urban perception. The street-view and perception literature further emphasizes cultural contingency, viewpoint distortion in panoramas, and the risk of stigmatizing places when subjective scores are operationalized. These constraints imply that UrbanPulse systems are strongest when they quantify uncertainty, report coverage, aggregate to privacy-preserving spatial units, and fuse multiple modalities rather than relying on a single signal (Kondor et al., 2015, Muller et al., 2022, Collins et al., 2024, Teledjieu et al., 15 May 2025, He et al., 26 Sep 2025).

Current research directions indicate continued expansion rather than convergence to one canonical architecture. The literature points toward multimodal fusion of street view, mobility, social media, and satellite signals; uncertainty-aware sequence reasoning for temporal urban change; edge and cloud deployment for resource-constrained environments; map-based dynamic overlays for live traffic systems; temporal forecasting with LSTM or TCN in streaming mobility analytics; reinforcement-learning-based adaptation for cross-city flow prediction; and human-perception-centered modeling that combines scene content with attention patterns. Taken together, these directions suggest that UrbanPulse is evolving from a descriptive visualization framework into a general city-scale inference layer that links urban rhythms, built-environment change, and subjective experience within a unified spatiotemporal analytic vocabulary (Huang et al., 2024, Teledjieu et al., 15 May 2025, Yang et al., 23 Jul 2025, He et al., 26 Sep 2025, Che et al., 1 May 2026).