Background

Urban parks provide essential ecosystem services and recreational opportunities that contribute to physical health, mental well-being, and social interaction in dense metropolitan areas. Evidence-based park planning requires accurate and fine-grained measurements of visitor numbers, spatial distribution patterns, and movement trajectories. Conventional data collection approaches—such as manual observation, surveys, and GPS tracking experiments—are labor-intensive, costly, and difficult to implement continuously at scale. Although commercial mobility datasets derived from smartphones provide large-scale behavioral information, they are typically proprietary and financially restrictive. Meanwhile, geotagged social media data have been widely used to estimate recreational visitation; however, these data are often sparse and insufficient for capturing microscale spatio-temporal and trajectory-level behavior patterns.

With the proliferation of Web 2.0 and volunteered geographic information (VGI), volunteered street view imagery (VSVI) platforms such as Mapillary offer a novel open-data source with global coverage. Unlike traditional social media posts that consist of isolated image points, VSVI data are organized into sequential, geotagged photo trajectories, inherently embedding movement information. This structural advantage suggests potential for behavioral analysis at fine spatial and temporal resolutions. Nevertheless, VSVI contributions are generated by self-selected individuals and often dominated by a small number of highly active users. Consequently, it remains unclear whether VSVI contribution behavior reflects general park-use behavior or primarily captures the habits of specific contributors. Furthermore, given the spatial heterogeneity of VSVI contributions and the potential role of data volume in improving data quality, we hypothesize that the statistical validity of such data depends on contribution volume.

Methods

This study empirically evaluates whether VSVI can serve as a reliable proxy for park-use behavior and investigates whether its effectiveness increases with higher levels of contribution. The analysis was conducted in 49 metropolitan parks located in Tokyo’s 23 wards, one of Japan’s major hotspots of Mapillary activity. Three dimensions of park-use behavior were examined: (1) estimation of park visitor counts, (2) spatio-temporal distribution patterns, and (3) travel-path characteristics. Two reference datasets were employed. First, official annual park visitor counts (2019) published by the Tokyo Metropolitan Government were used to validate visitation estimation. Second, commercial Real People Flow Data (May 2023) provided anonymized GPS trajectories from approximately 869,840 users (about 6% of Tokyo’s population), generating around 65 million records per day. After preprocessing—restricting to walking trips within park boundaries and correcting for temporal gaps and cross-park trajectories—523,057 PeopleFlow trajectories were retained for analysis. VSVI metadata were collected from Mapillary (2014–March 2024), yielding 242,534 cleaned image points organized into 2,007 sequences contributed by 45 users within park areas.

Results

To assess visitation estimation capacity, Spearman’s rank correlation analyses were conducted between official visitor counts and four VSVI contribution indicators: number of image points, number of sequences, number of photo-user-days (PUD), and number of creators. When considering all parks collectively, none of the indicators showed statistically significant correlations with official visitor counts. However, when parks were grouped based on mean PUD into low- and high-contribution categories, moderate and statistically significant correlations emerged in the high-PUD group for PUD and number of creators (p < 0.05), whereas low-PUD parks showed no significant relationships. These results indicate that VSVI data may not universally estimate visitation levels but can approximate visitor counts in parks with sufficiently dense contribution activity.

Spatio-temporal distribution patterns were compared at a 20 m × 20 m grid resolution using three indicators derived from both datasets: point counts, dwell time per trip, and average speed. When pooling all parks, statistically significant but weak positive correlations were observed between VSVI and PeopleFlow data for all indicators (p < 0.001). However, park-specific analyses revealed substantial heterogeneity. The proportion of parks exhibiting significant positive correlations was markedly higher in the high-PUD group, particularly for spatial indicators such as point density and dwell time. Improvements for average speed were comparatively moderate. These findings suggest that increased contribution volume mitigates individual behavioral bias and enhances representativeness at the aggregated spatial level.

Travel-path characteristics were further compared at the park level, including temporal features (start time, end time, duration), standardized movement range features (travel distance, standard deviation ellipse (SDE) area, and SDE perimeter normalized by park area), and behavioral features (average speed and deviation rate). Strong and highly significant correlations were identified for standardized travel distance, SDE perimeter, and SDE area, indicating a high degree of consistency between VSVI-derived trajectories and reference mobility data in terms of movement range characteristics. Average speed showed moderate positive correlations. In contrast, temporal features exhibited no significant relationships, suggesting fundamental differences between photo-sharing behavior and general recreational timing patterns. Group-specific analyses indicated that correlations were generally stronger in parks with a higher number of VSVI trips, particularly for movement range indicators, although variability remained across parks and metrics.

Conclusions and implications

Overall, the results demonstrate that VSVI can serve as a meaningful proxy for selected aspects of park-use behavior, especially spatial distribution and movement range characteristics. However, its reliability is strongly conditioned by contribution volume. Parks with higher contribution density—often scenic or tourism-oriented sites—show substantially stronger alignment with commercial mobility data. Low-contribution parks exhibit weaker or inconsistent relationships, underscoring the importance of sufficient data accumulation in VGI-based behavioral inference. For the FOSS4G community, this study demonstrates how crowdsourced open street-level imagery can complement proprietary mobility datasets in large-scale urban behavioral analysis. By empirically examining the relationship between contribution volume and analytical validity, the study contributes to ongoing discussions on data-quality growth, representativeness, and the practical integration of open geospatial data into reproducible urban analytics workflows. While limitations related to GPS accuracy, timestamp uncertainty, and contextual heterogeneity remain, the results highlight the potential of open, crowd-sourced street-level imagery as a scalable complement to traditional data sources in spatial planning and geospatial research.