Emotional Landscapes: Mapping Urban Park Sentiments Using Natural Language Processing

Xia Gao (Penn) and Eric Delmelle (Lehigh, Penn)

August 21, 2025

Motivation:

  • Role of Urban Parks:
    • promoting physical activity
    • reducing mental health issues
    • offering a space for community engagement


  • Public Health Impact:
    • green spaces reduce stress
    • mitigate urban heat
    • improve air quality

example of two extreme parks (from ChatGPT)

Motivation:

  • Limitations of Traditional Surveys
    • resource-intensive, time-consuming (Dony & Feteke 2020)
    • results can become quickly outdated


  • Leveraging Social Media Data
    • real-time (Cui et al. 2021)
    • cost-effective
    • extensive data


  • However,
    • can we capture public sentiment on park use and perceptions (?)
    • is social media representative (!?)


Relevant Work

  • Social Media in Urban Studies
    • Google Reviews, Twitter, and Instagram are valuable data sources in urban research
    • in the context of parks –> visitor behavior and preferences
    • Marti et al. (2019), Chen et al. (2018) ,


  • sentiment and emotion analysis in urban spaces (Cui et al. 2021)
  • natural language processing [NLP] –> interpret text data
  • derive public sentiment and emotional insights
  • emotional cartography (Griffin and McQuoid 2012, Caquard and Griffin 2018, Acedo et al. 2022, Feng et al. 2024,Shukla and Pujara 2025)


Research Questions

  • What are the temporal, functional and spatial trends in emotional sentiments and visitation patterns at urban parks?

  • What are the features frequently mentioned in reviews for different park visiting emotions?

  • Methodology: NLP algorithm to detect emotions

  • Data: Google Reviews scraped for all parks in Philadelphia

Methodology

  • Sentiment Analysis
    • Hugging Face’s NLP transformer models to classify sentiment as positive or negative
  • Emotion Analysis
    • Categorize emotions into six types Inspiration, Relaxation, Engagement, Discovery, Frustration & Annoyance, and Sorrow
  • Data Processing
    • Emotions categorized using k-means clustering to identify meaningful groups

Data

  • Source
    • Google Reviews for Philadelphia parks, covering data from 2016 to 2024
  • Data Preparation
    • Scraping with Outscraper, data cleaning and integration with PPR to remove duplicates and ensure accuracy
  • Review Filtering
    • Only parks with over 25 reviews were analyzed, resulting in a sample of 92 parks and over 37,000 reviews

Data: spatial distribution of reviews

Data: temporal distribution (1)

Data: temporal distribution (2)

Data: What do these reviews look like?

snapshot of reviews for Wissahickon Valley Park

Data: example from a frequent (?) contributor

Review from Anna K.

Data: example from a less frequent (?) contributor

Review from Miranda Burg

Sentiment Analysis

  • Sentiment Scoring Approach
    • each review assigned a sentiment score (0-1 scale) and a sentiment label (positive / negative)
    • scores close to 0 indicating more neutral sentiment and scores near 1 indicating stronger sentiment
    • (Hauthal & Burghardt 2016)
  • Example
    • “Unfortunately this park is not what it used to be the area has become a known high traffic drug area but there is police getting the areas trying to do what they can but has a lot of work to go would not recommend taking your dog are children here lot of used syringes everywhere especially on the park in Grass area” - negative 0.97

Distilbert

Temporal variation of sentiment scores

Aggregated scores per park

Emotion Analysis (1)

  • Plutchik model


  • Emotion Categorization Process
    • pretrained model from Hugging Face (“roberta-base-go_emotions”)
    • classify reviews into a broader range of emotions

Emotion Analysis (2)

Emotion Analysis (3)

  • Choice of colors based on Kushkin et al. (2023)

Emotion breakdown (counts)

  • variation per year (left) and month (right)

Emotion breakdown per park type

Emotion mapping

Emotion mapping

Emotion mapping

Additional Research

  • Latent Dirichlet Allocation (LDA) — a topic modeling in NLP

Discussion and Limitations (1)

  • Representativeness
    • reviews may not capture the full demographic spectrum (Marti et al. 2019)
    • overrepresentation of younger or tech-savvy individuals


  • Temporal Bias
    • reviews are not always posted immediately after visits, potentially diluting real-time sentiment


  • Single Platform Limitation
    • using only Google Reviews could introduce bias
    • combining with data from other platforms (e.g., Instagram) may yield a broader picture

Discussion and Limitations (2)

  • Ethical Considerations
    • Data was collected with respect to privacy, without identifying individual reviewers


  • Contributor Weight/Reliability
    • reliability of sentiment may vary between platform “super contributors,” who post frequently and may have a more balanced perspective, versus casual users who leave only one or two reviews

References (1)

  • Acedo, A., González-Pérez, A., Granell, C., & Casteleyn, S. (2022). Emotive facets of place meet urban analytics. Transactions in GIS, 26(7), 2954–2974.
  • Shukla, S., & Pujara, T. (2025). Emotion mapping methods in urban design and planning: a systematic review. Cities & Health, 1–20.
  • Yang, C., & Zhang, Y. (2024). Public emotions and visual perception of the East Coast Park in Singapore. Urban Forestry & Urban Greening, 94, 128285.
  • Caquard, S., & Griffin, A. (2018). Mapping emotional cartography. Cartographic Perspectives, 91, 4–16.
  • Chen, Y., Liu, X., Gao, W., Wang, R. Y., Li, Y., & Tu, W. (2018). Emerging social media data on measuring urban park use. Urban Forestry & Urban Greening, 31, 130–141.
  • Cui, N., Malleson, N., Houlden, V., & Comber, A. (2021). Using VGI and social media data to understand urban green space. ISPRS IJGI, 10(7), 425.
  • Dony, C. C., & Fekete, E. (2020). Leveraging social media to track urban park quality. In Geospatial Technologies for Urban Health, 157–177. Springer.

References (2)

  • Feng, J., Zhao, B., & Shaw, S. L. (2024). The study of emotion in GIScience: status and prospects. Cartography and GIScience, 1–16.
  • Griffin, A. L., & McQuoid, J. (2012). At the intersection of maps and emotion: representing experience. Kartographische Nachrichten, 62, 291–299.
  • Hauthal, E., & Burghardt, D. (2016). Mapping space-related emotions from photo metadata. The Cartographic Journal, 53(1), 78–90.
  • Kushkin, A., Giordano, A., Griffin, A., & Savelyev, A. (2023). Cognitively congruent color palettes for mapping spatial emotional data. Cartographic Perspectives, 102, 38–62.
  • Martí, P., Serrano-Estrada, L., & Nolasco-Cirugeda, A. (2019). Social Media data in urban studies: challenges and limitations. Computers, Environment and Urban Systems, 74, 161–174.
  • Mullenbach, L. E., Baker, B. L., Benfield, J., Hickerson, B., & Mowen, A. J. (2019). Community engagement and perceived ownership of an urban park. Journal of Leisure Research, 50(3), 201–219.

🙏