Social media analysis reveals environmental injustices in Philadelphia urban parks

Abstract

The United Nations Sustainable Development Goal (SDG) target 11.7 calls for access to safe and inclusive green spaces for all communities. Yet, historical residential segregation in the USA has resulted in poor quality urban parks near neighborhoods with primarily disadvantaged socioeconomic status groups, and an extensive park system that addresses the needs of primarily White middle-class residents. Here we center the voices of historically marginalized urban residents by using Natural Language Processing and Geographic Information Science to analyze a large dataset (n = 143,913) of Google Map reviews from 2011 to 2022 across 285 parks in the City of Philadelphia, USA. We find that parks in neighborhoods with a high number of residents from historically disadvantaged demographic groups are likely to receive lower scores on Google Maps. Physical characteristics of these parks based on aerial and satellite images and ancillary data corroborate the public perception of park quality. Topic modeling of park reviews reveal that the diverse environmental justice needs of historically marginalized communities must be met to reduce the uneven park quality-a goal in line with achieving SDG 11 by 2030.

Figure 1

Relationships between three demographic groups (education, income, and race) and their park score at the census tract level, measured as the average Google Map score of all parks within 800 m of a tract’s residential areas. These bivariate maps can be used to identify areas in need of park improvements due to having low park scores and high concentrations of certain demographic groups. A complete set of bivariate maps for all the significant demographics included in this study can be found in Supplementary Information Fig. 1.

Figure 2

Correlation coefficients between socioeconomic variables and park score determine which variables are positively or negatively associated with park scores and the strength of those relationships. P-values of the correlation coefficients are shown as measures of statistical significance, with * representing a p-value of less than 0.1, ** a p-value of less than .05, and *** a p-value of less than 0.01.

Figure 3

Eight review topics representing common park features that influence overall park score. (a) Displays the topic most likely to be mentioned by users in all of the parks within 800 m of a census tract. These dominant topics for each census tract can be compared to (b) that displays the average park score based on the number of stars given by users to all parks within 800 m of a census tract.

Figure 4

The probability that a topic is mentioned in reviews of parks that are accessible to census tracts with both advantaged and disadvantaged demographic compositions. Advantaged demographics tracts are tracts in which demographics positively correlated to park score (over 65, White, median income, bachelor’s degree, and graduate degree) are in the upper 90th percentile (n = 99). Disadvantaged demographics tracts are tracts in which demographics negatively correlated to park score (receive supplemental nutritional assistance, less than a high school degree, high school degree, under 9, disability, no health insurance, Black, and some college) are in the upper 90th percentile (n = 175).

Figure 5

Heat maps showing the relationships between all of the demographic variables in a census tract and how likely a topic is to be mentioned by parks accessible to that census tract. Panels show how different demographic groups are mentioning a review topic in (a) a negative review or (b) a positive review.

Figure 6

A heatmap of correlation coefficients between the likelihood of each of the eight perceived topics being mentioned for a park and the occurrence of each observed variable within a park.