Using Computer Vision Approaches for Identifying Urban Greenspaces that Support Child Development

• In Event: Lightning Session: Methods and Tools of Analysis

Abstract

Past research has shown that urban greenspaces correlate with improved air quality, better mental health outcomes, physical activity, and cognitive development among children and adolescents. Urban greenspaces can be particularly relevant for the development of young children as they may provide various developmentally promotive affordances. Affordances have functional possibilities that children realize when they use them in ways that align with their goals and developmental stage. For example, for a child engaged in symbolic play, a bench can be a car, for a child being active, it may be a climbable structure, and for other children, it may be a place to sit and reflect. While developmentally promotive affordances may be found in a variety of urban spaces, there is almost no research that documents how their presence varies across urban parks and playgrounds. Extant research has documented persistent socioeconomic and racial inequalities in access to high-quality greenspaces, but there is less work that has shown whether this is also the case when we look within the landscape to examine its potential to promote child development.

In this work, we develop a new methodological workflow that characterizes urban parks and playgrounds based on the presence and types of developmentally promotive affordances. We begin by reviewing interdisciplinary literature around greenspaces, playspaces, and child outcomes spanning the fields of urban planning, landscape architecture, environmental psychology, and child development. This review identifies the affordances that should be present in developmentally supportive spaces. Then, to operationalize this theoretical framework, we develop a novel street-view-based computer-vision pipeline that bridges coarse Geographic Information System (GIS) park inventories with micro-scale amenity counts, which is a new methodological application for this type of research question. Using Google Street View (GSV) panoramas, our methodology first establishes a spatial baseline via a GIS audit that clips official shapefiles against target-park boundaries. We then deploy a systematic sampling approach to query GSV metadata at park centroids, capturing four orthogonal headings per panorama to provide comprehensive visual coverage of each site. Next, we translate our identified affordances into a hierarchical annotation schema comprising 20 fine-grained amenity classes. We utilize a zero-shot labeling approach combining two different object detection systems: YOLOv8, allowing us to detect pre-defined object classes and OWL-ViT, that accommodates text-based descriptions. We use these labelled data to fine tune a YOLO11s object detection model. To provide robust performance and eliminate near-duplicate data leakage, we implement a novel panorama-grouped validation split. Our model is cross-validated to quantify and mitigate known limitations such as panorama double-counting. Finally, we apply our approach to classify the urban greenspaces of various North Carolina cities.

Our work has theoretical and practical applications. Theoretically, it contributes to the development of new applications in urban data science for policy analysis and how they can be used to understand the relationship between the built environment and child development. Practically, our method can be applied by urban policy experts, urban planners, and landscape designers to advance the development of child-friendly cities and urban spaces.

Authors

• Agustina Laurito, University of Illinois, Chicago
  Presenting Author

• Anton Rozhkov, New York University
  Non-Presenting Co-Author

• Shu Yang, New York University
  Non-Presenting Co-Author