Same Greenery, Different Green View Index: Inconsistent Greenery Assessment in Semantic Segmentation and Street View Imagery

Abstract

Urban greenery is paramount to well-being across physical, psychological, and societal functions. Advances in large street view imagery (SVI) datasets and semantic segmentation models have enabled mapping greenery close to the human experience. However, there are significant inconsistencies in the semantic class combinations (e.g., tree, grass, bush) used to define greenery, leading to a lack of standardization and comparability across studies. To address this variability, our study compared different semantic class combinations using an SVI dataset covering 143,899 locations in Helsinki, Finland. We identified semantic combinations from recent studies that use classes from ADE20K and Cityscapes to map street-level greenery. Next, we segmented the image dataset for Helsinki, and compiled greenery values (Green View Index, GVI) for every semantic combination. Finally, we analyzed GVI differences at both city-wide and local levels, by using rank tests, local Moran’s I, spatial regression, and land-use data. We found significant differences between the semantic combinations we tested. These resulted in varied mean GVI values for Helsinki ranging from 25.3 to 40.7. There was also significant spatial variation and clustering, with larger discrepancies in green areas of the city and smaller ones in denser urban areas. The findings show variations in urban greenery assessments, which can be traced to differing definitions of greenery. We call for critical assessment and increased attention to detail when mapping street-level urban greenery. Highlighting potential incomparability between studies, the findings aid understanding of methodological consequences of decisions for researchers and stakeholders working on greenery exposure and street view imagery topics.