Using Google Street View imagery, we explore the capabilities of a multimodal large language model and semantic segmentation to construct neighborhood measures that accurately: (i) identify the co-location of sustainability deficits; and (ii) recover effects of place-based interventions. We derive poverty and tree canopy measures using GPT-4o in a reason-then-estimate pipeline, and a conventional semantic-segmentation model, comparing both to authoritative benchmarks. We then estimate spatial autoregressive models to quantify associations between historical redlining and our geospatial measures. Both approaches reproduce the expected pattern–higher poverty and lower canopy in historically redlined areas relative to non-redlined areas–with GPT-4o estimates statistically indistinguishable from benchmark results. The findings highlight MLLMs’ potential to expand measurement capacity toward an urban policy-intelligence framework for sustainability monitoring and place-based evaluation by city agencies and community partners.
Howell et al. (Mon,) studied this question.