Estimating visible deprivation through the analysis of street view image embeddings

Material deprivation is often visible in the physical fabric of neighbourhoods in the form of litter, graffiti, abandoned buildings, etc. Recent advances in vision transformers and the widespread availability of street-level imagery make it possible to analyse these features systematically at scale. This paper assesses the extent to which area-level material deprivation, measured using the 2025 English Indices of Multiple Deprivation (IMD), can be estimated from visual imagery alone. Using Greater Manchester, UK, as a case study, CLIP embeddings are computed for approximately 75,000 Street View images and AlphaEarth embeddings are obtained for the corresponding satellite imagery. XGBoost models are trained on each source to predict LSOA-level IMD rank, and the street-level embeddings are further partitioned by k -means clustering to examine whether particular kinds of scenes carry stronger predictive signal. The street-level model explains approximately 68% of the variation in IMD rank, substantially outperforming the satellite model (33%). Importantly, the predictive signal is highly uneven across image types: residential scenes are considerably more informative than images of roads, greenery, or commercial and industrial areas. With the exception of the Barriers to Housing and Services domain, the signal is broadly similar across individual IMD domains, consistent with deprivation being spatially concentrated and the visual character of a neighbourhood appearing to act as a proxy for multiple dimensions of disadvantage simultaneously. Together, these results provide a quantitative measure of how much of an administrative, multi-domain deprivation index is recoverable from imagery alone, and identify the kinds of scenes – and the dimensions of deprivation – for which this visual signal is strongest.