Overview

Aim: estimate how much neighbourhood material deprivation is recoverable from street-level imagery alone

Method: Street view image embeddings (CLIP) and Alpha Earth embeddings used to predict deprivation

Results: Surprisingly good predictions! Pictures of houses seem to have the strongest signal

Paper (in review)

Asher, Comber, Golzari, Bui Quang, Kieu & Malleson (2026), Estimating Visible Deprivation through the Analysis of Street View Image Embeddings

Deprivation is partly visible

Deprivation = disadvantage relative to wider societal norms (Townsend, 1987)

Material deprivation is measured from administrative data – e.g. the English Indices of Multiple Deprivation (IMD)

Many signs are physically visible in the streetscape:

litter, graffiti, disrepair, housing condition … these co-vary with the conditions that produce deprivation (Sampson & Raudenbush, 1999)

But how much deprivation is visible?

Could we recover the IMD signal from images of the neighbourhood, rather than from socio-economic data?

GeoAI & image embeddings

Vision transformers turn an image into an embedding: a vector where visually/semantically similar images sit close together

Pre-trained once, reused across tasks – no hand-crafted features or task-specific training needed

We use CLIP (Contrastive Language-Image Pretraining) (Radford et al., 2021), a vision–language model that aligns images with text

In further work we start to examine the text associated with images of deprivation

Prior work links street imagery to crime, income, perceived safety, poverty and disorder – but mostly in US/China, on proxy indicators, rarely isolating the contribution of images alone

Research questions

Using Greater Manchester, UK as a case study:

RQ1. To what extent can area-level deprivation be predicted from visual imagery?

RQ2. Does the strength of the predictive signal vary across different kinds of images?

RQ3. Does the predictive signal vary by deprivation domain?

Two complementary sources tested for RQ1: street-level photography vs satellite imagery

Study area & data

Greater Manchester, UK – 3M+ people, ten conurbations

Deprivation: 2025 IMD at LSOA level (1,702 neighbourhoods), as national rank

Street View: points sampled on the OSM road network; 4 images (N/E/S/W) per point

18,897 points → ~75,600 images

Satellite: AlphaEarth Foundations embeddings (Brown et al., 2025) – 10 m, multi-sensor

Methods

CLIP embeddings calculated for each Street View image (512-dim)

AlphaEarth provides satellite embeddings (64-dim)

Embeddings are median-pooled to the LSOA

XGBoost models predict IMD rank from the embeddings

Street View embeddings are also clustered to compare the predictive strength of different types of scenes

RQ1: Street view v.s. satellite

How much IMD-rank variance does each source explain?

Ground-level cues are not readily visible from above

Remaining analysis uses street view only

RQ2: Do some scenes carry more signal?

Partition the image embeddings into k = 5 clusters of visually similar scenes (k-means)

Then fit a separate model per cluster. Clusters look like:

Cluster 1: greenery / rural

Cluster 2: commercial / industrial

Cluster 3: rural roads

Clusters 4 & 5: houses / residential

RQ2: 'Residential' scenes appear to carry the strongest signal

Residential scenes are far more informative (left).

Holds when controlling for sample size (right).

RQ3: Which deprivation domains are visible?

Separate models per IMD domain.

Most domains predicted similarly well (R² ~0.56–0.65)

Strong correlation between most domains, even if not 'visible'

Exception: Barriers to Housing & Services (R² = 0.39).

Possibly because houses more affordable in deprived areas?

Why so consistent across domains?

The IMD domains are highly correlated

Deprivation is spatially concentrated & mutually reinforcing

Visual character acts as a proxy for many domains

The model needn't ‘see’ educational attainment – it co-occurs with visually distinctive built environments

Housing & Services is the outlier: affordability may run inversely to the usual visual cues of deprivation

Conclusions

Off-the-shelf CLIP embeddings of Street View explain ~68% of IMD-rank variance, satellites only ~33%

Signal is uneven: scenes that appear residential carry far more than roads, greenery or industry

Broadly consistent across domains (bar Housing & Services)

Future work

Finer spatial/temporal resolution than administrative indices allow

Multimodal foundation model (street + satellite + POIs + …)

Linking embedding dimensions to visual features via CLIP text prompts