Is a (satellite) image worth a thousand data points? Comparing machine learning approaches to predict environmental and social inequalities in England

Accurately mapping environmental and social inequalities at fine spatial scales is critical for urban policy, yet the data required are often costly and infrequently updated. Vision foundation models can extract information directly from satellite imagery, offering a rapid and scalable alternative. We compare three modelling pipelines for predicting two contrasting indicators of urban inequality - air pollution and house prices - across England on a fine hexagonal grid: regression models trained on structured features of form and function (census, land cover and urban morphology), models trained on 128-dimensional image embeddings from a geospatial foundation model and a hybrid of the two, each evaluated with and without coarse regional context. Structured features achieve the best overall accuracy, but image embeddings become competitive once regional context is added - most clearly for air pollution, where image-only models reach a median R2 of 0.78 (0.85 with regional context), indicating that the embeddings capture genuine image signal. For house prices the picture is more cautious: image-only models achieve a median R2 of around 0.58; however, much of the embeddings' apparent gain reflects coarse spatial location rather than image content. Off-the-shelf satellite embeddings, while not yet surpassing data-intensive approaches, are a promising low-cost complement, particularly for rapid, large-scale, or repeated analyses and in settings where traditional data are limited.