Public Street View Images Predicted Property Neglect Before Code Violations

A DirtSignal preprint found that freely available aerial and Street View imagery can flag residential neglect months before local code-enforcement records are filed. The study also suggests smaller, task-specific models and cross-metro transfer can outperform larger foundation-model approaches for this use case.

Why it matters: - Public imagery can reveal property decline before city records do, which could help researchers, investors, and local governments spot risk earlier. - The finding matters because imagery is widely available across U.S. parcels, while digital code-enforcement records are not. - The work suggests some property-risk signals can be detected without local training data.

What happened: - DirtSignal published a preprint titled “Seeing Neglect: Measuring the Predictive Value of Public Aerial and Street-Level Imagery for Property Code Violations.” - The paper studied whether freely available aerial and Street View imagery could predict residential code-enforcement violations. - The study used dated code-enforcement records from Jacksonville, Florida. - The paper is archived as a SocArXiv preprint at 10.31235/osf.io/4u83q_v1. - DirtSignal also released an anonymized benchmark dataset and reproducible code on Zenodo, DOI: 10.5281/zenodo.20672748.

The details: - A Street View image of a property showed a statistically significant signal of a future first citation. - The signal was strongest when the image was taken within six months before the citation. - The signal remained significant up to one year before the citation. - Some properties showed visible neglect for years before any code-enforcement record appeared. - A four-number greenness measure out-predicted 768-dimensional foundation-model embeddings on the aerial signal. - A 2-billion-parameter vision model matched a larger model on the task. - The paper said the gains came from matching the question to the camera angle and combining views. - The study compared each violating property against a non-violating neighbor about 90 meters away. - Predictive accuracy plateaued around ROC-AUC of roughly 0.6 to 0.7. - The authors said that ceiling likely reflects noise in complaint-driven, uneven code-enforcement records. - A street-level detector trained with no local data transferred to Broward County without retraining.

Between the lines: - The results point to a practical advantage for simple, task-specific signals over bigger general-purpose AI models. - The cross-metro transfer result hints that public imagery may be useful where local enforcement data are sparse or missing. - The modest accuracy range suggests the method is useful for ranking and triage, not for making final enforcement decisions. - The findings depend on a single-source preprint, so the results should be viewed as early research rather than settled consensus.

What's next: - Other researchers can now test the benchmark dataset and code on new markets and property types. - DirtSignal said it will continue offering custom property-risk studies for regions outside its standard coverage. - The company’s broader product set includes lead feeds, dashboards, weekly briefs, and exports built from code-enforcement cases, liens, clerk filings, hearings, demolition records, and related public records.

The bottom line: - Public Street View and aerial imagery can surface property neglect before cities formally document it, and the clearest gains came from matching the model to the view, not from scaling up the model.