I'm working on an undergraduate thesis about building an image classifier that identifies potentially uninhabitable housing from visible features such as damaged roofs, missing materials, major wall cracks, neglected exteriors, or partial structural collapse. The planned result is a website with a dot-density map showing houses classified as potentially uninhabitable within a selected area, similar to interactive mapping projects that visualize urban vegetation or other features.
I'm considering two possible workflows. The first would have the program automatically explore street-level imagery from an imagery provider, inspect the roads and surrounding buildings, and classify each property. The second would be to first obtain the coordinates of available street-view images, download imagery from those locations, and then run the classifier on the resulting images.
Are either of these approaches practical? Which imagery providers offer suitable APIs or data licenses? How can I obtain the coordinates of available street-level images, and can the imagery vehicle or camera positions be retrieved? With 360-degree images, how would the system select views that point toward houses? Finally, since the camera coordinates represent the vehicle rather than the property, how could the detected house locations be estimated accurately?
I'm also concerned about creating a useful training dataset. My local government defines moderate damage as features such as large visible cracks, damaged or missing roofing, and badly deteriorated walls, while severe damage includes partial collapse, leaning walls, or a mostly collapsed or missing roof. I'd appreciate advice about overlooked technical, data, and licensing challenges.
3 Answers
The acquisition plan is plausible, but make sure you account for image dates. Street-level imagery for nearby properties may have been captured months or years apart, so a building’s apparent condition could reflect when it was photographed rather than a meaningful difference between properties. You’ll also need to handle camera perspective, occlusion, duplicate views, and the fact that one panorama can contain several buildings. Property footprints or building centroids from a separate map dataset can help associate detections with approximate house locations instead of using the vehicle’s position directly.
The imagery retrieval is probably the easier part, provided you use a provider with an API and follow its licensing and usage rules. The difficult part is defining and labeling “uninhabitable” consistently. A model needs a substantial set of representative examples, and the features visible from a street image may not be enough to determine whether a building is truly uninhabitable. Government records or locally collected images could help create a labeled dataset, but you’ll need clear criteria and careful validation.
For an undergraduate proof of concept, consider framing the output as “less likely to show visible damage” versus “more likely to show visible damage,” rather than claiming a definitive habitability judgment. You could use the model as a first-pass screening tool that highlights images for human review. Features such as broken edges, irregular shapes, missing roof sections, and inconsistent surfaces may provide useful signals, although the model will still be fooled by decorations, construction, vegetation, and unusual architecture.
I’m narrowing the labels to conditions that can actually be observed in an exterior image, using my local government’s damage criteria. I’m also trying to obtain reference photos from the government; otherwise I may need to build the dataset myself.