When places look alike but communities disagree: an empirical guide to transferring urban livability models across cities
This paper introduces a Source Applicability Typology and a stepwise framework to guide the transfer of urban livability models across diverse cities by demonstrating that cross-city transfer failures stem primarily from differing community perceptions of similar environments rather than unfamiliar urban features, thereby enabling more scalable and context-aware AI planning.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Cities are more than collections of buildings and streets; they are living spaces defined by how the people who live there feel about them. Urban planners and policymakers have long sought to measure "livability," a concept that blends objective factors like air quality and access to parks with subjective feelings of safety, comfort, and belonging. In deprived neighborhoods, where data is often scarce and conditions are harsh, understanding these local perceptions is vital for creating fair and effective improvements. For years, researchers have used a method called citizen science to map these feelings: they show residents pairs of images from their own neighborhoods and ask which one looks like a better place to live. By aggregating thousands of these simple choices, they can build a map of perceived livability. However, this process is slow and expensive because it requires gathering new data and training new computer models for every single city. The natural question arises: if a computer learns what makes a neighborhood livable in one city, can it simply be applied to another?
A team of researchers from the University of Twente and the Universidad Pública de Navarra set out to answer this by testing whether models trained in one city could successfully transfer to others. They focused on seven cities across Africa and Asia, including Nairobi, Accra, and Mumbai, using satellite radar images and thousands of resident comparisons. Their goal was to determine the conditions under which a model learned in one place could be adapted to another without losing the specific, local insights that make the data meaningful. They discovered that the answer is not a simple yes or no, but depends entirely on why a model might fail.
The researchers found that the most common reason a model fails to work in a new city is not because the new city looks different. It is not that the buildings, vegetation, or street layouts are so unfamiliar that the computer cannot recognize them. Instead, the failure usually happens because the people in the new city value familiar-looking places differently. Two neighborhoods might look identical in a satellite image, yet residents in one city might consider them highly livable while residents in another city find them undesirable. This difference in human judgment, which the authors call "cultural disagreement," is the primary barrier to transferring these models. A computer model trained on the first city's preferences will confidently predict the wrong livability score for the second city, even if the physical environment is perfectly familiar to the algorithm.
This finding challenges a common assumption in artificial intelligence: that simply feeding a model more data from more diverse locations will make it smarter and more adaptable. The researchers tested this by combining data from multiple cities to create a massive, multi-city source model. They expected this larger pool of information to help the model generalize better to new targets. Instead, they found the opposite. A model trained on a single, well-matched city consistently outperformed models trained on large aggregates of data from many cities. When data from different cities with conflicting preferences are mixed together, the model becomes confused, diluting the specific insights needed for any single location. In the context of urban livability, more data does not mean better results; a single, carefully chosen source is far more effective than a broad, generic one.
To help planners navigate this complexity, the team developed a step-by-step guide for deciding when and how to transfer models. The first step is to collect a small amount of local data in the new city—just a few hundred comparisons—to see if the source model agrees with the local residents. If the model and the residents disagree on how to judge familiar-looking places, the model is not ready to be used, no matter how much it "knows" about the physical environment. If the initial agreement is good, the next step is to look at the physical coverage: does the source model have experience with the types of buildings and landscapes found in the new city? If the physical features are familiar, the researchers found that adding more local data can significantly improve the model's accuracy. However, if the physical features are completely new to the model, adding local data will not help much, and a full new study would be necessary.
The study also identified the most efficient way to collect that necessary local data. When updating a model, it is better to ask residents to compare specific types of places where the model is likely to be wrong, rather than asking random questions. This targeted approach, known as informed sampling, allows planners to get the most value from a limited number of comparisons. The researchers found that this strategy is most effective when the number of comparisons is moderate—enough to provide a signal, but not so many that random sampling would eventually cover the same ground.
Ultimately, this work provides a practical framework for scaling up participatory urban planning. It shows that artificial intelligence can make these assessments more efficient, but only if it respects the unique cultural context of each community. The key is not to assume that what works in one city works everywhere, nor to rely on massive datasets to solve the problem. Instead, it requires a careful, diagnostic approach: check if the local community agrees with the model's initial instincts, and if they do, use that agreement to guide where new data is needed. By doing so, planners can use technology to extend the reach of citizen science without losing the essential, local knowledge that defines a livable city.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.