Introduction
Measuring "liveability"—the alignment between a city’s provisions and its residents' needs—is traditionally a slow, resource-intensive process involving large-scale social surveys. While remote sensing using satellite and aerial imagery has shown promise in creating liveability "snapshots," the true value for urban planners lies in longitudinal monitoring: the ability to track how neighborhoods improve or decline over decades.
However, a significant challenge arises: cities are not the only things that change over time; the technology used to photograph them evolves as well. This study investigates whether a deep learning model trained on data from a single year (2016) can reliably predict liveability in the past (2012) and the future (2020), or if the technical "noise" of image acquisition—known as domain shift—renders such comparisons unreliable.
Methods and Data
The researchers utilized a Convolutional Neural Network (CNN) optimized for urban feature extraction. The model was trained using 25cm resolution, cloud-free aerial imagery from 2016, paired with ground-truth scores from the Leefbaarometer (LBM). The LBM is a sophisticated Dutch government dataset providing liveability scores for 100m x 100m grid cells based on over 100 indicators across five domains: Environment, Housing, Amenities, Social Cohesion, and Safety.
Following training, the model was tasked with predicting liveability for 2012 and 2020 without further modification. Evaluation was conducted using Kendall’s $\tau$ (tau), a coefficient that measures how well the model ranks neighborhoods relative to one another, which is often more critical for policy decisions than absolute numerical accuracy. The study focused on two cities: Amsterdam (part of the training set) and Eindhoven (an unseen test city).
Key Results and the Challenge of Domain Shift
The study revealed a phenomenon termed the "Familiarity Trap." Surprisingly, the model performed worse at generalizing over time in Amsterdam—the city it was trained on—than in Eindhoven. In Amsterdam, the correlation (Kendall’s $\tau$) plummeted from 0.656 in the training year (2016) to approximately 0.33 in 2012 and 2020. Conversely, in Eindhoven, the model remained more consistent, maintaining a correlation between 0.47 and 0.53 across all years.
This discrepancy is attributed to domain shift. Differences in sensors, sun angles, and image processing between years confused the AI. For instance, the 2020 model for Amsterdam severely underpredicted the liveability of the city center, showing "unrealistic" drops in quality that do not reflect actual urban decay but rather changes in how the images were captured. The research confirms that while absolute liveability scores are difficult to predict across time, maintaining a stable ranking of neighborhoods is more achievable in cities the model has not previously "memorized."
Conclusion and Future Directions
The primary hurdle for AI-driven urban monitoring is distinguishing between actual physical change and variations in image acquisition. To transition from static snapshots to true time-series monitoring, the authors suggest several future paths:
- Domain Adaptation: Implementing techniques like adversarial training to force models to ignore sensor-specific features.
- Image Normalization: Developing sophisticated methods to align the visual style of images across different years.
- Temporal Modeling: Utilizing architectures that inherently process sequences of images over time rather than single snapshots.
Summary updated