Correlations of Socioeconomic and Urban Patterns Observed via an Interpretable Deep Learning Model
Urbanization is a great challenge for modern societies, promising better access to economic opportunities, but widening socioeconomic inequalities. Accurately tracking this process as it unfolds has been challenging for traditional data collection methods, but remote sensing information offers an alternative way to gather a more complete view of these societal changes. By feeding neural networks with satellite images, the socioeconomic information associated with that area can be recovered. However, these models lack the ability to explain how visual features contained in a sample trigger a given prediction.
Associate Professor Márton Karsai and his PhD student Jacobo Levy Abitbol (ENS Lyon/Inria) propose a solution to this problem in their paper published in Nature Machine Intelligence recently. They close this gap by predicting socioeconomic status from aerial images in five French metropolitan territories using Convolutional Neural Networks. Based on the trained models they propose an interpretable class activation mapping in terms of urban topology using gradient-weighted class activation mapping. Beyond this methodological leap, interestingly, they show that trained models disregard the spatial correlations existing between urban class and socioeconomic status to derive their predictions.
This contribution is timely and important as decision-makers are increasingly turning to technological solutions to inform or even devise their policies. The surge in demand for artificial intelligence (AI)-powered tools can be met as long as the models they rely on provide some accountable insights into their behaviour. This paper addresses some of these issues by building a joint dataset and deep-learning framework where SES predictions from aerial images can be interpreted in terms of urban planning. In doing so, it provides firm ground on which to study how a city’s local structure can be informative about the distribution of wealth in it. Furthermore, as new deep-learning architectures continue to emerge, it paves the way to build more advanced systems for SES inference, which can then be applied, analysed and interpreted using a common framework. Similarly to advances in medical imaging community, this work could enable the design of more sophisticated urban solutions that are poised to integrate expert knowledge from urban planners as useful prior knowledge to feed to deep-learning models.
Post by Márton Karsai