Cross-City Generalization of Air Quality Prediction: A Domain-Adversarial Learning Approach

Recent studies on urban air quality modeling have mostly prioritized predictive accuracy within individual cities, often overlooking the need for generalizable solutions. However, building and maintaining dedicated models for every city is unrealistic due to data limitations and resource constraints. To overcome this challenge, this work proposes a transferable modeling framework that integrates a Domain-Adversarial Neural Network with machine learning regressors. The goal is to extract domain-invariant features from a source city and apply them to structurally similar target cities without extensive retraining. This framework is evaluated using realworld sensor data from Paris, Madrid, Berlin and Helsinki, focusing on three key pollutants: NO2,PM2.5 and PM10. Model performance is assessed across various source-target city combinations to evaluate adaptability. Results show that the proposed approach transfers well among modest-similar cities even if using limited data.