SmartCamAirDetect: A Contrastive Approach for Probabilistic Ambient Air Pollution Estimation With Limited Images for Smart City Development

Effective air pollution monitoring is crucial for data-driven smart city initiatives. However, the sparse distribution of air quality monitoring stations (AQMSs) often results in localized measurements, which may not accurately represent broader air quality conditions. The cost of extensive AQMS expansion is prohibitive, whereas widely distributed stationary cameras offer a cost-effective alternative by providing high-density imagery for ambient air pollution estimation. Nonetheless, cameras distant from AQMSs or newly installed often lack sufficient labeled data, and conventional models trained on limited datasets tend to overfit, leading to substantial estimation errors. To address these challenges, we introduce a novel contrastive learning approach to enhance air quality estimation performance, tailored to stationary ground-level camera imagery under sparse labeling conditions. Unlike previous contrastive methods developed for satellite-based data, our model is specifically designed to handle the fixed scenery and high temporal variability unique to ground-level images. The model estimates pollution differentials between images and generates probabilistic pollution assessments by comparing a target image from a limited dataset to multiple images captured by other cameras with ample labels. To refine feature extraction, we synthesize scenery images to eliminate static background features, enabling the model to focus exclusively on pollution-related information such as sky appearance. The proposed method is evaluated across various pollutants using multiple performance metrics. Compared to the baseline deterministic model, our approach reduces the mean absolute error by 11.61%, decreases the symmetric mean absolute percentage error by 14.20%, and improves the Pearson correlation coefficient by 50.97%. Incorporating synthetic scenery images further reduces the pinball loss and Winkler score for probabilistic estimations by 5.38% and 5.81%, respectively. These results underscore the efficacy and robustness of our model for image-based air pollution estimation, highlighting its potential for fine-grained, AI-driven air quality monitoring in smart cities. In future work, we aim to improve the interpretability and robustness of the scenery synthesis and removal module under dynamic scenes, enhance contrastive pair selection via optimal sampling strategies, incorporate additional probabilistic baselines to better contextualize uncertainty estimation performance, and strengthen generalizability across diverse urban environments.