Abstract
Air quality is a major public health concern, while particulate matter (PM) measurements remain limited by the spatial coverage and high operational costs of reference monitoring stations. This scoping review synthesises the literature on image-based PM estimation from outdoor photos as a potential low-cost approach to increase PM concentration sampling coverage. Methods are classified by input data requirements and modelling strategy, with summaries of performance and implementation resources. Results suggest that deep learning models generally exhibit the strongest predictive capability, although cross-study comparisons remain indicative, as they are limited by the availability of comparable metrics and by differences in geographic application contexts. Machine learning and physics-based methods are generally less accurate but more interpretable and computationally efficient, while hybrid approaches offer a promising trade-off between accuracy and operability. Nevertheless, most applications remain below reference regulatory accuracy levels. Training datasets are often limited in size and coverage, benchmarks are lacking, and code and data are not always available. Future work should prioritise multi-site datasets, improved hybrid models, and open resources to support reproducible model assessment and reuse.