A multi-view machine learning approach for estimating PM_2.5 concentrations from smartphone photographs

Accurate and efficient estimates of fine particulate matter (PM_2.5) concentrations and associated exposure from smartphone photographs can provide the public with personalized risk information, helping to raise environmental risk awareness and reduce health risks. Existing studies are limited to estimating PM_2.5 concentrations from single-view smartphone photographs, where the spatiotemporal mismatch between photo capture and PM_2.5 measurements during data collection often introduces systematic bias that could otherwise be avoided. To address this, we construct a rigorously spatiotemporally registered benchmark dataset of PM_2.5 measurements and multi-view smartphone photographs for two Chinese provinces (Hebei and Fujian) between 2022 and 2024. For each province, we use the XGBoost model to relate PM_2.5 measurements to image features that we extract from multiple semantic segmentation regions within the multi-view smartphone photographs. The proposed method achieves R² values of 0.96 and 0.89, and RMSE values of 6.82 and 3.63 μg/m³ on the independent test sets for the Hebei and Fujian provinces, respectively, which consistently outperforms alternative methods that rely on single-view smartphone photographs or overall image features. Key predictive features include image color extracted from vegetation regions, as well as color and local binary pattern extracted from sky regions. Overall, our work advances image-based air quality monitoring by providing a multi-view machine learning approach and the first rigorously spatiotemporally registered dataset of PM_2.5 measurements and smartphone photographs, which enables more reliable PM_2.5 estimates and highlights the role of semantic image features in predictive performance.