Crowdsourced Image Driven PM2.5 Estimation based on Hybrid 3-Channel Feature Map

Industrialization has resulted in a relatively high airborne PM2.5 concentration in most developing areas, causing severe consequences due to its physico-chemical properties. In this paper, we propose a crowdsourced image driven PM
2.5
concentration estimation approach based on meteorological en-hanced 3-channel feature map. To fulfill this framework, we first use dark channel prior and Rayleigh's law of atmospheric scattering to correct the proportion of the sky area in the crowdsourced images and then construct hybrid 3-channel feature maps. We then use deep learning models to perform feature extraction on hand-engineered feature maps and provide fine-grained PM2.5 concentration estimation. We also conduct a series of experiments to evaluate the performance of our model. Evaluation on dataset collected at 8 sites over nearly 2 years demonstrates that, our system achieves an MAE of 23.27
$\mu\mathrm{g}/\mathrm{m}^{3}$
, outperforming baseline solutions.