MARVAir: Meteorology Augmented Residual-based Visual Approach for Crowdsourcing Air Quality Inference

Air pollution has become a prominent problem in citizens’ everyday life. Since the weather stations are not densely distributed, it is difficult to get fine-grained PM2.5 (Particulate Matter <2.5 μm) data. We propose MARVAir to address limitations of traditional PM2.5 prediction mechanism. First, we use crowdsourcing and spiderbots to fetch visual and meteorological dataset respectively. Second, a ResNet based visual core is designed to learn the image data, and an 1D-CNN based meteorology core is deployed to tune the inference. Besides, we use decision-level fusion mechanism to unite the sub-models and provide precise yet everywhere available fine-grained PM2.5 inference. In addition, cloud-side model training is also proposed to restrict local energy consumption. Evaluation on dataset collected at 8 sites over nearly 2 years suggests that, MARVAir achieves a precision of 98.8 %, a recall of 99.0 %, and an F1 score of 98.9 % under various air conditions, which notably exceed baseline solutions.