Model-Data Jointly Driven Method for Airborne Particulate Matter Monitoring

Abstract

In this paper we propose a novel model-data jointly driven (MDJD) method from a single picture for airborne particulate matter (APM) monitoring, towards assisting the decision-making for government and reducing the health risks for individuals. The MDJD method is mainly composed of three steps. First, we create a vector of .distance. as the model driven natural scene statistic (NSS) features through comparing the sparsity features that are extracted from one picture in five transform domains with their corresponding benchmark features that are derived by using a huge number of pictures with the extremely low APM concentrations in advance. Second, we produce a vector of .distance. as the data-driven NSS features through comparing the contrast-sensitive features that are chosen from hundreds of deep features with their associated benchmark features that are derived based on the same feature generation method as used in model-driven NSS features. Lastly, we fuse the aforesaid model- and data-driven NSS features by introducing a nonlinear regressor to estimate the APM concentration. Extensive experiments conducted on two large-size APM picture datasets validate the superiority of our proposed MDJD method over the state-of-the-art model-driven methods and data-driven methods by a sizable gain of 7.4% in terms of peak signal to noise ratio. Via a series of ablation studies, we can observe that fusing model- and data-driven NSS features is beneficial to improving the model's generalization and fitting abilities and leads to the gains of over 15.1% compared with using either type of features in isolation.