Remote sensing of air pollution is essential for air quality management and health risk assessment. Many machine-learning-based retrieval models have been established for estimating various kinds of air pollutants. These methods mainly aimed to estimate a single pollutant (single-pollutant approach). However, different air pollutants interact with each other and are highly correlated. Building a unified model and conducting a joint retrieval of multiple pollutant can make a better use of these connections and improve the model efficiency. This study proposed a physics-informed multi-task deep neural network (phyMTDNN) for the joint retrieval of six main air pollutants, i.e., PM2.5, PM10, SO2, NO2, CO, and O3. The relationships among these pollutants were used to design the physics-informed network structure and loss function. Top-of-atmosphere reflectance which can generate retrieval results at ultrahigh resolution was used as model input. Experiments for mainland China in 2019 showed that the proposed model successfully achieved simultaneous estimation of six air pollutants, with the cross-validated R2 for the six pollutants varying from 0.72 to 0.90. The contrast experiments proved that physics-informed network structure contributed to the most of the model performance improvement. Compared to the single-pollutant approach, phyMTDNN ameliorated the model accuracy on traces gases retrieval. Furthermore, the modeling efficiency was largely improved in that a lot of repetitive work was avoided and modeling method was optimized. The proposed new multiple-pollutant retrieval frame can be applied to various fields for multi-variate retrieval or estimation.
Keywords: Air pollutants; Deep learning; Joint retrieval; Multi-task learning; Physics-informed machine learning.
Copyright © 2022 Elsevier B.V. All rights reserved.