Estimating Ground‐Level PM_2.5 by Fusing Satellite and Station Observations: A Geo‐Intelligent Deep Learning Approach

Abstract

Fusing satellite observations and station measurements to estimate ground‐level PM2.5 is promising for monitoring PM2.5 pollution. A geo‐intelligent approach, which incorporates geographical correlation into an intelligent deep learning architecture, is developed to estimate PM2.5. Specifically, it considers geographical distance and spatiotemporally correlated PM2.5 in a deep belief network (denoted as Geoi‐DBN). Geoi‐DBN can capture the essential features associated with PM2.5 from latent factors. It was trained and tested with data from China in 2015. The results show that Geoi‐DBN performs significantly better than the traditional neural network. The out‐of‐sample cross‐validation R2 increases from 0.42 to 0.88, and RMSE decreases from 29.96 to 13.03 μg/m3. On the basis of the derived PM2.5 distribution, it is predicted that over 80% of the Chinese population live in areas with an annual mean PM2.5 of greater than 35 μg/m3. This study provides a new perspective for air pollution monitoring in large geographic regions.

Estimating Regional Ground‐Level PM2.5 Directly From Satellite Top‐Of‐Atmosphere Reflectance Using Deep Belief Networks

“…Meanwhile, the site‐based CV results of the AOD‐PM modeling report a larger decrease (from 0.79 to 0.72 for R 2 ) than the Ref‐PM modeling. It is worth noting that the site‐based CV approach, which uses a spatial hold‐out validation strategy, can reflect the spatial predictive power more adequately (T. Li, Shen, Yuan et al, ). Hence, the results indicate that the Ref‐PM modeling approach has a superior spatial predictive power than AOD‐PM modeling.…”

“… where f () means the estimation function. S ‐ PM 2.5 , T ‐ PM 2.5 , DIS denote the geographical correlation of PM 2.5 (see our previous study for details of their calculation, T. Li, Shen, Yuan et al, ). S ‐ PM 2.5 and T ‐ PM 2.5 take the spatial and temporal autocorrelation of PM 2.5 into consideration, and DIS is incorporated to reflect the spatial heterogeneity of unevenly distributed stations.…”

“…In this study, two RBM layers with 15 neurons in each RBM layer were chosen (Text S2). For details of the DBN model, we refer the readers to T. Li, Shen, Yuan et al (). The procedure of this model for Ref‐PM modeling is divided into three steps, which are shown in Figure .…”

“…Hence, the estimation of ground‐level PM 2.5 directly from TOA reflectance is highly complex, and the conventional models may encounter some challenges. Deep learning, which is a further development of neural networks, has been used for time series prediction (Ong et al, ) and satellite AOD‐based estimation (T. Li, Shen, Yuan et al, ) of ground‐level PM 2.5 . The deep learning models have shown the ability to effectively predict/estimate ground‐level PM 2.5 , which can be attributed to their ability to fit nonlinear and complicated relationships (Hinton et al, ; LeCun et al, ).…”

REAL-TIME AND SEAMLESS MONITORING OF GROUND-LEVEL PM2.5 USING SATELLITE REMOTE SENSING

“…Additionally, the geographical correlation of PM2.5 were incorporated into the DBN model (Geoi-DBN) (Li et al, 2017a). Because the nearby PM2.5 from neighbouring stations and the PM2.5 observations from prior days for the same station are informative for estimating PM2.5.…”