About IGSNRR
News
Research
Key Labs
People
International Cooperation
Education & Training
Publications
Facilities
Journals
Library
Scientific Database
Hosted Societies
Eye on Chinese Geography
Join Us
Links
Location: Home > Research > Research Progress
Inversion of Missing Satellite Aerosol Parameters and Spatiotemporal Reconstruction of Ground NO2 Using Full Residual Deep Learning
Update time: [January 11, 2021]
Close
Text Size: A A A
Print

Although remote sensing provides useful data products for global monitoring of air pollutants, the extensive non-random missingness due to cloud contamination or high surface reflectance severely limits their applicability. As a critical traffic-related air pollutant, nitrogendioxide (NO2) also plays an important role in ecological environment, it is considerably affected by traffic emissions and land-use. For mainland China, it is challenging to reconstruct the grids of ground NO2 with sufficient local gradients, which requires input of complete satellite NO2, traffic and land use data.  

Recently, an international team led by Associate Professor LI Lianfa at the Institute of Geographic Sciences and Natural Resources Research (IGSNRR), Chinese Academy of Sciences developed a new method on full residual deep learning to inverse missing satellite parameters such as Multiangle Implementation of Atmospheric Correction (MAIAC) Land Aerosol Optical Depth (AOD) and Ozone Monitoring Instrument (OMI) NO2, and reconstruct the grids of surface NO2 covering mainland China.

In the method, full residual connections are introduced in their encoder-decoder architecture to improve learning efficiency for deep layers. Depending on the large space of parameters and constrained optimization, the method well captured spatiotemporal variability between the variables, and achieved high generalization in the independent tests of the imputation (R20.90) of missing MAIAC AOD and OMI-NO2, and estimation (R2=0.82) of ground NO2.

The inversed missing satellite parameters presented consistent transmission with the original values, and the reconstructed NO2 surfaces had a reasonable spatiotemporal distribution with local gradients. The bootstrap aggregating of full residual network generated the coefficient of variation as an uncertainty metric.  

The inversed complete AOD can be accessed at https://doi.org/10.7910/DVN/RNSWRH. The full residual deep learning method was published at IEEE Transactions on Neural Network and Learning System, and the spatiotemporal estimation of NO2 for mainland China was published in Remote Sensing of Environment.   

1)    Lianfa Li, Ying Fang, Jinfeng Wang, Jun Wu, Ge Yong, 2020, Encoder-Decoder Full Residual Deep Networks for Robust Regression and Spatiotemporal Estimation, IEEE Transactions on Neural Network and Learning System.

2)    Lianfa Li, Jiajie Wu, 2021, Spatiotemporal Estimation of Satellite-Borne and Ground-Level NO2 Using Full Residual Deep Networks, Remote Sensing of Environment, 254.

 

 

Copyright © Institute of Geographic Sciences and Natural Resources Research, CAS
Email: weboffice@igsnrr.ac.cn Address: 11A, Datun Road, Chaoyang District, Beijing, 100101, China