CAS Scientists Propose Deep Learning Method for Atmospheric Aerosol Retrieval


Small particles suspended in the Earth’s atmosphere, scientifically known as aerosols, pose a major environmental problem because they degrade visibility, affect human health and influence the climate. Fine mode fraction (FMF) as a crucial parameter describing aerosol properties can be used to distinguish human-caused and natural aerosol types. Aerosol Optical Depth (AOD) as a quantitative estimate of the aerosol amounts in the atmosphere, combined with FMF, can be used as a proxy for surface Particulate Matter PM2.5, an air quality parameter.

A research team led by Prof.LI Zhengqiang from the Aerospace Information Research Institute (AIR), Chinese Academy of Sciences (CAS), together with their cooperators proposed an artificial Neural Network method for AEROsol retrieval (NNAero) to jointly retrieve FMF and AOD derived from MODIS data. The research was published in Remote Sensing of Environment.

As we have known, the technology of satellite remote sensing inversion to extract AOD information is relatively mature, while FMF inversion is more difficult. Therefore, in studies such as estimating PM2.5 through satellite remote sensing, there is a lack of key parameter to distinguish the size of aerosol particles. FMF over land is difficult to retrieve due to complex remote sensing mechanisms and lack of observation information.

In this study, scientists used the MODIS spectral reflectance of solar radiation at the top of the atmosphere and at the surface, together with ground-based Aerosol Robotic Network (AERONET) measurements of AOD and FMF to train a Convolutional Neural Network (CNN) for the joint retrieval of FMF and AOD.

The NNAero results over northern and eastern China were validated against an independent reference AERONET dataset. The results show that 68% of the NNAero AOD values are within the MODIS expected error envelope (EE) over land of ±(0.05 + 15%), which is similar to the results from the MODIS Deep Blue (DB) algorithm (63% within EE), and both are better than the Dark Target (DT) algorithm (31% within EE).

According to the study, the validation of the NNAero FMF versus AERONET data shows a significant improvement with respect to the DT FMF, with Root Mean Squared Prediction Errors (RMSE) of 0.1567 (NNAero) and 0.34 (DT). The NNAero method shows the potential of improved retrieval of the FMF.

As shown in Fig. 1, the neural network combines a fully connected neural network (FCNN) and a convolutional neural network (CNN). The retrieved FMF has obvious accuracy promotion compared with previous studies (Fig. 2, 3). 

Fig. 1.The multi-input neural network architecture of MODIS FMF and AOD prediction.

Fig. 2. Accuracies of NNAero, Deep Blue and Dark Target algorithms validated using AERONET ground-based observations.

Fig. 3. Image product examples to compare the DB AOD vs NNAero AOD (up) and DT FMF vs NNAero FMF (down).

The research results can provide basic remote sensing product supporting PM2.5 remote sensing and climate change research. This research is supported by the National Natural Science Foundation of China (No. 41501399) and the Chinese National Scholarship Fund (No. 201804910115).