An improved deep learning network for AOD retrieving from remote sensing imagery focusing on sub-pixel cloud

ABSTRACT

Following the success of MODIS, several widely used algorithms have been developed for different satellite sensors to provide global aerosol optical depth (AOD) products. Despite the progress made in improving the accuracy of satellite-derived AOD products, the presence of sub-pixel clouds and the corresponding cloud shadows still significantly degrade AOD products. This is due to the difficulty in identifying sub-pixel clouds, as they are hardly identified, which inevitably leads to the overestimation of AOD. To overcome these conundrums, we proposed an improved deep learning network for retrieving AOD from remote sensing imagery focusing on sub-pixel clouds especially and we call it the Sub-Pixel AOD network (SPAODnet). Two specific improvements considering sub-pixel clouds have been made; a spatial adaptive bilateral filter is applied to top-of-atmosphere (TOA) reflectance images for removing the noise induced by sub-pixel clouds and the corresponding shadows at the first place and channel attention mechanism is added into the convolutional neural network to further emphasize the relationship between the uncontaminated pixels and the ground measured AOD from AERONET sites. In addition, a compositive loss function, Huber loss, is used to further improve the accuracy of retrieved AOD. The SPAODnet model is trained by using ten AERONET sites within Beijing-Tianjin-Hebei (BTH) region in China, along with their corresponding MODIS images from 2011 to 2020; Subsequently, the trained network is applied over the whole BTH region and the AOD images over the BTH region from 2011 ~ 2020 are retrieved. Based on a comprehensive validation with ground measurements, the MODIS products, and the AOD retrieved from the other neural network, the proposed network does significantly improve the overall accuracy, the spatial resolution, and the spatial coverage of the AOD, especially for cases with sub-pixel clouds and cloud shadows.

KEYWORDS:

• Aerosol optical depth
• sub-pixel cloud
• deep learning
• channel attention
• CNN

Acknowledgments

The authors acknowledge MODIS and AERONET groups for the satellite and ground-based remote sensing data. The authors would like to acknowledge the SONET, AERSS for sharing the data and make them available for the community.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available upon request by contact with the corresponding author, or accessed through https://ladsweb.modaps.eosdis.nasa.gov/and https://aeronet.gsfc.nasa.gov/cgi-bin/webtool_aod_v3.

Additional information

Funding

The work was supported by the  National Key Research and Development Program of China [2019YFE0197800]; National Key Research and Development Program of China under Grant [2020YFE0200700].