Synthesis of multi-sensor top of atmosphere and ground level reflectances to support high-resolution AOD estimation with machine learning

In complex urban environments, the information of high-resolution Aerosol Optical Depth (AOD) is of great importance for effective air pollution control, air navigation, public health assessment, and meteorological forecasting. High resolution AOD may be produced by merging and/or fusing existing AOD products or reproduced by merging and/or fusing the reflectance data at the top of atmosphere (TOA) and ground levels through the deep blue method. However, the former can only lead to the production of AOD with 500 m∼1 km spatial resolution at best. To overcome this barrier, it is necessary to fuse the reflectance values of Landsat and MODIS imageries at the TOA level to be in concert with the fused land surface reflectance values for advanced synthesis. Such a collective endeavor can lead to the production of AOD with daily 30m spatial resolution via the deep blue method. This paper thus presents such a synthetic effort that synergizes the spatial and temporal advantages of two satellite sensors (MODIS Terra and Landsat 8) to reach the goal with the aid of machine learning and high-performance computing. Based on the deep blue method, the practical implementation of the synthetic image processing was assessed by a case study of the downtown Atlanta area in the United States. 10-fold cross validation was applied stepwise to control the uncertainty via machine learning. The predictions of AOD at the ground level were calibrated using the AErosol RObotic NETwork (AERONET) AOD data and finally validated by the AERONET) AOD data too.