A Deep-Neural-Network-Based Aerosol Optical Depth (AOD) Retrieval from Landsat-8 Top of Atmosphere Data

Abstract

The 30 m resolution Landsat data have been used for high resolution aerosol optical depth (AOD) retrieval based on radiative transfer models. In this paper, a Landsat-8 AOD retrieval algorithm is proposed based on the deep neural network (DNN). A total of 6390 samples were obtained for model training and validation by collocating 8 years of Landsat-8 top of atmosphere (TOA) data and aerosol robotic network (AERONET) AOD data acquired from 329 AERONET stations over 30°W–160°E and 60°N–60°S. The Google Earth Engine (GEE) cloud-computing platform is used for the collocation to avoid a large download volume of Landsat data. Seventeen predictor variables were used to estimate AOD at 500 nm, including the seven bands TOA reflectance, two bands TOA brightness (BT), solar and viewing zenith and azimuth angles, scattering angle, digital elevation model (DEM), and the meteorological reanalysis total columnar water vapor and ozone concentration. The leave-one-station-out cross-validation showed that the estimated AOD agreed well with AERONET AOD with a correlation coefficient of 0.83, root-mean-square error of 0.15, and approximately 61% AOD retrievals within 0.05 + 20% of the AERONET AOD. Theoretical comparisons with the physical-based methods and the adaptation of the developed DNN method to Sentinel-2 TOA data with a different spectral band configuration are discussed.