Global bias adjustment for MODIS aerosol optical thickness using neural network

Abstract

Large uncertainties in calculating radiative forcings from aerosols due to their location, loading, and types pose a great challenge to global climate modeling. Trying to improve retrievals in a statistical manner normally requires detailed knowledge of uncertainty statistics and bias due to possible error sources such as different measurement viewing geometries, instrument calibration, and dynamically changing atmospheric and earth surface conditions. However, a priori estimates of these error sources are not, in general, available. The use of a neural network (NN) approach to compensate for biases and systematic errors of aerosol optical thickness (AOT) from the Moderate Resolution Imaging Spectrometer (MODIS) operational retrieval algorithm is explored. By utilizing the NN as an estimator, we can compensate against unknown sources of errors, nonlinearity in the data sets, and the presence of non-normal distributions. In this study, the highly accurate ground-based Aerosol Robotic Network (AERONET) measurements are used as the ground truth (GT). Our results show that the adjusted AOT with NN has decreased root mean square errors, improved correlations with GT data by 4% to 6%, and increased the number of NN-adjusted data falling within the published expected error envelope by ∼10% .