Evaluation of a Neural Network-Based Approach for Aerosol Optical Depth Retrieval and Uncertainty Estimation

In many applications of the neural networks, predicting the conditional average of the target variable is not sufficient. Often, real life problems also require estimation of the uncertainty. In this study, uncertainty analysis is applied on a remote sensing problem of Aerosol Optical Depth (AOD) estimation. AOD is one of the most important properties of the atmosphere that indicates the amount of depletion that a beam of radiation undergoes as it passes through the atmosphere. To predict AOD, we used a data-driven approach based on training neural networks. Several techniques for uncertainty estimation which are tractable for large amounts of high-dimensional remote sensing data are considered. Under the assumption that the noise in targets is input-dependent, the uncertainty of AOD predictions is computed as the variance of the conditional distribution of targets given the input data. Several methods for uncertainty estimation were applied to a real data set with 67,907 observations collected over the whole Earth during three years (2005-2007) with the attributes derived from MODIS satellite instrument and with the targets obtained from ground-based AERONET instruments. Knowledge discovered from the uncertainty analysis of this data set can potentially be very useful for better understanding of aerosol properties.