Ensemble Estimation of Aerosol Optical Depth by Feature Selections from Remote Sensing Data

Abstract

Aerosol optical depth (AOD) is an important quantity parameter to study the Earth’s radiation balance, climate change and environment protection. For estimating AOD by a data mining method, the synchronized records by combing satellite observed information from MOderate Resolution Imaging Spectroradiometer (MODIS) equipment with the ground-based accurate measurements of AOD from Aerosol Robotic NETwork (AERONET) work as driving attributes and prediction targets, respectively. However, compared with the number of high-dimensional remote sensing attributes, the total number of spatial-temporal collocated MODIS-AERONET observations during a couple of years is relatively not large enough for estimation modeling. It leads to unstable feature selection subsets and drops the AOD estimation accuracy. In this paper, we propose a novel ensemble approach by aggregating multiple AOD estimators. Each estimator is modeled based on features selected from remote sensing attributes by using a subsampling strategy with instance perturbation. The ensemble approach provides aggregated retrievals of AOD with higher accuracy, while also providing an estimation of retrieval uncertainty. We conducted experiments to evaluate the empirical performance of the proposed approach on two years (2009-2011) of MODIS data over 197 global AERONET sites. The encouraging results clearly showed that aggregation of estimators modeled by multiple feature selection subsets leads to accuracy improvements and uncertainty reduction in AOD retrievals.