A Method for Unbiased High-Resolution Aerosol Retrieval from Landsat

摘要

Because the land surface reflectance varies spatially, the atmospheric radiative transfer over land in clear-sky conditions is essentially three-dimensional. This is manifested through horizontal radiative fluxes that blur satellite images. It is important that the atmospheric blurring systematically increases the apparent brightness of the dark pixels. As a consequence, there are systematic biases in the satellite products of aerosol optical thickness and surface albedo over dark targets based on 1D theory, which may have a negative impact on climate research. Below, a new dark target method is presented for unbiased simultaneous retrieval of the aerosol model and optical thickness over land from Landsat Enhanced Thematic Mapper Plus (ETM+) data, based on 3D radiative transfer theory. The method automatically selects an aerosol model from a large set of candidate models using a statistical approach of the probability distribution function. The dark target method of aerosol retrieval in the blue and red bands relies on prediction of the surface reflectance in these bands from the shortwave infrared region (2.1–2.2 μm) based on the linear regression. In the Moderate Resolution Imaging Spectroradiometer (MODIS) algorithm, the regression coefficients are constants, whereas different studies indicate that they have seasonal and geographic variations. The work here shows that the accuracy of aerosol retrieval over land can be significantly increased based on ancillary information on the regional and seasonal distribution of the regression coefficients. This information, which is called surface climatology, can be derived globally around Aerosol Robotic Network (AERONET) sites, using AERONET aerosol and water vapor information for accurate atmospheric correction. This paper describes the developed method in application to Landsat data and its initial validation with AERONET measurements for a set of ETM+ images of the Washington–Baltimore area, and studies biases of 1D retrievals.