PACO: Python-Based Atmospheric COrrection

Abstract

The atmospheric correction of satellite images based on radiative transfer calculations is a prerequisite for many remote sensing applications. The software package ATCOR, developed at the German Aerospace Center (DLR), is a versatile atmospheric correction software, capable of processing data acquired by many different optical satellite sensors. Based on this well established algorithm, a new Python-based atmospheric correction software has been developed to generate L2A products of Sentinel-2, Landsat-8, and of new space-based hyperspectral sensors such as DESIS (DLR Earth Sensing Imaging Spectrometer) and EnMAP (Environmental Mapping and Analysis Program). This paper outlines the underlying algorithms of PACO, and presents the validation results by comparing L2A products generated from Sentinel-2 L1C images with in situ (AERONET and RadCalNet) data within VNIR-SWIR spectral wavelengths range.

Keywords: DESIS; Landsat-8; Sentinel-2; aerosol optical thickness; atmospheric correction; remote sensing; surface reflectance; water vapor.

Figure 1

PACO algorithm workflow. The upper and bottom rows contain the input and output products, respectively. In the central row (blue box), the workflow of the algorithms described in Section 2 is displayed. The output products are explained in Section 3, and all of them might be combined in a final Output product if required by the mission.

Figure 2

Distribution of the AERONET stations (green circles) used in the Sentinel-2 (left) and Landsat-8 (right) validation study. The RadCalNet sites used for the Sentinel-2 surface reflectance validation are shown in the left plot (red stars): RailroadValley Playa (RVUS), La Crau (LCFR), and Gobabeb (GONA).

Figure 3

AOT (left) and WV (right) correlation plots between AERONET measurements and values retrieved by the remote sensing algorithm in a total of 494 Sentinel-2 scenes.

Figure 4

Accuracy (red star), precision (green diamond) and uncertainty (blue square) for calculated AOT (left) and WV (right) for Sentinel-2 as a function of the reference values (AERONET). The number of scenes included in each bin are shown in the shadow histogram and in the right axis. The accuracy and uncertainty fits are displayed as dotted red and blue lines, respectively.

Figure 5

(Left) Accuracy (red star), precision (green diamond), and uncertainty (blue square) as a function of the reference value of Sentinel-2 surface reflectance for the La Crau RadCalNet site. The number of pixels included in each bin are shown in the shadow histogram and on the right axis. Uncertainty fit is displayed as a dotted blue line. (Right) Uncertainty of the surface reflectance versus the reference value and the band central wavelength. The size of the points scales with the amount of pixels analyzed in each point. The amount of pixels considered for the smaller point is displayed in the legend. In both plots, the swir band (2250 nm), cirrus band (1380 nm), and water absorption bands at 940 nm have been discarded.

Figure 6

(Left) Distribution of the accuracy, precision, and uncertainty of Sentinel-2 surface reflectance as a function of the reference surface reflectance for the Gobabeb RadCalNet site. The number of pixels included in each bin are shown in the shadow histogram and on the right axis. Uncertainty fit is displayed as a dotted blue line. (Right) Uncertainty of the surface reflectance versus the reference value and the band central wavelength. The size of the points scales with the amount of pixels analyzed in each point. The amount of pixels considered for the smaller point is displayed in the legend. In both plots, the swir band (2250 nm), cirrus band (1380 nm), and water absorption bands at 940 nm have been discarded.

Figure 7

(Left) Distribution of the accuracy, precision, and uncertainty of Sentinel-2 surface reflectance as a function of the reference surface reflectance for the Railroad Valley Playa RadCalNet site. The number of pixels included in each bin are shown in the shadow histogram and on the right axis. Uncertainty fit is displayed as the dotted blue line. (Right) Uncertainty of the surface reflectance versus the reference value and the band central wavelength. The size of the points scales with the amount of pixels analyzed in each point. The amount of pixels considered for the smaller point is displayed in the legend. In both plots, the swir band (2250 nm), cirrus band (1380 nm), and water absorption bands at 940 nm have been discarded.

Figure 8

Surface reflectance for Sentinel-2B, DESIS and the Railroad Valley Playa RadCalNet (RCN) site of 28 June 2019. The spectra are convolved with Sentinel-2 spectral response function and the absolute and relative differences (middle, bottom plots) are calculated with respect to the DESIS spectrum.