IEEE Geoscience and Remote Sensing Magazine - June 2019 - 42

small and medium field-of-view (FOV) satellite imagery
with a reasonably flat terrain; and it is limited to multispectral and thermal imagery. ATCOR3 is similar to ATCOR2 but with support for imaging spectrometers in space
and for rugged terrain. ATCOR4 is the fullest featured of all
current atmospheric compensation models, and it is generally used for the compensation of small and wide FOV
airborne sensors with flat or
rugged terrain. It includes an
empirical method to account
IT IS ESSENTIALLY
for bidirectional reflectance
IMPOSSIBLE FOR A FIRSTdistribution function (BRDF)
PRINCIPLES PHYSICSeffects and options for the
BASED METHOD TO MODEL
user to input a digital elevation model (DEM). Outputs
THE ATMOSPHERE AND
include ground reflectance;
SENSOR PROPERTIES WITH
an aerosol optical thickness
COMPLETE ACCURACY.
map; a water-vapor map; a
dense dark vegetation map;
a haze-cloud-water map; and
quality indicators for cloud, snow, and water probabilities.
With precomputed look-up tables, ozone adjustment is
performed as part of the atmospheric compensation for
channels in the 0.45- to 0.80-nm range [66]. Haze detection and removal are based on an improvement to the DOS
technique by Chavez [34], searching the image locally for
dark objects and using the blue and green channels. Cirrus cloud detection and compensation can be performed
using the 0.66- and 1.38-nm channels [67]. Water vapor is
determined using the atmospheric precorrected differential absorption technique and precalculated look-up tables
[68]. For rugged or mountainous terrain, a DEM is used
to adjust for BRDF effects before generating the groundleaving reflectance product.
CHARACTERIZATION OF ILLUMINATION
IN 3D ENVIRONMENTS
For the most part, atmospheric compensation algorithms
assume that the surfaces in the scene are horizontal and

Surface Reflectance

0.6
0.5

Without Adjacency Correction
With Adjacency Correction

0.4
0.3
0.2
Chlorophyll Edge
(i.e., Vegetation)
Contaminates Soil

0.1
0

0.5

0.6

0.7
0.8
Wavelength (µm)

0.9

FIGURE 10. A plot of retrieved soil with and without adjacency
compensation.

have an unobscured view of the full sky. Inaccurate reflectance retrievals occur for those pixels that are in
shadow, are tilted, or have an occluded view of the sky.
Several studies have considered how to accurately model
the illumination in such 3D environments. A rigorous,
but fairly complicated, approach uses the adjoint radiosity method proposed by Borel et al. [69]-[72], which was
originally introduced in the computer graphics community. This approach includes illumination contributions
from the surroundings (i.e., scattered multiply from
nearby surfaces) in either the presence or absence of direct solar illumination; the latter pertains to objects in
shadow. The approach is modeled after the full versions
of (1) or (12), found in "Forward Sensor-Reaching Radiance Model," and requires an accurate 3D description of
the scene.
A similar approach [73] leverages lidar data to build
the 3D geometry and determine the illumination of
the pixel of interest. Directional and spectrally dependent downwelling radiance information is computed
along with shadow maps to improve the overall forward at-sensor model. This geometrically based approach was found to greatly improve the reflectance retrieval in shadows.
COMPARISONS AND EVALUATIONS
It is not the intention of this review to rigorously compare and contrast all of the compensation algorithms to
determine which is the best. Although some algorithms
are more sophisticated and, hence, generally more accurate than others, the results can be very scene, and sometimes user, dependent. That said, there have been a handful of review-type articles on atmospheric compensation
that do compare and contrast algorithms. Here, we briefly
examine, chronologically, those articles that deal with hyperspectral data, and we introduce some recent compensation results.
LAND-BASED IMAGERY
In a technical report [45] from 1999, Griffin et al. evaluate
ATREM and FLAASH in terms of operational characteristics,
input requirements, and algorithm limitations. This report
is similar to [10], published only four years earlier. Watervapor estimates for both algorithms are presented and
compared using data from AVIRIS and HYDICE. Surfacereflectance comparisons are presented for a variety of signatures, including soil, road, and vegetation. Sensitivity
studies were performed on input parameters, such as visibility, aerosol model type, atmospheric model, and solar zenith angle. Finally, computational complexity was studied in
terms of basic atmospheric compensation computation costs
and enhanced atmospheric compensation costs.
In 2003, Griffin and Burke presented a 24-page overview article [10] that provides an evaluation of ATREM and
FLAASH using AVIRIS and HYDICE data. A discussion of
scene-based methods was also presented (i.e., ELM, IAR,
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

JUNE 2019

IEEE Geoscience and Remote Sensing Magazine - June 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2019