IEEE Geoscience and Remote Sensing Magazine - June 2019 - 38

where the gain, gt (m), and bias, bt (m), estimates are found
by performing a per-band regression between the fieldmeasured panel reflectance spectra and the corresponding
radiance (or digital count) image spectra of the same panels.
This technique assumes that there are no topographic effects
and that the computed radiance-to-reflectance mapping is
valid for every pixel in the
scene (i.e., the atmosphere is
constant and spatially invariDESPITE ITS SIMPLIFYING
ant), not just over the bright
ASSUMPTIONS, THE ELM
and dark objects used to dePRODUCES ESTIMATES OF
termine the gain and bias.
When this is not met (i.e., senREFLECTANCE THAT ARE
sors with large fields of view),
USUALLY VERY CLOSE TO
corrections or additional
THOSE MEASURED IN THE
bright/dark objects need to be
FIELD OR LABORATORY.
placed elsewhere in the scene.
Despite its simplifying assumptions, the ELM produces
estimates of reflectance that are usually very close to those
measured in the field or laboratory [38]. Unlike compensation methods that use physical models, scene-based methods are quite insensitive to errors and uncertainties in radiometric and spectral calibration of the sensor.
QUICK ATMOSPHERIC CORRECTION
A more recent approximate scene-based compensation
technique is the QUAC algorithm [39], [40]. A commercial version is available in the Atmospheric Correction
Module for Harris Corp.'s ENVI software. Like the ELM,
QUAC does not require radiometrically calibrated data

and assumes a linear relationship between spectral reflectance and measured radiance. However, unlike the ELM, it
does not require pixels of known reflectance in the image
and determines the necessary parameters from the image
itself. Again, with the assumptions that t b(m) S(m) is small
and that the diffuse or adjacency term can be combined
with the upwelled term, (2) reduces to the same linear relationship as (4), or
L Lamb
quac (m) = g (m)·t t (m) + b(m),

(6)

where we assume the offset, b(m), is constant over the image. The offset is derived via the DOS method by using the
minimum signal at each wavelength, assumed to correspond to zero reflectance. A median filter helps filter out
data dropouts from the minimum signals. The gain factor
is derived by assembling spectrally diverse, nonvegetated
pixels with an endmember method, taking the average, and
comparing the result with an average of similarly selected
spectra measured in the laboratory. Hence, images are assumed to contain at least tens of spectrally diverse materials
(i.e., roads, cars, dirt, concrete, roof, and so on). Scenes that
are all one material (i.e., all water, all vegetation, or all sand)
are not well suited for this approach. QUAC works best in
scenes containing dark materials or shadows and some
amount of cultural clutter to provide the spectral diversity.
Typical reflectance results are within approximately 25% of
ground truth measurements. Despite this limited accuracy,
QUAC typically performs very well in rare target detection,
as illustrated in the "Atmospheric Compensation and Target
Detection" section.

Raw Digital Number Data

Calibrate to Radiance
Apply Gain/Offset

Estimation of
Column Water
Vapor

Estimation of
Optical Depth for
Aerosol Scattering
Retrieval

Wavelength
Calibration

Spectral Smoothing
Physics Based

Other Adjustments:
- Adjacent Scatter
- BRDF Effects
- Cloud Masking
- Deshadowing
- Pressure Depth

Known Objects
in Scene

No a Priori
Information

Surface Reflectance
Scene Based

FIGURE 5. The general process of atmospheric compensation in the VNIR-SWIR region using either a scene-based approach or a more

rigorous physics-based approach. The dotted-line boxes are optional steps. BRDF: bidirectional reflectance distribution function.
38

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