RESEARCH ARTICLE Atmospheric Correction of Multispectral VNIR Remote

10.1029/2019EA000710
Sensing Data: Algorithm and Inter-sensor Comparison
Key Points: of Aerosol and Surface Reflectance Products
• An algorithm for aerosol retrieval
and atmospheric correction of
visible and near-infrared sensor is
Manoj K. Mishra1 , Pradyuman S. Rathore1 , Arundhati Misra1 , and Raj Kumar1
presented
1 Space Applications Center, Indian Space Research Organization (ISRO), Ahmadabad, India
• The algorithm significantly
removes the haze from the
high-resolution images making
surface feature visible
• Comparison of derived aerosol
Abstract Optical imaging satellites, such as SPOT and Cartosat-2S, provide visible/near infrared
and surface reflectance with (VNIR) multispectral data at very high spatial resolution. The applications of these data sets are associated
MODIS-terra products shows good with precise mapping, monitoring, and change detection of Earth's surface, given that the measurements
agreement can be compensated for atmospheric effects. Existing atmospheric correction (AC) algorithms use visible
and shortwave infrared channels and therefore cannot be used for AC of data from VNIR sensors. This
Correspondence to: article describes an algorithm for aerosol optical depth (AOD) retrieval and AC of VNIR imaging data. The
M. K. Mishra, AOD algorithm relies on the fact that for vegetated surfaces there exists a visible/NIR surface reflectance
manoj8187@sac.isro.gov.in
relationship due to the absorption of solar radiation by photosynthetic pigments in visible bands, while
high reflectance in NIR bands governed by structural discontinuities in the leaves of healthy vegetation.
Citation: We then describe how retrieved AOD is used to derive surface reflectance. To test the algorithm, the
Mishra, M. K., Rathore, P. S.,
Misra, A., & Kumar, R. (2020). aerosol and surface reflectance products generated from 106 Cartosat-2S data sets are compared with
Atmospheric correction of MODIS-terra products. The algorithm significantly removes the haze from the images making surface
multispectral VNIR remote sensing feature visible. The comparison of Cartosat-2S and MODIS-terra AOD involving >1,500 data points shows
data: Algorithm and inter-sensor
comparison of aerosol and surface good correlation of 0.95 with a relative difference of ≤25%. Similarly, the comparison of surface reflectance
reflectance products. Earth and Space, involving >4,500 data points shows good correlation ranging from 0.75 to 0.86 with a relative difference
7, e2019EA000710. https://doi.org/10. ranging from 24% to 37%. The normalized difference vegetation index shows a correlation of 0.89, with a
1029/2019EA000710
relative difference of ≤18%. Results show that the given algorithm may be useful for AC of data from VNIR
sensors.
Received 15 MAY 2019
Accepted 19 APR 2020
Accepted article online 26 APR 2020
1. Introduction
The space-borne multispectral visible/shortwave infrared (VSWIR) imaging spectrometers such as Moderate
Resolution Optical Imaging Spectrometer-terra/aqua (MODIS-terra/aqua) (Salomonson et al., 1989), Visible
Infrared Imaging Radiometer Suite (VIIRS) onboard the Suomi National Polar-Orbiting Partnership (Cao
et al., 2014), and Advanced Baseline Imager (ABI) onboard GOES-16 (Schmit et al., 2005) map radiance
in multiple narrow spectral bands of solar reflective spectrum (400–2,500 μm). Surface reflectance features
in the solar reflective spectrum are a good source of information about the chemistry and composition of
Earth's terrestrial domains (Asner et al., 2017; Jetz et al., 2016; Ustin et al., 2004). However, since space-borne
spectrometer measures signal at the top of atmosphere (TOA) that contains coupled information from both
atmosphere and surface, therefore, atmospheric correction (AC) of TOA measurement is necessary to derive
the surface reflectance before using it for surface studies. In general, if atmospheric state parameters like
aerosol and water vapor concentration are known, the AC of any optical data set can be achieved with the aid
of radiative transfer (RT) calculations (Berk et al., 2014; Mayer & Kylling, 2005; Ricchiazzi et al., 1998; Tanre
et al., 1986; Vermote et al., 1997). For land surface, the existing ready-to-use AC software and algorithms such
as Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) and Atmospheric Removal Program
(ATREM) (Kruse, 2004; Perkins et al., 2012; Richter & Schläpfer, 2002) include a method for retrieving the
aerosol amount using a dark pixel visible and SWIR reflectance ratio method (Kaufman et al., 1997; Levy
et al., 2007). Thus with these models, the aerosol retrieval and therefore AC are only possible when the image
©2020. The Authors.
This is an open access article under the contains bands in visible as well as in SWIR wavelengths. While for water surface, the measurements in two
terms of the Creative Commons near-infrared bands are used to retrieve aerosol information (Ahmad et al., 2010; Antoine & Morel, 1999;
Attribution License, which permits Gordon, 1997; Sahay et al., 2014; Sarangi et al., 2015). Thus, the AC of data sets from VSWIR sensors (such
use, distribution and reproduction in
any medium, provided the original as MODIS, VIIRS, and ABI) is within the reach of the present technology; consequently, these sensors have
work is properly cited. a rich history of applications in various fields, such as mineral exploration, vegetation, atmospheric, and

MISHRA ET AL. 1 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 1. Top-of-atmosphere natural color composite image acquired from Cartosat-2S over a part of New Delhi, India,
on 01 November 2016. The left part of the image is hazy with less visibility due to aerosol scattering, while right part of
the image looks much darker.

ocean color studies. However, due to coarser spatial resolution (0.25–1.0 km), the data from these sensors
have limited utility in certain fields such as urban developments, precision agriculture, bathymetry, coastal
zone mapping, and other strategic applications where high spatial resolution is required.
Other kinds of sensors like SPOT-7 (Satellite Pour l-Observation de la Terre) and Cartosat-2S that provide
images covering only visible and near-infrared (VNIR) part of the solar reflective spectrum contain a sig-
nificant amount of spatial and spectral information about the Earth's surface. The disadvantage of these
sensors is the lesser number of spectral bands (typically 3–5) in the VNIR region and no SWIR bands, mak-
ing AC a challenging task with existing technology, but these sensors have the advantage of lower costs
and a very high spatial resolution of the order of sub-meters to few meters. The high aerosol loading and
adverse imaging condition (large sun or sensor zenith angle) magnify the atmospheric perturbation in TOA
radiance, thus making AC a necessary task. For example, Figure 1 shows the Cartosat-2S natural color com-
posite (NCC) image of a part of New Delhi, India. Due to significant aerosol scattering in the atmosphere,
the NCC image (Figure 1) appears hazy making the Earth's surface features unrecognizable in some parts of
the image. Figure 2 shows the TOA reflectance spectrums for some vegetated pixels taken from data shown
in Figure 1 along with typical vegetation spectrum convolved for Cartosat-2S bands. It is clear that due to
atmospheric effects the shapes of TOA reflectance spectrums are completely different from the typical vege-
tation spectrum. Thus, to get useful information (qualitative and quantitative) about Earth's surface, AC of
these data sets is required. Note that the typical spectrum shown in Figure 2 is derived using in situ surface
reflectance spectrums acquired elsewhere, and magnitude wise may not represent the shown region. To the
best of our knowledge, AC researches mainly remained focused on VSWIR sensors while the problem of
AC of data acquired by VNIR sensors (SPOT-7 and Cartosat-2S) over land remained unaddressed. High spa-
tial resolution data from VNIR sensors, if corrected for atmospheric effects, will be very useful for various
applications involving precise mapping, monitoring, and change detection of Earth surface.
In this paper, we address the problem of AC of data acquired from VNIR sensors such as Cartosat-2S and
SPOT-7. Using Cartosat-2S data sets, we first described an algorithm for aerosol retrieval and then described

MISHRA ET AL. 2 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 2. Typical spectrum for vegetation and Cartosat-2S MX TOA apparent reflectance spectrum at various known
sites falling in region shown in Figure 1. (a) Sites with dense vegetation cover. (b) Sites with pure soil cover. Band
numbers B1, B2, B3, and B4 refer to the corresponding wavelengths described in Table 1.

AC to derive “surface reflectance” assuming horizontal surface having Lambertian reflectance. After
brief introduction in Section 1, Section 2 describes the features of VNIR instrument onboard Cartosat-2S.
Section 3 describes the data sets and the study area used in this study. Section 4 describes the aerosol and
AC algorithm for aerosol optical depth (AOD) and the surface reflectance retrieval, respectively. Section 5
presents the aerosol and atmospherically corrected Cartosat-2S products and shows their comparison with
MODIS-terra products. Finally, Section 6 concludes the salient features of work.

2. Instrument Description
Cartosat-2S is an Indian satellite meant for applications like mapping, monitoring, and change detection of
Earth's surface (Cartosat-2 brochure, n.d; Sharma et al., 2016). It contains one multispectral (MX) and one
panchromatic (PAN) sensor having a high signal-to-noise ratio. The space-
craft is in a polar sun-synchronous orbit (orbital altitude of about 500 km)
with approximately 9:30 hr of equatorial cross over time. The spatial reso-
lution of MX and PAN data is around 2.0 and 0.65 m, respectively, with a
Table 1 swath of about 10 km. The salient features of Cartosat-2S MX payload are
Salient Features of Cartosat-2S MX Sensor provided in Table 1. Figure 3 shows the spectral response functions (SRF)
Parameters Values for MX bands and PAN band along with extraterrestrial irradiance at mean
Spatial resolution at nadir (m) 2.0 Earth-Sun distance.
Swath (km) 10.0
3. Data Used and Study Area
Spectral coverage (μm) B1: 0.45–0.52; B2: 0.52–0.59
B3: 0.62–0.68; B4: 0.77–0.86 The Cartosat-2S MX data sets acquired over different parts of the world
Saturation radiance B1: 53.0; B2: 53.0 covering parts of the Indian subcontinent (India, Nepal, Pakistan, and
Bangladesh), China, and the United States are used in the present study. A
(mW cm−2 nm−1 sr−1 ) B3: 47.0; B4: 31.5
total of 106 Cartosat-2S data sets acquired in January, July, November, and
Quantization (bits) 11
December months in the years 2016 and 2017 are used. Figure 4 shows the
Signal-to-noise at saturation >300
scene-center location of all 106 data sets. Each Cartosat-2S data set con-
Note. B stands for band. tains radiometrically calibrated radiance for all four bands. The Cartosat-2S

MISHRA ET AL. 3 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 3. Normalized spectral response function for Cartosat-2S MX Bands 1, 2, 3, and 4 and for Cartosat-2S PAN
band. Average extraterrestrial irradiance (Fo ) reaching at the top of the Earth's atmosphere at mean Earth-Sun distance
is also shown.

data are routinely calibrated using ground measurements for reference targets by the data processing team
(Sharma et al., 2016). The metadata file of each data set contains information about acquisition date/time,
solar zenith angle, sensor zenith angle, azimuth angles, satellite altitude, mean elevation of the scene, and
scene center latitude/longitude coordinates. As an example, Table 2 shows metadata information of two
Cartosat-2S data sets acquired on 01 November 2016 and 03 October 2016 over the parts of New Delhi and
Ahmadabad region, respectively. These two data sets are used to display the visual effect of AC on the TOA
apparent reflectance image. The reason for choosing these two data sets among 106 data sets is due to sig-
nificant aerosol loading in the atmosphere on these days of observation. To test the validity of our algorithm
and the accuracy of Cartosat-2S products, the Cartosat-2S aerosol and surface reflectance products at all 106
globally distributed locations (Figure 4) are compared with MODIS-terra products collocated in space and
time domain. The MODIS-terra AOD products (DT aerosol retrievals, Collection 6.1), namely, “MOD04_3K
MODIS/terra Aerosol 5-Min L2 Swath 3 km” at spatial resolution of 3 km, encompassing all 106 Cartosat-2S
data set locations are used for comparison with Cartosat-2S AOD. For comparison of surface reflectance
product, MODIS-terra surface reflectance product, namely, “MOD09GA MODIS/terra Surface Reflectance

Figure 4. Scene-center location of 106 Cartosat-2S MX images used in present work for comparison of Cartosat-2S
aerosol optical depth (AOD) and surface reflectance product with MODIS-terra products.

MISHRA ET AL. 4 of 21
 Earth and Space Science 10.1029/2019EA000710

Table 2
Details of Data Sets 1 and 2
Parameter Value (Data Set 1) Value (Data Set 2)
Region Delhi Ahmedabad
Acquisition date 01 November 2016 03 October 2017
Acquisition time 4:34 UTC 4:57 UTC
Scene center latitude 28.5388 23.02370
Scene center longitude 77.1017 72.47861
Sensor altitude 501.74 km 511.62 km
Solar azimuth angle 143.79◦ 128.26◦
Solar zenith angle 51.12◦ 40.54◦
Sensor azimuth angle 193.27◦ 192.46◦
Sensor zenith angle 20.05◦ 3.34◦
Mean elevation 0.232 km 0.045 km

Daily L2G Global 1 km and 500 m SIN Grid” at 500-m resolution, is used. All required MODIS-terra data sets
(MOD04_3K and MOD09GA) are available free from the website of “The Level-1 and Atmosphere Archive
& Distribution System (LAADS) Distributed Active Archive Center (DAAC)” at the Goddard Space Flight
Center (Kaufman et al., 1997; Levy & Hsu, 2015; Levy et al., 2007; Vermote & Wolfe, 2015). The equatorial
passing time of Cartosat-2S and MODIS-terra are not the same but fall within 1-hr time window. Moreover,
due to the agile platform of Cartosat-2S, the observation time difference may vary within a time window of
few minutes to hours. In the present study, after careful analysis of acquisition times of all 106 data sets,
it is found that the time difference between Cartosat-2S and MODIS-terra observations ranges from 30 min
to 2 hr, out of which >90% observation falls within time difference of 1.5 hr and >80% of observation fall
within 1-hr range. Due to the difference in time of observation, the precise spatiotemporal collocation of
MODIS-terra and Cartosat-2S aerosol retrieval is not possible which may introduce error in comparison
results particularly when the difference in observation time is very large. However, as mentioned earlier,
most of the observation (>80% of total observations) fall within 1-hr time difference; therefore, no significant
diurnal variation of AOD is expected in normal atmospheric conditions. The uncertainty in MODIS-terra
AOD varies location wise because of location-dependent validity of visible/SWIR surface reflectance rela-
tionship used in dark-target aerosol algorithm (Kaufman et al., 1997; Levy et al., 2007). However, global
validation studies using in situ AOD from Aerosol Robotic Network (Holben et al., 1998) have shown that
about 67% of MODIS 3 km AOD retrievals fall within an error envelope (EE) of ±0.05 ± 0.20𝜏 (Remer et al.,
2013) while about 66% of MODIS 10 km AOD retrievals fall within an EE of ±0.05 ± 0.15𝜏 (Levy et al., 2010).
On average, the overall uncertainty in MODIS AOD is up to 20–30% (Levy et al., 2007) for the AOD value of
0.3 to 0.5. Thus in the absence of in situ AOD observations collocated with Cartosat-2S AOD retrievals, the
comparison with MODIS-terra AOD product will give a good assessment of the performance of the presented
aerosol algorithm.

4. Algorithm Description
4.1. Radiative Transfer
Based on RT theory of coupled planetary atmosphere-surface system (Chandrasekhar, 1960; Lenoble, 1985),
an approximate but reasonably accurate vector RT code called Simulation of the Satellite Signal in the Solar
Spectrum (6SV) has been developed in recent past (Kotchenova et al., 2006; Vermote et al., 1997). The
relation expressing the TOA apparent reflectance for Lambertian surface is (Vermote et al., 1997)
[ ]
T(𝜃s , 𝜆)T(𝜃v , 𝜆)𝜌s (𝜆)
𝜌∗ (𝜃s , 𝜃v , Δ𝜙, 𝜆) = Tg (𝜃s , 𝜃v , 𝜆) × 𝜌ra (𝜃s , 𝜃v , Δ𝜙, 𝜆) + . (1)
1 − S(𝜆)𝜌s (𝜆)

Here 𝜃 s , 𝜃 v , Δ𝜙, and 𝜆 refer to solar zenith, sensor zenith, relative azimuth angles, and wavelength, respec-
tively. The various quantities in equation (1) are expressed in terms of equivalent reflectance defined by
relation 𝜌 = 𝜋 L∕𝜇s F0 , where L refers to measured radiance, F0 is the extraterrestrial irradiance at TOA, and
𝜇 s is the cosine of solar zenith angle. In equation (1), 𝜌ra (𝜃 s , 𝜃 v , Δ𝜙, 𝜆) refer to the reflectance due to molecule

MISHRA ET AL. 5 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 5. Gas transmittance, aerosol reflectance, and Cartosat-2S spectral response functions (SRF). Blue, green, red,
and cyan curves show the Cartosat-2S SRF of Bands 1, 2, 3, and 4, respectively. Gas transmittance is calculated for
𝜃 s = 400 , 𝜃 v = 00 , O3 = 0.250 cm − atm and H2 O = 10 mm. Aerosol reflectance is computed for aerosol optical depth
(AOD) of 0.5 at 0.55 μm and continental aerosol model.

and aerosols, and T(𝜃 s , 𝜆) and T(𝜃 v , 𝜆) refer to the downward and upward transmittance of the atmosphere,
respectively. Symbol S(𝜆) and Tg refer to the spherical albedo of the atmosphere and gaseous transmittance
(downward and upward direction both), respectively. Rearrangement of equation (1) gives
[ ]
𝜌s (𝜆) = ℜ∕ Tg (𝜃s , 𝜃v , 𝜆)T(𝜃s , 𝜆)T(𝜃v , 𝜆) − ℜ , (2)
where
ℜ = 𝜌∗ (𝜃s , 𝜃v , Δ𝜙, 𝜆) − Tg (𝜃s , 𝜃v , 𝜆)𝜌ra (𝜃s , 𝜃v , Δ𝜙, 𝜆). (3)

Given the satellite measurement, equation (2) can be solved for surface reflectance 𝜌s (𝜆) if state of the atmo-
sphere (gaseous transmittance and aerosol parameters) is known. Thus, AC of data from VNIR sensors
such as Cartosat-2S requires accurate gas information and most essentially an algorithm to derive aerosol
information from the data itself.
4.2. Gaseous Absorption
Out of many atmospheric gases, water vapor H2 O, ozone O3 , and oxygen O2 are the main gases that show
observable features in VNIR wavelengths under typical atmospheric conditions. Figure 5 shows atmospheric
transmittance simulated for H2 O = 10 mm and O3 and O2 gases content taken from standard tropical atmo-
sphere. It is seen that Cartosat-2S MX bands are mainly affected by absorption due to O3 and H2 O while the
PAN band is additionally affected by O2 . In the present work, we used the standard tropical atmospheric
model to calculate O2 transmittance. Since the atmospheric column content of O3 changes slowly with
latitude and season, therefore, we used climatology data to calculate O3 transmittance.
4.3. Aerosol Optical Depth Retrieval Algorithm
4.3.1. Physical Basis
It is a well-known fact that in the visible spectrum the vegetated surface shows low reflectance particu-
larly at blue and red wavelengths due to absorption by photosynthetic pigments (mainly chlorophyll and
carotenoids); therefore, reflectance in red and blue channels is generally less than 15% for vegetated sur-
faces (Peñuelas & Filella, 1998). However, from Figure 5, it is clear that for moderate aerosol loading the
apparent reflectance due to aerosols at TOA may be up to 20% in blue wavelengths. For a very thick aerosol
layer, it can be more than 20% depending on the absorption and scattering nature of aerosols. The major
challenge in any aerosol retrieval algorithm is to decouple the TOA reflectance into the contribution from
the surface and atmospheric constituents and then separating the molecular scattering contribution from

MISHRA ET AL. 6 of 21
 Earth and Space Science 10.1029/2019EA000710

atmospheric contribution giving the aerosol contribution. For water surfaces due to the absorption of solar
radiation, the surface reflectance in red and NIR wavelengths are nearly zero, therefore, assuming the entire
signal in these wavelengths as the atmospheric contribution is a good approximation for deriving aerosol
information. However, land surfaces show significant non-zero reflectance depending on the surface types.
MODIS-aqua/terra dark-target aerosol algorithm estimates the visible bands (blue and red wavelength) sur-
face reflectance using the measurement in SWIR band (Kaufman et al., 1997; Levy & Hsu, 2015). However,
due to the unavailability of SWIR measurements in VNIR sensors like Cartosat-2S, the dark-target aerosol
algorithm cannot be used. Observing that the effect of aerosol scattering is very small (excluding the case
of dust events) on TOA measurements in NIR wavelengths (see Figure 5), we propose the use of NIR chan-
nel to characterize surface reflectance in visible wavelengths for reference target types while the process
of aerosol retrieval. For vegetated targets, in visible channel (blue and red wavelengths) reflectance is low
due to absorption from photosynthetic pigments, while in NIR, where there are no significant absorption
features by pigments, the magnitude of reflectance is governed by structural discontinuities encountered
in the leaves of healthy vegetation (Peñuelas & Filella, 1998). This forms the physical basis of visible/NIR
surface reflectance relationship over the green vegetated target. Moreover, Lorenzen and Jensen (1988),
by conducting the field measurements of canopy reflectance of the vegetation in the blue, green, red,
and NIR wavebands, have shown that the surface reflectance in these wavebands is a strong function of
green biomass. In fact, the surface reflectance in blue and red wavelengths decreases with increasing green
biomass, while for NIR wavelength the nature of relation gets inverted (Lorenzen & Jensen, 1988). Based
on this discussion, the physical bases of our aerosol retrieval algorithm are as follows: (a) Solar radiation in
NIR wavelength is very less sensitive to aerosols, and (b) for vegetated target there exists a visible/NIR sur-
face reflectance relationship that depends on greenness of the target defined by quasi-normalized difference
vegetation index (NDVI′ ):
((RCR0.808 − RCR0.649 ))
NDVI ′ = , (4)
(RCR0.808 + RCR0.649 )

where RCR refers to Rayleigh corrected reflectance. This method is based on the vegetated land surface
modeling used in the NASA Deep Blue aerosol retrieval algorithm (Hsu et al., 2013, 2017, 2019).
4.3.2. Visible/NIR Surface Reflectance Relationship
To determine the spectral surface reflectance relationships between visible and NIR channels of Cartosat-2S,
the best strategy is to use the atmospherically corrected Cartosat-2S surface reflectance data generated using
in situ aerosol measurements. However, in situ aerosol observations collocated with Cartosat-2S data were
not available; therefore, hyperspectral data from Airborne Visible/Infrared Imaging Spectrometer Next Gen-
eration (AVIRIS-NG) acquired during late 2015 and early 2016 under Indian campaign are used (Mishra et
al., 2019; Thompson et al., 2018, 2019). The in situ AODs were measured using sun-photometers during the
AVIRIS-NG flights. The AVIRIS-NG data sets were corrected for atmospheric effects using measured AOD,
and then some of the pure vegetated and soil spectra were chosen. We selected six different types of dense
vegetation spectrum and six different soil type spectra. By mixing these soil and vegetation spectra, we syn-
thesized a total of 630 spectra (Figure 6a) with different values of NDVI (normalized difference vegetation
index). Then by convolution of these spectra with Cartosat-2S MX SRF (Figure 6b), we estimated the TOA
reflectance for NIR and red/blue channels using RT calculations (i.e., using equation (1)) assuming two val-
ues of AOD at 0.55 μm equal to 0.1 and 0.5 and continental aerosol type. Based on these results, we developed
an empirical relationship that estimates the surface reflectance at visible channels (0.485 and 0.649 μm) as
a function of the reflectance in the NIR band (i.e., 0.808 μm) and NDVI′ values. The empirically developed
relationships are given by

ESR0.485 = RCR0.808 (a ∗ NDVI ′ + b) + c,

(5)
ESR0.649 = RCR0.808 (d ∗ NDVI ′ + e) + 𝑓 .

Here RCR refers to Rayleigh corrected reflectance. The model coefficients a, b, c, d, e, and f are calculated
according to two NDVI′ classes. In addition, aerosol retrieval is only done if NDVI′ ⟩0.1 and ESR0.485 ⟨0.15.
Table 3 shows the resulting coefficients for each class.
To assess the performance of relationships in equation (5), we used four AVIRIS data sets acquired over
Kalaburgi and Muddur in Karnataka province, Patna in Bihar province, and Zawar in Rajasthan province

MISHRA ET AL. 7 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 6. (a) The 630 surface reflectance spectra obtained by mixing of pure vegetation and soil spectra with overlaid spectral response function (SRF) of
Cartosat-2S MX bands. (b) Equivalent 630 spectra convolved with Cartosat-2S MX bands SRF.

of India. The AVIRIS-NG data at 5.0-nm spectral resolution are convolved to Cartosat-2S MX SRF. Figure 7
shows the NCC image of all four data sets. Kalaburgi and Muddur are mainly urban and mixed-agriculture
sites, respectively. Patna is a city situated at the bank of Ganga River, and the region shown in the image
is an urbanized area with some portion of the Ganga River catchment area. Zawar is a geological site
mainly of mineralogical interest. Figure 7 shows the scatter plots between atmospherically corrected sur-
face reflectance (SR0.485 and SR0.649 ) and the estimated surface reflectance (ESR0.485 and ESR0.649 ) for blue
and red channel using NIR measurements (RCR0.808 ) in visible/NIR relationship (equation (5)). The linear
regression statistics of comparison are provided in Table 4. High correlation coefficients, r, are found at all
four stations with r ranging between 0.88 to 0.96 and 0.98 to 0.99 for the blue and red bands, respectively.
The slope of the linear regression lines shows variation about unity by 21% and 24%, respectively, for blue
and red bands. The relative error of estimates for all four sites ranges from 10% to 20% and 7% to 16% for the
blue and red bands, respectively. The maximum root mean square error observed is 0.012 and 0.019 for the
blue and red bands, receptively. An error in surface reflectance of the order of 0.01 leads to an error of up to
20% in inverted AOD (Mishra, 2018).
4.3.3. Aerosol Look-Up Tables
The inversion of atmospheric contribution to AOD requires RT calculations of various radiative quantities,
namely, path reflectance, aerosol transmittance, and spherical albedo, for different aerosol types and aerosol
loading. Since on-line RT simulations are computationally very time consuming, therefore to avoid on-line
RT simulations, pre-computed look-up tables (LUTs) of radiative quantities are used in operational aerosol
retrieval code. For RT calculations, the optical properties of seven aerosol models are used. Out of these, five
aerosol models are generated by external mixing of three basic aerosol types, namely, water-soluble aerosol,
dust-like aerosol, and shoot aerosol, assuming number density as given in Table 5. The parameters required
in Mie calculation to derive optical properties of the basic aerosol types are provided in Table 6 with an
assumption of log-normal particle-size distribution function:
[ 2 ]
1 ln r∕rm
n(r) = exp . (6)
rln𝜎 2ln2 𝜎

Here r is the particle radius, rm is the mean radius, and 𝜎 is the width of distribution function. It is to be noted
that the physical properties of the basic aerosol components described in Table 6 are from WMO-WCP112

Table 3
Surface Reflectance Coefficients Over Vegetated Regions for Cartosat-2S MX Blue and
Red Bands Centered at 0.485 and 0.649 μm, Respectively
Coefficients 0.0⟨NDVI′ ⟨0.5 NDVI′ ≥ 0.5
a, b, c −0.595, 0.450, −0.0020 −0.033, 0.324, −0.00220
d, e, f −1.258, 0.988, −0.0304 −0.571, 0.559, −0.00375

MISHRA ET AL. 8 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 7. Comparison of estimated surface reflectance (ESR0.649 and ESR0.485 ) using relations in equation (5) with actual surface reflectance (SR0.649 and
SR0.485 ) over AVIRIS-NG four flight paths. The natural color composite images of AVIRIS-NG data are also shown.

MISHRA ET AL. 9 of 21
 Earth and Space Science 10.1029/2019EA000710

Table 4
The Linear Regressions Statistics for the Comparison of Estimated Surface Reflectance Using
Equation (5) and Actual Surface Reflectance Including Linear Regression Slope s, Linear Regression
Offset i, Correlation Coefficient r, Root Mean Square Error 𝜎 , Mean of Actual Surface Reflectance ⟨SR⟩,
and Relative Error 𝜎∕⟨SR⟩
Band AVIRIS-NG site m i r 𝜎 ⟨SR⟩ 𝜎 ∕⟨SR⟩
Blue Kalaburgi 0.91 −0.002 0.94 0.011 0.067 0.16
Muddur 1.05 0.0001 0.88 0.012 0.059 0.20
Patna 0.79 0.0030 0.96 0.011 0.063 0.17
Zawar 1.00 0.0050 0.96 0.008 0.081 0.10
Red Kalaburgi 0.76 0.0160 0.98 0.008 0.091 0.09
Muddur 0.92 −0.008 0.98 0.019 0.120 0.16
Patna 0.88 0.0010 0.98 0.012 0.090 0.13
Zawar 0.92 0.0040 0.99 0.010 0.150 0.07

(1986); however, updated physical properties of these components as well as of many more aerosol models
can be referred from Hess et al. (1998), to generate improved LUTs which may closely represent real aerosols.
The parameters for the rest two models, namely, biomass burning and desert dust, are taken from Dubovik
et al. (2002) and from D'Almeida et al. (1991). The dust model represents mineral dust particles blown into
the atmosphere by the wind from arid or semi-arid regions and may have transported far away from the
source. The single scattering albedo (asymmetry parameter) of desert dust model for Cartosat bands with
central wavelength of 485, 555, 650, and 815 nm are 0.949, 0.965, 0.973, and 0.985 (0.675, 0.665, 0.653, and
0.632), respectively (D'Almeida et al., 1991). The increasing trend of single scattering albedo with wavelength
is due to the decrease in magnitude of imaginary part of refractive index. The biomass burning aerosols are
dominated by strongly absorbing fine mode black carbon particles. The single scattering albedo (asymmetry
parameter) of biomass burning model for wavelength of 485, 555, 650, and 815 nm are 0.938, 0.932, 0.922,
and 0.903 (0.651, 0.623, 0.578, and 0.517), respectively (Dubovik et al., 2002). In the present version of the
algorithm, 6SV RT code is used for generating LUTs. For every aerosol model, LUTs of path reflectance,
transmittance, and spherical albedo corresponding to AOD = 0, 0.2, 0.4, 0.6, 0.8, 1.0, 1.5, 2.0 at 0.55 μm are
generated. LUTs for path reflectance are generated for discrete values of solar zenith, sensor zenith, and
relative azimuth, while LUTs for transmittance are generated for discrete values of solar and sensor zenith.
The spherical albedo for every value of AOD is also tabulated. Four sets of these LUTs corresponding to each
Cartosat-2S waveband are generated and saved for operational use.
4.3.4. Retrieval Method
For aerosol retrieval, we first divide the whole image into 25 × 25 pixel boxes (for Cartosat-2S MX data it is
approximately 50 m by 50 m box). For each box, we first classify pixels based on NDVI′ , and then we retrieve
aerosol properties by taking advantage of the spectral relationship (equation (5)) of surface reflectance
between visible bands centered at 0.485 and 0.649 μm and NIR band centered at 0.808 μm to account for the
effect of the ever-changing dynamics of vegetation phenology on the surface reflectance. Very bright and
thick cloud contaminated pixels are avoided using a cloud mask generated using the threshold test in visi-
ble band reflectance. The remaining pixels are sorted based on blue channel reflectance, and 50% brightest

Table 5
The Optical Properties (Single Scattering Albedo [SSA] and Asymmetry Parameter (g)) of the Aerosol Models Used in Aerosol Retrieval Algorithm
Aerosol model SSA (485/555/650/815) g (485/555/650/815) Number density (WS/DL/SC)
Rural-SF79 (SSA and g similar to model by Shettle & Fenn, 1979) 0.954/0.947/0.940/0.917 0.653/0.647/0.638/0.62 0.9999/8.8612e-5/0.0
Tropospheric-SF79 (SSA and g similar to model by Shettle & Fenn, 1979) 0.968/0.963/0.957/0.936 0.646/0.637/0.626/0.609 1.0000/0.0/0.0
Rural-6S (D'Almeida et al., 1991; Lenoble & Brogniez, 1984; 0.899/0.893/0.886/0.864 0.662/0.657/0.652/0.647 0.4900/1.1820e-4/0.5099
WMO-WCP112, 1986)
Clean rural-6S (less soot content) 0.916/0.909/0.902/0.879 0.663/0.659/0.653/0.648 0.6615/1.5687e-4/0.3383
Polluted rural-6S (high soot content) 0.839/0.835/0.829/0.808 0.655/0.651/0.646/0.643 0.2362/5.8176e-5/0.7637
Note. The number density of basic aerosol components (water soluble [WS], dust-like [DL], and soot [SC]) used in external mixing to generate different aerosol
models. The number density is normalized to unity.

MISHRA ET AL. 10 of 21
 Earth and Space Science 10.1029/2019EA000710

Table 6
Characteristics of Basic Aerosol Components Used for Generating Aerosol Models in Table 5 (WMO-WCP112,
1986)
Basic type rm 𝜎 mreal (0.49/0.45/0.64/0.86) mimg (0.49/0.45/0.64/0.86)
Water soluble 0.005 2.99 1.53/1.53/1.53/1.52 5e-3/5e-3/6e-3/1.2e-2
Dust-like 0.500 2.99 1.53/1.53/1.53/1.52 8e-3/8e-3/8e-3/8e-3
Soot 0.0118 2.00 1.75/1.75/1.75/1.75 0.45/0.44/0.43/0.43

and 20% darkest values are also masked to avoid cloud contamination. It is to be noted that the pixels with
NDVI′ ⟨0.1 are also excluded to avoid pure soil or any inland water body pixel. The remaining pixels are
used in the aerosol retrieval algorithm to derive AOD value for the 25 × 25 pixel box. After estimating the
surface reflectance at 0.485 and 0.649 μm for the remaining pixels using relations in equation (5) for vege-
tated pixels in each box, we estimate mean surface reflectance for each box. Using these surface reflectance
values and pre-computed LUTs, modeled TOA reflectance at 0.485 and 0.649 μm are computed for given
sun-sensor geometry and for different values of indexed AOD, and single scattering albedo. These 0.485
and 0.649 μm modeled TOA reflectance are then compared to Cartosat-2S MX measured TOA reflectance
to match the appropriate values of AOD at 0.55 μm and aerosol model (i.e., single scattering albedo) for
each box. Since AOD at 0.55 μm is retrieved at a spatial resolution of 1/25 of the original resolution, the
cubic interpolation is performed to re-sample derived aerosol optical thickness map at original spatial
resolution.

4.4. Atmospheric Correction

The method for deriving surface reflectance from Cartosat-2S MX TOA radiance data is shown in Figure 8,
which consists of following major steps:
(a) One combined transmittance spectrum for O3 and O2 is calculated using sun-sensor geometry and cli-
matological values for O3 and O3 content. One H2 O transmittance spectrum is generated using total
column H2 O content from forecast reanalysis data. Both transmittance spectra are convolved for four
MX bands.
(b) The TOA 11-bit digital count image for each MX band is converted to TOA apparent reflectance using
relation

𝜋(Saturation Radiance∕2047) ∗ DN
𝜌TOA = , (7)
F0 𝜇s

where DN, 𝜇 s , and F0 refer to digital count, cosine of solar zenith, and extraterrestrial irradiance
corrected for earth-sun distance, respectively.
(c) The TOA reflectance is corrected for gaseous absorption by dividing 𝜌TOA of all bands with gaseous
transmittance computed in step (a).
(d) The entire image is divided into 25 × 25 pixel boxes. AOD for each 25 × 25 box of the scene is retrieved
by the method described in Section 4.3, and re-sampling of AOD map to original resolution using cubic
interpolation is done.
(e) Using estimated AOD and aerosol model, the radiative quantities, namely, the atmospheric reflectance,
spherical albedo, and the total upward and downward transmissions, are estimated with the aid of
pre-computed LUTs for all MX bands.
(f) Finally, the surface reflectance for each MX band is derived based on equation (2). In our computer
program for AC, the first three steps need to be performed only once at the beginning of the pro-
gram execution, and the last three steps are performed for each pixel. The output consists of the AOD
map at 0.55 μm and surface reflectance for all MX bands. It is important to note that the two-step
approach for AC (first estimate AOD and then use that to estimate surface properties) described
above introduces inconsistency in that the eventual AC surface reflectance may not be consistent
with the surface reflectance assumed in the AOD retrieval part even if they are done at the same
resolution.

MISHRA ET AL. 11 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 8. Flowchart to derive surface reflectance and aerosol properties using Cartosat-2S data.

5. Results and Discussion

5.1. Visual Improvement in Images After Atmospheric Correction
Using aerosol retrieval and AC algorithm, the AOD and surface reflectance products are generated for all
106 Cartosat-2S acquired over locations shown in Figure 4. Figures 9–14 show the results (aerosol and sur-
face reflectance) obtained from the Cartosat-2S data sets acquired over Delhi and Ahmedabad (see Table 1).
Figure 9b shows the retrieved AOD map from Cartosat-2S MX data acquired on 01 November 2016 over a part
of Delhi. It is clear from the AOD map that the left part of the image shows a high AOD (>1.2) relative to the
right part of the image. On average, whole image shows heavy aerosol loading with mean AOD > 1.2. Heavy

Figure 9. (a) MODIS-aqua NCC image and hot spot layer (red dots) showing dust storm event and active fires, respectively, affecting the visibility of region
imaged by Cartosat-2S on 01 November 2016 (red box). (b) Aerosol optical depth (AOD) map Cartosat-2S on 01 November 2016.

MISHRA ET AL. 12 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 10. Visual comparison of TOA and derived surface reflectance from data acquired by Cartosat-2S MX on
01 November 2016 over part of Delhi. (a, b) TOA and derived surface reflectance NCC image of full scene. (c, d) Same
as (a) and (b) but for Subset 1 represented by red square shown in (a). (e, f) Same as (a) and (b) but for Subset 2
represented by red square shown in (a).

aerosol loading on 01 November 2016 may have occurred due to the transportation of dust aerosols from the
regions of Afghanistan-Pakistan border over which a heavy dust storm event was detected in MODIS NCC
image (Figure 9a) on 01 November 2016. In addition to this, the practice of agriculture residue burning in
November month in adjoining northern states (Haryana and Punjab) of Delhi also has introduced a large
amount of carbonaceous aerosol as shown by MODIS-aqua hot spots (red dots) in Figure 9a. Similar qualita-
tive information is seen from Cartosat-2S MX NCC image shown in Figure 10a, where appreciable haziness,
especially in the left part of the image, is apparent. Figures 10a and 10b show TOA apparent reflectance and
retrieved surface reflectance NCC images, respectively, derived from Cartosat-2S MX data using the present
algorithm. Figures 10c and 10d show TOA and surface reflectance, respectively, for the Subset 1 (red square
in Figure 10a), while Figures 10e and 10f show TOA and surface reflectance, respectively, for the Subset 2
(red square in Figure 10a). From Figure 10, it is clear that derived surface reflectance (right panels) is free
from haze as compared to TOA NCC images (left panels) and therefore the surface features especially in the

MISHRA ET AL. 13 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 11. (a) Standard spectra for soil targets and Cartosat-2S MX derived surface reflectance (left) and TOA reflectance (right) spectra over soil targets on 01
November 2016 over part of Delhi. (b) Same as (a) but for vegetated targets.

left region that was obscured in TOA reflectance image became visible after AC. Figure 11 shows a quanti-
tative comparison of TOA apparent reflectance and derived surface reflectance from Cartosat-2S MX data
acquired on 01 November 2016 (shown in Figure 10) over some of the vegetated and soil type pixels. The
standard vegetated and soil spectra are also shown in Figure 11 to compare the shapes of the derived spectra.
From Figures 11a and 11b, it is clear that over soil and vegetated targets, respectively, the shapes of derived
surface reflectance spectra are in good agreement with standard spectra. While TOA apparent reflectance
spectra show very high reflectance values and shapes are completely different from the standard spectra.
Figure 12 shows the histogram analysis of TOA reflectances and the derived surface reflectances for Bands
1, 2, and 3 of Cartosat-2S MX data acquired on 01 November 2016. It is clear from the figure that after AC the
mean values (standard deviation values) of derived surface reflectance are significantly lower (larger) than
that of TOA reflectance for all three bands. It is clear that AC has appreciably increased (by a factor of >2.4)
the standard deviation (increased contrast) compared to TOA reflectance which shows low standard devi-
ation due to homogeneity introduced by heavy atmospheric aerosol layer. The maximum change in mean
reflectance value is observed for Band 1 (blue waveband), which is expected as blue wavelengths are most
sensitive to aerosol scattering.

Figure 12. The histogram of TOA reflectances and the derived surface reflectances for Bands 1, 2, and 3 of Cartosat-2S MX data acquired on 01 November 2016
over Delhi.

MISHRA ET AL. 14 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 13. Visual comparison of TOA and derived surface reflectance from Cartosat-2S MX data acquired on
03 October 2017 over Ahmedabad. (a, b) TOA and derived surface reflectance NCC image of full scene. (c, d) Same as
(a) and (b) but for Subset 1 represented by red square shown in (a). (e, f) Same as (a) and (b) but for Subset 2
represented by red square shown in (a).

Similarly, we processed the Data Set 2 acquired by Cartosat-2S on 03 October 2017 that covers part of Ahmed-
abad. Figure 13 shows a visual comparison of TOA reflectance and derived surface reflectance NCC images.
From Figure 13a we can see that significant region of the image is contaminated by bright haze and thick
clouds and therefore the surface features underneath the thick haze layer are almost invisible (see region
encompassed by redlined boxes and adjoining region in Figure 13a). On the other hand, in derived surface
reflectance NCC image (Figure 13b) the haze is significantly cleared and the surface features became visi-
ble. Figures 13c and 13d, and 13e and 13f, show the effect of AC for very hazy regions represented by Subsets
1 and 2, respectively. It is clear from these figures that except the regions with a very thick cloud where no

MISHRA ET AL. 15 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 14. The histogram of TOA reflectances and the derived surface reflectances for Bands 1, 2, and 3 of Cartosat-2S MX data acquired on 03 October 2017
over Ahmedabad.

information about surface reaches TOA, the surface features become visible and clear from the haze. It is well
known that the effect of high aerosol scattering on TOA data is to decrease the heterogeneity of the surface
features and generally increase the brightness of the scene. Figure 14 shows the histogram analysis of TOA
reflectances and the derived surface reflectances for Bands 1, 2, and 3 of Cartosat-2S MX data acquired on 03
October 2017. Similar to the previous case, here also after AC, the mean values (standard deviation values) of
derived surface reflectance are significantly lower (larger) than that of TOA reflectance for all three bands.
The AC has appreciably increased (by a factor of >1.7) the standard deviation compared to TOA reflectance
which shows low standard deviation due to homogeneity introduced by heavy atmospheric aerosol layer.
Here also the maximum change in mean reflectance value is observed for Band 1 (blue waveband), which
is expected. Therefore, these statistical parameters (decreased mean and increased standard deviation after
atmospheric correction) and visuals presented in Figures 9–14 prove that our algorithm effectively removes
atmospheric effects (haze due to aerosols), thereby increasing the contrast of the surface features.

5.2. Inter-sensor Comparison of Aerosol and Surface Reflectance Product

The qualitative analysis in the previous subsection clearly shows the improvement in Carosat-2S MX
image quality, the shape of the spectrum, and the statistical parameters (mean and standard deviation of
reflectance), after AC. However, the validation of aerosol and surface reflectance products with other sources
is needed to establish the validity of Cartosat-2S products and the presented algorithms. Due to lack of in situ
measurements of AOD and surface reflectance collocated in time and space with Cartosat-2S observations,
in the present work, MODIS-terra aerosol product at the spatial resolution of 3 km and surface reflectance
product at a spatial resolution of 500 m are used to compare Cartosat-2S products. For more information on
MODIS-terra products used in this work, see Section 3.
5.2.1. Comparison Methodology
The Cartosat-2S AOD and surface reflectance products are at different and very fine spatial resolution of 50
and 2 m, respectively, as compared to the MODIS-terra product. Moreover, the projection of Cartosat-2S
products is in UTM, while MODIS-terra aerosol and surface reflectance products are provided with per
pixel latitude and longitude information as separate layers for a given tile. Due to all these differences,
proper collocation, down-sampling, and data selection criteria of Cartosat-2S data must be defined for faith-
ful comparison with MODIS-terra product. The reason for selecting the MODIS-terra product is because
Cartosat-2S equatorial pass time is very near to that of MODIS-terra. As the first step, MODIS-terra AOD is
geo-referenced to geographic Lat/Lon projection using the latitude and longitude layers provided with each
data set. It is to be noted that before geo-referencing the AOD layer, the pixels with quality flag equal to 2 or
3 (i.e., very good or best quality retrievals) are only retained; all other pixels are masked with a fill value. In
the second step, the Cartosat-2S data are converted from UTM to geographic Lat/Lon projection. For every
MODIS Lat/Lon grid box, the corresponding Cartosat-2S AOD pixels are identified based on latitude and lon-
gitude information. The pixels having valid AOD retrieval among all of the identified pixels are counted. If
valid AOD pixel count is over ≥50% of total pixels, the average value of valid pixel values is filled. Doing so the

MISHRA ET AL. 16 of 21
 Earth and Space Science 10.1029/2019EA000710

down-sampled Cartosat AOD at 3 km and surface reflectance at 500 m

spatial resolution is obtained which are perfectly collocated with
MODIS-terra geo-reference AOD and surface reflectance products,
respectively. Finally, pixel by pixel values of AOD (at 3-km resolution)
and surface reflectance (at 500-m resolution) from 106 down-sampled
Cartosat-2S products and corresponding values from MODIS-terra
geo-referenced product are tabulated, and linear regression statistics
are estimated. The normalized difference vegetation index (NDVI) from
collocated surface reflectance of Cartosat-2S and MODIS-terra is also
calculated for comparison.
5.2.2. Comparison of Aerosol Optical Depth
The comparison of Cartosat-2S AOD (𝜏 c ) with MODIS-terra AOD (𝜏 m ) is
plotted in Figure 15. The linear regressions statistics of these validations
are provided in Table 7. High correlation coefficients, r, are found with a
value of r ≥ 0.95. Slopes of the linear regression lines show small vari-
ation about 1 by 3–6%, probably caused by the use of slightly different
aerosol optical properties in the Cartosat-2S aerosol algorithm as com-
pared to MODIS-terra aerosol algorithm where a dynamic mixture of fine
and coarse mode aerosols is used. For a similar reason or due to uncer-
Figure 15. The comparison of Cartosat-2S AOD (𝜏 c ) with MODIS-terra
AOD (𝜏 m ) at spatial resolution of 3 km. tainty in empirical surface reflectance (due to the BRDF effect) model,
the Cartosat-2S AOD is slightly negatively biased; however, the observed
bias is within the expected uncertainty of 25–30% for such algorithms (Levy et al., 2007). The slope of regres-
sion curve and negative bias of Cartosat-2S AOD relative to MODIS-terra AOD retrieval may be improved
by using the location-based aerosol model selection scheme with dynamic mixing of course and fine mode
aerosols with different scattering and absorbing properties. The relative difference of Cartosat-2S AOD esti-
mates (𝜎∕⟨P⟩m ), that is, the standard error of regression (𝜎 ) compared to the average observed AOD from
MODIS-terra (⟨P⟩m ), is found to be 25%. The result shown in Figure 15 and in Table 7 established that 𝜏 c
have an average deviation of around 25% with respect to 𝜏 m at 3-km spatial resolution; however, from scat-
ter plots in Figure 15, it is evident that the deviation of 𝜏 c relative to 𝜏 m is a function of 𝜏 m itself. Therefore,
scatter plot in Figure 15 is somewhat heteroscedastic in nature which may lead to bias in linear regression
fit parameters shown in Figure 15 and in Table 7. This makes necessary to derive uncertainty in 𝜏 c given by
variable Δ𝜏 c as a function of 𝜏 m . For this all 1,500 data points are first sorted in increasing order of 𝜏 m and
then are divided into 15 bins with equal number of data point (approximately 100 data points in each bin)
entries. For each bin, the modulus of difference of AODs |𝜏 c − 𝜏 m | is calculated. It is found that |𝜏 c − 𝜏 m |
shows near-normal distribution; therefore, the first standard deviation of this value (i.e., the value of |𝜏 c −𝜏 m |
below which 68% of retrieval falls) is defined as the uncertainty value Δ𝜏 c for the corresponding bin (Sayer
et al., 2013). Doing so 15 values of Δ𝜏 c are estimated corresponding to average value of 𝜏 m for each bin. The
results are plotted in Figure 16, where from the linear regressions analysis shows that Δ𝜏 c = 0.03 + 0.14𝜏 m ,
which means the uncertainty in Δ𝜏 c increases with increasing value of 𝜏 m . The extensive global validation

Table 7
The Linear Regressions Statistics for the Comparison of Various Geophysical Parameters Retrieved From Cartosat-2S and MODIS-terra Data Sets Including Linear
Regression Slope s, Linear Regression Offset i, Correlation Coefficient r, Root Mean Square Error (rmse) 𝜎 , Mean of Actual Surface Reflectance ⟨SR⟩, and Relative
Error 𝜎∕⟨SR⟩
Parameter (P) Spatial resolution N s i r 𝜎 ⟨P⟩c ⟨P⟩m 𝜎∕⟨P⟩m
𝜏 3 km 1,530 1.03 −0.005 0.976 0.088 0.350 0.358 0.246
SRblue 500 m 4,663 0.80 0.02 0.75 0.014 0.057 0.048 0.29
SRgreen 500 m 4,663 0.75 0.001 0.76 0.021 0.066 0.086 0.24
SRred 500 m 4,663 0.81 −0.01 0.86 0.034 0.076 0.091 0.37
SRnir 500 m 4,663 0.72 0.01 0.76 0.082 0.24 0.31 0.26
NDVI 500 m 4,663 1.05 0.01 0.89 0.1 0.59 0.55 0.18
Note. Here variables 𝜏 , SR, and NDVI refer to aerosol optical depth (AOD), surface reflectance, and normalized difference vegetation index, respectively.

MISHRA ET AL. 17 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 16. Uncertainty in Cartosat-2S AOD (Δ𝜏 c ) as a function of MODIS AOD (𝜏 m ) at 3-km spatial resolution.

of MODIS-terra aerosol retrievals with in situ AOD (𝜏 i ) shows an uncertainty of Δ𝜏 m = ±(0.05 + 0.15𝜏 i )
in 𝜏 m (Levy et al., 2010; Remer et al., 2013). Using the equation for Δ𝜏 m and Δ𝜏 c , it is found that the error
in retrieved AOD at low values (𝜏 i = 0.25) may be up to 65% of 𝜏 i . However, the error in retrieved AOD
decreases to 30% for moderate AOD values (𝜏 i = 0.5) and then again starts increasing with an error of around
40% for high AOD values (𝜏 i = 1.0). These observations show that the Cartosat AOD retrievals may have
high errors when actual AOD is too low (<0.25) and too high (>1.0). It is to be noted that these error esti-
mates hold valid when Δ𝜏 m and Δ𝜏 c are of the same sign; however, if Δ𝜏 m and Δ𝜏 c are of opposite sign, the
actual error will be less.
5.2.3. Comparison of Surface Reflectance
The Cartosat-2S surface reflectance (SRc,blue , SRc,green , SRc,red , SRc,nir ) comparison with MODIS-terra (SRm,blue ,
SRm,green , SRm,red , SRm,nir ) at spatial resolution of 500 m are plotted in Figure 16. The linear regressions statis-
tics of these validations are provided in Table 7. High correlation coefficients, r, for all bands are found with
values in range 0.75 to 0.86. The red band shows maximum correlation with r equal to 0.86, which is expected
as red channels of both sensors have similar center wavelengths (for SRF and wavelengths of MODIS-terra
bands, see Xiong et al., 2006). Variation of slopes of the linear regression lines about unity by 19–28% may
be associated with the difference in full width half maximum (FWHM), center wavelengths, and the SRF of
Cartosat-2S and MODIS-terra bands. The relative error of estimates (𝜎∕⟨P⟩m ), that is, the standard error of
regression (𝜎 ) compared to the average observed surface reflectance from MODIS-terra (⟨P⟩m ), is found to
be in range 24% to 37%. The high relative differences in MODIS and Cartosat-2S surface reflectance are also
associated with the difference in FWHM and the shape of SRFs of the two sensors. Moreover, this discrep-
ancy will be surface-type dependent. The comparison of NDVI obtained from Cartosat-2S and MODIS-terra
surface reflectance is plotted in Figure 17. The linear regressions statistics of NDVI comparison are provided
in Table 7. Very good correlation r with value 0.89 and slope near unity is found. The relative error of esti-
mates for NDVI is found to be ≤18%, which is due to the large difference in the center wavelength of NIR
channels.

6. Summary
Ability of aerosol retrieval and accurate AC of data acquired from low cost and high spatial resolution
( 1–2 m) optical VNIR sensors such as Cartosat, World-view-4, and SPOT-7 will keep these sensors relevant
in remote sensing of Earth's surface and atmosphere. In fact, it will make high-resolution VNIR sensors
complementary to advanced but coarser VSWIR multi/hyperspectral instruments. Here we described an
aerosol retrieval and AC algorithm for multispectral VNIR data acquired over land in fewer spectral bands
(three to five bands). The AOD, surface reflectance, and NDVI obtained by processing Cartosat-2S MX data
at 106 globally distributed locations have been compared with corresponding MODIS-terra products col-
located in time and space to test the performance of our algorithms. The comparison of Cartosat-2S and

MISHRA ET AL. 18 of 21
 Earth and Space Science 10.1029/2019EA000710

Figure 17. (a–d) The comparison of Cartosat-2S surface reflectance (SRc,blue , SRc,green , SRc,red , SRc,nir ) with
MODIS-terra (SRm,blue , SRm,green , SRm,red , SRm,nir ) for blue, green, red, and NIR bands, respectively, at spatial
resolution of 500 m. (e) Comparison of NDVI derived from Cartosat-2S and MODIS-terra surface reflectances at 500-m
spatial resolution.

MISHRA ET AL. 19 of 21
 Earth and Space Science 10.1029/2019EA000710

MODIS-terra derived AOD, surface reflectance, and NDVI shows a very good correlation. The relative dif-
ference in Cartosat-2S AOD with MODIS-terra AOD is within 25%. It is to be noted that the presented
algorithm requires measurements in at least one NIR band; therefore, similar algorithm with sensor-specific
modifications can be applied for AC of other high-resolution VNIR sensors such as Worldview-4 and SPOT-7.

Acknowledgments References
We would like to acknowledge
Dr. Tapan Misra and Dr. Prakash Ahmad, Z., Franz, B. A., McClain, C. R., Kwiatkowska, E. J., Werdell, J., Shettle, E. P., & Holben, B. N. (2010). New aerosol models for the
Chauhan for providing critical retrieval of aerosol optical thickness and normalized water-leaving radiances from the SeaWiFS and MODIS sensors over coastal regions
suggestions during the algorithm and open oceans. Applied Optics, 49, 5545–5560.
development. The MODIS-terra data Antoine, D., & Morel, A. (1999). A multiple scattering algorithm for atmospheric correction of remotely sensed ocean colour (MERIS
used here are freely available from instrument): Principle and implementation for atmospheres carrying various aerosols including absorbing ones. International Journal
LAADS DAAC - NASA website of Remote Sensing, 20(9), 1875–1916. https://doi.org/10.1080/014311699212533
(https://ladsweb.modaps.eosdis.nasa. Asner, G., Martinand, R., Knapp, D., Tupayachi, R., Anderson, C., Sinca, F., et al. (2017). Airborne laser-guided imaging spectroscopy to
gov/). The Cartosat-2S data are not map forest trait diversity and guide conservation. Science, 355, 385–389.
freely available to the user community. Berk, A., Conforti, P., Kennett, R., Perkins, T., Hawes, F., & van den Bosch, J. (2014). MODTRAN6: A major upgrade of the MODTRAN radia-
tive transfer code. In M., Velez-Reyes, & F. A. Kruse (Eds.), Algorithms and technologies for multispectral, hyperspectral, and ultraspectral
imagery XX (Vol. 9088, pp. 113–119). SPIE. https://doi.org/10.1117/12.2050433
Cao, C., De Luccia, F. J., Xiong, X., Wolfe, R., & Weng, F. (2014). Early on-orbit performance of the visible infrared imaging radiometer
suite onboard the Suomi National Polar-Orbiting Partnership (S-NPP) satellite. IEEE Transactions on Geoscience and Remote Sensing,
52(2), 1142–1156. https://doi.org/10.1109/TGRS.2013.2247768
Cartosat-2 brochure (n.d). National Remote Sensing Centre, Indian Space Research Organization. https://www.nrsc.gov.in/sites/default/
files/pdf/brcartosat2.pdf
Chandrasekhar, S. (1960). Radiative transfer: Dover Publications.
D'Almeida, GA., Koepke, P., & Shettle, E. P. (1991). Atmospheric aerosols: Global climatology and radiative characteristics: A. Deepak Pub.
Dubovik, O., Holben, B., Eck, T. F., Smirnov, A., Kaufman, Y. J., King, M. D., et al. (2002). Variability of absorption and optical proper-
ties of key aerosol types observed in worldwide locations. Journal of the Atmospheric Sciences, 59(3), 590–608. https://doi.org/10.1175/
1520-0469(2002)059h0590:VOAAOPi2.0.CO;2
Gordon, H. R. (1997). Atmospheric correction of ocean color imagery in the Earth Observing System era. Journal of Geophysical Research,
102(D14), 17,081–17,106. https://doi.org/10.1029/96JD02443
Hess, M., Koepke, P., & Schult, I. (1998). Optical properties of aerosols and clouds: The software package OPAC. Bulletin of the American
Meteorological Society, 79(5), 831–844. https://doi.org/10.1175/1520-0477(1998)h0831:OPOAACi.0.CO;2
Holben, B. N., Eck, T. F., Slutsker, I., Tanré, D., Buis, J. P., Setzer, A., et al. (1998). AERONET—A federated instrument network and data
archive for aerosol characterization. Remote Sensing of Environment, 66(1), 1–16. https://doi.org/10.1016/S0034-4257(98)00031-5
Hsu, N. C., Jeong, M. J., Bettenhausen, C., Sayer, A. M., Hansell, R., Seftor, C. S., et al. (2013). Enhanced deep blue aerosol retrieval
algorithm: The second generation. Journal of Geophysical Research: Atmospheres, 118, 9296–9315. https://doi.org/10.1002/jgrd.50712
Hsu, N. C., Lee, J., Sayer, A. M., Carletta, N., Chen, S. H., Tucker, C. J., et al. (2017). Retrieving near-global aerosol loading over land and
ocean from AVHRR. Journal of Geophysical Research: Atmospheres, 122, 9968–9989. https://doi.org/10.1002/2017JD026932
Hsu, N. C., Lee, J., Sayer, A. M., Kim, W., Bettenhausen, C., & Tsay, S. C. (2019). VIIRS Deep Blue aerosol products over land: Extend-
ing the EOS long-term aerosol data records. Journal of Geophysical Research: Atmospheres, 124, 4026–4053. https://doi.org/10.1029/
2018JD029688
Jetz, W., Cavender-Bares, J., Pavlick, R., Schimel, D., Davis, F. W., Asner, G., et al. (2016). Monitoring plant functional diversity from space.
Nature Plants, 2, 16,024.
Kaufman, Y. J., Wald, A. E., Remer, L. A., & Flynn, L. (1997). The MODIS 2.1-/spl mu/m channel-correlation with visible reflectance for use
in remote sensing of aerosol. IEEE Transactions on Geoscience and Remote Sensing, 35(5), 1286–1298. https://doi.org/10.1109/36.628795
Kotchenova, S. Y., Vermote, E. F., Matarrese, R., & Frank Jr, J. K. (2006). Validation of a vector version of the 6S radiative transfer code for
atmospheric correction of satellite data. Part I: Path radiance. Applied Optics, 45(26), 6762–6774. https://doi.org/10.1364/AO.45.006762
Kruse, F. (2004). Comparison of ATREM, ACORN, and FLAASH atmospheric corrections using low-altitude AVIRIS data of Boulder, CO.
13th JPL Airborne Geoscience Workshop. Jet Propulsion Laboratory Publication.
Lenoble, J. (1985). Radiative transfer in scattering and absorbing atmospheres: Standard computational procedures. Hampton, VA: A. Deepak
Publishing.
Lenoble, J., & Brogniez, C. (1984). A comparative review of radiation aerosol models. Beitraege zur Physik der Atmosphaere, 57, 1–20.
Levy, R., & Hsu, C. (2015). MODIS atmosphere L2 aerosol product. USA: NASA MODIS Adaptive Processing System, Goddard Space Flight
Center. https://doi.org/10.5067/MODIS/MOD04L2.061
Levy, R. C., Remer, L. A., Kleidman, R. G., Mattoo, S., Ichoku, C., Kahn, R., & Eck, T. F. (2010). Global evaluation of the Collection
5 MODIS dark-target aerosol products over land. Atmospheric Chemistry and Physics, 10(21), 10,399–10,420. https://doi.org/10.5194/
acp-10-10399-2010
Levy, R. C., Remer, L. A., Mattoo, S., Vermote, E. F., & Kaufman, Y. J. (2007). Second-generation operational algorithm: Retrieval of aerosol
properties over land from inversion of moderate resolution imaging spectroradiometer spectral reflectance. Journal of Geophysical
Research, 112, D13211. https://doi.org/10.1029/2006JD007811
Lorenzen, B., & Jensen, A. (1988). Reflectance of blue, green, red and near infrared radiation from wetland vegetation used in a
model discriminating live and dead above ground biomass. New Phytologist, 108(3), 345–355. https://doi.org/10.1111/j.1469-8137.1988.
tb04173.x
Mayer, B., & Kylling, A. (2005). Technical note: The libRadtran software package for radiative transfer calculations—Description and
examples of use. Atmospheric Chemistry and Physics, 5(7), 1855–1877. https://doi.org/10.5194/acp-5-1855-2005
Mishra, M. K. (2018). Retrieval of aerosol optical depth from INSAT-3D imager over Asian landmass and adjoining ocean: Retrieval
uncertainty and validation. Journal of Geophysical Research: Atmospheres, 123, 5484–5508. https://doi.org/10.1029/2017JD028116
Mishra, M. K., Gupta, A., John, J., Shukla, B. P., Dennison, P., Srivastava, S. S., et al. (2019). Retrieval of atmospheric parameters and
data-processing algorithms for AVIRIS-NG Indian campaign data. Current Science, 116(7), 00113891.
Peñuelas, J., & Filella, I. (1998). Visible and near-infrared reflectance techniques for diagnosing plant physiological status. Trends in Plant
Science, 3(4), 151–156. https://doi.org/10.1016/S1360-1385(98)01213-8

MISHRA ET AL. 20 of 21
 Earth and Space Science 10.1029/2019EA000710

Perkins, T., Adler-Golden, S. M., Matthew, M. W., Berk, A., Bernstein, L. S., Lee, J., & Fox, M. (2012). Speed and accuracy improvements
in FLAASH atmospheric correction of hyperspectral imagery. Optical Engineering, 51(11), 1–8 –8. https://doi.org/10.1117/1.OE.51.11.
111707
Remer, L. A., Mattoo, S., Levy, R. C., & Munchak, L. A. (2013). MODIS 3 km aerosol product: Algorithm and global perspective. Atmospheric
Measurement Techniques, 6(7), 1829–1844. https://doi.org/10.5194/amt-6-1829-2013
Ricchiazzi, P., Yang, S., Gautier, C., & Sowle, D. (1998). SBDART: A research and teaching software tool for plane-parallel radiative transfer
in the Earth's atmosphere. Bulletin of the American Meteorological Society, 79(10), 2101–2114.
Richter, R., & Schläpfer, D. (2002). Geo-atmospheric processing of airborne imaging spectrometry data. Part 2: Atmospheric/topographic
correction. International Journal of Remote Sensing, 23(13), 2631–2649. https://doi.org/10.1080/01431160110115834
Sahay, A., Mishra, M. K., Chauhan, P., & Ajai (2014). Absorption characteristics of ocean water in the Arabian Sea during winter bloom
from in situ measurements and Oceansat 2 OCM and MODIS data. International Journal of Remote Sensing, 35(9), 2982–2995. https://
doi.org/10.1080/01431161.2014.894659
Salomonson, V. V., Barnes, W. L., Maymon, P. W., Montgomery, H. E., & Ostrow, H. (1989). MODIS: Advanced facility instrument for
studies of the Earth as a system. IEEE Transactions on Geoscience and Remote Sensing, 27(2), 145–153. https://doi.org/10.1109/36.20292
Sarangi, R. K., Mishra, M. K., & Chauhan, P. (2015). Remote sensing observations on impact of Phailin cyclone on phytoplankton distribu-
tion in northern Bay of Bengal. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 8(2), 539–549. https://
doi.org/10.1109/JSTARS.2014.2347036
Sayer, A. M., Hsu, N. C., Bettenhausen, C., & Jeong, M. J. (2013). Validation and uncertainty estimates for MODIS Collection 6 “Deep Blue”
aerosol data. Journal of Geophysical Research: Atmospheres, 118, 7864–7872. https://doi.org/10.1002/jgrd.50600
Schmit, T. J., Gunshor, M. M., Menzel, W. P., Gurka, J. J., Li, J., & Bachmeier, A. S. (2005). Introducing the next-generation Advanced
Baseline Imager on GOES-R. Bulletin of the American Meteorological Society, 86, 1079–1096. https://doi.org/10.1175/BAMS-86-8-1079
Sharma, S., Prajapati, RP, & Mathur, AK (2016). Radiometric calibration of Cartosat-2 panchromatic sensor using artificial targets:prelude
to Cartosat-2c. https://www.mosdac.gov.in/filebrowser/download/416
Shettle, E. P., & Fenn, R. W. (1979). Models for the aerosols of the lower atmosphere and the effects of humidity variations on their optical
properties Edited by Cohen, I. D., & I. D. Cohen: Environmental Research Paper Air Force Geophysics Lab, Hanscom AFB, MA. Optical
Physics Div.
Tanre, D., Deroo, C., & Duhaut, P. (1986). Simulation of the satellite signal in the solar spectrum (5s), user's guide: Laboratoire d'Optique
Atmospherique, U.S.T. de Lille, 59655 Villeneuve dáscq, France.
Thompson, D. R., Babu, K. N., Braverman, A. J., Eastwood, M. L., Green, R. O., Hobbs, J. M., et al. (2019). Optimal estimation of spectral
surface reflectance in challenging atmospheres. Remote Sensing of Environment, 232, 111,258. https://doi.org/10.1016/j.rse.2019.111258
Thompson, D. R., Boardman, J. W., Eastwood, M. L., Green, R. O., Haag, J. M., Mouroulis, P., & Gorp, B. V. (2018). Imaging spectrometer
stray spectral response: In-flight characterization, correction, and validation. Remote Sensing of Environment, 204, 850–860. https://doi.
org/10.1016/j.rse.2017.09.015
Ustin, S. L., Roberts, D. A., Gamon, J. A., Asner, G. P., & Green, R. O. (2004). Using imaging spectroscopy to study ecosystem processes and
properties. BioScience, 54, 523–534.
Vermote, E. F., Tanre, D., Deuze, J. L., Herman, M., & Morcette, J. (1997). Second simulation of the satellite signal in the solar spectrum,
6S: An overview. IEEE Transactions on Geoscience and Remote Sensing, 35(3), 675–686. https://doi.org/10.1109/36.581987
Vermote, E. F., & Wolfe, R. (2015). NASA GSFC and MODAPS SIPS-NASA. MYD09GA MODIS/aqua surface reflectance daily L2G Global
1km and 500 m SIN Grid. NASA LP DAAC, https://doi.org/10.5067/MODIS/MOD04_L2.061
WMO-WCP112 (1986). A preliminary cloudless standard atmosphere for radiation computation, vol. WMO/TD-No 24. Geneva: World
Climate Programme, World Meteorological Organization.
Xiong, X., Che, N., & Barnes, W. L. (2006). Terra MODIS on-orbit spectral characterization and performance. IEEE Transactions on
Geoscience and Remote Sensing, 44(8), 2198–2206. https://doi.org/10.1109/TGRS.2006.872083

MISHRA ET AL. 21 of 21