Remote Sensing Applications: Society and Environment 4 (2016) 158–166

Contents lists available at ScienceDirect

Remote Sensing Applications: Society and Environment

journal homepage: www.elsevier.com/locate/rsase

Atmospheric correction assessment of SPOT-6 image and its influence

on models to estimate water column transparency in tropical reservoir
Luiz H.S. Rotta n, Enner H. Alcântara, Fernanda S.Y. Watanabe, Thanan W.P. Rodrigues,
Nilton N. Imai
São Paulo State University, Department of Cartography, 19060-900 Presidente Prudente, SP, Brazil

art ic l e i nf o a b s t r a c t

Article history: Remote sensing images have been increasingly used by its ability to collect data from extensive areas in a
Received 3 March 2016 short time and with relatively low cost. Studies conducted in aquatic environments require great at-
Received in revised form tention in relation to atmospheric correction, since the signal leaving water bodies is strongly attenuated.
12 May 2016
The present work aimed to assess the atmospheric correction of SPOT-6 image based on the variation of
Accepted 9 September 2016
Available online 10 September 2016
initial visibility parameter in FLAASH and analyze its influence on models to estimate Secchi depth (SD)
and diffuse attenuation coefficient (Kd). The study was carried out in Nova Avanhandava Reservoir, which
Keywords: belongs to the chain of the Tietê River reservoirs (São Paulo, Brazil). The models calibration was based on
FLAASH remote sensing reflectance (Rrs) of simulated SPOT bands from data collected in the field. The best
Visibility
models were obtained using the band ratio Rrs(560 nm)/Rrs(660 nm) for SD (R2 ¼92%, RMSE ¼11.45%)
Diffuse attenuation coefficient (Kd)
and the band Rrs(660 nm) for Kd (R2 ¼ 92%, RMSE ¼11.98%). Maps of the spatial distribution of SD and Kd
Secchi depth (SD)
Nova Avanhandava Reservoir were made by applying the models on atmospherically corrected images. The main problem was the high
amount of negative pixels when the suitable initial visibility value was not adopted in the atmospheric
correction, which prevents the use of bio-optical models to retrieve limnological variables of the re-
servoir.
& 2016 Elsevier B.V. All rights reserved.

1. Introduction because the signal from the water column is small (Hu et al.,
2001). For Wang (1999), the atmospheric correction is a funda-
The radiation reflected or emitted by the Earth’s surface passes mental procedure in water color imagery data processing and re-
through the atmosphere, which interacts with several gases, water move about 90% of the signal measured by the sensor in the visible
vapor and particulates. Thus, the radiation is influenced by the bands. Chen et al. (2007) ignored the atmospheric correction of
atmospheric scattering, absorption, reflection and refraction be- TM-Landsat data because only one scene was used; however, the
fore being recorded by the remote sensing system (Jensen, 2009). results are not comparable temporally or spatially. Usually, atmo-
Even when the sky is clear, the intensity of the solar beam is sig- spheric correction is necessary. An algorithm to perform atmo-
nificantly reduced during its passage through the atmosphere spheric correction of SeaWiFS data for turbid coastal waters has
(Kirk, 2011). In studies of aquatic environments, application of been described and tested by Ruddick et al. (2000). According to
atmospheric corrections is recommended (Wang et al., 1999; Hu Hadjimitsis and Clayton (2009), the atmospheric correction must
et al., 2001; Jamet et al., 2011), since the signal of the water bodies be done for the assessment of temporal variations of water quality.
are significantly lower than vegetation, soil or anthropogenic tar- Many studies of aquatic environments have shown the importance
gets, demanding greater accuracy in the correction process. of atmospheric correction using different satellite images such as
The use of satellites to monitor the color of the water requires Landsat (Hu et al., 2001; Giardino et al., 2001; Zheng et al., 2015),
effective removal of the contribution of the atmosphere to the SPOT (Doxaran et al., 2002), MODIS (He and Chen, 2014), ENVISAT/
total signal measured by the remote sensor by atmospheric cor- MERIS (Guanter et al., 2010), and SeaWiFS (Hu et al., 2000).
rection process (Jamet et al., 2011). Atmospheric correction over Several studies have been conducted in order to evaluate dif-
aquatic environments is generally more demanding than over land ferent atmospheric correction methods, especially for aquatic en-
vironments. Bonansea et al. (2015) evaluated the potential of 6S
n
radiative transfer model to improve the reliability for estimating
Corresponding author.
E-mail addresses: luizhrotta@gmail.com (L.H.S. Rotta),
water clarity in Río Tercero reservoir (Argentina). They demon-
enner@fct.unesp.br (E.H. Alcântara), fernandasyw@gmail.com (F.S.Y. Watanabe), strated the usefulness of 6S code for atmospheric correction in
twalesza@gmail.com (T.W.P. Rodrigues), nnima@fct.unesp.br (N.N. Imai). Landsat data. The water clarity algorithm using surface reflectance

http://dx.doi.org/10.1016/j.rsase.2016.09.001
2352-9385/& 2016 Elsevier B.V. All rights reserved.
 L.H.S. Rotta et al. / Remote Sensing Applications: Society and Environment 4 (2016) 158–166 159

was more reliable than the top of atmosphere (TOA) reflectance a source of uncertainty. Normally, the researchers do not test the
model. visibility in the atmospheric correction, and this can lead to error
According to Goyens et al. (2013), studies comparing algo- in the correction. Zhang et al. (2014) used initial visibility of 40 km
rithms of atmospheric correction are of great interest for further in atmospheric correction of EO-1 hyperion using FLAASH. They
improvement in water leaving radiance retrievals from satellite aimed to estimate the leaf area index in the forested area of
images. Four atmospheric correction algorithms for MODIS-Aqua Yongan county, Fujian province (China). The 20 km initial visibility
were tested and compared: (1) the standard NIR algorithm of was used by Yuan and Niu (2008) to evaluate the capability of
NASA, (2) the NIR similarity spectrum algorithm, (3) the NIR-SWIR FLAASH to make the atmospheric correction for Hyperion and ALI
algorithm and (4) an Artificial Neural Network algorithm. The image in Heihe River valley of Gansu province (China). Wicaksono
standard NIR algorithm performs better for water dominated by (2012), aimed remove sunglint from Quickbird image of an area
phytoplankton and the artificial neural network algorithm shows
located in the National Park of Ujung Kulon, Banten Province (In-
the best results for water mainly dominated by detrital and mi-
donesia). For the initial visibility, they used 40 km considering
neral material.
visibility during clear sunny cloudless day. Landsat images from
There are several atmospheric correction algorithms available,
1978 to 2013 (MSS, TM, ETMþ and OLI sensors) were used in
nevertheless, algorithms based on MODTRAN - FLAASH (Fast Line-
Zheng et al. (2015); initial visibility was chosen as 30–40 km for
of-sight Atmospheric Analysis of Spectral Hypercube) (Adler-
the aerosol retrieval, depending on the image quality.
Golden et al., 1999) have been widely used with success in studies
The visibility is a parameter required to minimize the effect of
of aquatic environments for both atmospheric corrections of hy-
perspectral (Brando and Dekker, 2003; Kutser, 2004; Moses et al., the atmosphere on the remote sensor signal. Being the initial
2012) and multispectral images (Kutser, 2012; Tebbs et al., 2013; visibility required for atmospheric correction in FLAASH, what is
Watanabe et al., 2015). the effect of setting different visibility values for the estimation of
In some cases, initial visibility value must be specified by user water quality parameters using satellite images? Therefore, this
when using FLAASH. The researchers can use appropriated study aimed to assess the atmospheric correction of SPOT-6 image
equipment to measure the local visibility. Initial visibility of based on the variation of initial visibility parameter in FLAASH and
33.86 km, obtained by a sun photometer (CE318), was used in analyze its influence on models to estimate Secchi depth (SD) and
FLAASH for atmospheric correction of Landsat image by Song et al. diffuse attenuation coefficient (Kd). It is noteworthy that atmo-
(2012). However, the authors commonly must insert an initial spheric correction is necessary mainly when temporal analysis is
visibility based on their personal perception of the day, and this is required, thereby initial visibility is needed in several procedures.

Fig. 1. (a) Study area – Nova Avanhandava Reservoir. (b) Sampling stations collected in Bonito River: deep water (yellow dots) and shallow water (white dots). Image: SPOT-
6; B0G1R2; July 9th, 2013; WGS-84; UTM Zone 22S.
160 L.H.S. Rotta et al. / Remote Sensing Applications: Society and Environment 4 (2016) 158–166

2. Study area L u( λ ) t
R rs(λ )= Fi
Ed( λ ) n2 (2)
Nova Avanhandava Reservoir (Fig. 1a) belong to Tietê River's
cascading system, fully contained in São Paulo State (SP) - Brazil. where,
The Nova Avanhandava hydroelectric plant is located in Buritama
city - SP, in the lower course of Tietê River, between Promissão 1. Lu(λ): is the upwelling radiance at nadir just below-surface;
(upstream) and Três Irmãos (downstream) dams. Nova Avanhan- 2. Ed(λ): is the downwelling irradiance;
dava began its operation in 1982 and is managed by AES Tiete 3. t: is the transmittance at air-water interface (0.98);
Company (http://www.aestiete.com.br/). It has a flooded area of 4. n: is the refractive index of water relative to air (1.33);
210 km2 and volume of 2830  106 m3. The dam is the type 5. Fi: is the spectral immersion coefficient.
ground/concrete, with 2038 m length.
Those 20 samples were collected in Bonito River, tributary of
Tietê River, as follow: 10 sampling stations in optically deep water,
3. Methodology i.e. regions with reflectance without bottom signal; and 10 sam-
pling stations in optically shallow water, i.e. regions where the
3.1. Field data collection water-leaving reflectance is the sum of water column and bottom
reflectance. The Rrs was calculated for every 20 sample stations,
Secchi depth (SD) and diffuse attenuation coefficient (Kd) were however, the SD and Kd were measured only at 10 stations in deep
collected in a field campaign carried out between 28th and 30th water (Fig. 1b – yellow points). This procedure was adopted in
June 2013. Nowadays, the Secchi disk method still widely used in order to avoid bottom reflectance contribution.
water body studies, due to the low cost and simplicity: a circular
disk about 30 cm (Fig. 2a) must be submerged up to the point of its 3.2. Atmospheric correction of SPOT-6
disappearance – the distance between the water surface and that
point is called SD. Kd is defined as the attenuation rate of irradiance SPOT-6 satellite sensor built by AIRBUS Defense & Space was
with depth, which diminishes almost as an exponential function successfully launched on September 9th, 2012. SPOT-6 satellite
(Kirk, 2011). Kd is derived using the downwelling irradiance just and SPOT-7 ensure service continuity of SPOT- 4 and SPOT-5 sa-
below the water surface, Ed (z;λ), measured in field as shown in tellites, which have been operating since 1998 and 2002. The re-
Eq. (1) (Mobley, 1994): visit time is 1d with SPOT 6 and SPOT 7 operating simultaneously
(Astrium, 2013). The Table 1 shows the main characteristics of
Ed( z; λ ) = Ed( 0−; λ )e−Kd( z; λ)z (1)
SPOT-6 image acquired on study area.
where, The atmospheric correction of SPOT-6 image was made using
FLAASH which is an atmospheric correction tool based on MOD-
1. Ed(z;λ): downwelling irradiance at depth z; TRAN4 (MODerate spectral resolution atmospheric TRANsmittance
2. Ed(0-;λ): downwelling irradiance just below the water surface; algorithm and computer model) (Adler-Golden et al., 1999). Users
3. Kd(z;λ): diffuse attenuation coefficient at depth z. have limited control over the choice and setting of input para-
meters in FLAASH. This program is simple to run, but the user has
A vertical profile of the downwelling irradiance (Ed) was ac- to specify appropriate input parameters that characterize the at-
quired using a spectroradiometer RAMSES/TriOs (http://www. mospheric conditions and illumination/viewing geometry at the
trios.de; Rastede, Germany) through the water at every 1.0 m time of image acquisition. The default values are based on theo-
depth (Fig. 2b). The hyperspectral UV/VIS irradiance sensor has retical estimates and information from the literature are used as
190 channels ranging from 320 to 950 nm. However, the Ed values input parameters when actual measurements are not available
between 400 and 700 nm were integrated to obtain Ed of Photo- (Moses et al., 2012).
synthetically Active Radiation (PAR), Ed(PAR). The scene visibility value must be inserted in FLAASH for the
Besides Ed, upwelling radiance (Lu) data were also collected and atmospheric correction. According to ITT (2009), for clear atmo-
therefore, the remote sensing reflectance above surface (Rrs) (see sphere, i.e. without haze, the initial visibility should be approxi-
Fig. 1b for sampling location) was calculated at 20 sampling sta- mated to the values from 40 to 100 km. Due to uncertainty about
tions. Rrs was calculated using Eq. (2), based on Dall’Olmo and the adequate initial visibility; three different values
Gitelson (2005) and Gitelson et al. (2008): were tested: 40, 70 and 100 km (Table 2). Therefore, three

Fig. 2. (a) Secchi disk and (b) hyperspectral sensor TriOS.

 L.H.S. Rotta et al. / Remote Sensing Applications: Society and Environment 4 (2016) 158–166 161

Table 1
SPOT-6 image characteristics.

Acquisition date 2013-07-09

Acquisition time 13:08:43.5
Number of spectral bands 4 (B0 B1 B2 B3)
Wavelength B0 (455–525 nm)
B1 (530–590 nm)
B2 (625–695 nm)
B3 (760–890 nm)
Across angle  6.60018867713°
Along angle 19.1336359344°
Coordinate reference system WGS  84
Resampling distance 6.7 m

Fig. 3. Relative spectral response of SPOT-6. Adapted from ASTRIUM (2003).

Table 2
Input parameters in FLAASH. correction procedure was assessed based on RMSE, image histo-
gram and amount of negative values, using Kd and SD models. The
Sensor type SPOT 6 Atmospheric model Tropical
atmospherically corrected SPOT-6 image, which presented the best
Sensor altitude 695 km Aerosol model Rural
Ground elevation 362 m Initial visibilitya 40, 70 and 100 km results in their assessment based on Kd and SD models, was se-
Flight date July 9th, 2013 Zenith angle 161.5° lected. The chosen models were applied in the same SPOT image
Flight time GMT 13:08:43 Azimuth angle  165° (or 195°) producing thematic maps which show the spatial distribution of
a
SD and Kd.
Initial visibilities tested in FLAASH considering the recommendation in ITT
(2009) for clear weather condition (between 40 and 100 km).

atmospherically corrected images were generated. 4. Results and discussion

To simulate the Rrs(λi) signals that would be recorded by the
satellite sensor at each channel centered at wavelength λi, 4.1. Field data
weighted averages of each Rrs spectrum were calculated by using
as weights the spectral band responses of SPOT-6, (Fig. 3; Astrium, Based on Lu and Ed the Rrs of 20 sampling stations were cal-
2013) as given below (Eq. 3). culated and its spectra data are shown in Fig. 4a. The SD and Kd of
the 10 sample stations located in deep water are shown in Fig. 4b.
∑λ R rs(λ )*S(λ ) The P04 presented the major discrepancy among the sample
R rss(λ )=
∑λ S(λ ) (3) points, showing Rrs values much greater than the average at wa-
velengths between 520 nm and 720 nm. Low transparency was
where S(λ) is the SPOT-6 spectral response function, and Rrs_s(λ) is also observed at this sample, i.e. high Kd and low SD. A sand
the simulated Rrs. The spectral response characteristics of the mining company had been observed near the P04, therefore, the
SPOT-6 instrument, similar to any remote sensing instrument, are sand extraction from the bottom of the river could be accelerating
the key to understand and utilize the data. the sediment re-suspension which increases the Kd and decreases
transparency values for the surrounding locations, as showing in
3.3. Model calibration and validation Fig. 4b. In other words, the higher sediment concentration in P04
causes an increasing in scattering, so the reflectance is increased.
The simulated bands (Rrs_s(λ)) were used for calibration and the
models for SD and Kd were tested based on each individual band, 4.2. Atmospheric correction of SPOT-6
i.e. linear regression of SD and Kd with blue (Rrs(485 nm)), green
(Rrs(560 nm)) and red (Rrs(660 nm)) bands. Furthermore, other Three images were generated as result of the atmosphere cor-
models were tested empirically based on all possible band ratios, rection of SPOT-6, one image for each initial visibility adopted in
i.e. regression of SD and Kd with Rrs(560 nm)/Rrs(485 nm), FLAASH, i.e. visibility of 40 km (Fig. 5b), 70 km (Fig. 5c) and
Rrs(560 nm)/Rrs(660 nm) and Rrs(660 nm)/Rrs(485 nm). The best 100 km (Fig. 5d) used in FLAASH. The assessment of the correc-
model for each variable (SD and Kd) was selected based on the tions was conducted considering simulated SPOT-6 bands (Fig. 5a)
coefficient of determination (R2) and the Root-Mean-Square-Error in order to compare the atmospherically corrected images with a
(RMSE). Both selected models were used to evaluate the atmo- reference, that is, data without atmospheric influence.
spheric correction of the image, analyzing the influence of input The atmospherically corrected images using visibility of 70 and
“visibility” data in FLAASH in estimation of Kd and SD. 100 km presented similar reflectance values, mainly in the green
The 10 samples acquired in deep water were used in models band (560 nm). The image using visibility of 40 km presented Rrs
calibration (see Fig. 1 for sampling stations). The samples in deep values lower than simulated bands based on in situ data. The at-
water avoid signal from the bottom. Due to the low number of mospheric correction of NIR (Near Infrared) band did not show
samples for model calibration, LOOCV (Leave-One-Out Cross Va- satisfactory result for all tested visibilities. These results may be
lidation) method was used to evaluate the models, so it is not associated with the strong radiation absorption by the water at
necessary to remove samples for validation. In LOOCV, the model NIR wavelength, which interferes in atmospheric correction pro-
is calibrated using all data except one point, then the mean cess. The absorption by water in the red end region is quite sig-
squared error is computed based on this point (Stone, 1974; Arlot nificant (about 35% of incident light at 680 nm is absorbed in one-
and Celisse, 2010). The main advantage of LOOCV is that all the meter thick of pure water), but it is much stronger in the infrared
data can be used for training the model. This procedure was re- (Kirk, 2011).
peated n times, where n is the sample number. Thus, the final Negative values on water body were also obtained; especially
RMSE is calculated. when we used initial visibility of 40 km. Table 3 shows the per-
The adoption of different initial visibility data in atmospheric centage of negative pixels for each one of the three
162 L.H.S. Rotta et al. / Remote Sensing Applications: Society and Environment 4 (2016) 158–166

Fig. 4. Remote sensing reflectance spectra (a) and SD and Kd at each sampling stations (b).

Fig. 5. Rrs from simulated SPOT-6 based on in situ data (a). Rrs for SPOT image using visibility of 40 km (b), 70 km (c) and 100 km (d) in the atmospheric correction process.

Table 3 showed an underestimation on the corrected image values when

Percentage of negative values for water body in Blue, Green and Red bands based compared with field data. This underestimation explains the high
on different initial visibility data.
number of negative pixels when using visibility of 40 km in the
Negative values atmospheric correction. Adopting a low value for visibility, such as
40 km, we may exceedingly remove the haze effect, so negative
Visibility 40 km 70 km 100 km values may happen. Furthermore, a high RMSE was noted, parti-
Blue (485 nm) 54.92% 0.90% 0.02%
cularly for blue (95.9%) and green (38.2%) bands. A slight over-
Green (560 nm) 2.68% 0.01% 0.00% estimation in Rrs was observed on image corrected by visibility of
Red (660 nm) 6.63% 0.01% 0.00% 100 km when compared to field data (Fig. 6c). Therefore, com-
paring the SPOT image with the field data (simulated SPOT), image
corrected using visibility of 70 km presented the best performance
atmospherically corrected images. A high number of negative
for all bands (Fig. 6b). In this case, we can see the points well
pixels was observed in the corrected image using visibility of
distributed for three bands along the line y¼x (dashed line), i.e.
40 km for all bands; however, the blue band (485 nm) was the
there was not apparent overestimation or underestimation in
most compromised, with the majority of negative pixels (about
corrected pixels. RMSE of 44.6%, 22.6% and 42.8% for the blue,
55%). The images corrected using visibility of 70 km and 100 km
green and red bands, respectively, were calculated based on visi-
did not exhibit a significant number of negative pixel, so presented
better correction. bility of 70 km, the lowest values among the atmospherically
Fig. 6 shows the dispersion plot of atmospherically corrected corrected images.
SPOT image versus simulated SPOT bands based on field data. The
20 sampling points (Fig. 1b) were plotted for each band. The dis- 4.3. Models calibration and validation
persion graphics of images corrected by visibility of 40 km, 70 km
and 100 km are shown in Fig. 6a, b and c, respectively. The models calibration was performed based on Rrs of simu-
The best atmospheric correction performance was observed for lated SPOT bands from the field data. The model calibration for SD
green band in all cases, followed by red and blue bands. Fig. 6a presented the best results when ratio Rrs(560 nm)/Rrs(660 nm)
 L.H.S. Rotta et al. / Remote Sensing Applications: Society and Environment 4 (2016) 158–166 163

Fig. 6. Atmospheric correction validation for visible bands according to the inserted initial visibility of 40 km (a), 70 km (b) and 100 km (c).

was used (Fig. 7a). The best model for Kd was obtained using only for Kd. There was an offset to lower values in histograms of images
the band Rrs(660 nm) (Fig. 7c). The SD (Eq. 4) and Kd (Eq. 5) models corrected for visibility of 40 km (SD and Kd – Fig. 8a and d, re-
were selected based on the coefficient of determination (R2), RMSE spectively), i.e. the frequency is higher for lower SD and Kd values,
and then presented in graphs confronting the measured and cal- which corroborates with the previously discussion (Fig. 5) about
culated values. We observed an adjustment for both models with the underestimation of reflectance when using this visibility
high accuracy. The R2 of 92% and RMSE less than 12% were ob- (40 km). Furthermore, there was a high frequency in the class “No
tained for both models, SD and Kd. Data”, i.e. the model was not able to calculate the value for SD
R rs(560nm) approximately in 6% of the pixels, probably because of negative
SD=2. 0709× −1. 2697 values that occurred using visibility of 40 km in the atmospheric
R rs(660nm) (4)
correction. Finally, it was observed higher frequency in values
greater than 4.0 m for image based on visibility of 40 km, what is
Kd=67.4155×R rs( 660nm)+0.3978 (5) not consistent with the data observed in the field work.
The atmospherically corrected image based on visibility of
4.4. Atmospheric correction assessment based on SD and Kd models 40 km should not be used due to the high percentage of negative
pixels, which could disturb the results in models calibration. Fur-
The best models based on simulated SPOT bands (Fig. 7) were thermore, the histogram using visibility of 40 km showed in-
selected and applied to each corrected image, using visibility of compatibility with findings in the field. Based on Fig. 5 and Table 3,
40 km, 70 km and 100 km. Further, the RMSE based on in situ data the image corrected using visibility of 70 km presented better
were calculated (Table 4). Therefore, the spatial distribution of SD results when compared to visibility of 100 km and 40 km, with
and Kd for the three atmospheric conditions were obtained, re- RMSE of 53.9% for SD and 21.6% for Kd. Fig. 9 shows RMSE values
sulting in six different histograms (Fig. 8). and scatter plot between measured and calculated values for SD
The frequency was quite similar to the images corrected by (a) and Kd (b) based on selected models and image corrected by
visibility of 70 km and 100 km, with the highest occurrences in visibility of 70 km. It is noteworthy that these results (RMSE,
values between 1.0 and 2.0 m for SD and between 0.6 and 1.0 m  1 measured versus calculated graphs, and histogram graphs) were
164 L.H.S. Rotta et al. / Remote Sensing Applications: Society and Environment 4 (2016) 158–166

Fig. 7. Model calibration (a) and validation (b) for SD and model calibration (c) and validation for Kd.

Table 4 4.5. SD and Kd maps

RMSE for SD and Kd base on atmospheric corrected images using visibility of 40 km
(Vis40), 70 km (Vis70) and 100 km (Vis100).
The selected models for SD and Kd were applied to atmo-
RMSE spherically corrected image based on visibility of 70 km. Therefore,
SD (Fig. 10a) and Kd (Fig. 10b) maps were produced. The range of
Vis40 Vis70 Vis100 the legend were defined based on sample values collected in the
SD 207.2% 53.9% 56.3%
field, i.e. using the minimum and maximum values observed in the
Kd 20.1% 21.6% 27.1% field.
Higher SD values were observed in the tributaries of the Tietê
River, with values exceeding 3.0 m. In Tietê River, the SD values
were mostly between 1.0 and 2.0 m (Fig. 10a). According to vali-
obtained using the calibrated model based on simulated SPOT
dation, (Fig. 9) Kd provided a more accurate model, thereby the Kd
bands.
map (Fig. 10b) represented with more truth the horizontal dis-
The SD model when applied to SPOT image did not show points
well distributed in measured versus calculated graph (Fig. 9a). The tribution of water transparency in Nova Avanhandava Reservoir.
bad results for SD model validation may have occurred because Overall, lower Kd values were observed in the tributaries, corro-
atmospheric correction for the green band (560 nm) have been borating with higher transparency. However, there was a region in
more effective than for red band (660 nm). Therefore, when the the Bonito River with high Kd. This result is consistent taking into
ratio of these bands (560 nm/660 nm) is used, the model may not account the data observed in field, due to the presence of a sand
have been able to retrieve accurately the SD, with RMSE of 53.9%. mining company near P04 (see Fig. 1). The sand extraction from
In contrast, the validation for Kd (Fig. 9b) showed well-distributed the bottom of the river could accelerate the sediment re-suspen-
points along the line y ¼x, besides a satisfactory RMSE (21.6%) was sion which increases Kd values for the surrounding locations.
obtained. Therefore, the Kd map is consistent with the field observation.
 L.H.S. Rotta et al. / Remote Sensing Applications: Society and Environment 4 (2016) 158–166 165

Fig. 8. Histogram of SD images based on atmospheric correction with visibility of 40 km (a), 70 km (b) and 100 km (c); and histogram of Kd images based on atmospheric
correction with visibility of 40 km (d), 70 km (e) and 100 km (f).

Fig. 9. Validation of atmospherically corrected image with visibility of 70 km based on models for SD (a) and Kd (b).

Fig. 10. Thematic maps showing the spatial distribution of SD (a) and Kd (b) using SPOT-6 image (Acquisition data: July 9th, 2013) in Nova Avanhandava Reservoir.
166 L.H.S. Rotta et al. / Remote Sensing Applications: Society and Environment 4 (2016) 158–166

5. Conclusion Chen, Q., Zhang, Y., Hallikainen, M., 2007. Water quality monitoring using remote
sensing in support of the EU water framework directive (WFD): a case study in
the Gulf of Finland. Environ. Monit. Assess. 124, 157–166.
FLAASH proved to be a satisfactory tool for atmospheric cor- Dall’Olmo, G., Gitelson, A.A., 2005. Effect of bio-optical parameter variability on the
rection of SPOT-6 images. The challenge faced was to establish an remote estimation of chlorophyll-a concentration in turbid productive waters:
appropriate initial visibility for atmospheric correction of SPOT experimental results. Appl. Opt. 44 (3), 412–422.
Doxaran, D., Froidefond, J.-M., Lavender, S., Castaing, P., 2002. Spectral signature of
image. The main problem was the high amount of negative pixels
highly turbid waters. Application with SPOT data to quantify suspended par-
when the proper visibility was not adopted in the atmospheric ticulate matter concentrations. Remote Sens. Environ. 81, 149–161.
correction. Therefore, using the suitable initial visibility in the Giardino, C., Pepe, M., Brivio, P.A., Ghezzi, P., Zilioli, E., 2001. Detecting chlorophyll,
atmospheric correction procedure is essential to avoid such pro- Secchi disk depth and surface temperature in a sub-alpine lake using Landsat
imagery. Sci. Total Environ. 268 (1–3), 19–29.
blems, which would prevent the use of bio-optical models to es- Gitelson, A.A., Dall’Olmo, G., Moses, W., Rundquist, D.C., Barrow, T., Fisher, T.R.,
timate limnological variables of the water body. Gurlin, D., Holz, J., 2008. A simple semi-analytical model for remote estimation
The influence of the atmospheric correction on the accuracy of of chlorophyll-a in turbid waters: validation. Remote Sens. Environ. 112,
3582–3593.
the models to retrieve parameters about water transparency was
Goyens, C., Jamet, C., Schroeder, T., 2013. Evaluation of four atmospheric correction
shown in this study. It was observed an optimal value for initial algorithms for MODIS-Aqua images over contrasted coastal waters. Remote
visibility in atmospheric correction process that minimize the er- Sens. Environ. 131, 63–75.
rors in inverse models. However, the visibility used in atmospheric Guanter, L., Verdú, A.R., Odermatt, D., Giardino, C., Simis, S., Estellés, V., Heege, T.,
Gómez, J.A.D., Moreno, J., 2010. Atmospheric correction of ENVISAT/MERIS data
correction may be different in other places and with other climate over inland waters: validation for European lakes. Remote Sens. Environ. 114,
conditions. The visibility of 70 km provided the most appropriate 467–480.
correction on the satellite image, although attention should be He, Q., Chen, C.A., 2014. New approach for atmospheric correction of MODIS ima-
gery in turbid coastal waters: a case study for the Pearl River Estuary. Remote
paid mainly to the negative values from non-appropriate adoption
Sens. Lett. 5 (3), 249–257.
of the visibility parameter. In this study, the use of visibility of Hu, C., Muller-Karger, F.E., Andrefouet, S., Carder, K.L., 2001. Atmospheric correction
40 km in atmospheric correction created high amount of negative and cross-calibration of LANDSAT-7/ETM þ imagery over aquatic environ-
pixels, which prevented the adoption of this image. Adopting a ments: a multiplatform approach using SeaWiFS/MODIS. Remote Sens. Environ.
78 (1–2), 99–107.
low value for visibility, such as 40 km, the haze effect may be re- Hu, C., Carder, K.L., Muller-Karger, F.E., 2000. Atmospheric correction of SeaWiFS
moved exceedingly, so negative values can happen. The use of imagery over turbid coastal waters: a practical method. Remote Sens. Environ.
visibility of 70 km showed better results than when used visibility 74, 195–206.
ITT Visual Information Solutions, 2009. Atmosphere Correction Module: QUAC and
of 100 km, however there was no significant improvement, with
FLAASH User's Guide. Version 4.7.
RMSE of 53.9% (visibility of 70 km) and 56.3% (visibility of 100 km) Jamet, C., Loisel, H., Kuchinke, C.P., Ruddick, K., Zibordi, G., Feng, H., 2011. Com-
for the SD and RMSE of 21.6% (visibility of 70 km) and 27.1% (vis- parison of three SeaWiFS atmospheric correction algorithms for turbid waters
ibility of 100 km) for Kd. using AERONET-OC measurements. Remote Sens. Environ. 115 (8), 1955–1965.
Kirk, J.T.O., 2011. Light and Photosynthesis in Aquatic Ecosystems, third ed. Cam-
For future works is suggested to test similar procedures used in bridge University Press, New York.
this study for estimating optically active components with ecolo- Kutser, T., 2004. Quantitative detection of chlorophyll in cyanobacterial blooms by
gical significance in water bodies, such as chlorophyll-a, dissolved satellite remote sensing. Limnol. Oceanogr. 49 (6), 2179–2189.
Kutser, T., 2012. The possibility of using the Landsat image archive for monitoring
organic matter, and suspended solids concentrations. Therefore,
long time trends in coloured dissolved organic matter concentration in lake
the influence of atmospheric correction on these variables could waters. Remote Sens. Environ. 123, 334–338.
be assessed. Mobley, C.D., 1994. Light and Water: Radiative Transfer in Natural Waters. Aca-
demic Press, San Diego.
Moses, W.J., Gitelson, A.A., Perk, R.L., Gurlin, D., Rundquist, D.C., Leavitt, B.C., Bar-
row, T.M., Brakhage, P., 2012. Estimation of chlorophyll-a concentration in
Acknowledgment turbid productive waters using airborne hyperspectral data. Water Res 46,
993–1004.
Ruddick, K.G., Ovidio, F., Rijkeboer, M., 2000. Atmospheric correction of SeaWiFS
The authors thank to São Paulo Research Foundation - FAPESP imagery for turbid coastal and inland waters. Appl. Opt. 39 (6), 897–912.
(Project number: 2012/19821-1 and 2013/09045-7) and National Song, Y., Wang, J., Yang, K., Ma, M., Li, X., Zhang, Z., Wang, X., 2012. A revised surface
Counsel of Technological and Scientific Development - CNPq resistance parameterisation for estimating latent heat flux from remotely
sensed data. Int. J. Appl. Earth Obs. Geoinf. 17, 76–84.
(Projects numbers: 400881/2013-6, 472131/2012-5 and 482605/
Stone, M., 1974. Cross-validatory choice and assessment of statistical predictions. J.
2013-8) for financial support and CAPES for scholarship. R. Stat. Soc. Series B (Methodological) 36 (2), 111–147.
Tebbs, E.J., Remedios, J.J., Harper, D.M., 2013. Remote sensing of chlorophyll-a as a
measure of cyanobacterial biomass in Lake Bogoria, a hypertrophic, saline–al-
kaline, flamingo lake, using Landsat ETM. Remote Sens. Environ. 135, 92–106.
References Wang, M., 1999. A sensitivity study of the SeaWiFS atmospheric correction algo-
rithm: effects of spectral band variations. Remote Sens. Environ. 67 (3),
Adler-Golden, S.M., Matthew, M.W., Bernstein, L.S., Levine, R.Y., Berk, A., Richts- 348–359.
meier, S.C., Acharya, P.K., Anderson, G.P., Felde, G., Gardner, J., Hike, M., Jeong, L. Watanabe, F.S.Y., Alcântara, E., Rodrigues, T.W.P., Imai, N.N., Barbosa, C.C.F., Rotta, L.
S., Pukall, B., Mello, J., Ratkowski, A., Burke, H.-H., 1999. Atmospheric correction H.S., 2015. Estimation of chlorophyll-a concentration and the trophic state of
for shortwave spectral imagery based on MODTRAN4. SPIE Proc. Imaging the Barra Bonita hydroelectric reservoir using OLI/Landsat-8 images. Int. J.
Spectrom. 3753, 61–69. Environ. Res. Public Health 12 (9), 10391–10417.
Arlot, S., Celisse, A., 2010. A survey of cross-validation procedures for model se- Wicaksono, P., 2012. The effect of sunglint on satellite-based benthic habitat
lection. Stat. Surv. 4, 40–79. identification. Int. J. Adv. Res. Comput. Commun. Eng. 1 (6), 364–370.
Astrium, 2013. SPOT 6 & SPOT 7 Imagery – User guide. Si/DC/13034-v1.0. J., Yuan, Z., Niu, 2008. Evaluation of Atmospheric Correction using FLAASH. Int.
Bonansea, M., Ledesma, C., Rodríguez, C., Pinotti, L., Antunes, M., 2015. Effects of Workshop on Earth Obs. Remote Sens. Appl. IEEE, Beijing.
atmospheric correction of Landsat imagery on lake water clarity assessment. Zhang, Z., He, G., Dai, Q., Jiang, H., 2014. Leaf area index estimation using MESMA
Adv. Space Res. 56, 2345–2355. based on EO-1 hyperion satellite imagery. Int. J. Inf. Electron. Eng. 4 (1), 11–15.
Brando, V.E., Dekker, A.G., 2003. Satellite hyperspectral remote sensing for esti- Zheng, Z., Li, Y., Guo, Y., Xu, Y., Liu, G., Du, C., 2015. Landsat-based long-term
mating estuarine and coastal water quality. IEEE Trans. Geosci. Remote Sens. 41 monitoring of total suspended matter concentration pattern change in the wet
(6), 1378–1387. season for Dongting Lake, China. Remote Sens. 7, 13975–13999.