ISPRS Journal of Photogrammetry and Remote Sensing 105 (2015) 120–127

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

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

Variance-preserving mosaicing of multiple satellite images for forest

parameter estimation: Radiometric normalization
Anna Eivazi a,⇑, Alexander Kolesnikov b, Virpi Junttila a, Tuomo Kauranne a
a
Department of Mathematics and Physics, Lappeenranta University of Technology, Lappeenranta, Finland
b
Arbonaut Ltd., Joensuu, Finland

a r t i c l e i n f o a b s t r a c t

Article history: Measuring, Reporting and Verification (MRV) systems of the United Nations programme on Reducing
Received 19 March 2014 Emissions from Deforestation and forest Degradation (REDD+) aim to provide robust and reliable data
Received in revised form 14 March 2015 on carbon credits over large areas. Multitemporal satellite mosaics are often the only cost-effective
Accepted 16 March 2015
remote sensing data that allow such coverage. Although a number of methods for producing mosaics
has been proposed, most of them are dependent on the order in which tiles to normalized are presented
to the algorithm and suffer from loss of input scenes’ variance which can substantially reduce the carbon
Keywords:
credits. In this study we propose a variance-preserving mosaic (VPM) algorithm that considers all images
Relative normalization
Image mosaics
at the same time, minimizes overall error of the normalization and aims to preserve average variance of
Pseudo-invariant features input images. We have compared the presented method with a popular relative normalization algorithm
Multispectral imagery commonly used nowadays. The proposed algorithm allows to avoid iterative pair-wise normalization,
Landsat time series results in visually uniform mosaics while maintaining also the original image variance.
REDD + MRV Ó 2015 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier
B.V. All rights reserved.

1. Introduction to-wall’’ observations of the target area. Secondly, the price of

In model-based estimation of geographical quantities based on
satellite images, regression models are built on normalized band
values of image pixels. There are many applications where not only
the fit of estimates to ground truthing is important, but also the fit-
ting of variance. One such application is the Measuring, Reporting
and Verification (MRV) of the United Nations programme on
Reducing Emissions from Deforestation and forest Degradation
(REDD+). REDD+ aims to create a financial value (carbon credits)
for the carbon stored in forests, especially those of developing
countries, in order to reduce greenhouse gas emissions.
One of the important steps within REDD+ is to develop a cost-
effective and accurate methodology for carbon monitoring over
large areas. Such methodology requires an approach that combines
together ground measurements and remote sensing technologies
(Angelsen, 2008). Possible remote sensing technologies that can
be employed are satellite imagery, LiDAR, aerial images and radar
data.
For REDD + MRV, satellite imagery has several advantages over
other remote sensing technologies. Firstly, satellites provide ‘‘wall-
⇑ Corresponding author.
E-mail address: anna.eivazi@lut.fi (A. Eivazi).
satellite images is considerably cheaper, and even more, some of
the satellite imageries, such as Landsat, can be acquired completely
free of charge. And lastly, satellites offer reliable historical data. For
instance Landsat delivers global images for the last four decades
(Gibbs et al., 2007).
Despite many benefits satellite imagery provides, there are
also challenges that should be addressed. One of them is radio-
metric differences between adjacent multitemporal scenes. Due
to variation in acquisition conditions (e.g. solar illumination,
atmospheric scattering and atmospheric absorption) the same
ground object on two overlapping images can result in different
spectral values (Yuan and Elvidge, 1996). Because of this, radio-
metrically uniform mosaics using multitemporal scenes should
be created before employing satellite imagery into carbon assess-
ment. Another challenge is the variance suppression that likely
occurs during mosaicing of multiple images, whenever it is based
on averaging pixel values of overlapping parts of images.
REDD + MRV’s credits are based on measuring the change in car-
bon captured in forests. If regression estimates for carbon capture
are built using satellite images, suppressing the true variance of
band values gets translated into suppression of change in carbon
captured, and hence into a reduction of the corresponding carbon
credit.

http://dx.doi.org/10.1016/j.isprsjprs.2015.03.007
0924-2716/Ó 2015 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.
 A. Eivazi et al. / ISPRS Journal of Photogrammetry and Remote Sensing 105 (2015) 120–127
To overcome radiometric differences between multitemporal
scenes there are two approaches commonly in use: absolute and
relative normalization (also called correction). The first one aims
to convert pixel intensity values to true surface reflectance using
in situ ground measurements of atmospheric properties. The main
disadvantage of the method is that it is very difficult and often not
possible at all to obtain atmospheric measurements (Du et al.,
2002). The second one (relative normalization) uses an assumption
of a linear relationship between overlapping regions of multitem-
poral images (Song et al., 2001). The linear relationship is con-
structed using pseudo-invariant features (PIFs) which represent
areas of temporally constant reflectance. Substantial research has
been done on this topic in past decades and researchers propose
various methods for automatic PIF identification, as well as for
modeling linear relationships (Hall et al., 1991; Yuan and Elvidge,
1996; Song et al., 2001; Du et al., 2002; Gibbs et al., 2007; Canty
and Nielsen, 2008; Zhang et al., 2008; Liu et al., 2012).
All proposed relative radiometric normalization methods are
following a common routine to normalize two scenes: identify
PIFs, run regression analysis on the overlapping region and apply
the coefficients found to the whole scene. If the mosaic needs to
be built for a large area that is covered by many satellite images,
the same routine is applied recursively pairwise, where a normal-
ized scene becomes the reference for the next one (Du et al., 2001;
Furby and Campbell, 2001; Olthof et al., 2005).
Such an approach raises several concerns. Firstly, the resulting
mosaic depends on the order of the normalization and may vary
significantly (Furby and Campbell, 2001). Secondly, due to the
recursive manner of the approach, error propagation brings a high
level of uncertainty (Olthof et al., 2005). Additionally, as regression
analysis is commonly used for the normalization step, a change
(loss or gain) in variance of the normalized image values is
expected. The change of variance over mosaic scenes will in turn
cause error to the forest parameters estimated from them (e.g. bio-
mass, height, etc.).
To resolve these problems, we propose variance-preserving
mosaic (VPM) algorithm that considers all images at the same
time, minimizes overall error of the normalization and aims to pre-
serve average variance of input scenes. In order to validate the pro-
posed method, we compute two independent mosaics – one using
an existing radiometric algorithm and the other one with the pro-
posed algorithm. We assess the mosaics by visual inspection, by
computing the variance of each scene, and by leave-one-out
cross-validation of the forest parameters estimated from them.
2. Study area and data
2.1. Study area
The study area is located over the Terai Arc Landscape (TAL),
along the foothills of the Himalayas in the southernmost part of
Nepal, with altitude ranging from less than 100 m up to 2200 m.
Influenced both by tropical and subtropical climate about half of
the study area is covered by subtropical mainly deciduous forests.
The dominating forest types are sal (Shorea robusta) terai mixed
hardwood, khair-sisau (Acacia catechu/Dalbergia sissoo) and chir-
pine (Pinus roxburghii). TAL is one of the priority landscapes in
Nepal, both for the conservation of its biodiversity and the protec-
tion of the ecological services (e.g. greenhouse gas mitigation, pur-
ification of air and water) it provides.
2.2. Ground-truth data
The field data consists of 738 plots (12.6-m radius) collected in
the spring of 2011. Systematic cluster sampling was used to design
121
the locations of sample plots. As this was done for the so-called
LAMP, or LiDAR-Assisted Multi-source Process for MRV, the plots
are clustered as in three-stage sampling. In the first stage, a set
of LiDAR blocks of 5000 hectars is randomly allocated over the
whole area. In the second stage, six plot clusters are assigned in
a systematic fashion on those blocks. In the third stage, eight circu-
lar sample plots are systematically placed within plot clusters.
Possible spatial correlation of ground truth data resulting from plot
clustering is ignored.
Diameter at breast height (DBH), height and species were mea-
sured for those trees in the plot that had DBH more than five
meters. The following forest attributes for each plot were then
derived from the tree-level measurements: stem count (1/ha),
mean diameter at breast height weighted by basal area (cm), basal
area (m2/ha), mean tree height weighted by basal area (m), stem
volume (m3/ha), and above-ground biomass (tons/ha). For more
detailed information about field measurements and equations in
use reader is referred to Gautam et al. (2013).
For applying the LAMP method, LiDAR data was collected from
about five percent of the whole study area in blocks. Each block
was scanned with full coverage from a height of 2200 m above
ground. Raw LiDAR data was classified into three categories:
ground returns, vegetation returns, and errors. Digital Terrain
Model was built from the ground returns and using it LiDAR data
was converted from absolute elevation into distance-to-ground.
Set of 10000 circular-shaped one hectare size ‘‘surrogate plots’’
were calculated using original field and LiDAR data. Forest attri-
butes were estimated for this set. Locations of the surrogate plots
were selected through weighted random sampling using the
inverse of the block weights applied in LiDAR block sampling
(Gautam et al., 2013).
2.3. Satellite imagery
Medium and high resolution satellite images were used in this
study, specifically Landsat 5 and RapidEye. The scenes were chosen
so that they have as little as possible clouds and are acquired in the
same growth season. All bands, but thermal infrared, were
employed in this study. Table 1 represents detailed information
on the images, and Fig. 1 depicts the location of the scenes over
the study area. As a pre-processing step clouds and snow were
identified within satellite scenes and masked out by setting their
pixel values to nodata.
3. Method
3.1. VPM method
We are given a set of images In ; n ¼ 1; . . . ; N that should be
normalized in order to construct a uniform, variance-preserving
mosaic. Each image Ii overlaps at least one neighbor Ij . We will
define Iij as the subset of Ii corresponding to the no-change pixels
for the overlap with image Ij , and mij as the corresponding size of
Iij . The pixels can be identified as no-change ones in all pairwise
overlaps simultaneously using well established technique (e.g.
Canty and Nielsen, 2008). Obviously, mij is always equal to mji ,
and mji is always zero when the images Ii and Ij do not overlap.
We can describe how different two neighbor images Ii and Ij are
by analyzing pixel values of their overlapped regions. We define
the squared difference dij as the measure of similarity between
two images over the same overlapped area:
mij
X
dij ¼ ðpm  qm Þ2 ; ð1Þ
m¼1
where pm and qm are pixel values of areas Iij and Iji respectively.
122 A. Eivazi et al. / ISPRS Journal of Photogrammetry and Remote Sensing 105 (2015) 120–127

mij
Table 1 X X 2
Satellite images employed in the study. Dða; bÞ ¼ ððai þ bi pm Þ  ðaj þ bj qm ÞÞ : ð4Þ
i;j21;...;N m¼1
Satellite Path/row Date (dd.mm.yy) Resolution (m) j>i

Landsat 5 143/40 20.10.2009 30 We aim to find such a set of coefficients a ¼ ða1 ; . . . ; aN Þ and
Landsat 5 143/41 26.10.2011 30
b ¼ ðb1 ; . . . ; bN Þ that minimizes total difference D(a, b) of the nor-
Landsat 5 144/40 11.10.2009 30
Landsat 5 144/41 01.10.2011 30 malized mosaic. Using above formula we can express our task as
RapidEye 4452115 26.03.2010 5 an optimization problem:
RapidEye 4452116 14.04.2010 5
RapidEye 4452215 26.03.2010 5 minimize Dða; bÞ: ð5Þ
RapidEye 4452216 26.03.2010 5 a;b
RapidEye 4452311 12.04.2010 5
RapidEye 4452312 12.04.2010 5 Problem (5) has a trivial solution when ai ¼ 0 and bi ¼ 0 for all i.
RapidEye 4452313 18.03.2010 5 To avoid the trivial solution we introduce an additional constraint
RapidEye 4452411 18.03.2010 5 Dr ðbÞ and define it as the squared difference between the
RapidEye 4452412 12.04.2010 5
standard deviation of the original image (r) and the transformed
RapidEye 4452413 18.03.2010 5
one ð r
e Þ:

X
N X
N X
N

The smaller dij is, the more similar the corresponding neighbor
Dr ðbÞ ¼ ri Þ2 ¼ ðri  ri bi Þ2 ¼ r2i ð1  bi Þ2 :
ðri  f ð6Þ
i¼1 i¼1 i¼1
images Ii and Ij are. The total difference D over the whole mosaic is
then defined as a sum of overlapped regions squared differences: As we are aiming to preserve the average variance of the images
X after normalization, we introduce a tolerance  2 ½0; 1 that
D¼ dij : ð2Þ describes how much we allow the average variance to be changed:
i;j21;...;N
j>i
X
N
Dr ðbÞ 6  r2i :
Relative radiometric normalization is based on the assumption
i¼1
that the difference between two satellite images from different
dates but from the same area can be explained by a linear relation- Thus, our task can be written as the following constrained
ship (Vicente-Serrano et al., 2008). Thus, for each pixel p of the optimization problem:
image Ii there should be found a linear operator T i with coefficients
ai and bi : minimize Dða; bÞ
a;b
T i ðpÞ ¼ ai þ bi p; i ¼ 1; . . . ; N: ð3Þ X
N X
N ð7Þ
subject to r2i ð1  bi Þ2 6  r2i :
The total difference for the mosaic after normalization can be i¼1 i¼1
written as:

Fig. 1. Satellite scenes employed in the study.

 A. Eivazi et al. / ISPRS Journal of Photogrammetry and Remote Sensing 105 (2015) 120–127
We solve our constrained optimization problem (7) using a pen-
alty function method. Hence, we get the following unconstrained
optimization problem:
minimize Dða; bÞ þ ak Dr ðbÞ
a;b
To find the solution to the problem (8) we used the Limited-
memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) algorithm.
L-BFGS is an iterative method for solving unconstrained nonlinear
optimization problems. As it belongs to the family of quasi-Newton
methods, the Hessian matrix does not need to be computed, since
successive estimates to it are updated repeatedly by analyzing suc-
cessive gradient vectors instead. We have employed libLBFGS
which is a C implementation of the L-BFGS1 method.
The ak are the Lagrange multipliers that correspond to the con-
straint (7) on maximum allowable variance reduction  that the
user can define. The parameter  should reside in the half-open
interval 0 <  6 1. Restricting variance reduction strongly, i.e.
choosing   1, results in a rough mosaic with speckle-like noise.
Allowing stronger variance reduction by choosing   0 will turn
the mosaic into a standard least squares interpolation on the over-
lapping areas. There are methods for biomass estimation, such as
Lidar-assisted Multi-source Program (LAMP) that require an expli-
cit variance to be imputed to the mosaic that require such explicit
control (Gautam et al., 2013).
The pseudo-code of the proposed method is presented below.
For each band do:
1. Collect statistics for each image Ii ; i ¼ 1; . . . ; N.
2. Collect statistics for each overlap area Iij ; i; j 2 1; . . . ; N j > i.
3. Construct a1 ; . . . ; aK as an arithmetic sequence started with
zero.
4. For each ak from constructed sequence find ðak ; bk Þ by solving
optimization problem (8) using L-BFGS.
5. Among solutions ðak ; bk Þ find the one that causes the smallest
average variance change, in accordance with:
X
N X
N estimates computed using the mosaics.
r2i ð1  bi Þ2 6  r2i ;
i¼1 i¼1
where  2 ½0; 1 is defined by the user.
6. Apply transformation T i with found coefficients ðai ; bi Þ on each
image Ii .
3.2. Mosaic validation
In order to validate VPM method we need to compare it with
existing radiometric normalization methods. According to recent
reviews (Xu et al., 2012; Schroeder et al., 2006) the iteratively re-
weighted Multivariate Alteration Detection (IR-MAD) algorithm
(Canty and Nielsen, 2008) is a very good mosaicing method.
Thus, we built two mosaics using IR-MAD and the proposed
method and assess them in three ways: (1) visual comparison (2)
standard deviation analysis and (3) validation of forest parameter
estimates (height, volume, biomass, etc.) computed using the
corresponding normalized mosaics.
To estimate forest parameters over the area of interest we
employed LiDAR Assistant Multisource Program (LAMP) (Gautam
et al., 2013) that was developed specifically for tropical forest con-
ditions. The LAMP takes as an input collected ground field and
LiDAR data, produces ‘‘surrogate plots’’ that provides better cover-
1
http://www.chokkan.org/software/liblbfgs/.
123
age of the area, builds corresponding models and as an output pro-
duces estimates for various forest parameters over the satellite
image’s coverage.
For our analysis we estimated six forest parameters: the num-
ð8Þ ber of stems (N), mean diameter (D), mean height (H), basal area
(G), volume (V) and above ground biomass (AGB). We validate
model results using leave-one-out cross-validation and assess
them in terms of root mean square error (RMSE), relative RMSE,
bias and relative bias:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn 2
^
i¼1 ðyi  yi Þ RMSE
RMSE ¼ ; RMSErel ¼ Pn   100; ð9aÞ
n i¼1 yi n
Xn
^
yi  yi Bias
Bias ¼ ; Biasrel ¼ Pn   100; ð9bÞ
n i¼1 yi n
i¼1
where n is a number of sample plots, i observation index, y^i pre-
dicted value for sample plot i; yi observed value for sample
plot i.
When we applied LAMP, all input data but mosaics themselves
were the same for both methods. Thus, if any difference in
RMSE and bias is found, it is due to a difference between the two
mosaics.
4. Results
In this section we evaluate the mosaic produced using the
VPM algorithm and compare it with the corresponding IR-MAD
mosaic. We run three independent test cases with different set-
ups: Landsat satellite with four tiles, RapidEye satellite with four
tiles, and RapidEye satellite with six tiles. We choose such set-
ups that satellite scenes overlap in a circular rather than linear
manner, as this is often the most challenging case.
We assess the mosaics in three ways: (1) visual comparison (2)
standard deviation analysis and (3) validation of forest biomass
4.1. Visual comparison
Based on our experience it is important that the mosaic looks as
uniform as possible over the area of interest. When the mosaic is
used for producing various forest feature maps by regression mod-
els (e.g. biomass maps and forest carbon maps), discontinuous bor-
ders on the mosaic are transferred on the corresponding feature
maps, and this in turn raises a lot of questions and doubts about
the results.
IR-MAD is a pair-wise algorithm and thus, requires the user to
choose a processing chain: (1) a starting tile and (2) pair-wise
path of processing. Fig. 2 shows how IR-MAD mosaics are affected
by the choice of starting tile and the direction of processing.
These examples reveal that one can get a much more uniform
mosaic if an optimal starting tile and path of normalization is
chosen.
In order to decide which processing chain to use, we run all of
the possible combinations and compare output mosaics. Fig. 3
shows the original mosaic and comparison of IR-MAD and the
proposed algorithm for Landsat data with four tiles. Mosaics are
presented in false color using the near infrared, red and green
spectral bands. We present here the best IR-MAD mosaic, which
is achieved when tile 143=41 is chosen as the starting tile and
the normalization process is done in clockwise direction. We then
compared the mosaics produced by both algorithms using two
more setups of RapidEye images with four and six tiles. A similar
124 A. Eivazi et al. / ISPRS Journal of Photogrammetry and Remote Sensing 105 (2015) 120–127

Fig. 2. IR-MAD mosaics with various processing paths (roman numerals define the processing path): (a) the same direction but different starting tile and (b) the same starting
tile but different direction (clockwise/counterclockwise).

Fig. 3. Landsat mosaic: (a) original; (b) IR-MAD; and (c) VPM.

results to Landsat data is achieved when testing with RapidEye
(see Fig. 4 as an example of results).
Compared to original tiles both methods significantly improve
the visual appearance of the mosaics for all three cases, however
discontinuous borders still remain.
4.2. Standard deviation analysis
To ensure an accurate estimation of carbon credits any mosaic-
ing method need to preserve the true variance of band values. We
analyzed the standard deviation of satellite scenes before and after
 A. Eivazi et al. / ISPRS Journal of Photogrammetry and Remote Sensing 105 (2015) 120–127
(a) original
(c) IR-MAD
(e) VPM
Fig. 4. RapidEye mosaics.
radiometric normalization for both Landsat and RapidEye images.
Fig. 5 illustrates that the proposed method results in more stable
variance after normalization. Similar results were observed on
RapidEye mosaics.
4.3. Estimation of forest parameters at the plot level
For each test case mosaic we run the LAMP method to estimate
forest parameters. Table 2 shows the statistics of leave-one-out
cross-validation for six forest parameters for the test case with
Landsat images. The VPM algorithm improves relative RMSE of for-
est parameters by less then one percent compare to IR-MAD algo-
rithm. No bias was observed in the results with either of the
mosaics, which is a property of the LAMP method. Test cases with
RapidEye images produced similar outputs with one-two percent
difference of relative RMSE.
IR-MAD mosaic greatly depends on choosing optimal process-
ing chain. In the best case IR-MAD and the proposed algorithm per-
form similarly in terms of visual comparison and RMSE of
estimated forest parameters. Finding the right starting tile and
125
(b) original
(d) IR-MAD
(f) VPM
processing path for the IR-MAD algorithm requires substantial
additional computation, the amount of which grows with the num-
ber of scenes included in the mosaic. On the contrary, the VPM
method is straightforward to run, saving time and additional
calculations.
5. Discussion and conclusions
In this study we presented a novel relative normalization proce-
dure for creating uniform mosaics using multitemporal satellite
images over large areas. The method is designed specifically for
applications where change of mosaic variance could adversely
affect results. One of such application is MRV systems of the UN
REDD+ where loss of variance will be translated into suppression
of estimated forest carbon change and thus into underestimating
corresponding carbon credits. The VPM method is independent of
the order of the normalization procedure, minimizes overall nor-
malization error, and preserves average variance of input scenes.
We validated the VPM approach against a commonly used rela-
tive normalization procedure, namely IR-MAD. In this study we
126 A. Eivazi et al. / ISPRS Journal of Photogrammetry and Remote Sensing 105 (2015) 120–127

Std 143_41 Std 144_41

20
15

15
10

10
5

5
0

0
1 2 3 4 5 7 1 2 3 4 5 7

Band Band

Std 144_40 Std 143_40

25
20

20
15

15
10

10
5

5
0

1 2 3 4 5 7 1 2 3 4 5 7

Band Band

Original tile MAD algorithm Proposed algorithm

Fig. 5. Comparison of standard deviations.

Table 2
Leave-one-out cross-validation using Landsat mosaic. choosing a processing chain. In order to decide which processing
Target Bias Bias_rel (%) RMSE RMSE_rel (%) chain to use for IR-MAD algorithm, one would need to run all of
them and compare output mosaics, which results in a great num-
IR-MAD algorithm
N 0.02 0.01 148.74 21.9 ber of additional calculations. In a simple linear tiles set-up, where
DGW 0.01 0.01 8.55 25.1 each tile overlaps maximally with two neighboring tiles, the deci-
HGW 0.01 0.01 3.48 21.8 sion how to choose processing chain might be straightforward.
G 0.01 0.01 4.96 25.6 However, when tiles are arranged in a circular set-up there are
V 0.1 0.06 59.32 37.5
multiple possible paths, especially when there are more than four
AGB 0.13 0.06 73.37 36.8
tiles used.
The proposed algorithm
Furthermore, in IR-MAD algorithm it is often difficult to say
N 0.02 0.01 148.74 21.3
DGW 0.01 0.01 8.55 24.3 which one of the resulting mosaics is better, as they might appear
HGW 0.01 0.01 3.48 21.4 both quite uniform but in a different manner. On contrary, the vari-
G 0.01 0.01 4.95 25.6 ance-preserving mosaic method takes into account all possible
V 0.1 0.06 59.31 37.3
overlapping areas at the same time, and thus, does not require
AGB 0.13 0.06 73.37 36.6
any decision making step on the processing path. Therefore, the
mosaic produced by VPM method will always assure that esti-
tested the proposed method with medium and high resolution mated forest parameters will not have abrupt changes and will
satellite images that overlap in a circular manner, specifically four appear as a uniform map. This is an important issue as users of
Landsat 5, and four and six RapidEye scenes. such maps consider it as a proof that the results are reliable.
According to previous research, scene-to-scene approaches For this study our aim was to develop a robust relative normal-
such as IR-MAD algorithm vary greatly only if a large number of ization method for producing satellite imagery mosaics. However,
input images is in use, otherwise the difference is not significant the long-term goal is to develop a reliable procedure for forest
(Furby and Campbell, 2001). However, our results show that the parameter estimations in terms of both RMSE and estimated vari-
resulting mosaic depends on the order of normalization even with ance, which in turn, will provide reliable estimates of carbon cred-
a four tile set-up. We demonstrated that IR-MAD algorithm pre- its for MRV REDD+. For this reasons, we will address the following
forms as well as VPM only if an optimal processing chain is used issues in the future.
by trying all possible processing combination. Otherwise, the pro- First, the visual comparison of results for both methods shows
posed method performs better than IR-MAD in terms of visual that there are still some noticeable discontinues. These scenes
comparison and RMSE of estimated forest parameters, while pre- are originally very different, which makes it almost impossible to
serving also the original variance well. remove borders completely using a radiometric normalization pro-
The proposed method allows users to build a uniform and cedure only. In such cases adding an additional post-normalization
visually seamless mosaic (Fig. 3) without being involved in step that employs elaborate blending of overlapped images, such
 A. Eivazi et al. / ISPRS Journal of Photogrammetry and Remote Sensing 105 (2015) 120–127
as for example described by Bielski et al. (2007), may improve the
resulting mosaic.
Second, the RMSE values calculated based on IR-MAD and our
method are relatively high. One way to further reduce RMSE is to
employ a topographic correction method before starting the nor-
malization procedure, e.g. C-correction (Teillet et al., 1982). That
is also indicated in the study of Vicente-Serrano et al. (2008),
where the authors argue that it is highly recommended to apply
topographic correction before the relative radiometric normaliza-
tion step.
Third, the proposed method is designed to be used for normal-
ization of multitemporal images. In the current study we used sce-
nes collected in a relatively short time period (within two years),
however, the method can also be used for applications where
images have been obtained from a wide time span and need to
be compared. One of the difficulties in such applications is to take
into account the real ground change that is normally addressed by
using PIFs. The proposed normalization procedure employs a
method to identify PIFs proposed by Canty and Nielsen (2008)
and run the method using statistics only of the corresponding
pixels.
References
Angelsen, A., 2008. Moving ahead with REDD: issues, options and implications.
Cent. Int. Forest. Res. (CIFOR), <http://www.cifor.org/publications/pdf_files/
Books/BAngelsen0801.pdf>.
Gibbs, H.K., Brown, S., Niles, J.O., Foley, J.A., 2007. Monitoring and estimating
tropical forest carbon stocks: making REDD a reality. Environ. Res. Lett. 2 (4),
045023, <http://stacks.iop.org/1748-9326/2/i=4/a=045023>.
Yuan, D., Elvidge, C.D., 1996. Comparison of relative radiometric normalization
techniques. {ISPRS} J. Photogramm. Remote Sens. 51 (3), 117–126. http://
dx.doi.org/10.1016/0924-2716(96)00018-4 (ISSN 0924-2716), <http://
www.sciencedirect.com/science/article/pii/0924271696000184>.
Du, Y., Teillet, P.M., Cihlar, J., 2002. Radiometric normalization of multitemporal
high-resolution satellite images with quality control for land cover change
detection. Remote Sens. Environ. 82 (1), 123–134. http://dx.doi.org/10.1016/
S0034-4257(02)00029-9 (ISSN 0034-4257), <http://www.sciencedirect.com/
science/article/pii/S0034425702000299>.
Song, C., Woodcock, C.E., Seto, K.C., Lenney, M.P., Macomber, S.A., 2001.
Classification and change detection using landsat {TM} data: when and how
to correct atmospheric effects? Remote Sens. Environ. 75 (2), 230–244. http://
dx.doi.org/10.1016/S0034-4257(00)00169-3) (ISSN 0034-4257), <http://
www.sciencedirect.com/science/article/pii/S0034425700001693>.
Hall, F., Strebel, D., Nickeson, J., Goetz, S., 1991. Radiometric rectification: toward a
common radiometric response among, multidate, multisensor images. Remote
Sens. Environ. 35 (1), 11–27. http://dx.doi.org/10.1016/0034-4257(91)90062-B
(ISSN 0034-4257), <http://www.sciencedirect.com/science/article/pii/00344
2579190062B>.
127
Canty, M.J., Nielsen, A.A., 2008. Automatic radiometric normalization of
multitemporal satellite imagery with the iteratively re-weighted MAD
transformation. Remote Sens. Environ. 112 (3), 1025–1036. http://dx.doi.org/
10.1016/j.rse.2007.07.013 (ISSN 0034-4257), <http://www.sciencedirect.com/
science/article/pii/S0034425707003495>.
Zhang, L., Yang, L., Lin, H., Liao, M., 2008. Automatic relative radiometric
normalization using iteratively weighted least square regression. Int. J.
Remote Sens. 29 (2), 459–470. http://dx.doi.org/10.1080/
01431160701271990, <http://www.tandfonline.com/doi/abs/10.1080/0143116
0701271990>.
Liu, S.-H., Lin, C.-W., Chen, Y.-R., Tseng, C.-M., 2012. Automatic radiometric
normalization with genetic algorithms and a Kriging model. Comput. Geosci.
43 (0), 42–51. http://dx.doi.org/10.1016/j.cageo.2011.12.016 (ISSN 0098-3004),
<http://www.sciencedirect.com/science/article/pii/S0098300411004407>.
Du, Y., Cihlar, J., Beaubien, J., Latifovic, R., 2001. Radiometric normalization,
compositing, and quality control for satellite high resolution image mosaics
over large areas. IEEE Trans. Geosci. Remote Sens. 39 (3), 623–634. http://
dx.doi.org/10.1109/36.911119 (ISSN 0196-2892).
Furby, S., Campbell, N., 2001. Calibrating images from different dates to like-value
digital counts. Remote Sens. Environ. 77 (2), 186–196. http://dx.doi.org/
10.1016/S0034-4257(01)00205-X (ISSN 0034-4257), <http://www.
sciencedirect.com/science/article/pii/S003442570100205X>.
Olthof, I., Pouliot, D., Fernandes, R., Latifovic, R., 2005. Landsat-7 ETM+ radiometric
normalization comparison for northern mapping applications. Remote Sens.
Environ. 95 (3), 388–398. http://dx.doi.org/10.1016/j.rse.2004.06.024 (ISSN
0034-4257), <http://www.sciencedirect.com/science/article/pii/
S003442570500026X>.
B. Gautam, J. Peuhkurinen, T. Kauranne, K. Gunia, K. Tegel, P. Latva-Kyr, P. Rana, A.
Eivazi, A. Kolesnikov, J. Hmlinen, S.M. Shrestha, S. Gautam, M. Hawkes, U.
Noecker, A. Joshi, T. Suihkonen, P. Kandel, S. Lohani, G. Powell, E. Dinerstein, D.
Hall, J. Niles, S. Nepal, U. Manandhar, Y. Kandel, C. Joshi, Estimation of forest
carbon using LiDAR-assisted multi-source program (LAMP) in Nepal. In:
International Conference on Advanced Geospatial Technologies for
Sustainable Environment and Culture, At Pokhara, Nepal, 2013.
Vicente-Serrano, S.M., Prez-Cabello, F., Lasanta, T., 2008. Assessment of radiometric
correction techniques in analyzing vegetation variability and change using time
series of Landsat images. Remote Sens. Environ. 112 (10), 3916–3934. http://
dx.doi.org/10.1016/j.rse.2008.06.011 (ISSN 0034-4257), <http://
www.sciencedirect.com/science/article/pii/S0034425708002186>.
Xu, Q., Hou, Z., Tokola, T., 2012. Relative radiometric correction of multi-temporal
{ALOS} AVNIR-2 data for the estimation of forest attributes. {ISPRS} J.
Photogramm. Remote Sens. 68 (0), 69–78. http://dx.doi.org/10.1016/
j.isprsjprs.2011.12.008 (ISSN 0924-2716), <http://www.sciencedirect.com/
science/article/pii/S0924271612000275>.
Schroeder, T.A., Cohen, W.B., Song, C., Canty, M.J., Yang, Z., 2006. Radiometric
correction of multi-temporal Landsat data for characterization of early
successional forest patterns in western Oregon. Remote Sens. Environ. 103
(1), 16–26. http://dx.doi.org/10.1016/j.rse.2006.03.008 (ISSN 0034-4257),
<http://www.sciencedirect.com/science/article/pii/S0034425706001179>.
Bielski, C., Grazzini, J., Soille, P., 2007. Automated morphological image composition
for mosaicing large image data sets, in: Geoscience and Remote Sensing
Symposium, 2007. IGARSS 2007. IEEE International, pp. 4068–4071 (http://
dx.doi.org/10.1109/IGARSS.2007.4423743).
Teillet, P.M., Guindon, B., Goodenough, D.G., 1982. On the slope-aspect correction of
multispectral scanner data. Can. J. Remote Sens. 8 (2), 84–106.