Professional Documents
Culture Documents
net/publication/235244169
CITATIONS READS
48 1,014
4 authors, including:
Dengsheng Lu
Fujian Normal University
148 PUBLICATIONS 15,476 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Human and Physical Dimensions of Land Use/Cover Change In Amazonia: Towards Sustainability View project
Comparative Analysis of Modeling Algorithms for Forest Aboveground Biomass Estimation in a Subtropical Region View project
All content following this page was uploaded by Dengsheng Lu on 30 May 2014.
a Minnaert constant k. However, the k value depends on To date, topographic correction is still a challenge,
the nature of land-cover, topographic factor, and wave- and no universal approach is available yet. The Minnaert
length. In areas with large incident angles, the k value correction method is most frequently used in previous
can prevent a division by small values, thereby evading research, but a single global Minnaert value has demon-
overcorrection as in the cosine model (Meyer et al., 1993). strated its poor ability in reducing the topographic effects.
Different ways for computing k value have been devel- Another challenge is the assessment of topographic correc-
oped. A simple way is to use a single global k value for an tion results. Hence, the objective of this paper is to develop
entire image, which is used in most previous research (Colby a pixel-based Minnaert coefficient image for topographic
and Keating, 1998; Garcia-Haro et al., 2001; Mitri and Gitas, correction of Landsat ETM image in a mountainous region
2004; Gitas and Deverux, 2006) based on the assumption in Zhejing, China, and to explore a texture based approach
that the anisotropic nature of reflectance is homogenous for evaluation of the topographic correction effects.
over the study area. However, this assumption is not valid
because the topographic effects and land-cover variations
generate k variation spatially (Bishop and Colby, 2002).
In reality, a global k value cannot result in accurate correc- Study Area and Dataset
tion for all slopes and aspects (Ekstrand, 1996; Bishop and The study area is located in Lin’An County, northwest
Colby, 2002) because of differences in topographic impacts Zhejiang province, China (Figure 1). The terrain undulates
on the land surface reflectance. The second approach for greatly, ranging from a highest elevation of 1,787 m in
k computation is related to land-cover classes, assuming that northwest to a lowest elevation of 9 m in east of Lin’An
land covers have strong influences on anisotropic reflectance County. The elevations in northwest and southwest moun-
(Bishop and Colby, 2002; Bishop et al., 2003). Three land- tainous areas are often greater than 1,000 m, but they are less
cover classes (snow, vegetation, and non-vegetation) are than 50 m in the plain in the east part of this county. In the
first classified with an unsupervised method based on selected study area, the highest elevation is 1,580 m with an
the normalized difference vegetation index (NDVI) image. average value of 360 m. The majority of slope is less than
The classes are then used as stratifications, and k values are 50 degrees with an average value of 21.5 degree (Figure 2).
developed based on each class (Bishop and Colby, 2002; The complex topographic condition in this study area is
Bishop et al., 2003). A similar approach for computation of ideal for understanding the topographic impacts on land-
multiple k values is based on the stratification of such zones cover reflectance and for exploring topographic correction
as relative flat regions and mountainous regions (Blesius and methods to reduce the topographic effects.
Weirich, 2005). Digital elevation model (DEM) data with 25 m spatial
In addition to the Minnaert correction method, other resolution and a Landsat-7 enhanced Thematic Mapper
models have also been explored in previous research. For Plus (ETM) image (path/row: 119/39), which was acquired
example, the statistical-empirical approach assumes that on 26 May 2003 cloud-free were used in this research. The
differential illumination is the same for all cover types and a solar azimuth and elevation angles were 133.49 and 52.20
linear relationship exists between illumination and spectral degrees, respectively, when the ETM image was captured.
reflectance (Teillet et al., 1982; Allen, 2000). Teillet et al. The ETM image was rectified with 37 control points
(1982) introduced an additional adjustment to the cosine which were collected from 1:50 000 topographic maps.
function called the C-correction model, which the constant The root mean square error (RMSE) was 0.4802 (x: 0.3447,
C was used to weaken overcorrection of faintly illuminated y: 0.3343) pixels. A nearest-neighbor algorithm was used
pixels. Riano et al. (2003) reviewed the major approaches for to resample the ETM image into a pixel size of 25 m by
topographic normalization related to Lambertian and non- 25 m in order to match the spatial resolution of selected
Lambertian assumption. A combined correction model of DEM data. After geometric rectification, the ETM digital
atmospheric and topographic effects (Conese et al., 1993; numbers were converted to at-sensor reflectance with an
Sandmeier and Itten, 1997) has been also used for topo- apparent reflectance model (Markham and Barker, 1987;
graphic correction, such as the atmospheric and topographic Chavez, 1989) in order to eliminate the impacts of sensor
correction (ATCOR) model. A detailed description of this instruments and solar zenith angle.
model is provided by Richter (1997).
An evaluation of the topographic correction result
is often required in order to understand the performance Methods
of the methods which are used for reducing topographic The Minnaert correction method is frequently used in
effects. Riano et al. (2003) summarized four possible methods, previous research, and thus is tested in this research. The
i.e., analysis of changes in spectral characteristics, graphic Minnaert correction model can be expressed as Equation 1:
analysis of reflectance before and after topographic correction,
standard deviation for each class, and the accuracy compari- LH LT cos e /(cos e cos i)k (1)
son of classification images. Most of previous research in
evaluating the topographic correction effects is based on the where LH is the equivalent reflectance on a flat surface with
comparison of classification accuracies between before and incident angle of zero, LT is the measured radiance in the
after topographic correction. Some research has indicated that remotely sensed data, k is a Minnaert constant, e is slope,
topographic correction is helpful in improving classification and i is the solar incident angle in relation to the normal
accuracy (Civco, 1989; Meyer et al., 1993; Tokola et al., 2001), of a pixel. The cosine of the incident solar angle (cos i),
but other research did not show the capability in improving referred to as illumination, is calculated using Equation 2
the classification accuracy (Blesius and Weirich, 2005). In (Holben and Justice, 1980; Smith et al., 1980):
reality, many factors may affect the classification accuracy.
Thus, classification accuracy is not a good indicator for the cos i cos u cos e sin u sin e cos(wm ws), (2)
evaluation of topographic correction effects (Bishop and Colby, where u and fm are solar zenith angle and azimuth, and e
2002). Another approach is based on semivariogram analysis and fs are slope and aspect of the terrain. In order to solve
(Bishop and Colby, 2002; Bishop et al., 2003) based on the k, Equation 1 can be reorganized as Equation 3:
assumption that spectral variation within the same land-cover
should be reduced after topographic correction. log(LT cos e) log LH k log(cos e cos i). (3)
Figure 1. Location of the study area: Lin’An County, Zhejiang Province, China.
This formula indicates that the k value is equivalent to variable. The coefficient of determination (R2) is used as an
the slope of a regression line. Thus, the k value can be indicator for the selection of best models.
found from the regression analysis for each band based on When the slope is less than one degree, no topographic
the selected sample data (Meyer et al., 1993; Jensen, 1996). effect on land surface reflectance is assumed; thus, no
Equations 2 and 3 indicate slope, aspect, solar zenith topographic correction is conducted. When the slope is
angle, and azimuth are the required parameters used in greater than 50 degrees, no specific k values are developed
the Minnaert correction model. The DEM data are used to because of the limited number of samples. So with pixels
compute slope and aspect images. The solar zenith angle with slopes of greater than 50 degrees, the k value derived
and azimuth are from the ETM metadata. The above four from slope of 50 degrees is used with the consideration of
parameters are used to compute illumination (cos i) with model limitation in extrapolation. After the model for each
Equation 2. The ETM image, illumination, and slope images band is developed, this model is then used to estimate the
are then stacked into one file. Sample data are extracted k image for each band. Equation 1 is finally used to correct
at every ten pixels from the entire image on the stacked the topographic effects for each band.
file. After the samples with slopes less than one degree are The evaluation of topographic correction results is
removed, a total of 67,008 samples are used for the develop- conducted based on texture images, which are derived
ment of Minnaert coefficients for each band. The selected with the homogeneity texture measure and a window size
samples are separated into 10 groups based on slope ranges, of 5 pixels by 5 pixels for each corrected image. The homo-
as shown in Table 1. Because very limited samples have geneity texture measures the spectral variation within the
slopes greater than 50 degrees, the samples with slopes selected window size. The higher homogeneity values repre-
greater than 45 degrees were merged as the tenth group. sent less spectral variation within the window size. Because
A regression analysis based on Equation 3 is conducted for topographic correction reduces the topographic effects on
each slope group in order to develop the k value for each the surface reflectance, the spectral variation within the same
ETM band corresponding to each slope group. land-cover in the topographic correction result should be
A scatterplot between computed k values and the middle decreased. The homogeneity is calculated with Equation 4:
values in the slope ranges is first explored to examine their
relationships (Figure 3). Because of the nonlinear relation- n1
Pi,j Vi,j
ships illustrated in Figure 3, nonlinear algorithms, such as HOM
i, j0 1(i j)
2
, and Pi,j n1
(4)
logarithmic, exponential, and polynomial, are tested where k Vi,j
is used as a dependent variable, and slope as an independent i, j0
Results
The Minnaert k value is wavelength-dependent and related
to slopes. As Figure 3 shows, the k value has a similar trend
for each band, i.e., the k values decrease as slope increases
Figure 2. (a) elevation, and (b) slope images of the with nonlinear trends. The k values decrease rapidly in the
study area. slope ranges between 1 and 15 degrees, and then gradually
become stable when slopes are in the ranges of 15 to
35 degrees, and finally decrease slightly when slopes are
greater than 35 degrees. This trend confirms that a single
global k value is not suitable for an entire image, especially
where Vi,j is the value in the cell i, j (row i and column j) for the areas with very low or very high slopes, because of
of the moving window, and n is the number of rows or different impacts of topographic conditions on the remotely
columns. sensed data. Generally, k value increases as wavelength
In order to explore the effects of slope on topographic increases, as band 1 has lowest k values and band 7 has
correction results, a mean homogeneity value for each slope highest k values at the same slope.
is calculated with Equation 5: A comparison of different nonlinear regression models,
such as logarithmic, polynomial, and exponential, indicates
N
that the polynomial regression models can successfully
Ms HOMsi/N (5)
i1 simulate the trend of k values as the slope changes for each
ETM band, which the coefficient of determination (R2) are
where Ms is the mean value of homogeneity at the slope s, greater than 0.99 for each band (Figure 4). Therefore, a pixel-
and N is the number of samples in the slope s. Higher Ms based k image is developed with the polynomial regression
values indicate more homogenous. A scatterplot illustrating model for each band. Figure 5 provides an example of a
the relationship between Ms and slope is then used to assess k image for ETM band 5. The flat areas, such as urban,
agricultural lands, and water have high k values, but steeper
slopes result in lower k values. A lower K value can reduce
the topographic effects, especially for those areas with steep
TABLE 1. SUMMARY OF SELECTED SAMPLES FOR EACH SLOPE GROUP
slopes. With the developed k image for each band, as well as
No. Slope Range Median Slope No. of Samples the images of cos e and cos i, the topographic correction for
each ETM band is then conducted with Equation 1. Figure 6
1 1–5 2.5 3,433 provides part of the study area with color composites (i.e.,
2 5–10 7.5 4,776 ETM bands 4, 5, and 3 as red, green, and blue) as a compar-
3 10–15 12.5 6,455 ison of before and after topographic correction. Visual
4 15–20 17.5 8,438 examination of the images indicates that the topographic
5 20–25 22.5 10,695 effects have been significantly reduced with the pixel-based
6 25–30 27.5 10,853
Minnaert correction model.
7 30–35 32.5 9,926
8 35–40 37.5 7,411 Figure 7 provides a quantitative evaluation of topo-
9 40–45 42.5 3,722 graphic effects for each ETM band through a comparison
10 45–65* 47.5 1,299 of the mean homogeneity values between before and after
topographic correction. For ETM bands 1, 2, 3, and 7, the
Note: *very limited pixels have slopes greater than 50 degrees. mean homogeneity values increases gradually as slopes
Figure 4. Polynomial regression models simulating the trend of k values for each band: (a) TM
band 1, (b) TM band 2, (c) TM band 3, (d) TM band 4, (e) TM band 5, and (f) TM band 7.
Figure 6. A comparison of ETM+ images between (a) before, and (b) after topographic correction
(bands 4, 5, and 3 as red, green, and blue). A color version of this figure is available at the
ASPRS website: www.asprs.org.
the mean homogeneity values are higher in the ETM reflectance (thus, also implicitly on derived surface
bands 1, 2, 3, and 7 after topographic correction than the reflectance) because these bands usually have less spectral
corresponding bands before topographic correction. This variations. In contrast, the mean homogeneity values in
indicates that topographic correction is effective in reduc- bands 4 and 5, especially band 4, decreased after topo-
ing the topographic influences on the derived apparent graphic correction comparing with the before topographic
Discussion
In the Minnaert correction method, one critical element is
to determine the k value. In previous research, a single
global k value is the common way used for topographic
correction for an entire image. This method is simple and
easy to implement, but a single global k value has demon-
strated a problem in correcting the topographic effects,
especially in those areas with steep slopes, because the
k value is dependent on the land-cover class and topo-
graphic conditions. Thus, multiple k values corresponding
to land-cover classes have been used (Bishop and Colby,
2002; Bishop et al., 2003). However, different topographic
slopes have various impacts on the land-cover reflectance.
Hence, a single k value corresponding to a land-cover type
cannot effectively reduce the topographic effects either.
A pixel-based k image, which is developed based on the
relationships between k values and slopes, has shown
more advantages in improving the topographic correction
performance than a single k value or multiple k values.
Considering the interactions of k, land-cover, and topo-
graphic conditions, the development of a pixel-based
k image based on a combination of land covers and slopes
may be the future direction in the topographic correction.
The selection of algorithms for development of k values
is also important. The polynomial regression analysis used
in this research can closely simulate the k trend within the
data range, but has problems in extrapolation for the areas
with high slopes when the slopes are beyond the range of
selected samples. In this case, different models may com-
bine, for example, use of polynomial model for estimation of
k values with relative low slopes and use of exponential
models for high slopes.
The evaluation of topographic correction results is an
important part, but no universal approaches have been
used in previous research. Many researchers evaluated
Figure 7. Evaluation of topographic effects for each the topographic correction success with the assistance of
ETM+ band (a) before, and (b) after topographic classification accuracies. The improvement of classification
correction (b1…7 and pb1…7 are ETM+ bands and accuracy can be attributed to different factors, such as the
topographically corrected bands with the pixel-based selected classification algorithm, the image processing
Minnaert k approach). procedure used, the complexity of vegetation types and
structures, and topographic impacts. An important reason
that topographic correction can improve classification
accuracy is that topographic correction reduces the topo-
graphic effects on the vegetation reflectance, and highlights
the differences among different vegetation types or struc-
correction image. This is because the topographic impact tures. Some previous research for evaluation of topographic
in gentle slopes (less than 15 degrees) is not an important correction effects using semivariogram or standard devia-
factor in affecting land-cover reflectance. Thus, topo- tion is based on the assumption that topographic correction
graphic correction for gentle slopes does not reduce the reduces the spectral variations. This assumption may be
heterogeneity in bands 4 and 5. As slopes increase, the true for visible bands, but may be not true for NIR and SWIR
impacts of topographic slopes result in reduced vegetation images, especially in the vegetated areas with steep slopes
reflectance and decrease the heterogeneity, even between and vegetation covers of different types.
different vegetation types. Therefore, one consequence of
the topographic effects in mountainous areas is the
difficulty in quantitative analysis of the vegetation types, Conclusions
such as vegetation classification. After topographic correc- Previous research has indicated the drawback of using a
tion, the topographic impacts on the vegetation reflectance single global k value in topographic correction, but no
are decreased, thus highlighting the reflectance differences effective approaches have been developed. This research
among different vegetation stand structure and species develops a new approach for calculation of a pixel-based
composition, and resulting in the improvement of classifi- k image based on the relationship between k and slope
cation performance of vegetation types after topographic and uses a homogeneity approach to evaluate the topo-
correction. So the mean homogeneity values after topo- graphic correction result. This research has demonstrated
graphic correction are decreased in ETM bands 4 and 5. the promising in using the pixel-based approach for
When slopes are greater than 50 degrees, the mean homo- topographic correction in a mountainous region. The
geneity values become unstable because the k values with homogeneity-based approach used in this research can be
slopes of greater than 50 degrees are replaced with the used to explain the topographic effects on the land-cover
k value at slope equal to 50 degrees. Higher slopes result reflectance. More research is needed to develop the pixel-
in more severe topographic effects on the land-cover based k coefficient image for each band based on a combi-
reflectance. nation of land covers and topographic slopes.
Acknowledgments Holben, B.N., and C.O. Justice, 1980. The topographic effects on
The authors wish to thank the National High Technology spectral response from nadir-point sensors, Photogrammetric
Engineering & Remote Sensing, 46(10):1191–1200.
Research and Development Program of China (863 Program:
#2006AA12Z109) for the financial support. We also wish to Itten, K.L., and P. Meyer, 1993. Geometric and radiometric correc-
tion of TM data of mountainous forest areas, IEEE Transactions
thank three anonymous reviewers for their constructive
on Geoscience and Remote Sensing, 31:764–770.
comments and suggestions.
Jensen, J.R., 1996. Introductory Digital Image Processing: A Remote
Sensing Perspective, Second edition, Prentice Hall, Upper
Saddle River, New Jersey, 318 p.
References Jones, A.R., J.J. Settle, and B.K. Wyatt, 1988. Use of digital terrain
Allen, T.R., 2000. Topographic normalization of Landsat Thematic data in the interpretation of SPOT-1 HRV multispectral imagery,
Mapper data in three mountain environments, Geocarto International Journal of Remote Sensing, 9:669–682.
International, 15:13–19.
Leprieur, C.E., J.M. Durand, and J.L. Peyron, 1988. Influence of
Bishop, M.P., and J.D. Colby, 2002. Anisotropic reflectance topography on forest reflectance using Landsat Thematic
correction of SPOT-3 HRV imagery, International Journal of Mapper and digital terrain data, Photogrammetric Engineering &
Remote Sensing, 23:2125–2131. Remote Sensing, 54(4):491–496.
Bishop, M.P., J.F. Shroder Jr., and J.D. Colby, 2003. Remote sensing Markham, B.L., and J.L. Barker, 1987. Thematic Mapper bandpass
and geomorphometry for studying relief production in high solar exoatmospheric irradiances, International Journal of
mountains, Geomorphology, 55:345–361. Remote Sensing, 8:517–523.
Belsius, L., and F. Weirich, 2005. The use of the Minnaert correc- Meyer, P., K.L. Itten, T. Kellenberger, S. Sandmeier, and R. Sand-
tion for land-cover classification in mountainous terrain, meier, 1993. Radiometric corrections of topographically induced
International Journal of Remote Sensing, 26:3831–3851. effects on Landsat TM data in alpine environment, ISPRS
Chavez, P.S. Jr., 1989. Radiometric calibration of Landsat Thematic Journal of Photogrammetry and Remote Sensing, 48:17–28.
Mapper multispectral images, Photogrammetric Engineering & Mitri, G.H., and I.Z. Gitas, 2004. A performance evaluation of a
Remote Sensing, 55(11):1285–1294. burned area object-based classification model when applied
Chavez, P.S. Jr., 1996. Image-based atmospheric corrections – to topographically and non-topographically corrected TM
Revisited and improved, Photogrammetric Engineering & imagery, International Journal of Remote Sensing,
Remote Sensing, 62(9):1025–1036. 25:2863–2870.
Civco, D.L., 1989. Topographic normalization of Landsat Thematic Riano, D., E. Chuvieco, J. Salas, and I. Aguado, 2003. Assessment of
Mapper digital imagery, Photogrammetric Engineering & Remote different topographic corrections in Landsat-TM data for
Sensing, 55(11):1303–1309. mapping vegetation types, IEEE Transactions on Geoscience
Colby, J.D., 1991. Topographic normalization in rugged terrain, and Remote Sensing, 41:1056–1061.
Photogrammetric Engineering & Remote Sensing, 57(5):531–537. Richter, R., 1997. Correction of atmospheric and topographic effects
Colby, J.D., and P.L. Keating, 1998. Land-cover classification using for high spatial resolution satellite imagery, International
Landsat TM imagery in the tropical highlands: The influence of Journal of Remote Sensing, 18:1099–1111.
anisotropic reflectance, International Journal of Remote Sensing, Sandmeier, S., and K.I. Itten, 1997. A physically-based model to
19:1479–1500. correct atmospheric and illumination effects in optical satellite
Conese, C., M.A. Gilabert, F. Maselli, and L. Bottai, 1993. Topographic data of a rugged terrain, IEEE Transactions on Geoscience and
normalization of TM scenes through the use of atmospheric Remote Sensing, 35:708–717.
correction method and digital terrain models, Photogrammetric Smith, J.A., L.L. Tzeu, and K.J. Ranson, 1980. The Lambertian
Engineering & Remote Sensing, 59(12):1745–1753. assumption and Landsat data, Photogrammetric Engineering &
Ekstrand, S., 1996. Landsat TM-based forest damage assessment: Remote Sensing, 46(10):1183–1189.
Correction for topographic effects, Photogrammetric Engineering Teillet, P.M., B. Guindon, and D.G. Goodenough, 1982. On the
& Remote Sensing, 62(1):151–161. slope-aspect correction of multispectral scanner data, Canadian
Franklin, S.E., D.R. Connery, and J.A. Williams, 1994. Classification Journal of Remote Sensing, 8:84–106.
of alpine vegetation using Landsat Thematic Mapper, SPOT HRV Tokola, T., J. Sarkeala, and M. van der Linden, 2001. Use of topo-
and DEM data, Canadian Journal of Remote Sensing, 20:49–56. graphic correction in Landsat TM-based forest interpretation in
Garcia-Haro, F.J., M.A. Gilabert, and J. Melia, 2001. Monitoring fire- Nepal, International Journal of Remote Sensing, 22:551–563.
affected areas using Thematic Mapper data, International Vermote, E., D. Tanre, J.L. Deuze, M. Herman, and J.J. Morcrette,
Journal of Remote Sensing, 22:533–549. 1997. Second simulation of the satellite signal in the solar
Gitas, I.Z., and B.J. Devereux, 2006. The role of topographic spectrum, 6S: An overview, IEEE Transactions on Geoscience
correction in mapping recently burned Mediterranean forest and Remote Sensing, 35:675–686.
areas from Landsat TM images, International Journal of Remote Walsh, S.J., J.W. Cooper, I.E. von Essen, and K.R. Gallagher, 1990.
Sensing, 27:41–54. Image enhancement of Landsat Thematic Mapper data and GIS
Gu, D., and A. Gillespie, 1998. Topographic normalization of data integration for evaluation of resource characteristics,
Landsat TM images of forest based on subpixel sub-canopy- Photogrammetric Engineering & Remote Sensing,
sensor geometry, Remote Sensing of Environment, 64:166–175. 56(10):1135–1141.
Gu, D., A. Gillespie, J.B. Admas, and R. Weeks, 1999. A statistical
approach for topographic correction of satellite images by using
spatial context information, IEEE Transactions on Geoscience (Received 01 August 2006; accepted 01 February 2007; revised
and Remote Sensing, 37:236–246. 20 February 2007)