a physically-based model to correct · pdf filesandmeier, st., and k.i. itten, 1997 1 a...
TRANSCRIPT
Sandmeier, St., and K.I. Itten, 1997 1
A PHYSICALLY-BASED MODEL TO CORRECT ATMOSPHERIC AND ILLUMINATION
EFFECTS IN OPTICAL SATELLITE DATA OF RUGGED TERRAIN
St. Sandmeier and K.I. Itten
Remote Sensing Laboratories
Department of Geography
University of Zurich-Irchel CH-8057 Zurich, Switzerland
ABSTRACT
A physically-based model to correct atmospheric and topographically induced illumination
effects in optical satellite data is developed and tested. Special emphasis is put on the impact of
rugged terrain. Ground reference data for various land use classes enables the assessment of
the corrections' influence on land use classifications. The estimation of surface reflectance is
achieved in a two-step procedure. First irradiance components and atmospheric parameters are
calculated for horizontal surfaces using the atmo-code 6S [1], then the influence of the
topography on the parameters is integrated using DEM data.
I. INTRODUCTION
After the launch of the first Landsat satellite in July 1972, scientific studies in remote
sensing primarily focused on land use classification and long-term changes in terrestrial land
cover. In general, a flat terrain was assumed in order to avoid difficulties caused by the
topography. A large fraction of the earth surface, however, consists of mountainous areas
where the impact of topography on remote sensing data has to be examined prior to any remote
sensing application. The interfering effect of topography is evident in a single satellite scene
and introduces even stronger distortions in multi-temporal approaches.
The apparent radiance measured by remote sensing systems in rugged terrain is affected
by (1) the intensity of solar irradiance, (2) the atmospheric effects, (3) the bidirectional
reflectance distribution function (BRDF) of the surface sensed, and (4) the spectral response
functions of the sensor bands. In rugged terrain special emphasis has to be put on the influence
of topography on solar irradiance and on atmospheric effects.
A first approach to correct atmospheric and illumination effects is based on the empirical
relationship between the at-satellite radiance from an object and the direct irradiance provided
Sandmeier, St., and K.I. Itten, 1997 2
by the cosine of the solar incidence angle [2], [3]. The evaluation revealed that these methods
are optimised on a specific satellite scene, test site and object class like forest. Another
drawback is the lack of a solid physical base, which prevents further development of the
empirical and semi-empirical approaches. The aim of a newly initiated ESA study therefore
focused on the physically-based retrieval of surface reflectance [4] following the conditions
listed below:
• atmospheric effects are corrected under consideration of horizontal and vertical variability
• topographically induced variations of the illumination are eliminated taking into account
direct and diffuse irradiance components from sky and terrain
• the methodology is independent from objects, satellite scene and test site; i.e. the
determination of the model parameters is physically-based and makes no use of empirical
data such as the content of a satellite scene
• the bidirectional effects of objects are neglected, assuming Lambertian reflectance
characteristics.
Consequently, objects with identical spectral properties have to reveal the same
reflectance in the satellite image. The three-dimensional relief effect and the atmospheric
blurring have to be suppressed. The satellite image shall appear 'flat' and contrast-enhanced.
The methodology should be applicable to any test site and is not restricted to a specific satellite
sensor. Based on the Lambertian assumption any object class shall be processed. The resulting
target reflectances represent object properties and have to be free from atmospheric and
illumination effects to the fullest possible extent. The radiometrically corrected data shall allow
for multitemporal studies on a multi-sensor basis. Time series of different sensors with similar
spectral bands can be compared because changes in atmosphere and illumination are eliminated,
and the processed imagery represents surface reflectances and not arbitrary digital numbers.
II. DATA BASE
A Landsat TM-scene (frame 194-27) of 11 July 1991, 9:40 am (UT), acquired under
rather hazy atmospheric conditions, is used. At the time of satellite overpass the sun's position
in relation to the test site centre at 47.05° north and 8.52° east is at 33.6° zenith, and 128.3°azimuth. All six reflective TM bands are used.
The test region covering an area of 36.0 km by 17.5 km in the centre of Switzerland
includes the three Swiss Federal Office of Topography Maps 1:25'000 "Zug", "Rigi", and
"Beckenried". The northern part of “Zug” is composed of mainly agricultural areas, water
(Lake of Zug) and settlements. It is characterised by low altitudes between 392 and 1174 m
and a moderate relief. "Rigi" contains large lake areas (Lake of Lucerne) and is dominated by
Sandmeier, St., and K.I. Itten, 1997 3
the Rigi Mountain. Situated in the mountainous pre-Alps, it reveals terrain elevations between
434 m and 1798 m. "Beckenried", in the southern part of the test site, is located partially in
the pre-Alps and in the alpine regions. Here terrain elevation varies from 434 to 2404 m.
Pronounced deep valleys and steep slopes offer a splendid test site for topography-oriented
radiometric corrections.
Radiosonde data at the highest available resolution measured in Payerne by the Swiss
radiosonde station is used to calibrate the atmospheric model for temperature, humidity,
pressure and ozone. Since no aerosol data was available for the time of the satellite overpass,
the 6S-continental aerosol model [1] is used. The estimation of the observed horizontal
visibility is based on meteo-stations "Luzern", "Pilatus", "Altdorf", "Schwyz" and "Engelberg"
which are operated by the Swiss Institute of Meteorology.
A digital elevation model (DEM) with a resolution of 25 m in x and y and 0.1 m for
elevation was available from the Swiss Federal Office of Topography1. It is based on the
1:25'000 Swiss Topographic Maps. Data sets for slope, cosine of incidence angle i, cast
shadow, and sky- and terrain-view factors were derived from this DEM.
Ground reference data for various land use classes was provided by the Swiss Land Use
Statistics 1979/85 of the Swiss Federal Statistical Office. They define land use in 69 different
classes for sample points in a resolution of 100 m for all of Switzerland. The data in the test
site was acquired in 1981. For the classification the original classes have been aggregated into
eight homogenous categories in accordance with the Swiss Federal Statistical Office.
Unfortunately, the Swiss Land Use Statistics do not contain any information on forest stands.
Thus maps of forest stands were digitised in order to assess the influence of radiometric
correction on the classification of forest stands. They had been produced by the Swiss
Sanasilva Project using colour-infrared aerial photographs at a scale of 1:10'000, taken on 25
July 1985 and 13 August 1987. Table 1 gives an overview of the ground reference data used in
the classification.
All data sets used in this study are georeferenced to the rectangular coordinate system of
the Swiss Topographic Maps preceding radiometric corrections. This rectification also includes
geometric correction of relief displacement due to variations in terrain elevation [3]. In order to
avoid introducing new Digital Numbers, a nearest neighbour resampling technique was applied
instead of a bilinear or cubic convolution interpolation.
III. MODEL DESCRIPTION
For a given satellite band b the surface reflectance ρ(b), assuming a Lambertian ground
reflectance, can be calculated by
1 DHM and map data courtesy: Swiss Federal Office of Topography, June 14, 1995
Sandmeier, St., and K.I. Itten, 1997 4
ρ(b) =π L(b) − Lp (b,z)( )E(b,z) Tu (b,z)
(1)
where E(b,z) is the total solar irradiance reaching a surface on altitude z in a given band b;
Tu (b,z) is the upward transmission from the surface to the sensor; Lp (b,z) is the path
radiance from the surface altitude z up to the sensor altitude and L(b) is the scene radiance in
satellite band b.
The derivation of surface reflectance requires firstly the conversion of the digital numbers
(DN) to quantitative physical values. For the six reflective bands of Thematic Mapper the at-
satellite radiances L [mW·cm-2·sr-1·µm-1] are calculated using the TM calibration constants a0
(offset) and a1 (gain) with
L = a0 + a1 ⋅ DN (2)
We used the updated in-flight calibration constants assessed by Slater [5] over the
gypsum sand area of White Sands, New Mexico to convert DN into radiance. To obtain at-
satellite radiance L(b) for a given TM band b, L, as a result of equation (2), has to be
convoluted with the relative spectral response function of each TM band b. They are taken here
from the 6S source code [1] and are based on Markham and Barker [6].
A. The atmosphere module (atmo-module)
The signal reaching a sensor depends on the surface reflectance ρ, but it is perturbed by
two atmospheric processes, the gaseous absorption and the scattering by molecules and
aerosols. The atmospheric code 6S [1] used in this study takes into account Rayleigh and
aerosol scattering, as well as gas absorption due to water (H2O), carbon dioxide (CO2), ozone
(O3), oxygen (O2), methane (CH4), nitrous oxide (N2O), and carbon monoxide (CO) between
0.25 and 4.0 µm in a spectral resolution of 2.5 nm. The input parameters for 6S can be
chosen from proposed standard conditions, or specified by the user. Air pressure, air
temperature, air humidity and ozone concentration in a vertical profile are derived from
radiosonde measurements. Unfortunately, information about type and concentration of aerosols
is usually not gathered in a field campaign although it is of essential importance. The type of
aerosol can be approximated by a standard aerosol model, but optical thickness giving the
concentration of aerosols has to be estimated by horizontal visibilities from meteo-stations or by
the user if no measurement data is available. Horizontal visibilities and optical thickness,
respectively, play a decisive role in atmospheric modelling, thus they have to be carefully deter-
mined.
The 6S code predicts the satellite signal reflected from a plane horizontal surface
assuming cloudless atmosphere. The altitude of the targets is considered. 6S provides all
Sandmeier, St., and K.I. Itten, 1997 5
parameters used in the physically-based model but only for horizontal surfaces. The
topography, except for its altitude, is not considered.
B. The topography module (topo-module)
The path radiance and the upward transmittance Lp (b,z) and Tu (b,z) in equation (3) are
dependent on altitude and spectral conditions only. At the same time however, the total solar
irradiance E(b,z) is strongly affected by the surface orientation. An object lying in shadow
obviously gets less solar irradiance than one exposed to the sun. Furthermore, the geometry
between the sun's position and the surfaces' orientation affects the ratio of direct and diffuse ir-
radiance components, and the amount of terrain reflected radiance reaching an adjacent surface.
Thus this study has to concentrate not only on the altitude dependence of atmospheric effects,
but even more on the impact of topography on the solar irradiance reaching an inclined surface.
The total solar irradiance E(b,z) in a band b for a tilted surface on altitude z consists of
three components: direct, diffuse, and terrain irradiance. Similar to [7] and [8], it is given by:
E(b,z) = Θ ⋅ Edh (b,z) ⋅ cos(i)
cos(sz)+ (direct irradiance)
Efh (b,z) ⋅ k(b,z) ⋅ cos(i)
cos(sz)+ (1 − k(b,z)) ⋅ Vd
+ (diffuse irradiance)
Eh(b,z) ⋅ Vt ⋅ρadj (terrain irradiance)
(3)
where:
E(b,z) = total irradiance on an inclined surface
Eh(b,z) = total irradiance on a horizontal surface
Edh (b,z) = direct component of irradiance on a horizontal surface
Efh (b,z) = diffuse component of irradiance on a horizontal surface
k(b,z) = anisotropy index
Vd = sky-view factor
Vt = terrain-view factorρadj = average reflectance of adjacent objects
Θ = binary coefficient to control cast shadow
i = angle of sun's incidence (cos (i) · 100 = illumination)
sz = solar zenith angle
The first term shows the cosine law applied to direct irradiance Edh (b,z) on a horizontal
surface and results in the amount of direct irradiance on a tilted target. Parameter i is the angle
between the normal on the surface and the sun's rays, thus the angle of incidence of direct
Sandmeier, St., and K.I. Itten, 1997 6
irradiance. Θ is a binary coefficient, and is set to zero for surfaces in cast shadow. Both
parameters i and Θ are derived from the DEM [9].
The second term represents the diffuse irradiance in a sloped terrain. As Proy et al. [10]
recommend, Efh (b,z), the diffuse irradiance on a horizontal surface, is separated into an
isotropic and a circumsolar (anisotropic) component. This becomes necessary as the diffuse
irradiance exhibits a fairly strong anisotropic circumsolar portion which has to be modelled
differently from the isotropic component. On a misty day, it is obvious that there is a peak of
diffuse irradiance in the sun's direction, otherwise the position of the sun could not be detected.
The values of the isotropic and circumsolar components are derived using Hay's [11]
anisotropy index k(b,z) . It is calculated from the ratio of direct irradiance on a surface normal
to the sun's rays Edn (b,z) and the top of the atmosphere radiance Ed
t (b):
k(b,z) = Edn (b,z)
Edt (b)
(4)
k(b,z) is related to the atmospheric transmittance for direct irradiance and values between
0 and 1. It seems to satisfy the wavelength dependence of the scattering process [8]. The lower
the atmospheric transmittance the stronger the isotropic component of the diffuse irradiance and
as a consequence, the lower is k(b,z) .
The circumsolar component of diffuse irradiance can be modelled for topography in the
same way as the direct irradiance Edh (b,z), though it is part of the diffuse irradiance. The
amount of isotropic diffuse irradiance on the other hand, is a function of the proportion of sky
hemisphere not obstructed by topography. Dozier and Marks [12] introduced a sky-view factor
Vd defined as the ratio of the sky portion seen from a specific surface to that on an
unobstructed horizontal surface, i.e. 0 < Vd ≤ 1. The total diffuse irradiance can therefore be
calculated for tilted surfaces by
Ef (b,z) = Efh (b,z) ⋅ k(b,z) ⋅ cos(i)
cos(sz)+ 1 − k(b,z)( ) ⋅ Vd
(5)
anisotropic isotropic
portion portion
where Efh (b,z) and Ef (b,z) are the diffuse irradiance on a horizontal and a tilted surface,
respectively.
The third term in equation (3) refers to terrain irradiance. Especially in the case of deep
valleys, radiance reflected from neighbouring slopes contributes to the irradiance on adjacent
surfaces. The amount of the terrain irradiance depends upon (1) the total irradiance E(b,z)
reaching the adjacent slopes, (2) the portion of adjacent terrain seen from a surface Vt , (3) the
surface reflectances of the adjacent objects ρadj and (4) the distance between the surface sensed
and the adjacent slopes. Thus terrain irradiance has to be accounted for above all in snow
Sandmeier, St., and K.I. Itten, 1997 7
covered rugged terrain. In shadowed areas, however, this effect cannot be neglected even for
dark objects as Vt is large and Ed (b,z) and Ef (b,z) are small [10], [12].
Two approaches are implemented and tested to obtain Vt and Vd , (1) a simplified
trigonometric approach described by Kondratyev [13], and (2) an analytical procedure
introduced by Dozier et al. [14], [15]. The approach of Kondratyev [13] approximates Vt and
Vd by trigonometric functions. The slope angle s of the surface considered is used as the only
parameter to estimate the amount of sky and terrain seen from a point. For a horizontal plane
with slope angle 0° the approach reveals a sky-view factor of 1 and a terrain-view factor of 0,
while for a vertical plane both Vd and Vt turn to 0.5:
Vd = 1 + cos s( )2
(6)
Vt = 1 − cos s( )2
(7)
This simple trigonometric approach can only be applied to a horizontal surface adjoining
an infinitely long slope with slope angle s. The impact of adjacent hills reducing the amount of
visible sky is not considered. The elevation angle s is extracted from the DEM.
The procedure described by Dozier [14], [15] determines Vd and Vt analytically. It
defines first the local horizon points H(i) for each DEM-point i over the entire azimuth circle in
a given resolution ∆θ. Then the local horizon angles h0 are calculated. They represent the
largest slope angle h(i,j) between a DEM-point i and any other DEM-point j in a given direction
θ. Vd and Vt are then obtained by the integration of h over the azimuth circle:
Vd = cos(γ )h 0
π / 2
∫ cos h θ[ ]( )dhdθ0
2π
∫ (8)
Vt = cos(γ )0
h 0
∫ cos h θ[ ]( )dhdθ0
2π
∫ (9)
where:
Vd = sky-view factor
Vt = terrain-view factor
γ = angle between normal on surface and vector with elevation h
θ = azimuth angle
h = elevation angle
h0 = horizon angle
Fig. 1 gives an example of the methodology applied to the sub test site "Beckenried",
pointing out the horizon pixels H(i) for an initial DEM point i with Swiss Map coordinates
676'625 | 201'250. The inset shows the azimuth projection of the corresponding horizon
Sandmeier, St., and K.I. Itten, 1997 8
angles h0 and represents the local horizon line of point i. A resolution of ∆θ = π/16 (32
directions over the azimuth circle) is chosen. The background depicts the integrated surface
within the horizon line of the inset for each DEM-point, which is the sky-view factor Vd .
All algorithms applied in the physically-based model were implemented in a commercially
available image-processing software [9].
IV. RESULTS
A. Visual analysis
In Fig. 2, the radiometrically raw image and the resulting images of the various
radiometric correction steps are shown. To enable a comparison, the images are not processed
by image enhancement techniques except for a linear histogram stretching, applied to all four
images. The radiometrically uncorrected image (Fig. 2a) appears blurred and demonstrates the
hazy atmospheric conditions at the time of satellite overpass. Details in the valley bottom cannot
be distinguished and the topographically induced illumination variations are small, due to the
large amount of blurring diffuse irradiance. In the atmospheric correction (Fig. 2b) the three
TM bands are processed solely using the atmo-module. Thus the only factor corrected is the
altitude dependent effect of the atmosphere. As no illumination correction was applied, the
topographically induced illumination variations are emphasised due to a reduction of the
atmospheric blurring effect. Thus the relief is pronounced. Moreover the spatial resolution
seems improved by a reduction of the atmospheric hazing. Details in the valley bottom as well
as in the alpine agricultural regions are enhanced as a result of the correction. All colours are
more saturated in comparison with the raw image, e.g. the blue of the lakes and the green of the
meadows. The image appears homogeneous over the various altitudes. No artefacts brought in
by the atmospheric correction can be detected. An impressive improvement of the satellite data
from a visual point of view could be obtained.
The correction of the illumination effects using the trigonometric approach of Kondratyev
[13] (Fig. 2c) is successful only to a certain degree. In the medium and highly illuminated
areas the illumination effect is corrected properly. The relief impression got lost and these parts
of the image appear flat, best seen in the little valley depicted in the zoom section. The faintly
illuminated surfaces, however, are overcorrected and expose artefacts, e.g. along the ridges
and in the left side of the zoom section. The correction of the illumination effects based on the
horizon line approach (Fig. 2d) proved to be the most successful. Most of the artefacts could
be eliminated, although along the ridges some overcorrected pixels remain. They are most
probably due to an insufficient spatial resolution of the DEM used in the study, because tests on
the exact location of the artefacted areas revealed an inadequacy in the data sets of cos(i) and of
the cast shadow.
Sandmeier, St., and K.I. Itten, 1997 9
The impact of the DEM inaccuracies is emphasised by the mixed signature problem.
Surfaces along ridges in "Beckenried" are often bare limestone with high reflectance properties.
A pixel of the region just 'behind' the ridge consists of dark shadowed areas and to some
extent illuminated and highly reflective limestone. The mixed signature of such a ridge pixel is
influenced by the brightening effect: with regard to the proportion of dark and bright parts
within the pixel, the surface appears to be bright, and as a consequence it is overcorrected.
B. Statistical analysis
Fig. 3 shows histograms of band 2 radiometrically raw (Fig. 3a) and corrected for atmo-
spheric (Fig. 3b) and illumination effects (Figs. 3c and 3d) in the Buochserhorn area, a
subsection within the test site "Beckenried" (Swiss Topographic Map coordinates
673'250 | 202'550 upper left and 677'660 | 199'250 lower right corner). The site was
chosen as it contains areas of predominantly forest and alpine agriculture in various illumination
conditions between 400 and 1800 m. In the spectral range of band 2 the correction of
illumination-effects should result in a bimodal histogram, the peaks representing forest and
alpine agricultural areas. In contrast to this, the histograms of the radiometric raw and the
atmospheric corrected image should appear non-bimodal, since they are influenced by the
impact of topographically induced illumination effects.
Indeed the non-bimodality can be seen in the radiometrically raw band 2 (Fig. 3a),
although the blurring influence of the atmosphere reduces the impact of illumination on the
histograms shape. The atmospheric correction reveals a contrast enhancement by reducing the
scattering effect of the atmosphere. Thus illumination effects are emphasised and cause a strong
heterogeneous appearance of the objects in the satellite imagery. In spite of the predominant
presence of two discriminant object classes the histogram of the atmospheric corrected image
appears non-bimodal (Fig. 3b). By the combination of illumination and atmospheric correction
using the trigonometric approach (Fig. 3c), however, the impact of illumination on the
appearance of the histogram can be eliminated successfully. The bimodality of the histogram
clearly shows the frequency-distribution of the two dominant object classes forest and agri-
culture. By considering the horizon lines, the bimodality can be impressively enhanced
(Fig. 3d).
Figs. 4a-c demonstrate that a correction of atmospheric and illumination effects does not
always lead to an enhancement of the bimodality in a frequency distribution of two object
classes. Fig. 4a shows the histogram of arable land, meadows and farm pastures combined
with alpine agricultural areas in test site "Beckenried" in the raw data of TM 5. A weak but still
obvious bimodality can be observed for both object classes. The correction of the atmospheric
effects leads to a smoothing of the histograms (Fig. 4b) and to an even better fit of the two
objects' histograms. The correction of illumination effects (Fig. 4c) finally results in a nearly
ideal Gaussian distribution. This proves that the two object classes reveal an identical spectral
Sandmeier, St., and K.I. Itten, 1997 10
behaviour in band 5, but are influenced by atmospheric and illumination effects. This can lead
to an improved separability of the classes in the raw data: arable land, meadows and farm
pasture located on a mean altitude of 686 m are stronger affected by atmospheric effects than
the alpine agricultural areas which are found on a mean altitude of 1486 m. Thus the impact of
the atmosphere on the spectral appearance improves the classification, but one discriminates
altitude dependent atmospheric effects rather than spectral varying surfaces. Also test site
specific influences of illumination effects can bias the classification results: an object
predominantly lying in shadowed areas is probably easier to classify before an illumination
correction takes place.
C. Classification analysis
The impact of the radiometric correction on an image classification is evaluated for
various objects. Based on the ground reference data, a classification of Swiss Land Use
Statistics aggregates and the forest stand classes coniferous, deciduous and mixed stands is
performed. For both classifications a maximum likelihood procedure with four bands is
applied. Clouds and cloud shadows are omitted. A Kolmogoroff-Smirnov test was applied to
each of the TM bands in order to test the prerequisite of a normality distribution. Except for
water, which is very easy to classify, all object classes fulfil the normality test. The use of an a-
priori value is neglected, and the threshold value is set to three standard deviations resulting in
classifying between 95% and 99% of the training area. The measures used to assess the
classification accuracy are the producer, user and overall accuracy. The producer accuracy is
defined as the total number of correctly classified pixels in a category divided by the total
number of pixels of that category in ground reference data. The total number of correct pixels in
a category divided by the total number of pixels that were classified in that category is called
user accuracy [16]. The overall accuracy is simply the number of correct pixels of all categories
divided by the total number of pixels in ground reference data. User and producer accuracies
are class specific and have to be referenced for each class under assessment. The overall
accuracy, as the name indicates, is a general measure for the classification in a test site.
In order to prevent from impacts due to the selection of training sets, the complete ground
reference data as given in Tab. 1 was used for training and verification. The results of the
classification are depicted in Fig. 5. The indices (a) to (d) correspond to: raw (a), atmo-
corrected (b), atmo-illu-corrected without horizon line (c), and atmo-illu-corrected data with
horizon line (d).
The interpretation of Fig. 5a is difficult as no clear tendency is obvious. The
classification accuracies of forests is almost the same in (a), (b), and (d), even though the
producer and user accuracies vary. In (c) the accuracy is clearly lower. The accuracies of the
settlement and urban areas perform best in (a), because the atmospheric correction in (b), (c),
and (d) seems to reduce the classification. Orchards, vineyards, and horticulture are not much
Sandmeier, St., and K.I. Itten, 1997 11
influenced by the radiometric correction steps and remain almost the same in (a) to (d).
Meadows, arable land and farm pastures are clearly improved, predominantly by the
atmospheric correction (b), but - as expected from the histogram analysis - the classification of
alpine agricultural areas is strongly weakened in (b), (c), and (d) since differences in
geoecological niche populations adapted to specific illumination conditions are smoothed out.
The classification of lakes and rivers, water shores, shore vegetation, wetlands, and other
unproductive areas remain almost untouched by the radiometric correction. Thus it must be
concluded that the radiometric correction has almost no effect on the classification accuracy of
all eight aggregates.
Fig. 5b illustrates the results of the forest stand classification in the Buochserhorn area,
showing the overall accuracy and the number of unclassified pixels. Here, the classification
accuracy can be enhanced considerably by the radiometric correction steps. The atmospheric
correction (b) reveals an improvement of 1 %, the illumination correction without horizon line
(c) more than 3 %, and the correction considering the horizon line (d) an improvement of
almost 7 %. In addition, the number of unclassified pixels rejected by the threshold is reduced
by the radiometric correction. Thus the spectral signatures became more distinctive after the
radiometric correction.
V. CONCLUSIONS
In this study a physically-based radiometric correction model is developed in order to
improve a land use classification. The methodology is non-empirical and therefore in principal
applicable to any test site, scene and sensor within a range of 0.25 and 4.0 µm. It is proved
that atmosphere and illumination variations have a crucial impact on the spectral appearance of
object classes in a satellite data set. Choosing training area samples and deciding on aggregating
object classes should therefore only be performed after a radiometric correction. Adjacent
slopes cause considerable additional irradiance in faintly illuminated areas. Thus an appropriate
calculation of the terrain irradiance is essential in rugged terrain. Also the isotropic diffuse
irradiance (sky irradiance) cannot be neglected in steep terrain. The simplified trigonometric
approach to calculate sky- and terrain-factors is insufficient in rugged terrain. It causes artefacts
and leads to misclassifications in the faintly illuminated areas. The more sophisticated method
to determine the local horizon line is computer time-intensive, but results in an impressive
improvement of the illumination correction in critical areas with a large amount of diffuse
irradiance.
The visual examination of the corrected images and statistical analysis clearly confirm the
effectiveness of the physically-based radiometric correction procedure. The three-dimensional
effect is enhanced in the atmospherically corrected image by improving the image contrast. The
image appears clear due to an impressive reduction of the atmospheric blurring effect. The
illumination correction reduces the relief impression and leads to a flat appearance of the image
Sandmeier, St., and K.I. Itten, 1997 12
particularly when the horizon line approach is used. Histogram analysis confirms the
elimination of the adverse effect of the atmosphere and topographically induced illumination
variations.
The assessment of the land use classification results is non-uniform. While the forest
stands discrimination could be improved markedly by the correction, other land use classes
were only slightly improved, remained unchanged or even worsened. This is in part due to
inadequacies in the ground reference data:
• the acquisition date of the ground reference data differs from the satellite overflight date:
the Swiss Land Use Statistics data set is 10 years older, and the forest stand maps are
about 5 years older
• the Swiss Land Use Statistics consists of sample points and does not contain surface
information
• the original categorisation of objects in the Swiss Land Use Statistics is mainly based on
land use and not on spectral homogeneity, e.g. class meadow consists of many different
kinds of species and even includes bare soil
• the forest canopy is not considered in the DEM data.
Furthermore the radiometric correction itself and the heterogeneity of the study area lead to a
test site specific decrease of the classification accuracy:
• as a result of atmospheric influence objects lying in low altitudes appear 'brighter' and
thus can be distinguished more easily from objects lying predominantly in higher alti-
tudes: the separation of meadows from alpine agricultural areas is easier before an
atmospheric correction is performed
• objects lying predominantly in steep slopes get a lower irradiance and appear 'darker' and
thus can be discriminated more easily from the surroundings before an illumination
correction takes place
• differences in geoecological niche populations adapted to specific illumination conditions
are smoothed out
• the heterogeneity of the surface land cover present in the study area can only be
inadequately registered by a 30 m pixel.
Unlike 'conventional' image enhancement techniques like histogram stretching or colour
look-up table manipulations, the physically-based radiometric correction takes vertical
variations of atmospheric effects into account. A still unsolved drawback in the physically-
based model, however, is the Lambertian assumption. The consideration of bidirectional effects
is beyond the scope of this study and will be treated within the Field-Goniometry and BRDF-
Sandmeier, St., and K.I. Itten, 1997 13
Research-Project at RSL [17]. It will help to further improve radiometric corrections and to
optimise land use classifications in rugged terrain.
VI. ACKNOWLEDGEMENT
This study was supported by the European Space Agency within ESA study no. 125487 and
by the Swiss National Science Foundation, grant no. 21-40539.94. The authors also wish to
thank M. Funk from VAW ETH, Zurich for the base of the horizon line determination
algorithm, and I. Leiss from RSL for help regarding classification procedures.
Sandmeier, St., and K.I. Itten, 1997 14
VII. REFERENCES
[1] E. Vermote, D. Tanré, J.L. Deuzé, M. Herman, and J.J. Morcrette, "Second Simulation
of the Satellite Signal in the Solar Spectrum (6S)," User Guide April 18, NASA GSFC,
Greenbelt MD, USA, p. 183, 1994.
[2] P.M. Teillet, B. Guindon, and D.G. Goodenough, "On the slope-aspect correction of
multispectral scanner data," Canadian J. of Remote Sensing, vol. 8 no. 2, pp. 84-106,
1982.
[3] K.I. Itten and P. Meyer, "Geometric and Radiometric Correction of TM-Data of Moun-
tainous Forested Areas," IEEE Trans. Geosci. Remote Sensing, vol. 31, no. 4,
pp. 764-770, 1993.
[4] St. Sandmeier, K.I. Itten, and P. Meyer, "Improvements of Satellite Land Cover
Classifications in Rugged Terrain Through Correction of Scene Related Effects," Final
Report ESA Study No. 125487, Dept. of Geography, University of Zurich, p. 51,
1994.
[5] P.N. Slater, S.F. Biggar, R.G. Holm, R.D. Jackson, Y. Mao, M.S. Moran, M. Palmer,
and B. Yuan, "Absolute radiometric calibration of the Thematic Mapper," SPIE,
vol. 660, pp. 2-8, 1986.
[6] B.L. Markham and J.L. Barker, "Spectral characterisation of the Landsat Thematic
Mapper sensors," Int. J. Remote Sensing, vol. 6, no. 5, pp. 697-716, 1985.
[7] C.R. Duguay and E.F. LeDrew, "Estimating Surface Reflectance and Albedo from
Landsat-5 Thematic Mapper over Rugged Terrain," Photogram. Eng. Remote Sensing,
vol. 58, no. 5, pp. 551-558, 1992.
[8] D.J. Gratton, P.J. Howarth, and D.J. Marceau, "Using Landsat-5 Thematic Mapper and
Digital Elevation Data to Determine the Net Radiation Field of a Mountain Glacier,"
Remote Sensing Envir., vol. 43, pp. 315-331, 1993.
[9] St. Sandmeier, "A Physically-Based Radiometric Correction Model - Correction of
Atmospheric and Illumination Effects in Optical Satellite Data of Rugged Terrain," Ph.D.
Thesis, Remote Sensing Series, Department of Geography, University of Zurich,
vol. 26, p. 140, 1995.
[10] C. Proy, D. Tanré, and P.Y. Deschamps, "Evaluation of Topographic Effects in
Remotely Sensed Data," Remote Sensing Envir., vol. 30, pp. 21-32, 1989.
[11] J.E. Hay, "Solar energy system design: the impact of mesoscale variations in solar
radiation," Atmos. Ocean, vol. 21, pp. 138-157, 1983.
[12] J. Dozier and D. Marks, "Snow mapping and classification from Landsat Thematic
Mapper data," Annals of Glaciology, vol. 9, pp. 97-103, 1987.
[13] K.Ya. Kondratyev, "Radiation in the Atmosphere," Academic Press, London, 1969.
[14] J. Dozier, J. Bruno, and P. Downey, "A faster solution to the horizon problem,"
Computers and Geosciences, vol. 7, pp. 145-151, 1981.
[15] J. Dozier and J. Frew, "Rapid Calculation of Terrain Parameters For Radiation Modeling
From Digital Elevation Data," IEEE Trans. Geosci. Remote Sensing, vol. 28, no. 5,
pp. 963-969, 1990.
Sandmeier, St., and K.I. Itten, 1997 15
[16] R.G. Congalton, "A Review of Assessing the Accuracy of Classifications of Remotely
Sensed Data," Remote Sensing Envir., vol. 37, pp. 35-46, 1991.
[17] St. Sandmeier, W. Sandmeier, K.I. Itten, M.E. Schaepman, and T.W. Kellenberger,
"The Swiss Field-Goniometer System (FIGOS)," in Proc. of IGARSS'95, Firenze, Italy,
pp. 2078-2080, 1995.
Sandmeier, St., and K.I. Itten, 1996 16
Tab. 1
ground reference class # of pixels
deciduous stands 31'443
mixed stands 35'922
coniferous stands 20'513
wooded areas 6'860
settlement and urban areas 741
orchards, vineyards, horticultures 254
arable land, meadows and farm pastures 3'202
alpine agricultural areas 1'731
lakes and rivers 5'354
water shores, shore vegetation, wetlands 25
other unproductive areas 500
An original of this Table is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 17
Fig. 1
45°
90°
135°225°
315°90°
0°
270°
180°
Vd45°
An original of Figure 1 is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 18
Fig. 2a-d
An original of the color plate is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 19
Fig. 3a
0
4000
8000
12000
16000
20000
12 17 22 27 32 37 42
Digital Numbers in TM 2
freq
uen
cy
An original of Figure 3a is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 20
Fig. 3b
0
1000
2000
3000
4000
5000
6000
0 10 20 30 40 50 60 70 80 90 100
rel. reflectance in TM 2
freq
uen
cy
An original of Figure 3b is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 21
Fig. 3c
0
1000
2000
3000
4000
5000
6000
0 10 20 30 40 50 60 70 80 90 100
rel. reflectance in TM 2
freq
uen
cy
An original of Figure 3c is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 22
Fig. 3d
0
1000
2000
3000
4000
5000
6000
0 10 20 30 40 50 60 70 80 90 100
rel. reflectance in TM 2
freq
uen
cy
An original of Figure 3c is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 23
Fig. 4a
0
20
40
60
80
100
120
10 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Digital Numbers in raw TM 5
freq
uen
cy
arable land, meadowsand farm pastures
alpine agriculturalareas
An original of Figure 4a is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 24
Fig. 4b
0
20
40
60
80
100
120
30 40 50 60 70 80 90 100 110 120 130 140 150
rel. reflectance in atmo-corr. TM 5
freq
uen
cy
arable land, meadowsand farm pastures
alpine agriculturalareas
An original of Figure 4b is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 25
Fig. 4c
0
20
40
60
80
100
120
30 40 50 60 70 80 90 100 110 120 130 140 150
rel. reflectance in atmo-illu-corr. TM 5
freq
uen
cy
arable land, meadowsand farm pastures
alpine agriculturalareas
An original of Figure 4c is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 26
Fig. 5a
0.0
10.0
20.0
30.0
40.0
50.0
60.0
70.0
80.0
90.0
100.0
a b c d a b c d a b c d a b c d a b c d a b c d a b c d a b c d
accu
racy
[%
]
user accuracy producer accuracy
woodedareas
settle-ment
orchardsetc.
mea-dows etc.
alpine agri-cultural areas
lakes andrivers
watershores etc.
other un-productive
An original of Figure 5a is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 27
Fig. 5b
42
44
46
48
50
52
54
a b c d
ove
rall
accu
racy
[%
]
0
1
2
3
4
5
6
un
classified p
ixels [%]
overall accuracy unclassified pixels
An original of Figure 5b is attached and should be used for publication
Sandmeier, St., and K.I. Itten, 1996 28
FIGURE AND TABLE CAPTIONS
Fig. 1: Local horizons for a point i. The inset depicts an azimuth projection of the horizon
angles of the same initial point i. The surface within the plot line corresponds to the
amount of sky seen from i. In the background the integrated sky-view factors Vd
can be seen in test site "Beckenried". Low values of Vd correspond to low bright-
ness and vice versa.
Fig. 2: TM imagery of 11 July 1991, bands 1 (blue), 2 (green) and 3 (red). On the right
side zoom sections (a') - (d') of the lower right corners in (a) - (d) are depicted.
(a) and (a'): Radiometrically uncorrected.
(b) and (b'): Atmospherically corrected.
(c) and (c'): Corrected for atmospheric and illumination effects using the
trigonometric approach to determine the sky- and terrain-view
factors.
(d) and (d'): Corrected for atmospheric and illumination effects using the
horizon line approach to determine the sky- and terrain-view
factors.
Fig. 3: Histogram of band 2 in the Buochserhorn area, a section of test site "Beckenried"
with predominantly forest and alpine agricultural areas.
top left (a): Radiometrically uncorrected.
bottom left (b): Atmospherically corrected.
top right (c): Corrected for atmospheric and illumination effects using the
trigonometric approach.
bottom right (d): Corrected for atmospheric and illumination effects using the
horizon line approach.
Fig. 4: Histogram of class arable land, meadows and farm pastures (in background) and of
class alpine agricultural areas (in foreground). Test site "Beckenried", TM band 5.
top (a): Radiometrically uncorrected.
center (b): Atmospherically corrected.
bottom (c): Corrected for atmospheric and illumination effects using the
horizon line approach to determine the sky- and terrain-view
factors.
Sandmeier, St., and K.I. Itten, 1996 29
Fig. 5: Maximum likelihood classification. The indices (a) to (d) in the Figures correspond
to: raw (a), atmo-corrected (b), atmo-illu-corrected without horizon line (c), and
atmo-illu-corrected data with horizon line (d).
top (a): User and producer accuracies of eight Swiss Land Use
Statistics aggregates in test site "Beckenried". The
classification is performed using bands 1, 3, 5, and 7.
bottom (b): Overall accuracy and unclassified pixels of forest stands
(deciduous, mixed, and coniferous) in the Buochserhorn area
of "Beckenried". The classification is performed using bands
2, 4, 5, and 7.
Tab. 1: Overview on ground reference data used in the classification. The forest stand
classes are derived from forest stand maps. The land use classes are based on
sample points provided by the Swiss Land Use Statistics.
Sandmeier, St., and K.I. Itten, 1996 30
Stefan Sandmeier received the M.Sc from University of Zurich
in 1991 and the Ph.D. in 1995. He is Research Scientist and
Project Manager at the Remote Sensing Laboratories. His major
interest are in the field of radiometric corrections and field-
goniometry / BRDF-research.
Klaus I. Itten received the M.Sc. and Ph.D. degrees in
geography from the University of Zurich, Switzerland, in 1969
and 1973, respectively.
In 1974 he joined NASA/GSFC for one year as a Research
Fellow of the European Space Agency. Since 1982 he was
Assistant Professor and acts since 1988 as Full Professor in
Geography at the University of Zurich-Irchel, where he is head of
the Remote Sensing Applications Division of the Remote Sensing
Laboratories. His research and teaching interests are remote
sensing and image processing for natural resources inventorying
and monitoring.
K.I. Itten is currently president of the Swiss Remote
Sensing Commission, member of the Swiss Commission on
Space Research, member of the Federal Commission of Space
Affairs, and acts as Swiss Delegate to the ESA Program Board
Earth Observation (PBEO).