A PHYSICALLY-BASED MODEL TO CORRECT ATMOSPHERIC AND ILLUMINATION EFFECTS IN OPTICAL SATELLITE DATA OF RUGGED TERRAIN. St. Sandmeier and K.I.

Transcription

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,

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 ) 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 north and 8.52 east is at 33.6 zenith, and 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,

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 Topography 1. 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 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 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,

4 ρ(b) = π ( L(b) L p(b,z) ) E(b,z) T u (b,z) (1) where E(b,z) is the total solar irradiance reaching a surface on altitude z in a given band b; T u (b,z) is the upward transmission from the surface to the sensor; L p (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 atsatellite radiances L [mw cm -2 sr-1 µm-1 ] are calculated using the TM calibration constants a 0 (offset) and a 1 (gain) with L = a 0 + a 1 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 atsatellite 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 (H 2 O), carbon dioxide (CO 2 ), ozone (O 3 ), oxygen (O 2 ), methane (CH 4 ), nitrous oxide (N 2 O), 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 determined. 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,

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 L p (b,z) and T u (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 irradiance 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 E d h (b,z) cos(i) cos(sz) + (direct irradiance) E h f (b,z) k(b,z) cos(i) cos(sz) + (1 k(b,z)) V d + (diffuse irradiance) E h (b,z) V t ρ adj (terrain irradiance) (3) where: E(b,z) = total irradiance on an inclined surface E h (b,z) = total irradiance on a horizontal surface E h d (b,z) = direct component of irradiance on a horizontal surface E h f (b,z) = diffuse component of irradiance on a horizontal surface k(b,z) = anisotropy index V d = sky-view factor V t = terrain-view factor ρ adj = 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 E h d (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,

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, E h f (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 E n d (b,z) and the top of the atmosphere radiance E t d (b): k(b,z) = E d n (b,z) E d t (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 E h d (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 V d defined as the ratio of the sky portion seen from a specific surface to that on an unobstructed horizontal surface, i.e. 0 < V d 1. The total diffuse irradiance can therefore be calculated for tilted surfaces by E f (b,z) = E h f (b,z) k(b,z) cos(i) cos(sz) + ( 1 k(b,z) ) V d (5) anisotropic portion isotropic portion where E h f (b,z) and E f (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 V t, (3) the surface reflectances of the adjacent objects ρ adj 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,

7 covered rugged terrain. In shadowed areas, however, this effect cannot be neglected even for dark objects as V t is large and E d (b,z) and E f (b,z) are small [10], [12]. Two approaches are implemented and tested to obtain V t and V d, (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 V t and V d 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 V d and V t turn to 0.5: V d = 1 + cos() s 2 V t = 1 cos() s 2 (6) (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 V d and V t 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 h 0 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. V d and V t are then obtained by the integration of h over the azimuth circle: 2π π /2 V d = cos(γ) 0 h 0 cos( h[ θ] )dhdθ (8) 2π h 0 V t = cos(γ) cos( h [ θ ] )dhdθ (9) 0 0 where: V d = sky-view factor V t = terrain-view factor γ = angle between normal on surface and vector with elevation h θ = azimuth angle h = elevation angle h 0 = 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' '250. The inset shows the azimuth projection of the corresponding horizon Sandmeier, St., and K.I. Itten,

8 angles h 0 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 V d. 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,

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 atmospheric (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' '550 upper left and 677' '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 agriculture. 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,

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), atmocorrected (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,

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,

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 altitudes: 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 physicallybased 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,

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 and by the Swiss National Science Foundation, grant no 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,

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, [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 , [3] K.I. Itten and P. Meyer, "Geometric and Radiometric Correction of TM-Data of Mountainous Forested Areas," IEEE Trans. Geosci. Remote Sensing, vol. 31, no. 4, pp , [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 , Dept. of Geography, University of Zurich, p. 51, [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, [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 , [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 , [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 , [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, [10] C. Proy, D. Tanré, and P.Y. Deschamps, "Evaluation of Topographic Effects in Remotely Sensed Data," Remote Sensing Envir., vol. 30, pp , [11] J.E. Hay, "Solar energy system design: the impact of mesoscale variations in solar radiation," Atmos. Ocean, vol. 21, pp , [12] J. Dozier and D. Marks, "Snow mapping and classification from Landsat Thematic Mapper data," Annals of Glaciology, vol. 9, pp , [13] K.Ya. Kondratyev, "Radiation in the Atmosphere," Academic Press, London, [14] J. Dozier, J. Bruno, and P. Downey, "A faster solution to the horizon problem," Computers and Geosciences, vol. 7, pp , [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 , Sandmeier, St., and K.I. Itten,

15 [16] R.G. Congalton, "A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data," Remote Sensing Envir., vol. 37, pp , [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 , Sandmeier, St., and K.I. Itten,

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,

17 Fig V d An original of Figure 1 is attached and should be used for publication Sandmeier, St., and K.I. Itten,

18 Fig. 2a-d An original of the color plate is attached and should be used for publication Sandmeier, St., and K.I. Itten,

19 Fig. 3a frequency Digital Numbers in TM 2 An original of Figure 3a is attached and should be used for publication Sandmeier, St., and K.I. Itten,

20 Fig. 3b frequency rel. reflectance in TM 2 An original of Figure 3b is attached and should be used for publication Sandmeier, St., and K.I. Itten,

21 Fig. 3c frequency rel. reflectance in TM 2 An original of Figure 3c is attached and should be used for publication Sandmeier, St., and K.I. Itten,

22 Fig. 3d frequency rel. reflectance in TM 2 An original of Figure 3c is attached and should be used for publication Sandmeier, St., and K.I. Itten,

23 Fig. 4a An original of Figure 4a is attached and should be used for publication arable land, meadows and farm pastures alpine agricultural areas frequency Digital Numbers in raw TM 5 Sandmeier, St., and K.I. Itten,

24 Fig. 4b An original of Figure 4b is attached and should be used for publication arable land, meadows and farm pastures alpine agricultural areas frequency rel. reflectance in atmo-corr. TM 5 Sandmeier, St., and K.I. Itten,

25 Fig. 4c An original of Figure 4c is attached and should be used for publication arable land, meadows and farm pastures alpine agricultural areas frequency rel. reflectance in atmo-illu-corr. TM 5 Sandmeier, St., and K.I. Itten,

26 Fig. 5a wooded areas orchards etc. settlement meadows etc. alpine agricultural areas lakes and rivers water shores etc. other unproductive accuracy [%] abcd abcd abcd abcd abcd abcd abcd abcd user accuracy producer accuracy An original of Figure 5a is attached and should be used for publication Sandmeier, St., and K.I. Itten,

27 Fig. 5b overall accuracy [%] unclassified pixels [%] a b c d 0 overall accuracy unclassified pixels An original of Figure 5b is attached and should be used for publication Sandmeier, St., and K.I. Itten,

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 V d can be seen in test site "Beckenried". Low values of V d correspond to low brightness and vice versa. Fig. 2: Fig. 3: Fig. 4: 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. 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. 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,

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,

30 Stefan Sandmeier received the M.Sc from University of Zurich in 1991 and the Ph.D. in He is Research Scientist and Project Manager at the Remote Sensing Laboratories. His major interest are in the field of radiometric corrections and fieldgoniometry / 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). Sandmeier, St., and K.I. Itten,