15arspc Submission 223

8/8/2019 15arspc Submission 223

1/22

1

ASSESSMENT OF ADEQUACY OF DEM LEVEL-1 AS A

SUBSTITUTE FOR DEM LEVEL-2 FOR RADIOMETRICCORRECTION OF SATELLITE IMAGERY FOR NATIONAL LAND

COVER MAPPING OF SAUDI ARABIA

A. I. ALOMRAN and M. J. MCCULLAGH

General Commission for Survey, Riyadh, Saudi ArabiaPhone number: +966-555205816Email: [email protected]

School of Geography, University of Nottingham, Nottingham, UK.Phone number: +44 (0)1332-874103

Email: [email protected]

AbstractThe Kingdom of Saudi Arabia has not yet had a full (national) land covermapping. The very limited areas in Saudi Arabia covered by the 30m resolutionDEM level-2 that is compatible with the fine and medium resolution (SPOT andTM) imagery makes it unreliable as a topographic source. The validity of thelower spatial resolution (100m) DEM level-1 of full coverage of the Kingdom asa substitute is tested for that purpose. Radiometric (atmospheric andtopographic) correction of satellite imagery for the desert bare soil study area isachieved by using Radiance and two Reflectance (Chavez (1996) COST andRadiative Transfer Code in ATCOR-3) based techniques, implementing fourLambertian and non-Lambertian topographic correction models. The results

show that the adequacy of DEM level-1 decreases with increase of terrain slopeand DEM level-1 can be an adequate alternative to DEM level-2 for areas of flatto gently sloping (0 to 5). The optimum performance of DEM level-1 will be onsun-facing slopes in phase angles 46 to 90 and its worst performance in thesun-facing-away slopes in phase angles 136 to 180. Moreover, for betterperformance of topographic correction using DEM level-1 for high sun angleimages in flat to moderate terrain (0 to 25), C-correction should be used as theoptimum model, followed by Minnaert and last by Cosine. But for rugged terrain(steeper than 25), the Cosine should be used, followed by C-correction and lastby Minnaert. Using low sun angle imagery in flat to gentle terrain (0 to 5), C-correction is the optimum, Minnaert comes as second and Cosine is the last.For moderate to rugged terrain (steeper than 5), C-correction should be used,followed by Cosine and last by Minnaert. Results also demonstrates preferenceof using fine resolution and high sun angle imagery with DEM level-1.

1. Introduction

The Kingdom of Saudi Arabia (about two million square kilometres area) hasnot yet had a full (national) land cover mapping survey. To perform accurateland cover classification, satellite imagery must be radiometrically corrected foratmospheric and topographic effects. Correction for those effects requires anaccurate DEM that is compatible with the satellite image resolution, such that

derived slope and aspect will match image pixel in size and location. The very


2/22

2

limited areas in Saudi Arabia covered by the 30m resolution DEM level-2 that iscompatible with the fine and medium resolution (SPOT and TM) imageryreceived by King Abdulaziz City for Science and Technology (KACST) in Riyadhmakes it unreliable as a topographic source. DEM level-1 of full coverage of the

Kingdom as a substitute makes testing its validity for that purpose essential. IfDEM level-1 is substitutable then land cover mapping for the Kingdom usingDEM level-1 instead of level-2 for radiometric correction can be carried out onan immediate operational basis. To reach this goal, two questions are to beanswered. The first is to what extent in terms of terrain slope and orientation isDEM level-1 an adequate alternative to DEM level-2 in radiometric correctionand consequently in land cover mapping? The second is what is the optimumcombination of temporal (sun angle) selections, sensor images and radiometriccorrection techniques that leads to the best performance of DEM level-1? Thus,performance of DEM level-1 against level-2 is tested under the effect of thevariable image sun angles and resolutions: SPOT-XI (representing finer

resolutions and higher sun elevation angles), TM8-5 (representing coarserresolutions and higher sun angles) and TM16-1 (representing coarserresolutions and lower sun angles) acquired on 14-5-2001, 8-5-2001 and 16-1-2001, respectively, at sun zenith angles at the image acquisition times of 17.4,26 and 54, respectively, and at sun azimuth angles of 118.7, 102, and 145,respectively. For same purpose, finding the optimum techniques for radiometriccorrection of atmospheric and topographic effects is investigated by comparingthree different atmospheric correction based techniques: Radiance, Chavez(1996) COST simplified reflectance and ATCOR-3 accurate reflectancecompiled using Radiative Transfer Code (RTC), and four different simplifiedtopographic and anisotropic correction methods: Lambert (Cosine) and non-Lambert Minnaert, C-correction and Cicone and Malilas (1972) ModifiedLambert.

2. Radiometric Correction Techniques used in this Study

Before radiometric (atmospheric and topographic) corrections, all satelliteimages involved in this study were ortho-rectified. Slope and aspect valuesused for topographic correction were derived from the DEM level-1 and level-2using Erdas Imagine software (its algorithm is similar to the widely usedSharpnak and Akin (1969) and Evans (1980) algorithms). Programs werewritten by the authors to perform all the radiometric corrections of image data,

except for ATCOR3 Lambert correction. Minnaert and C topographic correctionof the satellite images were performed using single (global) K and Ccoefficients.

Radiance based radiometric correction

The atmospheric effect was corrected by removing path radiance from apparent(i.e., at satellite) radiance using the Chavez (1988) improved Dark ObjectSubtraction (DOS) method. The main disadvantages of the improved DOSmethod are its low accuracy and its correction only for additive scattering effects(due to path radiance) and the assumption of full upward and downwardtransmission.


3/22

3

The Lambertian (Cosine) topographic correction model has the followingformula:

)icos(

LL

T

n =

where cos(i) is the cosine of the incidence angle between the sun and thenormal vector to the surface, calculated as follows:

cos(i) = cos(s) cos(n) + sin(s ) sin(n) cos(n - s)

where iis illumination (incidence) angle; s is the sun zenith angle; n is terrainslope angle; n is terrain aspect angle; ands is the sun azimuth angle;

nL is the normalised radiance (i.e., corrected for topographic effect) that would

be measured when i= e (exitance angle) = 0; and TL is the radiance at tilted

surface (i.e., uncorrected).

Developed by Smith et al. (1980), the Backward Radiance CorrectionTransformation (BRCT) employing the Minnaert law (Minnaert, 1941) can bedescribed as:

)e(cos)i(cos

LL 1kk

T

n =

where kis the Minnaert coefficient, and e is terrain slope angle.

In similar fashion to Minnaert, the C-correction is used for topographic andanisotropic correction. The formula of this model is:

)C)icos(

C)cos((LL

s

Tn +

+=

where s is the sun zenith angle.

The Modified Lambertian model (Cicone and Malila, 1977) can be described as:

RI

T

n F

LL =

where RIF is the relative inolation factor and can be computed as follows:

)ecos()cos()esin()tan(F sRI +=

Where is the phase angle (difference between sun azimuth and terrainaspect).

Simplified reflectance based radiometric correction

This technique, which is supposedly correct for atmospheric effects moreaccurately than the previously discussed radiance based technique, isimplemented using the Chavez (1996) COST model, in which atmosphericdownward transmittance is approximated by the cosine of the sun zenith angle,and upward transmittance is assumed 1.0 (full transmittance).


4/22

4

The Lambertian topographically corrected surface reflectance based on Chavez(1996) technique is expressed as:

)cos()icos(E

LL(d(

so

ps

2

Lamb =

Where Lamb is the target surface reflectance under the assumption of surface

Lambertian behavior; pL is the path radiance for TM and SPOT images

computed with an assignment of one-percent (0.01) reflectance to the dark

features (Song et al., 2001); oE is the sun exo-atmospheric irradiance; d is the

relative earth-sun distance to the mean distance in astronomical units at the

image acquisition day; and )cos( s is the cosine of sun zenith angle

representing downward transmittance.

Similarly, the Minnaert correction model integrated into the Chavez (1996)COST reflectance model is:

k

so

pathsat

2

Minnaert )]ecos()i)[cos(cos(E

)ecos()LL(d = ;

and the C-correction reflectance formula is expressed as follows:

]C)i)[cos(cos(E

]C))[cos(LL(d

so

spathsat

2

C +

+=

The reflectance formula of the Modified Lambertian model will be similar to the

Lambertian formula, except that the incidence angle is replaced by RIF :

)cos(FE

LL(d(

sRIo

pathsat

2

mod =

Accurate reflectance based radiometric correction

The third technique to radiometric correction used in this study is assumed to bethe most accurate technique in atmospheric correction, for that it uses RadiativeTransfer Code (RTC). This technique is implemented using the ATCOR-3program for atmospheric (compiled using the MODTRAN-4) and topographic

correction of rugged terrain, developed by the DLR-German Aerospace Centreand integrated in Erdas Imagine by Geosystems Gmbh.

ATCOR-3 was used to estimate visibility, adjusted sensor calibrationcoefficients and aerosol type and standard atmosphere and to perform theatmospheric and topographic correction for the three satellite images used inthis study.

The three images were corrected for topographic effects under the Lambertianassumption using ATCOR-3 and with Minnaert model implemented for ATCOR-3 by the present authors using the formula:


5/22

5

)e(cos)i(cos

1k1k

Lamb

Minnaert =

where Lamb is the output Lambertian reflectance from ATCOR-3.

3. Study area

The study area is 10 km by 10 km, located in the central part of Saudi Arabia inan area between two small towns: Malham and Huraimla, 65 km to theNorthwest of Riyadh, the capital city. It is composed of desert bare soil of gentleslope on top of plateau ruptured by areas of rock protrusions in steep to semi-orthogonal rock outcrops, and torn by wide waterways (wadis) and small gullies(figure 1). This study area was chosen to represent the overwhelming majorityof the Kingdom's terrains covered with same desert bare soil.

4. Results and Analysis

Visual results and analysis

DEM level-1 corrected images using the four topographic correction modelshave shown less shadow reduction and more artifacts compared to those ofDEM level-2. This finding is applicable to both coarse resolution (e.g. TM) andfiner resolution (e.g. SPOT). Figure 2, as an example, explains this finding bycomparing DEM level-2 corrected TM images with those corrected using DEMlevel-1. Shadow effect and artifact are more pronounced in the DEM level-1corrected images, due to the coarser resolution of level-1 DEM and its failure todetect wadis (valleys) short sides. This indicates initial less support to the ideaof DEM level-1 being an adequate alternative to DEM level-2.

Semi-variogram results and analysis

Researchers (e.g. Bishop and Shroder, 2000; Bishop et al., 2003) haveconducted semi-variogram analysis to measure topographic complexity andtopographic correction. As a part of the process of evaluating the performanceof DEM level-1 against level-2 for topographic correction, semi-variogramanalysis was conducted on the two DEMs corrected radiance for a selectedtransect of a length of about 1km (30 TM pixels) crossing a 500m wide wadi inthe study area.

The semi-variance, (h), measures the variance between points at successively

greater distances. Pixels (point values) at small lags will have lower semi-variance than at greater lags. Smi-variance can be calculated as (Treitz andHowarth, 2000; Bishop et al., 2003):

2ii

)h(m

1i

)]hx(Z)x(Z[)h(m2

1)h( +=

=

where )h( is the average semi-variance of several pixel pairs at each lag;

Z( ix ) is the value of the variable to be tested (e.g., DN, radiance) at position ix ;

Z( ix +h) is the value of the variable at lag distance h from ix ; and m (h) is the

number of data pairs separated by same lag h.


6/22

6

Figure 3 illustrates the resulting semi-variograms for the Green bands of TM8-5and TM16-1 uncorrected and corrected (with C model) radiance using the twoDEMs. The substantial decrease of semi-variance in both DEMs correctedradiance of TM8-5 compared with those of uncorrected radiance and the small

differences in semi-variance between DEM level-1 and level-2 correctedradiance indicates the effectiveness of either DEMs and the adequacy of DEMlevel-1 as an alternative to level-2 for topographic correction of high sun angleimages. However, the higher semi-variance for the TM16-1 corrected radianceusing either DEMs, especially DEM level-1 compared with those for the TM8-5indicates the low efficiency of topographic correction for low sun angle images,and consequently a preference to use high sun angle images instead. Incontrast with the high sun angle image, the considerable differences in semi-variance between DEM level-1 and level-2 corrected radiance for the low sunangle TM16-1 image and at the same time the close semi-variance values ofuncorrected radiance to those of DEM level-1 corrected radiance indicates the

inadequacy of DEM level-1 as an alternative to level-2 for topographiccorrection of low sun angle images.

Effect of terrain slope and phase angle on performance of DEM level-1against level-2 for radiometric correction

Number of studies (e.g. Cicone and Malila, 1977; Stohr and West, 1985;Thomson and Jones, 1990) have demonstrated the effect of terrain slope andaspect and phase angle (the difference between terrain and sun azimuths, onsatellite image data, and others (e.g. Smith et al. 1980; Justice et al., 1981;Chen et al., 2001; Falkenstrum and Ekstrand, 2002) consequently ontopographic and anisotropic corrections using techniques such as, Cosine,Minnaert, C-Correction, etc.

As the effect of terrain aspect will vary from one image to another according tothe image sun azimuth angle, the effect of phase is investigated instead in thisstudy.

Most earth surfaces reflects unequally in all directions (anisotropy). Surfaceanisotropy (BRDF effects), generally speaking, is strongly manifested inbackward and forward scattering, with the maximum in backward scattering.Considering desert areas, however, some studies (e.g. Holben and Justice,1980; Takemata et al., 2000) have suggested Lambertian behaviour whereas

others (e.g. Shoshany, 1993; Karnieli and Cierniewski, 2001) have suggestedanisotropic (BRDF) effects. Hence, correcting for topography and BRDF effectsin this study has involved Lambertian and simple empirical BRDF (non-Lambertian) models, such as Minnaert and the semi-empirical C-correction.

Owing to the major effects of terrain slope and phase angle on satellite imagedata and topographic correction mentioned above, performance of DEM level-1against level-2 in topographic correction has been investigated thoroughly inthis study as described below. This investigation includes the effect of terrainslope and phase angle in determining terrain slope and phase angle (aspect)limitations, optimum topographic correction models, optimum sun elevationangles and resolutions, and optimum radiometric (atmospheric) correction

techniques for optimum performance of DEM level-1 compared to level-2.


7/22

7

4.3.1 Effect of terrain slope on performance of DEM level-1 against level-2for radiometric correction

4.3.1.1 Effect of terrain slope on correlation between the two DEMs

radiometrically corrected dataPerformance of DEM level-1 against level-2 for radiometric (atmospheric andtopographic) correction is measured here by the degree of correlation betweenDEM level-1 and level-2 corrected radiance images. Higher correlation indicatesa greater likelihood of DEM level-1 to replace level-2. Hence, only the highcorrelation values (r>0.5) were considered. For that purpose, four monochromic(0-255; representing correlation value range of 0-1.0) images of correlationbetween DEM level-1 and level-2 corrected (using Minnaert, C, ModifiedLambertian and Cosine models) TM8-5 NIR were produced using MIPSsoftware (Mather, 1999). To investigate the effect of terrain slope onperformance of DEM level-1 against level-2, the study area slopes were

separated into four classes: 0-5 (flat to gentle terrain), 6-15 (moderateterrain), 16-25 (steep terrain) and 25+ (very steep terrain). DEM level-2 slopeimage was used as a source for its higher accuracy compared with that of level-1. For that purpose, Spatial Analyst Model Builder of ArcView was used toperform the slope and correlation images classifications and the arithmeticoverlay analysis. Figure 4 illustrates the results. Performance of DEM level-1against level-2 is expressed by what we called " Relative Area CoveredPercentage-RACP", which is the number of pixels of high correlation (i.e., r>0.5or grey values of 127 to 255) divided by the total number of pixels in that slopeclass. This is to provide unbiased results to slope classes occupying largerarea. Higher RACP indicates higher performance of DEM level-1 against level-2. As shown in figure 4, sharp decrease in RACP starts after the first terrainslope class 0-5 (a drop of about 40%), then decreases asymptotically withincrease of slope. This indicates that adequacy of DEM level-1 decreases withincrease in terrain slope and DEM level-1 can not be an adequate alternative toDEM level-2 for topographic correction of areas with terrain slope higher than5. Figure 4 also shows that the differences between the four topographiccorrection models are unsubstantial, except possibly for the ModifiedLambertian model for slopes higher than 5.

4.3.1.2 Effect of terrain slope on classification accuracy using the twoDEMs radiometrically corrected data

The effectiveness of radiometric correction in improving classification accuracyis still arguable in literature. Some researchers (e.g. Itten and Meyer 1993;Tokola et al., 2001; Gitas and Devereux, 2006) have reported noticeablesuccess in improving classification by radiometric correction, others (e.g. Teilletet. al., 1982; Song et al., 2001), however, have reported otherwise.

The DEM level-1 and level-2 radiometrically corrected images wereunsupervised classified (using Erdas ISODATA) into six spectral classes. Thisnumber of classes was chosen arbitrarily, owing to the limited number of desertbare soil types of study area and to that no limitation for the number of spectralclasses that can be derived spectrally using unsupervised classification.


8/22

8

The effect of terrain slope on performance of using DEM level-1 instead of level-2 in classification is investigated for low and high sun elevation angles (TM8-5and TM16-1, respectively), for the three level of atmospheric correctionaccuracies (Radiance, Chavez (1996) and ATCOR-3 techniques) and for the

four topographic correction models. Each of DEM level-1 and level-2 classifiedimages was separated into four images based on the four slope classes (0-5,6-15, 16-25 and >25) using programs written by the authors. Otherprograms were also developed to calculate the Khat values of the errormatrices, produced from comparing DEM level-1 classified images with those oflevel-2, as a reference, for the four slope classes. Classified DEM level-2radiometrically corrected images were used in the error matrices as referencefor the DEM level-1 corrected images due to the absence of ground truth (i.e.,reference data) and to the fact that the assessment of DEM level-1 performanceis relative to that of level-2. Khat (estimate of Kappa) is adopted against OveralAccuracy due to the fact that it is more informative, as it incorporates omission

and commission errors (Congalton and Green, 1999).

Performance of DEM level-1 against level-2 is evaluated through therelationship between classification accuracy (Khat values) of DEM level-1images (compared with those of level-2 as a reference) and terrain slope.Figures 5(a), (b) and (c) illustrate the relationship of Khat values for the fourtopographic correction models with terrain slope for Radiance, Chavez (1996)and ATCOR-3, respectively. A similar general trend to that found in thepreviously discussed correlation test (section 4.3.1.1) is also found here. Therate of decrease in Khat values for the high sun angle images TM 8-5 andSPOT-XI is rapid from about 80% for slope class 0-5 to about 20% for slope

class >25. However, the rapid decrease slows by the slope class 6-15 for thelow sun angle image (TM16-1), then decreases moderately until slope class>25. The continuous sharp drop of Khat values (classification accuracy) afterslope class 0-5 in the high sun angle images and the low Khat values in thelow sun angle image beyond slopes 0-5 indicates limitation of DEM level-1adequacy to replace DEM level-2 to flat to gentle terrain (0-5) only.

In figure 5 and based on the high sun angle SPOT-XI image using the Radiancetechnique, with the exception of Modified Lambertian model, C is the optimummodel at all slope classes, followed by Minnaert and last by Cosine. Similarresults apply for the other high sun angle TM8-5 image, but Cosine replaces Cas optimum at the slope class (>25). The higher performance of the ModifiedLambertian model compared with other models in high sun angle images and itscomparable performance with other models in low sun angle image TM16-1indicates to its sensitivity to sun angle rather than to topographic effect (slopeand aspect), which lowers its reliability. The Chavez (1996) technique showssimilar results for SPOT-XI, except Cosine replacing C as optimum at the slopeclass (>25). TM 8-5 is similar to SPOT-XI, except that C is the optimum only upto 15, whereas Cosine is the optimum above 15. For ATCOR-3 where onlyMinnaert and Cosine models are used, model ranking is similar to that ofChavez (1996), such that Minnaert is the optimum up to 25 and Cosine is theoptimum above. For the low sun angle image (TM16-1), and based on

Radiance and Chavez (1996) results, C is the optimum model, followed by


9/22

9

Minnaert and least by Cosine in the flat to gentle terrain (0 to 5). In moderateto rugged terrain (i.e., slopes higher than 5), C is optimum, followed by Cosineand last by Minnaert. Using ATCOR-3, model ranking is similar to those forRadiance and Chavez (1996), such that Minnaert is the optimum below 5 and

Cosine above 5. Thus, for optimum performance of DEM level-1, with fewexceptions, if high sun angle images are used in flat to moderate terrain (slopesof 0 to 25) using any radiometric (atmospheric) correction technique, oneshould use C as the optimum model, followed by Minnaert and last by Cosine.However, if the terrain is rugged (slopes higher than 25), one should useCosine, followed by C and last by Minnaert. For using low sun angle imagery inflat to gentle terrain (slope of 0 to 5), C is the optimum, Minnaert comes assecond and Cosine is the last. For moderate to rugged terrain (slopes higherthan 5), one should use C, followed by Cosine and last by Minnaert, whoseperformance with DEM level-1 deteriorates with the increase of terrain slope.

For optimum sun angle, figure 5 shows that the performance of DEM level-1against level-2 in TM8-5 is better (higher Khat values) than in TM16-1 for flat tomoderate slopes (i.e., 0-15), then the differences decrease with increase ofslope until for slopes higher than 25 where performance in TM16-1 is betterthan that for TM8-5 using C and Cosine models in Radiance technique (figure5a) and C in Chavez (1996) technique (figure 5b). The better performance ofDEM level-1 for TM16-1 in this slope class using Minnaert and Cosine inATCOR-3 confirms this (figure 5c). Thus, as found before, for betterperformance of DEM level-1 with any topographic correction model using anyradiometric (atmospheric) correction technique, high sun angle images shouldbe used instead of low sun angle images. However and oddly, low sun angle

images are better used in areas with terrain slopes higher than 25. The reasonis not clear to the authors and worthy future investigation.

For optimum image resolution, figure 6 shows that the performance of DEMlevel-1 in association with the fine resolution SPOT-XI image is better than thatwith coarser resolution TM8-5 image, except when using the Cosine model forslopes between 16 and 25 for Chavez (1996) and ATCOR-3 techniques. Thisexception can be ignored owing to the insignificant difference between the twoimages (Khat = 23.0% compared with 23.1% for Chavez (1996)) and to themany findings of weakness of the Cosine with DEM level-1. Hence, it can besaid that for better performance of DEM level-1, one should use finer resolution(e.g. SPOT-XI) instead of coarser resolution (e.g. TM8-5) images for all terraintypes. Differences in DEM level-1 performance (i.e., Khat values) between thetwo images, however, get smaller with the increase of terrain slope.

For optimum radiometric correction technique, figure 6 illustrates that thedifferences in DEM level-1 performance between the three radiometriccorrection techniques for the high sun angle image TM8-5 decrease withincrease of slope until slopes higher than 25, after which the performances arealmost identical. For the low sun angle image TM16-1, the differences increasecontinuously with the increase of slope for the favour of ATCOR-3. The reasonfor this for the hazy (visibility of 12km) TM8-5 may be the increase ofdominance of atmospheric effect compared to topographic effect with the

decrease of slope in flatter slopes, which lead to higher differences in


10/22

10

atmospheric correction between the three techniques for the favour of Radianceand Chavez (1996) compared with ATCOR-3. The dominance of topographiceffect in steeper slopes makes the differences smaller due to the use ofmarginally similar topographic correction methods by the three radiometric

correction techniques. For the clearer sky (visibility of 25km) TM16-1, theexplanation may be the accurate estimation of terrain contribution in radiationby ATCOR-3 (e.g. adjacency effect and terrain radiation) in the absence of orvery slight atmospheric effect in a clear sky compared with other two techniquesthat do not consider them. From the above, it can be stated that more accurateatmospheric correction using Radiative Transfer Code-RTC (ATCOR-3) will notadd substantial improvement in performance of DEM level-1 for rugged terrain(slopes higher than 25) for high sun angle high atmospheric effect images (likeTM-8-5 image). Despite that, the accurate atmospheric correction techniqueATCOR-3, which demonstrates the true but the weaker performance of DEMlevel-1, is preferred to be used, owing to that the apparent better DEM level-1

performances in the Radiance and Chavez (1996) are due to their pooreratmospheric effect removal and lower topographic effect influence. Theperformance of DEM lvel-1 compared with level-2 for all three techniques for theclear sky low sun elevation angle TM16-1 image is almost identical for flat togentle terrain (slopes of 0 to 5), and as slope increases, its performance usingATCOR-3 improves compared with the other two techniques. Moreover and forsame image, performance of DEM level-1 using Radiance and ATCOR-3techniques is considerably better than that for Chavez (1996).

Despite its higher accuracy and fidelity, using the RTC based ATCOR-3atmospheric correction technique for operational use for the anticipated national

land cover mapping of the Kingdom is questionable, considering theuncertainties involved in estimation of the atmospheric parameters (especiallyfor aerosols), the cost and practicality of collecting required information aboutatmospheric condition for uninhabited areas.

4.3.2 Effect of phase angle on performance of DEM level-1 against level-2for radiometric correction

4.3.2.1 Effect of phase angle on correlation between the two DEMsradiometrically corrected data

The effect of phase angle on the high correlation (i.e., r > 0.5) between DEM

level-1 and level-2 corrected (using "C" model) radiance of the NIR band of thehigh sun angle TM8-5 image data was investigated in a similar fashion to thatmade for terrain slope effect (section 4.3.1.1).

In figure 7, RACP values of TM8-5 for all four topographic correction modelsstart low, increase to their highest values (optimum performance of DEM level-1) in phase class 46-90, then they decrease with slopes facing away from thesun to lowest values in phase class 136-180. Thus, the optimum performanceof DEM level-1 (highest RACP) for high sun angle images is expected to befound in pixels located on sun-facing slope with phase angles 46-90, whereasworst performance will be in sun-facing-away slopes at 136-180. Similar to


11/22

11

terrain slope results, the differences between the four topographic correctionmodels are unsubstantial (figure 7).

4.3.2.2 Effect of phase angle on classification accuracy using the two

DEMs radiometrically corrected dataA similar procedure was used to that implemented for assessment of terrainslope effect on classification accuracy of classified images (section 4.3.1.2),except that phase angle was investigated instead of slope.

In figures 8(a), (b) and (c), relation of classification accuracy (Khat values) withphase angle for the three radiometric correction techniques follows an identicalpattern to that found in previous experiment where Khat increases with increaseof phase angle until the optimum phase class 46-90, then decreases with theincrease of phase angle to lowest values in phase angles 136-180. The onlyexception is when correcting SPOT-XI with Minnaert where Khat values at

phase class 91-135 are just little higher than those at 46-90 class, but canbe ignored. Thus, Optimum performance of DEM level-1 for classification isexpected with phase angles 46 to 90.

For optimum topographic correction model, figure 8(a) based on SPOT-XI usingRadiance technique shows that C is the optimum model at all phase angleclasses, followed by Minnaert and last by Cosine, except at 0-45 where itperformed equally to C and Minnaert. Performance of Cosine worsens asterrain faces away from the sun. Similar results apply for TM8-5. The higherperformance of the Modified Lambertian model compared with other models, asdiscussed earlier, is due to its sensitivity to sun angle rather than totopographic effect. For the low sun angle image (TM16-1), C is also theoptimum, followed by Minnaert, Modified Lambertian and last by Cosine.Modified Lambertian exceptionally performed better than C at 0-45. TheChavez (1996) technique in figure 8(b) shows similar results, except that theModified Lambertian model performed equally to C at 46-90. For ATCOR-3where only Minnaert and Cosine models are used, model ranking is similar toprevious results where Minnaert performed better than the Cosine, except withTM16-1 at 0-45. Thus, same conclusion derived in the terrain slope resultscan be adopted here where by using any radiometric correction technique andany image resolution and sun angle, one should use C as the optimum model,followed by Minnaert and last by Cosine.

For optimum sun angle, figure 8 shows that differences in Khat values betweenTM 8-5 and TM16-1 images are significant, which strongly recommend the usehigh sun angle images for better performance of DEM level-1 data.

For optimum radiometric correction technique, figure 9 shows that the Khatvalues of DEM level-1 for the high sun angle images using the Radiancetechnique are highest, followed by Chavez (1996), and last by ACTOR-3. Onceagain as discussed earlier, this should not mislead one to the extent of relyingon the two simplified techniques, as the more accurate atmospheric effectremoval using ATCOR-3 reveals the poor performance of DEM level-1.Therefore, If simplified atmospheric correction techniques are used with high

sun angle images containing high atmospheric effects (e.g. TM8-5 and SPOT-


12/22

12

XI), the performance of DEM level-1 using the Radiance technique will bemarginally better than Chavez (1996). However, if ATCOR-3 is used to removethe atmospheric effects more accurately, substantially lower performance ofDEM level-1 compared with the other two techniques is expected and the

adequacy of DEM level-1 to replace level-2 is questionable.Differences in Khat values in figure 9 between the three techniques for the lowsun angle image TM16-1 are small, especially in sun-facing-away slopes (lessthan 5%). The small differences here compared with those found previouslywith TM8-5 for instance would be due to the atmospheric effect for TM 8-5image was greater than that for TM 16-1 image (visibility of 12 km and 25 km,respectively), giving less atmospheric correction for TM16-1, especially withATCOR-3. Radiance and ATCOR-3 techniques using Minnaert have producedalmost identical Khat values, but these are higher than those for Chavez (1996),especially in the first three phase classes. No clear trend has been found using

Cosine model. Thus, if correcting low sun angle low atmospheric effect images(e.g. TM16-1), performance of DEM level-1 using one of the three techniqueswill not add major improvement compared with the others. Moreover,performance of DEM level-1 using Minnaert using the Radiance or ATCOR-3techniques will be better than using Chavez (1996).

results. (b) TM8-5 results. (c) TM16-1 results.

5. Conclusions

The validity of using the 100m resolution DEM level-1 of full coverage as an aadequate alternative to the 30m resolution DEM level-2 of very limited coveragefor radiometric correction of satellite imagery and consequently in land cover

mapping of the Kingdom of Saudi Arabia has been investigated in this study.Visual results have revealed the inadequacy of DEM level-1 to act as analternative to level-2, but semi-variogram results have shown that DEM level-1may be used for high sun elevation angle imagery only. Results based on theeffects of terrain slope and phase angle on performance of DEM level-1 againstlevel-2 for radiometric correction, which have been adopted for their morereliably as they have thoroughly considered the effects of the sun elevation andazimuth angles, terrain variables (slope and aspect) and image resolution, haveshown that that the adequacy of DEM level-1 as an alternative to DEM level-2for radiometric correction decreases with the increase of terrain slope for bothhigh and low sun elevation angle imagery. The results have also shown thatDEM level-1 can be an adequate alternative to level-2 for areas of flat to gentlysloping (slopes of 0 to 5). The optimum performance of DEM level-1 will be onsun-facing slopes in phase angles 46 to 90, and its worst performance in thesun-facing-away slopes in phase angles 136 to180.

The analysis of effects of terrain slope and phase angle on performance of DEMlevel-1 against level-2 have revealed insignificant differences in DEM level-1performance between the four topographic correction models and shown that ifhigh sun elevation angle images are used in flat to moderate terrain (slopes of0 to 25) using any atmospheric correction technique (Radiance, Chavez(1996) COST simplified reflectance and ATCOR-3 accurate reflectance

compiled using Radiative Transfer Code-RTC), one should use C as the


13/22

13

optimum model, followed by Minnaert and last by Cosine. But for rugged terrain(slopes higher than 25), one should use Cosine, followed by C and last byMinnaert. Using low sun elevation angle imagery in flat to gentle terrain (slopesof 0 to 5), C is the optimum, Minnaert comes as second and Cosine is the last.

For moderate to rugged terrain (slopes higher than 5), one should use C,followed by Cosine and last by Minnaert whose performance with DEM level-1has been found to deteriorate with the increase of terrain steepness. TheModified Lambertian model has been found more sensitive to sun elevationangle rather than to topography (i.e., DEM information), which lowers itsreliability.

Results have also demonstrated the low efficiency in topographic correction oflow sun elevation angle imagery, and high sun elevation angle imagery shouldbe used for better DEM level-1 topographic correction. Oddly, performance ofDEM level-1 with low sun elevation angle imagery for terrain slopes higher than

25 has been found better than that with high sun elevation angle imagery. Thereason is not known to the authors and worthy future investigation. In addition,for better DEM level-1 topographic correction, finer resolution imagery (e.g.SPOT-XS) should be used instead of coarser resolution (e.g. TM).

Use of accurate atmospheric correction implementing the Radiative TtransferCode in ATCOR-3 is recommended compared with the simplified modelsemployed in the Radiance and Chavez (1996) COST reflectance techniques, forits greater fidelity in revealing the actual but weaker performance of DEM level-1 compared with DEM level-2. However, for operational use for the anticipatednational land cover mapping of the Kingdom, the worth of using accurateatmospheric correction is questionable, considering the uncertainties involved in

estimation of the atmospheric parameters (especially for aerosols), the cost andpracticality of collecting required information about atmospheric condition foruninhabited areas. In case the use of the accurate atmospheric correctiontechnique in ATCOR-3 is not feasible, Radiance technique has more potentialitycompared with Chavez (1996) COST reflectance for better performance of DEMlevel-1 radiometric correction.

The authors here see the importance of performing these types of investigationson the new generation of high resolution satellite imagery, such as SPOT-5,IKONOS, GeoEye, WorldView, etc., and see if this study derived conclusionscan still be generalised and adopted.

Acknowledgments

The authors wish to acknowledge the support from the General Commission forSurvey for provision of data and staff and equipment necessary for the fieldwork. The acknowledgment is also extended to King Abdulaziz for Sciences andTechnology who provided the satellite imagery used in this study.

Diagrams:


14/22

14

Figure 1. The study area. (a) A picture of the desert bare soil covering the study area.

(b) A picture of one side of one of the waterways (Wadis) tearing the study areaplateau(same small Jeep in the waterway bed is illustrated for scale purpose.

Figure 2. Shadow effect reduction and artifact appearance after topographic correctionusing DEM level-1 and level-2. (a) the uncorrected low sun angle image (TM16-1) of

the study area. (b) and (c) same image corrected with Minnaert model using DEM

level-1 and level-2, respectively. (d) a subset (the red square in "a") of the uncorrectedhigh sun angle image (TM8-5). (e) and (f) same subset image corrected with "C" model

using DEM level-1 and level-2, respectively.

(a) (b)


15/22

15

Figure 3. Semi-variograms of the TM8-5 and TM16-1 uncorrected and corrected

radiance with the two DEMs, calculated at lag distance increments of TM pixel size(30m).

Figure 4. Effect of terrain slope on the high correlation (r>0.5) between DEM level-1and level-2 corrected radiance of NIR of TM8-5 for the four topographic correctionmodels. Performance of DEM level-1 against level-2 represented by Relative Area

Covered Percentage-RACP of high correlation values in the four slope classes.


16/22

16

Figure 5. Effect of terrain slope on classification accuracy of DEM level-1 classifiedradiometrically corrected high (TM8-5 and SPOT XI) and low (TM16-1) sun elevationangle images (with reference to their corresponding DEM level-2 classified images)

with the four topographic correction models using the three radiometric correctiontechniques. (a) Radiance results. (b) Chavez (1996) results. (c) ATCOR-3 results.


17/22

17

Figure 6. Performance of DEM level-1 in classification (compared with DEM level-2 asa reference) under the effect of terrain slope using the three radiometric correction

techniques. Only Minnaert and cosine corrections are included, as they are the only

two corrections implemented in ATCOR-3. (a) SPOT-XI results. (b) TM8-5 results. (c)TM16-1 results.


18/22

18

Figure 7. Effect of phase angle on the high correlation (r>0.5) between DEM level-1and level-2 corrected radiance of NIR band of TM8-5 for the four topographic correction

models. Higher RACP value indicates higher performance of DEM level-1.


19/22

19

Figure 8 Effect of phase angle on classification accuracy of DEM level-1 classifiedradiometrically corrected high (TM8-5 and SPOT XI) and low (TM16-1) sun angle

images using the four correction models using the three radiometric correction

techniqueses with reference to their corresponding level-2 classified images.Classification accuracy is represented by Khat values. (a) Radiance based results. (b)

Chavez (1996) based results. (c) ATCOR-3 based results.


20/22

20

Figure 9 Performance of DEM level-1 in classification (compared with DEM level-2 as areference) under the effect phase angle using the three radiometric correction

techniques. Only Minnaert and cosine corrections are included, as they are the onlytwo corrections implemented in ATCOR-3. (a) SPOT-XI results. (b) TM8-5 results. (c)

TM16-1 results.

References

Bishop, M.P., and Shroder Jr ,J.F, 2000, Remote sensing and geomorphometric

assessment of topographic complexity and erosion dynamics in the Nanga


21/22

21

Parbat massif. In: Khan, M.A., Treloar, P.J., Searle, M.P. and Jan, M.Q. (eds.)Tectonics of the Nanga Parbat Syntaxis and the Western Himalaya. GeologicalSociety, London, 181-200.

Bishop, M.P., Shroder Jr., J.F., and Colby, J. D., 2003, Remote sensing andgeomorphometry for studying relief production in high mountains.Geomorphology, 55:345-361.

Chavez, P.S., 1988, An improved dark-object subtraction technique foratmospheric scattering correction of multispectral data. Remote Sensing ofEnvironment, 24:459-479.

Chavez, P., 1996, Image-based atmospheric corrections-revisited andimproved. Photogrammetric Engineering and Remote Sensing, 62(9):1025-1036.

Chen, F., Muramoto, K., and Kubo, M., 2001, Improved topographic correction

for satellite imagery. IEICE Transactions on Information and Systems, E84-D12:1820-1827.

Cicone, R., and Malila, W., 1977, Investigation of techniques for inventoryingforested regions. Vol. II. Forestry Information System Requirements and JointUse of Remotely Sensed and Ancillary Data. Final Report NASA-CR-ERIM-122700-35-F2.

Congalton, R.G., and Green, K., 1999, Assessing the Accuracy of RemotelySensed Data: Principles and Practices. Lewis Publishers, Boca Raton, 131p.

Evans, I.S., 1980, An integrated system for terrain analysis for slope mapping.Zeitschrift fur Geomorphologie, 36:274-295.

Falkenstrom, H., and Ekstrand, S., 2002, Evaluation of IRS-1C LISS-3 satellitedata for defoliation assessment on Norway spruce and Scots pine. RemoteSensing of Environment, 82:208-223.

Gitas I.Z., and Devereux, B.J., 2006, The role of topographic correction inmapping recently burned Mediterranean forest areas from LANDSAT TMimages. International Journal of Remote Sensing, 27(1):41 - 54.

Holben, B., and Justice, C.O., 1980, The topographic effect on spectralresponse from nadir-pointing sensors. Photogrammetric Engineering andRemote Sensing, 46(9):1191-1200.

Itten, K.I, and Meyer, P., 1993, Geometric and radiometric correction of TM dataof mountainous forested areas. IEEE Transactions on Geoscience and RemoteSensing, 31(4):764-770.

Justice, C.O., Wharton, S.W., and Hoblen, B.N., 1981, Application of digitalterrain data to quantify and reduce the topographic effect of Landsat.International Journal of Remote Sensing, 2:213-230.

Karnieli A., and Cierniewski, J., 2001, Inferring the roughness of desert rockysurfaces from their bidirectional reflectance data. Advances in Space Research,28(1):171-176.


22/22

Mather, P., 1999, Computer Processing of Remotely Sensed Images. Wiley,Chichester, 292p.

Minnaert, M., 1941, The reciprocity principle in lunar photometry. AstrophysicalJournal, 93(2):403-410.

Sharpnak, D., and Akin, G., 1969, An algorithm for computing slope and aspectfrom elevations. Photogrammetric Engineering and Remote Sensing, 35:247-248.

Shoshany, M., 1993, Roughness-Reflectance relationship of bare desert soilterrain: An empirical study. Remote Sensing of Environment, 45:15-27.

Smith, J. A., Lin, T. L. and Ranson, K. J., 1980. The Lambertian assumptionand Landsat data. Photogrammetric Engineering and Remote Sensing, 46, pp.1183-1189.

Song, C., Woodcock, C., Seto, C., Lenney, M., and Macomber, S., 2001,Classification and change detection using Landsat TM data: When and how tocorrect atmospheric effects? Remote Sensing of Environment, 75(2):49-58.

Stohr, C.J., and West, T.R., 1985, Terrain and look angle effects uponmultispectral scanner response. Photogrammetric Engineering and RemoteSensing, 51(2):229-235.

Takemata, K., Izumiya, T., and Kawata Y., 2000, Analysis of ADEOS/ POLDERdata over land surfaces. Advances in Space Research, 26(7):1065-1068.

Teillet, P.M., Guindon, B., and Goodenough, D.G., 1982, On the slope-aspectcorrection of multispectral scanner data. Canadian Journal of Remote Sensing,

8(2):84-106.Thomson, A. G., and Jones, 1990, Effects of topography on radiance fromupland vegetation in North Wales. International Journal of Remote Sensing,11(5):829-840.

Tokola, T., Sarkela, J., and Linden, V., 2001, Use of topographic correction inLandsta TM-based forest interpretation in Nipal. International Journal of RemoteSensing, 22(4):551-563.

Date post:	10-Apr-2018
Category:	Documents
Upload:	reneebartolo
View:	215 times
Download:	0 times

15arspc Submission 223

Documents