An Assessment of the Discrimination of Iron- mineralized Latentes in the Amazon Region (Carajás Province) based on Textural Attributes from C-band Airborne SAR Data

M. C. de Morais, W. R. Paradella and C.C. Freitas National Institute for Space Research (INPE) São José dos Campos, São Paulo, 12227-010 - Brazil

Abstract

Imaging radar has potential for application in the tropics mainly due to the all-weather sensing capabilit■ This allows penetration through cloud cover, rain, haze, or smoke, and makes SAR unique in their ability to

recover information under conditions when optical sensors are unable to perform. As the topo graphy and

surface roughness are highlighted, SAR textural attributes can be used for rock alteration mapping. The NI

deposit is located in the Carajás Province, Brazilian Amazon Region. The deposit is related to a plateau with

a lateritic cover showing units related to iron mineralizations and to specific savanna-type vegetation. The

research was based on airborne C-band data with distinct illumination geometry and polarization. The

images were analyzed through textural classifications (GLCM, GLDV) aiming at the mapping of the latentes.

Differences were found when the classified maps and ground truth information were compared, but textural

attributes can be used for preliminary mapping (i.e., as a guide for field based verification). Textures were

sensitive to the sensor and the target parameters, and surface roughness measurements were important for the

evaluation of the results.

1. Introduction

Airborne and spaceborne C-band SAR data have played an important role in the acquisition of geological information in the Carajás Province, Amazon Region (Paradella et al., 1997; Paradella et al., 2000). This is the most important Brazilian mineral province with the world's largest iron deposits. The area has a mountainous terrain, covered by Ombrophilous Equatorial forest (Paradella et al., 1994).

Since 1967, when the iron deposits were discovered, a remarkable geobotanical control given by the iron-mineralized latentes and specific vegetation types has been recognized. The deposits are covered by thick, hard iron-rich crusts developed over volcanic rocks and ironstones. A specific low-density savanna-type vegeta-tion ("Campus Rupestres") is associated with the deposits, and shows a strong contrast ("clearing") with the dense equatorial forest. The Province was covered by airborne C-band SAR imagery during the SAREX' 92 campaign (Wooding et al., 1993).

The preliminary evaluation of these data has shown that the backscattered responses are sensitive to this geobotanical contrast (Morais, 1998). As changes in the lateritic compositions play an important role in the

expression of the macro and micro (roughness) topo-graphy, it was considered worthwhile to evaluate SAR texture classification aiming at the automatic mapping of the lateritic units. Due to the economic importance of this area, there is a practical need to provide accurate and up-to-date surface maps to support mineral exploration and environmental programs.

2. The Study Area

The Carajás Province is located on the southeast part of the Amazonian Craton (Figure 1). The N1 is the first of a series of similar plateaus related to metavolcanic and ilietasedimentary sequences of Early Proterozoic/Archean age, located in the northern border of the Province. It has an iron-ore deposit with an estimated reserve of 854 million metric tons with 66.4% iron concentration. The plateau area is 24 km2 with altitudes approximately 700 m. The Ni area is related to rocks of the Grão Pará Group, and has been subdivided into two units: volcanic rocks of the Parauapebas Formation (Meirelles et al., 1984), and the ironstones of the Carajás Formation (Beisiegel et al., 1973). The ironstones of the Carajás Formation are composed of several types of iron ore from oxide facies, mainly jaspelite and interlayered hematite and silica,

which are distinguished by hardness (soft hematite, hard hematite, etc.). Under the humid tropical climate, ferru-ginous duricrusts and latosoils are extensively developed in the plateau, with varying degrees of alteration that are responsible for the differences in composition, hardness and textures.

The N1 plateau was mapped during the economic evaluation ot the iron reserves in the Province (Resende

and Barbosa, 1972) and the following types of ferru-ginous crusts were identified (Figure 2): duricrust (in situ duricrust with limonite blocks), chemical crust (hematite fragments with goethitic pisolites), iron-ore duricrust (hematite ore blocks and subordinately specu-larite, cemented with hydrous ferric oxides), and hematite (mainly hematite outcrops).

50° 00'

0.04.9.1.

A I I

Ill■ 0,111 ,0

....

'E Z.)

..

'1 4

f •

1 ,

1 ;'; ■ ; ; , è .., .... ,-.•-••••

-

CARAJÁS REGION ,

...........

............

4%151521~

_

6° 00' 6° 00'

- -_-------_- _-_-_ - -..."---_- - •• .. •• . . . . .

:::::::::::::-:-:-:::-K- -------- - - - - . • • .. •. • . •. •. - ,, • • .• -

• • • • • • • - • . . _ - . ------- - - , • .• ...... .. . .„ .. .. • • ... • • . , , . „

_-_-----_-------- - : - . -:---:--_----

. . .

--- - --- - .' - - - - -

-1..•., ' • '' • " ° ' . .... .• .. • . Ns. :".. ;-..* -..;:'' . . -,...,,,,...

. . ....... . -'-' .... • . ... • . .

............ ......... , '• • .` . . . . ,

der • MIEI 4i~ ...........,

0 10 20 km ..........., 50° 00'

Andean Chain Braziliano belt

1111111111 Patagonian Plataform Anorogenic granitoids Phanerozoic covers . of the South America --------- Proterozoic sediments

, Precambrian basement of the South America Archean granitoids

Metasedimentary and metavolcanic sequences

II Study Area •;> Undifferentiated basement

Granulitic Complex

Figure 1: Main tectonic units related to the study arca (source: Dall'Agnol et ai., 1994)

i E

gg. 5.;...,

i e 8

8

34 —

i í , c. —

31—

z a 8

<,

575300In E. 77 1

78 1

80 1 58 I soow E.

a , ? "Z-

— 34

,-2 , — 1;,'

8 ' ?

—31

a i'á .? ..;1.

w

i

,-....

(

1,F=

• Lat~il

Duricrum tchemical )

Duricuct

Iron-are 1-)uricnr,t

ie llemat it e

fig Lake

.

,

r—r0 k)

\ i

"tili

1 575500m E. 377000m a

I 78

I 8r) 581500m E

Surficial Cover Map- NI (Carajás. Brazil) •50)1,,,,Ien

Figure 2: Lateritic units in the Ni plateau (adapted from Resende and Barbosa, 1972)

3. SAR Dataset

The SAR flown by the Canada Centre for Remote Sensing (CCRS) was mounted in a Convair 580 aircraft. The images were acquired in April 1992. The Nadir flight tine (Bra 2.3) was 6 km above the terrain and oriented parallel to the RADARSAT-1 ascending orbit (N12W). The Narrow flight une (Bra 2.2) was oriented parallel to the main structural trend in the Carajás Province (N54W). Table 1 provides details of the SAR data and Figure 3 shows the four images related to Narrow and Nadir modes, HH and VV polarisations.

4. Background and Methodological Approach

There have been several approaches to texture measurement with remote sensing images. They can be grouped into two main categories: statistical and structural

(Haralick, 1979). In the first one, the textural meaning is related to statistical properties of spatial distribution of grey levels. In the structural approach, the texture is a set of spatial sub-patterns on the image with spatial rules arrangement. The textura! analysis is based on a statistical approach known as the Grey Levei Co-occurrence Matrix

-(GLCM). It is a common technique applied to SAR data with few examples related to geological applications (Azzibrouck et al., 1997).

The gray tone co-occurrence can be expressed in a nmatrix of relative frequencies P.. i which two neigh- t)

bouring resolution cells separated by distance d occur on the image, one with gray tone i and the other with gray tone j. Such matrices are symmetric and are a function of the angular relationship between the neighbouring resolu-tion cells as well as a function of the distance between them (Haralick, 1979). Several statistical parameters can be extracted from the GLCM, which can be used as input

'Table 1: Characteristics of the SAR data

Platform Sensor Acquisition Date

Incident Angle (deg)

Spatial Resolution (m)

Pixel Size (m)

Number of looks

Image Format

Illumination Geometry

Convair 580 SAR-C (HH,VV) NatTow Mode

April 15, 1992

45/76 ° 4.8 x 6.1 4.0 x 4.31 7 8-bits Look Azimuth =

216 °

Convair 580 SAR-C (HH,VV)

Nadir Mode

April 15, 1992

08/76° 4.8 x 6.1 4.0 x 4.31 7 8-bits Look Azimuth =

78 °

Figure 3: SAR images of N1 plateau (Carajás) obtained from the CV580 in April 1992: Narrow mode (HH and VV) and Nadir mode (HH and VV)

data in automatic classification. In addition, a class of local properties based on absolute differences between pairs of gray leveis can be used. The Grey Levei Difference Vector (GLDV) is based on the absolute differences between pairs of grey leveis i and j at a distance d and at an angle O.

The investigation was based on first-order texture measures and on textural descriptors extracted from the GLCM and GLDV used as input for a supervised classification scheme. The SAR images were radiometri-cally corrected for antenna panem effects based on the APC package from the EASI-PACE software (PCI, 1997). No speckle filtering was applied to the images, since the radar images were originally multi-look processed. Representative samples of nine classes were chosen on the basis of field observations and the geologicaI map: Cl = Latosoil, C2 = Soil (anthropogenic effects), C3 = Duricrust (chemical), C4 = Iron-ore duricrust, C5 = Iron-ore duricrust (with shadow). C6 = Hematite, C7 = Hematite (with shadow), C8 = Duricrust, and C9 = Lake. The inclusion of some classes with shadow-effects was necessary since shadow effects were pronounced as a result of the large incidence angles of the airborne SAR. Based on these samples, first (mean, standard deviation) and second order measures derived from GLCM (homogeneity, contrast, dissimilarity, entropy, energy, correlation) and from GLDV (energy, entropy, mean, contrast) were analyzed. The second order measurements were computed with eight confianrations of distance (d), i.e., (-1,0), (1,1), (0,1), (-1,1), (-2.0), (2,2), (0, 2), and (-2,2). Since 82 measurements were made, it became impracti-cable to use such a large number of measurements in the classification.

Therefore, texture measurement selection was based on a decision mie based on the Discriminant Factor (DF), which evaluates the separability between classes (Rennó et al., 1998). Thus, for two hypothetical classes A and B, and one texture measurement k, the DF was computed according to the variation between and within these classes, given by:

nA na

n E (x - X ) + n E (X --£ ) 2 DF —

A' i=i

Ai,k B.k a' i...ri Bi,k A,k

AB,k nA ne

n E (X - 5? ) + n E (x - X J2 A.k B. i=I Bi,k 13,1,-

where X is the ith sample of class co for the measurement k' X(o,k is the mean value of measurement k of class (), and n is the number of samples of class 03. co

The texture measure chosen to separate classes A and B is the one with higher value of DFAB for ali k, indi-cating the better separability for these classes. DF AB k values near one denote that there is confusion betwee'n classes A and B for the texture measure k. Details of this method can be found in Rennó et al. (1998).

The next step was the generation of the selected textural channels through the Texture Analysis package, available in the EASI-PACE software. For a better control of the grey leveis, the textural channels were processed with 32-bits and scaled later to 8-bits. A cell size 7 x 7 pixels was selected in order to keep the GLCM sensitivity to the smallest details of the targets, while reducing both noise effects and computation time. A maximum like-lihood classifier was used for the classification. The classi-fications were based on the best sets of texture measures for each polarization isolated (VV, HH) and combined (VV and HH). In order to refine the results, a post-classification contextual approach based on the Iterated Conditional Modes (ICM) algorithm was also applied (Vieira et al., 1997). Finally, the classified maps were geometrically corrected in order to allow verifications in the field.

The classification results were analyzed through a confusion matrix based on the Kappa coefficient of agreement (Foody, 1992), which was evaluated through two ways: using test samples extracted from the Surficial Cover Map (see Figure 2), and field observations, con-sidering the whole Surficial Cover Map. For the first mode, test samples were collected in reQions different for those used for training samples. Both test and training samples were collected from SAR data that were not geometrically corrected, in order to preserve the radio-metric properties of the original SAR data. For the second method, the thematic maps and the Surficial Cover Map were geometrically corrected based on control points extracted from a detailed topographic map (scale 1:2000) and compared.

Roughness measurements were also collected at representative sites of the main classes. The height values from each unit were obtained, in RMS (root mean square) values, by inserting a thin plate into the surface, photographing it, and digitizing the profile. The roughness classification was based on the criterion proposed by Peake and Oliver (1971).

5. Results and Discussions

5.1. Measures Selection Table 2 shows, for each SAR configuration, the best

selected textural measures (first and second order statistics) used in the classifications, with the DF for each pair of classes of interest, and the distances d (distance in pixels at considered direction for measures related to GLCM). Thus, for the Narrow C-HH data, one first-order (mean) and three-second order statistics (correlation, GLDV entropy and GLDV mean) were indicated as the best set of the textural attributes. With the exception of C3 (chemical crust), and C7 (hematite with shadows), the remaining classes were sensitive to the textura! attributes

Table 2: Best selected teZtural measures (first and second order statistics) with the respective discriminate factors (DF), pairs of classes of interest, and the distances d for each SAR configuration

SAR data Measures Related Classes

d

Narrow Mode HH

Mean C5 x C9 - Con-elation Cl x C6 -1,0

GLDV entropy C4 x C8 -1,0 GLDV mean C2 x C5 -1,0

Narrow Mode VV

Mean C2 x C5 Correlation Cl x C2 0,1

Contrast C5 x C8 0,1 GLDV mean Cl x C9 0,1

Nadir Mode HH

Mean C4 x C9 - Correlation

GLDV energy Cl x C4 Cl x C9

0,1 0,1

Nadir Mode VV

Mean C4 x C9 _

Standard-Deviation C4 x C9 - GLDV enerRy Cl x C4 -2,2

Cl = Latosoil, C2 = Soil (anthropogenic effects), C3 = Duricrust (chemical), C4 = Iron-ore duricrust, C5 = Iron-ore duricrust (with shadow), C6 = Hematite, C7 = hematite (with shadow), C8 = Duricrust and C9 = lake.

in the selection. For the Narrow C-VV data, the selected attributes were mean, correlation, contrast and GLDV mean, respectively. In addition, three classes, i.e., C3, C6 (hematite) and C7 were not distinguished. For the Nadir C-HH data, the indicated textural attributes were mean, correlation and GLDV energy. It was observed that only three classes were sensitive to the textura! attributes, i. e., Cl (latosoil with arboreal vegetation), C4 (ore crust) and C9 (lake). When considering the Nadir C-VV data, three textural attributes were selected, i.e., mean, standard deviation and GLDV energy, and only three classes were distinguished (Cl, C4 and C9). According to these results, it is concluded that the Narrow image presents an overall better performance than the Nadir data. In addition, the HH polarization was superior to VV taking into account the number of discriminated classes.

5.2. Classification Accuracy using Test Samples

The set of selected measures previously discussed was used as input for the supervised classification. Figure 4 shows the textural classification for the Narrow (HH and VV) images. Table 3 shows the confusion matrix obtained by test samples for this classification based on eight textura! images (see Table 2). Table 4 shows the Kappa coefficients for the supervised classification. In the case of Narrow HH and VV, it was 0.620309. According to Table 3, the classification was successful for most of the classes. Two classes, latosoil and lake, were 100% correctly

classified. The hematite, which is the most relevant class in economic context, has presented arcas on the ground that were well classified (86.20%). The worst class was iron-ore duricrust (1.14%) that was incorrectly classified as latosoil (6.84%), as duricrust (59.88%) and as hematite (32.12%).

The application of the Kappa ranking, as proposed by Landis and Koch (1977), indicates the following: (1) the best results (very good in the ranking) were obtained when the classifications were based on both polarization under Narrow and Nadir modes, (2) the best classification with only one band occurred when using Narrow HH (good in the ranking); which has a Kappa coefficient statistically different from the Nadir VV, and (3) VV was considered better than the HH under the Nadir mode.

5.3 Classification Accuracy by Ground Truth Map

Table 5 shows the confusion matrix for the textural classification for Narrow (HH and VV) data, considering the Surficial Cover Map (see Figure 2) as ground truth. Table 6 shows the Kappa coefficients for the classifications. In the case of Narrow HH and VV, it was 0.10. According to Table 5, the latosoil was well classified (59.37%). However, for the remaining classes, the classification was not successful. Duricrust (chemical), duricrust, iron-ore duricrust, and hematite were the classes that presented most confusion, and lake was the worst classified.

O KilomAtr. 1 1 3 2

Figure 4: Textural classification for the Narrow (HH and VV) data

Table 3: Confusion Matrix for the classification based on textural attributes extracted from airborne SAR data (Narrow HH and VV). The rows present the results of the classification in percent and the columns are the truth obtained from test samples

Soil

(anthropogenic

effects)

Latosoil

Duricrust

(chemical)

Iron-ore

Duricrust Duricrust Hematite Lake Number of

pixels

Soil 65.94 O 13.46 0 O 1.41 O 267

Latosoil O 100 0 O "6.84 O O 428

Duricrust

(chemical) 24.14 O 84.24 0 O 8.11 O 412

Duricrust O O 2.29 76.63 59.88 1.21 O 576

Iron-ore

duricrust 0 O O 1.55 1.14 3.04 O 26

Hematite 10 O O 21.80 32.12 86.20 O 695

Lake O O O O O O 100 86

Number of

pixels 323 392 349 321 526 493 86 2490

Table 4: Kappa Coefficients for the supervised classifications obtairied by test samples

SAR Data Kappa (k) Variance of Kappa

HH and VV (Narrow mode) 0.620309 1.7634E-04

HH and VV (Nadir mode) 0.548816 7.71094E-05

HH (Narrow mode) 0.509043 1.32934E-04

VV (Nadir mode) 0.504346 8.27072E-05

HH (Nadir mode) 0.471216 8.06805E-05

VV (Narrow mode) 0.416334 1.33233E-04

Table 5: Confusion Matrix for the classification based on textural attributes extracted from airborne SAR data (Narrow HH and VV). The rows present the results of the classification in percent and the columns are the ground truth from Ni latente units map (see Figure 2)

Latosoil Duricrust (chemical)

Duricrust duricrust

Iron-ore Hematite Lake Number of

pixels

Latosoil 59.37 8:69 15.60 7.63 9.21 24.92 141874

Duricrust (chemical)

1.39 21.41 6.25 3.80 3.25 7.69 46529

Duricrust 20.61 30.52 42.66 38.27 31.89 27.48 340855

Iron-ore duricrust

9.68 24.36 20.74 19.09 21.72 19.24 178842

Hematite 7.27 15.00 14.23 30.83 33.66 15.56 180426

Lake 1.66 O 0.49 0.35 0.24 5.07 5470

Number of pixels

52511 9051 497096 145526 166097 23715 893996

Table 6: Kappa Coefficients for the supervised classifications obtained by ground truth map

SAR Data Kappa (k) Variance of Kappa

HH and VV (Narrow mode) 0.109011 4.291E-07

HH and VV (Nadir mode) 0.112675 4.376E-07

HH (Narrow mode) 0.095459 3.93E-07

VV (Nadir mode) 0.102997 3.93E-07

HH (Nadir mode) 0.081170 3.991E-07

VV (Narrow mode) 0,055995 3.796E-07

These results have shown that the accuracy classifi-cation decreased dramatically when the whole Surficial Cover Map was used as reference. This can be caused by: (1) the duricrust that presents a high occurrence in the Surficial Cover Map and therefore the Kappa coefficient was very affected by this class, which was not well classified; (2) the geometric distortions between the thematic classes from both sources of data (SAR images and Surficial Cover Map), particularly with the comparison of pixels belonging to the borders of the maps; (3) the geological map that may be not completely reliable; and (4) textura! C-band SAR attributes these may not be sufficient to express the physi-

'cochemical variations of latentes as expressed in the Surfi-cial Cover Map.

It was possible to observe from these classifications that the textural features are sensitive to the sensor pro-perties (polarization, illumination geometry) and the target (roughness) parameters. The best results were obtained when using both polarizations. In addition, HH polariza-tion was superior to VV under the Narrow configuration, while the opposite occurred with the Nadir mode.

In relation to the look azimuth, it was observed that narrow imagery, with an illumination almost orthogonal to the main geological trend in the area, was more sensitive

Table 7: Surface roughness classification for the main lateritic classes based on the criteria proposed by Peake and Oliver (1971)

SAR Data Class Height (rms) - cm Classification

Narrow mode (Smooth if h < 0.5109 cm

and Rough if h> 2.9033 cm)

Duricrust 2.9981179 Rough Iron-ore Duricrust 1.9489008 Intermediate

Duricrust (chemical) 0.63627008 Intermediate Latosoil 0.48070608 Smooth

Nadir mode (Smooth if h < 0.65496 cm

and Rough if h> 3.72112 cm)

Duricrust 3.2816216 Intermediate Iron-ore Duricrust 3.0415915 Intermediate

Duricrust (chemical) 1.0019746 Intermediate Latosoil 0.24670331 Smooth

in the enhancements of the terrain features than the Nadir image. The plateau was located on the far range for the Nadir imagery, showing a larger variation of incident angles than the Narrow imagery. As a consequence, Nadir mode exhibits stronger influence of the macro-topo-graphy, given by bright returns and shadows (front/back slopes).

On the other hand, the texture attributes were also sensitive to the surface roughness variations. Table 7 shows the roughness classification for the lateritic units derived from in situ measurements based on a scheme proposed by Peake and Oliver (1971). According to this table, Narrow was much more sensitive to the roughness varia-tions than the Nadir data, expressed by a larger variation in the crust classifications.

6. Conclusions

Airborne C-band SAR textures derived from the first and second order measures (GLCM, GLDV) can be used as a practical tool for preliminary mapping, i.e., as a guide for detailed mapping of iron mineralized latentes in similar areas of the Amazon. The classification accuracy evaluated through test samples has shown an increase in classifi-cation performance, as compared to the classification accuracy based on the Surficial Cover Map used as ground truth. The investigation has also shown that textural descriptors were sensitive to the (1) SAR polarization, (2) SAR illumination geometry (look azimuth, incidence angles), and (3) target parameters (macro-topography and roughness). The findings in this exploratory research are in accordance with the conclusions from Frost et al. (1984), showing that changes in the sensor parameters will affect the classification usefulness derived from the GLCM.

Acknowledgements

The authors wish to thank to Camilo D. Rennó (INPE) for the support with the Texture algorithm. This research

was partly funded by FAPESP. Thanks are due to CAPES for a MSc. grant for the first author and to CNPq for a research grant to the senhor authors. Special thanks to the mining company CVRD (SUMIC), particularly to the senior-geologist Sergio C. Guedes for the infrastructure in Carajás, and to Juscelino A. Nézio for providing assis-tance in the field for the latente characterization.

References

Azzibrouck, G. A., Saint-Jean, R., and Prévost, C., 1997, Analyse de la texture d'ane image RADARSAT pour la

cartographie géologique dans la Foret Équatoriale de

Ngoutou, est du Gabon. In: Geomatics in the era of RADARSAT (GER'97), Proceedings, Ottawa, Canada (CD-ROM).

Beisiegel, V. R., Bemadelli, A. L., Drummond, N. F., Ruff, A. W., and Tremaine, J. W., 1973, Geologia e Recursos Minerais da Serra dos Carajás. Brazilian Journal of

Geoscience, 3, 215-242. Dall'Agnol, R., Banos, C. E., Magalhães, M. S., Villas, R.

N. N., Rios, F. J., Nogueira, A. C., Silva, C. M. G., Soares, A. D. V., Vieira, E. A. P., and Martins, L. P. B., 1994, Estudo petrológico da borda oeste do Granito Central e dos corpos máficos associados á Formação Águas Claras, Internai Repor t, CVRD-UFPA, Belém, Brazil.

Foody, G.M., 1992, On the compensation for change agreement in image classification accuracy assess- ment. Photogrametric Engineering and Remote Sensing,

6 ,1459-1460. Frost, V. S., Shanmugan, K. S., and Holtzman, J. C., 1984,

The influence of sensor and flight parameters on texture in radar images. IEEE Transactions on Geo-

science and Remote Sensing, 22, 440-448. Haralick, R. M., 1979, Statistical and structural approaches

to texture. IEEE Transactions on Geoscience and

Remote Sensing, 67, 786-804.

Landis, J. R., and Koch, G. G., 1977, The measures of ob-Server agreement for categorical data. Biometrics, 33, 159-174.

Meirelles, E. M., Hirata, W. K., Amaral, A. F., Medeiros Filho C. A., and Gato, W. C., 1984, Geologia das

folhas Carajás e Rio Verde, Província Mineral dos

Carajás, Estado do Pará. In: XXXIII Brazilian Con-gress of Geology, Annals, Rio de Janeiro, Brazil. Sociedade Brasileira de Geologia, 2164-2174.

Morais, M. C., 1998, Discriminação de lateritas minerali-

zadas em ferro no Depósito Ni (Carajcis-PA) por radar

imageador: uma avaliaçã o através de classificação

textural. M.Sc. Dissertation in Remote Sensing, INPE, São José dos Campos, Brazil, 190 pp.

Paradella, W. R., Silva, M. F. F., Rosa, N. A., and Kushigbor, C. A., 1994, A geobotanical approach to the tropical rain forest environment of the Carajás Mineral Province (Amazon Region, Brazil), based on digital TM-Landsat and DEM data. International Journal of

Remote Sensing, 15, 1633-1648. Paradella, W. R., Bignelli, P. A, Veneziani, P., Pietsch, R.

W., and Toutin, T., 1997, Airborne and spaceborne Synthetic Aperture Radar (SAR) integration with Landsat TM and gamma ray spectrometry for geo-logical mapping in a tropical rainforest environment, the Carajás Mineral Province, Brazil. International

Journal of Remote Sensing, 18, 1483-1501. Paradella, W. R., Santos A. R., Veneziani, P., Sant'Anna, M.

V., and Morais, M. C., 2000, Geological investigation

uing RADARSAT-1 images in the tropical rainforest environment of Brazil. Canadian Journal of Remote

Sensing, 26, 1483-1501. PCI, 1997, Using PCI software. PCI Geomatics, Canada

(version 6,2 - EASI-PACE) Peake, W. M., and Oliver, T. L., 1971, The response of

terrestrial surfaces at microwave frequencies. Tech-

nical Report, Columbus, Ohio, pp. 2440-2447. Renno, C. D., Freitas, C. C., and Sant'Anna, S. J. S., 1998,

A system for region classification based on textural measures. In: IX Brazilian Remote Sensing Sym-posium, Armais, Santos, Brazil (CD -ROM).

Resende, N. P., and Barbosa, A. L. de M., 1972, Relatório de Pesquisa de Minério de Ferro, Distrito Ferrífero da Serra dos Carajás, Estado do Pará. Final Report,

AMZA, Belém, Brazil. Vieira P. R., Yanasse, C. C. F., Frery, A. C., and Sant'

Anna, S. J. S., 1997, Um Sistema de Análise e Classi-ficação Estatísticas para Imagens SAR. In: First Latino-American Seminar on Radar Remote Sensing, Proceedings, Buenos Aires, Argentina. ESA SP-407 (1), 179-185.

Wooding, M. G., Zmuda, A. D., Attema, E., 1993, An over-view of SAREX'92 data acquisition and analysis of the tropical forest environment. In: South American Radar Experiment (SAREX' 92) Workshop, Proceed-

ings, 6-8 December, Paris, France, 3-14 (ESA WPP-76).