ARTICLE IN PRESS

0098-3004/$ - se

doi:10.1016/j.ca

�Correspondfax: +5512 394

E-mail addr

Computers & Geosciences 32 (2006) 1169–1181

www.elsevier.com/locate/cageo

Modeling small watersheds in Brazilian Amazonia with shuttleradar topographic mission-90m data

Marcio M. Valeriano�, Tatiana M. Kuplich, Moises Storino, Benedito D. Amaral,Jaime N. Mendes Jr., Dayson J. Lima

Brazilian National Institute for Space Research (INPE), Remote Sensing Division, av. Astronautas, 1758, Jardim da Ganja,

12.227-010 Sao Jose dos Campos, SP, Brazil

Received 9 June 2005; received in revised form 3 August 2005; accepted 27 October 2005

Abstract

This work presents a methodology for the refinement of shuttle radar topographic mission (SRTM-90m) data available

for South America to enable detailed watershed studies in Amazonia. The original data were pre-processed to properly

map detailed low-order drainage features and allowed digital estimates of morphometric variables. Spatial-resolution

refinement (300 to 100, or �90 to �30m) through data kriging was found to be an interesting solution to construct digital

elevation models (DEMs) with more adequate presentation of landforms than the original data. The refinement of spatial

resolution by kriging interpolation overcame the main constraints for drainage modeling with original SRTM-90m, such

as spatial randomness, artifacts and unrealistic presentation due to pixel size. Kriging with a Gaussian semivariogram

model caused a smoothing of the resulting DEM, but the main features for drainage modeling were preserved. Canopy

effects on the modeled surface represented the main remaining limitation for terrain analysis after pre-processing. Data

regarding a small watershed in Amazonas (�38 km2), Brazil, were evaluated through visualization techniques,

morphometric analyses and plot diagrams of the results. The data showed limitations for use in the original form, but

could be applied for watershed modeling at relatively detailed scales after the described pre-processing.

r 2005 Elsevier Ltd. All rights reserved.

Keywords: Remote sensing; Geoprocessing; SRTM; Watershed; Amazonia

1. Introduction

Many studies dealt with the relation betweentopographic variables and the characterization oflandscape attributes as described in systems ap-proach (Dent and Young, 1981; Meijerink, 1988).Indeed, besides the topographic information itself,

e front matter r 2005 Elsevier Ltd. All rights reserved

geo.2005.10.019

ing author. Tel.: +5512 3945 6424;

5 6488.

ess: valerian@dsr.inpe.br (M.M. Valeriano).

the spatial characterizations of other landscapeaspects like soils (Moore et al., 1993), weather(Goovaerts, 2000) and vegetation (Florinsky andKuryakova, 1996) were shown to be improved whencoupled with relief modeling. The use of digitalelevation models (DEMs) through geographicalinformation systems (GIS) is a powerful approachin this matter, since automatic methods to analyzetopographic features are allowed, with both opera-tional and quality advantages. While GIS-basedfeature extraction are processed with all advantages

.

ARTICLE IN PRESSM.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–11811170

of digital resources (speed, repeatability and com-puter integration with other databases), the reduc-tion of manual interventions (and thus subjectivity)and the possibility of parametric approach representqualitative distinctions from interpretative-basedmethods.

In watershed studies, GIS modeling of DEMs isoften applied to erosion estimates (Molnar andJulien, 1998), in which the automatic extractionof its related variables consists as important fieldof research (Desmet and Govers, 1996). Directlyconcerned to the comprehension of watersheddrainage structure, watershed partition (Band,1986) and the identification of terrain units(Miliaresis, 2001) are important tasks to besupported by DEM analysis.

In developing countries, the lack of topographicdata for great extensions generally hinders thetopographic modeling. This deficiency is critical inareas like Amazonia, where inaccessibility andhistorical reasons kept large territory uncoveredby systematic mapping. The increasing use oforbital remote sensing methods represents a strate-gic alternative to overcome this lack of data.

In February 2000, the shuttle radar topographicmission (SRTM) collected radar data at C- andX-band. For the near-global coverage, however,only the C-band data were processed. Differentlyfrom the previous shuttle imaging radar missions, atthe SRTM the shuttle was equipped with a secondradar antenna, separated from the first antenna by amast, which allowed the use of interferometrictechniques. synthetic aperture radar (SAR) inter-ferometry is a technique that uses information onphase difference between two SAR images acquiredfrom slightly different antenna (sensor) positions.The phase difference between the two positionsindicates the average three-dimensional position ofthe scattering elements/objects at the Earth’s surfaceand, under certain conditions, allows the height ofobjects to be inferred (Baltzer, 2001). In addition totopographic mapping and DEM construction,interferometric data have been also used to estimateforest height (Wegmuller and Werner, 1995).

The recent release of SRTM-90m data for SouthAmerica suggested the direct application of theexisting GIS resources in terrain analysis to befeasible for the whole region, thus overcoming thelack of topographic information in a short period oftime. Nevertheless, SRTM is a remotely sensed dataset with the typical need for evaluations and pre-processing work before its insertion is GIS analysis

of the relief. The objective of this work was toevaluate and test the quality and possibilities ofusing SRTM data for small watersheds modelingwith emphasis on morphometry and drainagechannel description. This article describes themethodology used for the construction of a DEMfor a study area—the Asu river watershed inBrazilian Amazonia—with a refinement of SRTMdata from 90 to 30m. Data, methods and softwaresused and the first results obtained were outlined.

2. Material and methods

This study was partially carried over a whole150 km� 100 km sheet corresponding to 1:250,000quads, as shown in Fig. 1. This SRTM data set, a1800� 1200 (columns� lines) image, had 30� 30samples submitted to the geostatistical analysis andthe whole set was interpolated for the constructionof a DEM. After preliminary checks, a small subsetcomprising Asu river watershed was selected for adetailed analysis including controlled morphometricanalyses and further verifications.

Asu is a tributary of Cuieiras river, which istributary of Rio Negro, the large riverbed orientednorthwest–southeast in Fig. 1. It is a fourth-orderriver in a 38.34 km2 watershed with heights rangingfrom 40.28 to 125.70m. The site is characterized byplateaus gently dissected by small streams. Thedominant soil at the plateaus is a yellow Oxisol, richin kaolinite clay, while Spodosols, consisting ofsand with topsoil organic matter, prevail in thelowlands. Climate is classified as Ami (Koeppen),with annual rainfall levels averaging 2800mm.

The watershed is covered by undisturbed vegeta-tion named terra firme forest (in uplands) andcampinarana in the valley floor. In Fig. 2, a 1999Landsat/Thematic Mapper (TM) image showsundisturbed forest in the study area near theacquisition date of SRTM data. Landsat/TMquicklooks for dates up to 2005 were also checkedand confirmed the dense forest cover in the studyarea. Fig. 2 shows the two types of forest formationscovering the study site, according to RADAM-BRASIL (1978) project: In A, uplands are coveredwith dense forest with emerging trees combined withopen forest with palms and dense woody campinar-

ana, while flat valley floors may be covered withopen woody campinarana. In B, dense forest withemerging also prevails, but associated with openforest with palms in dissected valleys.

ARTICLE IN PRESS

Fig. 2. RADAM vegetation map (white vectors) overlaid on a Landsat/TM gray coded composition of bands TM3, TM4 and TM5. A

and B are forest formations occurring in watershed (delimited in black vectors), according to RADAMBRASIL classification (details in

text).

Fig. 1. Study area, with Asu watershed in box.

M.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–1181 1171

The main structural differences between thesetypes of forest formations are related to tree heightand density: dense terra firme forests are higher thancampinaranas, with decreasing density of trees andbiomass (Ayres, 1993). These differences are alsorelated to their topographic condition. Ferreira(1997) also relates similar structural differencesalong slope profile in actually non-flooded areas.Wittmann et al. (2002) explained early successional

stages (and lower forests) to occur next to river–channels, as an effect of flood height and duration.Nelson (1994) points the mortality of forests due toextreme flood episodes (together with wind and fireevents) as the main cause of forest succession indrainage channels, although occurring also in terra

firme forests.The processing consisted of modifying original

SRTM data into a new DEM with desirable

ARTICLE IN PRESSM.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–11811172

characteristics for watershed modeling. First appre-ciation of SRTM data showed, besides occasionaldata failures and artifacts, the data sensitivity fornon-terrain objects (trees, buildings, antennas),which were believed to contribute to short-rangeabrupt near-random variations. The intended mod-ifications were: improved resolution from 300 to 100

(�30m); removal of the data failures; reduction ofartifacts; and distribution of the spatial random-ness, estimated as nugget effect through thegeostatistical analysis.

SRTM data kriging followed basically a geosta-tistical approach originally designed to prepareDEM from contour lines, computing digitizedquotas (and their geographic positions) as samplesfor geostatistical analysis and interpolation. Themethod was applied to SRTM data after selecting(and detaching) squared sample areas for thegeostatistical analyses. This method uses the resi-dues obtained from trend analysis to assure geo-stationarity of the data set submitted to thegeostatistical analysis. Valeriano (2002b) showedthat variogram models calculated with linear trendresidues of topographic data showed adequate fitswith classical variogram models (Isaaks and Srivas-tava, 1989), which present a clear defined sill.Variogram analysis of these residues provided thegeostatistical coefficients applied in kriging inter-polation of the original SRTM elevation data set.Fig. 3 presents the whole data flow, from originaldownloaded SRTM tiff image until the final DEM.

The computational programs used were (i) forfailure correction, sampling and ASCII data ex-port—ENVI (Research Systems Inc., 2002); (ii) fortrend analysis and calculation of residues—MINI-TABs (MINITAB Inc., 2000); (iii) for geostatisticalanalysis—VarioWin (Pannatier, 1996) and (iv) for

Fig. 3. Processing of SRTM data.

kriging interpolation—Surfer (Golden Software,1995). After the generation of the new DEM, Idrisi(Eastman, 1995) was used for operations such astest applications and extraction of morphometricvariables.

Results were evaluated through visualization ofgraphical and statistical analyses, and observationson morphometric outputs. The shaded relief ren-dered by Surfer typically presents the relief in avertical exaggeration so as to balance illuminatedand shaded surfaces in each image, independently ofthe model actual relief. This property, in combina-tion with the possibility of controlling the illumina-tion geometry, allows detailed observations of thesurface characteristics, representing a very importanttool for visual evaluation of numerical surfacesbefore their insertion in GIS. A quick talvegue–di-vide delineation process was also applied to verifydrainage modeling potential of the DEM. The three-steps process shown in Fig. 4 was repeated in the fourfilter directions and the results were overlaid toproduce a talvegue–divide map for all directions.

Slope angle, direction and plan and profilecurvatures were calculated using a suite of algo-rithms developed in previous works. Slope angleimages were calculated using a mapping program(Valeriano, 2002a) based on the vector sum of slopeorthogonal components, as quantified throughmoving windows in ‘‘� ’’ and ‘‘+’’ orientationsystems. According to the best correlation inrelation to manual measurements, slope was calcu-lated choosing the maximum resultants between theorientations and the maximum height in eachwindowed direction. The method to map profilecurvature (Valeriano, 2003) is based on local 3� 3pixel windows and designed to perform geometri-cally the second-order derivative through the slopeprofile. The curvature calculations required theDEM spatial resolution as one of the inputs, so asto calculate a comparable absolute value, with theslope change rate per horizontal distance as unit, indegrees per meter (1/m). The classification of slopedirection in octants was used to control the over-lapping derivation results, calculated toward theeight neighbor pixels of each windowed position.Analogously, to map plan curvature, a simpleapproach was based on a similar application ofmoving windows on the slope direction imageinstead, providing the slope direction change rateper horizontal distance (1/m) as unit (Valeriano andCarvalho Jr., 2003). For slope direction, aspectfunction of the GIS (Idrisi) was used.

ARTICLE IN PRESS

Fig. 4. Simple three-steps process for talvegue and divide delineation.

M.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–1181 1173

3. Results and discussion

3.1. Canopy effects

Since the first examination of SRTM images, themodel height was observed to be sensitive to theforest canopy. More than simply increasing height,the forest canopy often works as a smooth blanketover the terrain, therefore hiding minor details. InFig. 5 (top), a detailed drainage network can beeasier observed in the deforested areas (darker) thanunder forest canopy. In the circled area, abruptinterruptions in drainage pattern can be seen in thedeforestation boundary. Besides the loss of detailedinformation under the canopy, this limitation willrequire two-fold analysis of drainage patterns whenworking on site of forested terrain. Similarly,canopy height irregularities are also able to affectthe calculation of morphometric variables, yet forslope angle. In Fig. 5 (bottom), clear-cut bordersrepresent an extreme canopy disturbance for mor-phometric modeling, as it causes the mapping ofunreal C-class slopes (above 8%) in the A/Bexemplified terrain.

When interpreting radar data, mainly in vege-tated terrains, the degree to which the radar beampenetrates through a forest canopy determines thetype of information available. Radar C-band refersto wavelengths of around 5.6 cm, which are able topenetrate dense forest canopies, although not reach-ing the ground (Le Toan et al., 1992). There aremany factors that contribute for the radar scatteringin a tropical forest, and the number and height ofindividuals along with closeness of the canopy aredeterminants of the radiation penetration into the

canopies. Assuming a thick tropical forest cover,information on SRTM data refers to the surfaceand near surface of the canopy. For the radiationthat penetrates the canopy, the typical sources ofscattering at C-band are branches and leaves andvolumetric scattering—as it occurs in the ‘‘volume’’of tree crowns—is the main scattering mechanismfor C-band in tropical dense canopies (Leckie andRanson, 1998).

Fig. 5 (top) shows attenuation of the drainagefeatures, although the opposite effect may beexpected where the decrease in elevations is relatedto a decrease in tree height and density. Due toforest height and tree density differences, each of theforest type occurring in the study site may result indifferent heights as modeled through C-bandaltimetry, but not in a predictable way. Otherdifficulty to overcome complex canopy effects is thestatus of vegetation mapping itself, which does notprovide information for canopy structure modelingin a desirable precision. Finally, the short-rangevariations and complex relief–canopy interactionsdeprive the establishment of a canopy heightcorrection of C-band altimetry.

Fig. 6a outlines exaggeration and attenuationeffects in a hypothetical transect. In the conditionpresented at left, tree growth compensates elevationaround drainage channels, obscuring their detection(as observed in Fig. 5). By the other effect (Fig. 6a,at the right), channel depth is exaggerated becausecanopy over enhances the modeled surface byfollowing the elevation changes at a higher ratethan the terrain gradient. Deforestation also playsan undesirable effect in drainage modeling (Fig. 6b),as canopy borders are expressed in the modeled

ARTICLE IN PRESS

Fig. 5. Gray level SRTM drainage differences between forested and deforested areas (top). Circled area presents boundary between forest

and deforested area, where drainage channels disappeared due to attenuation by forest cover. Bottom map shows clear-cut borders

affecting mapping of slope angle classes.

M.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–11811174

surface the same way as ridges and may lead to falsechannel detection where forest border faces slopeupward.

Although the canopy effects showed in Fig. 6 areimportant, the remaining terrain information isbelieved to support reasonably terrain and drainagemodeling within an accuracy level given by theforest condition. Since the local morphometricvariables are calculated through neighbor functions(operations between pixel values encompassed in amoving window position), their results dependenton relative precision rather than absolute accuracy.Thus, the assumption of canopy height uniformityenables the calculation of the local variables.

Despite the assumption of canopy height uni-formity, required for calculation purposes, canopyeffects must be considered when handling SRTMdata of forested areas. All available descriptions ofvegetation structure in Asu watershed suggest acanopy exaggeration of slope angle to occur some-where near the boundary of the valley floor. This

would modify slope angle scores locally, withpossible effects in the estimate of profile curvatures.Meanwhile, horizontal-related variables, such asslope direction and horizontal curvature, areexpected to result free from canopy effects.

3.2. Geostatistics

Though elevation data are generally expected topresent a high spatial dependence (data similarity atshort distances), the noisy patterns observed indetailed presentations suggest unexpected variationsto occur within the grid distance. In spatialstatistics, this can be evaluated through semivario-gram analysis, expressed in the rate of nugget effect.Theoretical semivariogram models present threemain coefficients that scales the fit to experimentalsemivariograms, namely range, sill and nuggeteffect. Range is a measurement of the curve’shorizontal scale (from origin to the beginning ofthe sill plateau) and corresponds to the maximum

ARTICLE IN PRESS

Fig. 6. Undesirable effects from undisturbed canopy (a) and from deforestation (b) on channel modeling with SRTM data.

Fig. 7. Semivariogram for SRTM elevation data in study area.

M.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–1181 1175

distance of spatial dependence. Nugget effect is they-intercept height and corresponds to a residualvariation at the shortest sampling interval, randomand not spatially correlated. Sill is the remainingheight of the curve above its y-intercept (nugget),and corresponds to the variance due to spatialstructure (Isaaks and Srivastava, 1989).

Of the theoretical semivariogram models, Gaus-sian curves present a region of low slope near thezero distance, which determines samples within thatdistance to have similar weights for the estimationof points nearby. Data that varies smoothly, such asground-water levels and landforms, often havesemivariograms with this inflection that can be bestmodeled by the Gaussian model (Burrough, 1987).Fig. 7 presents the semivariogram for SRTMelevation data in the study area.

The short-range effect of kriging interpolationwith Gaussian models corresponds to the smoothingeffect of a moving average window, which is super-imposed to the structure given by higher distancesamples. Very often, Gaussian modeled semivario-gram of Amazonian elevation data showed around10% of its height in the nugget effect. The use of such

coefficient in kriging interpolation leads to a locallycontrolled smoothing, where sites of short-distancevariations may be more intensely smoothed thangentle ones. As an effect of these smoothingmechanisms (Gaussian weight distribution and nug-get effect), the noisy pattern observed at detailedscales has its variance re-distributed in such a waythat terrain features hidden in the original databecome better perceptible when interpolated underhigher resolution (Fig. 8).

ARTICLE IN PRESSM.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–11811176

The exemplified shaded relief in Fig. 8 shows areasonably representative sample of Amazonianrelief under two scales: high-density drainage net-work, with short interfluves and low amplitude. At afirst glance, SRTM data seem to describe only themain features under this condition. As regarded todrainage and watershed features, only major (higher-order) channels are characterized, while minorstreams remain fuzzy. The improvements observedin the shaded relief presentation correspond to anorganization of the surface data in such a way thatangles (i.e., slope steepness and direction) becomecoherent between neighboring pixels. Since the samegeometrical features control surface matter trans-port, it can be said that the visual improvements arerelated to an increase in the data capability ofmodeling runoff and drainage features. Indeed, in theprocessed DEM, meaningful drainage channels andwatersheds were turned perceptible where originaldata showed a near-random pattern.

The above-related improvements are not expectedto overcome neither canopy effects nor resolutionlimitations. Intra-resolution dimensioned featuresare barely enhanced mainly as a result of noiseremoval and of the neighbor surface trends.Occasionally, small channels may be enhanced bythe trend of the neighboring pixels rather than thelocal height itself, allowing short interfluves (around200m) to be unambiguously characterized.

3.3. Morphometric modeling

Although Asu watershed drainage network pre-sents larger interfluves than the relief exemplified inFig. 8, original SRTM data seem to hardly support

Fig. 8. Shaded relief of original (left) a

digital analysis in acceptable performance due tonoisy information, while the processed modelresulted in a more easily interpretable set (Fig. 9).Computer analysis of DEM works like a verymeticulous interpreter, since the calculations occurmostly through 3� 3 pixel moving windows, depriv-ing the gross perception necessary to overcomepunctual surface irregularities.

The calculation of local morphometric variablesis a consistent way to evaluate DEM performance inrelation to drainage modeling. Local morphometryis a numerical description of landforms, which arerecognized as the topographic constraints acting onmatter (water included) transport through theterrain surface. Slope (steepness) and directiondetermine the local first-order vector of the poten-tial runoff flow. Plan and profile curvature areconsidered second-order constraints, since theycontrol flow acceleration and lateral forces ofconvergence/divergence. A simple talvegue–dividedelineation, analogous to second-order derivativemaximum and minimum detection, is overlaidon a synthetic hill shading, here called the azimuth,drainage and divide (ADD) process, completesthe collection (Figs. 10–12). A draft vectorof watershed limits and drainage channels wasdigitized on the gray level-coded elevation image(Fig. 10, left). The final vectors of the limits anddrainage channels were digitized on their precisepositions given by ADD image (Fig. 10, bottomright) after overlaying the draft vectors on it, andfurther precision is achieved adjusting the vectorlines to features enhanced by the zoomedslope direction map. This step-by-step digitizing isnecessary because slope direction, when coded

nd processed (right) SRTM data.

ARTICLE IN PRESS

Fig. 10. Morphometric maps of Asu watershed: elevation and talvegue–channel delineation, extracted from original (top) and refined

(bottom) DEM.

Fig. 9. Shaded relief from original (top) and processed (bottom) SRTM data. Detail extracted from Asu watershed uplands.

M.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–1181 1177

with a illumination from north, is a good guideto precisely locate the drainage and divide lines,but easily leads to mistaken interpretation ofthe major networks of a watershed. The defini-tive vectors were overlaid on the morphometricmaps of Figs. 10–12 on the maps where thesefeatures were not enhanced, with dashed line fordrainage channels and solid line for the watershedlimits.

The improvement after refinement of SRTM datais more perceptible in channel/divide with hill-shading image (right) than in gray level-codedelevation (Fig. 10). It can be seen that the processedSRTM data generated meaningful morphometricmaps of the watershed, since the variables aredistributed in a naturally expected arrangement,although numerically all results may be affected bycanopy. In the circled detail of Fig. 10 (bottom

ARTICLE IN PRESS

Fig. 11. Morphometric maps of Asu watershed: slope direction and angle, extracted from original (top) and refined (bottom) DEM.

Fig. 12. Morphometric maps of Asu watershed: profile and plan curvature, extracted from original (top) and refined (bottom) DEM.

M.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–11811178

right), first-order channels had their springs de-tected, but the junction with the main river couldnot be enhanced by the applied transversal second-order derivative enhancement. In other words,channel detection under this condition (i.e., under

the combination of tropical forest and processedSRTM data) is not possible by simply analyzingdrainage cross sections. To model drainage featuresas a functionally meaningful network, a morecomplex algorithm is required.

ARTICLE IN PRESSM.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–1181 1179

Slope angle and direction (Fig. 11) also showed anotable improvement due to pre-processing, withthe latter providing an unambiguous delineation ofthe watershed drainage structure in the refinedDEM (bottom left). Slope angle improvements wererelated to the surface smoothness rather than to itsdescriptive potential.

The plan curvature mapping was shown to bespecially improved by the smoothing effect onSRTM data after kriging refinement. OriginalDEM resulted in speckle-like pattern of plancurvature, due to abrupt variations which areunavoidable under coarse resolution. Otherwise,refined data resulted in a meaningful description ofthe surface hydrology, exhibiting a drainage net-work as extreme convergence areas (dark) and aclear set of divide lines as extreme divergent areas(bright). Profile curvature analysis of the refinedDEM was not improved the same way, showingmainly smoothing effects, undesirably strong, ifcompared to the perceptible details enhanced on theoriginal data (Fig. 12).

The attenuation effect of the canopy (enhanced inchannel delineation, Fig. 10) was also observed inthe channel network shown by horizontal curvature(Fig. 12). However, the resulted morphometricmaps showed that relief hydrological constraintscould be fairly characterized through DEM model-ing with the processed SRTM data. Scatter plots ofelevation, slope angle, plan curvature and profile

Fig. 13. Scatter plots between morphometric variables in Asu watershed

data in black).

curvature (Fig. 13) showed processed SRTM toresult in narrower ranges of these variables than theoriginal data and to reduce dispersion of pointclouds, essentially. The refinement of SRTM datachanged the interrelation between slope angle andprofile curvature (Fig. 13d), as compared to resultsfrom original SRTM-90m. For original data, thehigher slopes were related to higher convex andconcave curvatures, whereas processed data pro-duced independent data sets for these variables.

In the scatter plots in Fig. 13, it can be seen thatsteepness (a) and profile curvature (b) coherentlyindicate the slopes to be convex–concave (upward).Also shown in (b) is a slight predominance ofconvexities. Despite very few exceptions, extremeconvergent flow (c) occurs only in lowlands, while,reciprocally, extreme divergence is restricted touplands. Profile curvature was shown to decreaseat extreme (high and low) slope angles (d), thoughmost of its range varied regardless of steepness inrefined data, while original data showed a directrelation between slope angle and profile curvaturemodules, concave or convex. Otherwise, plancurvature was unambiguously restricted by steep-ness (e), since high angles exhibited near zero(planar) curvature. This particular relation mayindicate the relative morpho-pedogenic stability ofthe watershed, the absence of ravines and a lowdissection rate. Taking the distribution of planand profile curvatures (f), divergent-convex and

(results from original SRTM-90m data in white and from refined

ARTICLE IN PRESSM.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–11811180

convergent-concave pixels prevail over divergent-concave and convergent-convex ones. If no otherfactor is acting on water infiltration and retention(hydrologically uniform soil), the predominance ofthese highly contrasted landforms are expected tobring diverse soil-physic environments and conse-quent biotic expressions.

The above observations are remote inferencesabout the watershed environment, based on mor-phometric analysis. Nevertheless, the morphometricvariables showed meaningful distributions andcoherent interrelations. More than supporting suchtype of description, these morphometric relationsare expected to feed numerical analyses for ahydrological state-and-process-based classificationof watersheds in GIS, as well as to model erosion,sediment, transport and other related processes.

4. Concluding remarks

This research showed a methodology for the useof SRTM data to model small watersheds, withemphasis on morphometry and drainage channeldetection. Results lead to the following conclusions:

Original SRTM-90m data have a limited cap-ability to model detailed Amazonian drainagefeatures, visually or digitally. Only visual inter-pretation of major features was possible with thisdata.Canopy showed smoothing effects that obscurethe perception of low-order channels. In spite ofthis and of the possible opposite effect (channeldepth exaggeration), the canopy condition mustbe accounted for when modeling relief withSRTM in forested areas.The methodology of kriging SRTM-90 to 30mresulted in important improvements for bothvisual and digital analysis of the terrain, enablingthe application of digital morphometric map-ping.

Acknowledgements

M. M. Valeriano was supported by ConselhoNacional de Desenvolvimento Cientıfico e Tecno-logico (CNPq). The Brazilian National Institute forSpace Research (INPE) is greatly acknowledged forthe housing of TOPODATA project and thechallenging interface with GEOMA’s proposals.Thanks are due in special to all TOPODATA staffand to all SRTM-90m users, whose requests were

the starting point of this research. MINITABs is atrademark used with permission of Minitab Inc. inthe United States.

The complete affiliations of the authors are:Marcio de Morisson Valeriano, DSR/OBT/INPE,CNPq Scholarship; Tatiana Mora Kuplich, DSR/OBT/INPE; Moises Storino CAPTEA/IAC, CIR-AD, EMBRAPA Scholarship; Benedito Dominguesdo Amaral, IGCE/UNESP, CAPES Scholarship;Jaime Nogueira Mendes Junior, IG/UNICAMP;Dayson Jardim de Lima, PG-SER/DSR/INPE.

References

Ayres, J.M., 1993. As Matas de Varzea do Mamiraua. Medio Rio

Solimoes, second ed. Sindicato Nacional dos Editores de

Livros, Rio de Janeiro (123pp).

Baltzer, H., 2001. Forest mapping and monitoring with

interferometric synthetic aperture radar (InSAR). Progress

in Physical Geography 25 (2), 159–177.

Band, L.E., 1986. Topographic partition of watersheds with

digital elevation models. Water Resources Research 22 (1),

15–24.

Burrough, P.A., 1987. Spatial aspects of ecological data. In:

Jongman, R.H., ter Braak, C.J.F., Tongeren, O.F.R. (Eds.),

Data Analysis in Community and Landscape Ecology. Pudoc,

Wagenigen, pp. 213–251.

Dent, D., Young, A., 1981. Soil Survey and Land Evaluation.

George Allen & Unwin, London (278pp).

Desmet, P.J.J., Govers, G., 1996. A GIS procedure for

automatically calculating the USLE LS factor on topogra-

phically complex landscape units. Journal of Soil and Water

Conservation 51 (5), 427–433.

Eastman, J.R., 1995. Idrisi for Windows: User’s Guide. Clark

University, Worcester, MA (440pp).

Ferreira, C.A.C., 1997. Variac- ao florıstica e fisionomica da

vegetac- ao de transic- ao campina,campinarana e floresta de

terra firme na Amazonia Central, Manaus. Masters Thesis,

Universidade Federal Rural de Pernambuco, Brazil.

Florinsky, I.V., Kuryakova, G.A., 1996. Influence of topography

on some vegetation cover properties. Catena 27, 123–141.

Golden Software, 1995. Surfer Version 6.01—Surface Mapping

System. Golden Software Inc., Golden, CO.

Goovaerts, P.A., 2000. Geostatistical approaches for incorporat-

ing elevation into the spatial interpolation of rainfall. Journal

of Hydrology 228 (1–2), 113–129.

Isaaks, E.H., Srivastava, R.M., 1989. An Introduction to Applied

Geostatistics. Oxford University Press, New York, NY

(561pp).

Le Toan, T., Beaudoin, A., Riom, J., Guyon, D., 1992. Relating

forest biomass to SAR data. IEEE Transactions on

Geoscience and Remote Sensing 30, 403–411.

Leckie, D.G., Ranson, K.J., 1998. Forestry applications using

imaging radar. In: Henderson, F.M., Lewis, A.J. (Eds.),

Principles and Applications of Imaging Radar. Wiley, New

York, NY, pp. 435–509.

Meijerink, A.M.J., 1988. Data acquisition and data capture

through terrain mapping units. ITC Journal 1988 (1), 23–44.

ARTICLE IN PRESSM.M. Valeriano et al. / Computers & Geosciences 32 (2006) 1169–1181 1181

Miliaresis, G., 2001. Geomorphometric mapping of Zagros

Ranges at regional scale. Computers & Geosciences 27 (7),

715–728.

Minitab Inc., 2000. Meet MINITAB, Release 13 for Windows.

Minitab, State College, PA.

Molnar, D.K., Julien, P.Y., 1998. Estimation of upland erosion

using GIS. Computers & Geosciences 24 (2), 183–192.

Moore, I.D., Gessler, P.E., Nielsen, G.A., Peterson, G.A., 1993.

Soil attribute prediction using terrain analysis. Soil Science

Society of America Journal 57 (2), 443–452.

Nelson, B.W., 1994. Inventario florıstico da Amazonia: tendencia

de amostragem nos herbarios e potencial de Sensoriamento

Remoto. Doctorate Thesis, Universidade Federal do Amazo-

nas, Brazil.

Pannatier, Y., 1996. VarioWin: Software for Spatial Data

Analysis in 2D. Springer, New York, NY 91pp.

RADAMBRASIL, 1978. Levantamento de Recursos Naturais.

Folha SA.20—Manaus, vol. 18. DNPM/Projeto RADAM-

BRASIL, Rio de Janeiro, Brazil.

Research Systems Inc., 2002. Environment for Visualizing

Images—ENVI, Version 3.6. Research Systems Inc., Boulder,

CO (126pp).

Valeriano, M.M., 2002a. Programac- ao do calculo da declividade

em SIG pelo metodo de vetores ortogonais. Espac-o e

Geografia 5 (1), 69–85.

Valeriano, M.M., 2002b. Modelos digitais de elevac- ao de

microbacias elaborados com krigagem. Information and

Documentation Service (SID), INPE, Technical Report

INPE-9364-RPQ/736 (54pp).

Valeriano, M.M., 2003. Curvatura vertical de vertentes em

microbacias pela analise de modelos digitais de elevac- ao.

Revista Brasileira de Engenharia Agrıcola e Ambiental 7 (3),

539–546.

Valeriano, M.M., Carvalho Jr., O.A., 2003. Geoprocessamento

de modelos digitais de elevac- ao para mapeamento da

curvatura horizontal em microbacias. Revista Brasileira de

Geomorfologia 4 (1), 17–29.

Wegmuller, U., Werner, C.L., 1995. SAR interferometric

signatures of forest. IEEE Transactions on Geoscience and

Remote Sensing 33 (5), 1153–1161.

Wittmann, F., Anhuf, D., Junk, W.J., 2002. Tree species

distribution and community structure of central Amazonian

varzea forests by remote-sensing techniques. Journal of

Tropical Ecology 18, 805–820.