Use of sar satellites for mapping zonation of vegetation communities in the amazon floodplain

Use of SAR satellites for mapping zonation of vegetationcommunities in the Amazon floodplain

M. P. F. COSTA

National Institute for Space Research, Av. Dos Astronautas, 1758, Sao Josedos Campos, SP 12 227-010, Brazil; e-mail: [email protected] University of Victoria, Department of Geography, P.O. Box 3050,Victoria, BC, Canada V8W 3P5; e-mail: [email protected]

(Received 6 September 2001; in final form 29 November 2002 )

Abstract. Radarsat and JERS-1 imagery were used for mapping zonation ofvegetation communities in the Amazon floodplain. Imagery analysis indicatesthat at periods of minimum water level the backscattering values of both C andL bands are the lowest and as the water level rises, so do the backscatteringvalues. JERS-1 imagery exhibits a larger dynamic range of backscattering inresponse to the ground cover for the two extremes of water level (10 dB) com-pared to Radarsat imagery. The backscattering differences from different groundcover allowed the use of a region-based classification that produced seasonalmaps with accuracies higher than 95% for vegetated areas of the floodplain.These seasonal maps were used to estimate the spatial distribution and time ofinundation and the vegetation cover of the floodplain. It was possible to deter-mine that semi-aquatic vegetation, tree-like aquatic plants, and shrub-like treescolonize regions flooded for at least 300 days year21. Secondary colonizers, suchas tall well-developed floodplain forest, cover regions flooded for approximately150 days year21, and floodplain climax forest colonize regions inundated forapproximately 60 days year21.

1. Introduction

The most important plant communities of the Amazonian floodplains are algae,

aquatic and semi-aquatic herbaceous plants, and forest. These communities have

adapted to survive in an environment that changes by season, year, decade, and

century, because of the ‘flood pulse’ of the Amazon (Junk 1997). In the Amazonian

varzea (flooded by ‘white water’ rivers) the number of days that a region is flooded

controls the zonation of vegetation communities. Generally, grass communities

tolerate 300 days of flooded conditions, shrubs and low biomass trees tolerate 260

flooded days, and tall-high biomass climax forests tolerate 230 to 150 flooded days

(Junk and Piedade 1997, Worbes 1997).

The determination of the spatial distribution and time of inundation and,

therefore, the zonation of floodplain vegetation communities at the Amazon scale is

only possible using remotely sensed data. Optical satellites are limited because of

intense cloud cover over the Amazon region. For instance, Novo et al. (1997) had

to acquire 10 years of Landsat data to produce a cloud-free mosaic of part of the

Amazon floodplain. Therefore, Synthetic Aperture Radar (SAR) satellites are

International Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online # 2004 Taylor & Francis Ltd

http://www.tandf.co.uk/journalsDOI: 10.1080/0143116031000116985

INT. J. REMOTE SENSING, 20 MAY, 2004,VOL. 25, NO. 10, 1817–1835

currently the most suitable systems to study the Amazon floodplain because of their

all-weather functionality and their independence from the Sun as an illumination

source. Moreover, microwave radiation interacts differently with distinct plant

communities of the floodplain and surrounding areas (Costa 2000). The charac-

teristics of the plant (density, distribution, orientation, shape of the foliage, dielec-

tric constant, height, and branches), the ground (dry, moist, and flooded), and the

sensor (polarization, incidence angle, and wavelength) are important in determining

the radiation backscattered towards the radar antenna (Dobson et al. 1996).

Using both C and L band imagery and field observations of a specific site in the

lower Amazon floodplain, this study investigated (1) the temporal variability of

radar backscattering of different vegetation communities of the floodplain and

surrounding areas and (2) the use of radar imagery for mapping spatial distribution

and time of inundation, and zonation of vegetation communities in the floodplain.The spatial distribution and time of inundation associated with the vegetation

cover constitutes key information to understand annual autochthonous carbon

production (Junk and Piedade 1997, Melack and Forsberg 2000) and methane

emission (Wassmann and Martius 1997, Forsberg et al. 2000), to evaluate pre-

ferential fish habitat (Junk 1997), and to define ecological sound management of the

Amazon floodplain.

2. Test site and data set

2.1. Site description

The study area, at 2‡00’ S/54‡00’W to 2‡30’ S/54‡30’W (figure 1), is a sedi-

mentary basin located in the northeast of the Brazilian Amazon, at the border of

the central and the lower Amazon regions. The predominant floodplain vegetation

communities are large homogeneous stands of herbaceous semi-aquatic plants,

pioneer shrubs, and various forest types. The predominant upland vegetation

is savanna (‘cerrado’), secondary forest, grasslands (pasture), and dense forest

(RADAMBRASIL 1976). The northern part of this floodplain is located in a small

rural setting, where some areas of floodplain forest have been converted to pasture

land. In the high water period, the southern part of the floodplain is connected to

the Amazon main channel.

The Amazon River water levels at Obidos (200 km northwest of the study

region) are shown in figure 2. Generally, the Amazon River presents a monomodal

cycle in which high water levels occur in May/June and low water levels in October/

November. The curves of water level fluctuation show that a similar amplitude of

variations occurred during the three years of study. Dissimilarities higher than 1 m

are mostly observed in 1997 when water levels were lowest from September to

November but recovered to similar values in January. The similar timing and

amplitude of water level variation allowed the assumption that field and satellite

data from different hydrological cycles can accurately portray the system over one

hydrological cycle. This was an important assumption because timing and amp-

litude of water variation are of great influence on zonation of vegetation in the

floodplain (Junk and Piedade 1997).

2.2. Ground data

Five field campaigns were conducted at different phases of three hydrological

cycles: high water (May 1996), falling water (August 1996), low water (November

1996), rising water (April 1997), and, again, high water (June 1999). Ground data

1818 M. P. F. Costa

included structural characterization and photographic surveying (hand-held 35 mm

camera) of floodplain forest, aquatic and semi-aquatic vegetation, upland forest,

and agriculture areas. Aerial photographs of selected areas were acquired at high

Figure 2. Water level fluctuation of the Amazon River at Obidos. Arrows represent theacquisition dates of JERS-1 and RADARSAT imagery.

Figure 1. Study area.

Mapping vegetation communities 1819

and rising water stages. These materials were used as a reference for characteriza-

tion and identification of major habitats such as floodplain forest, aquatic vege-

tation, upland forest, savanna forest, and pasture/agriculture.

2.3. Satellite data

Radarsat and JERS-1 data were acquired in 1996, 1997, and 1999, coinciding

with the field campaigns. The characteristics of the imagery are: ground range,

16 bit unsigned, and geocoded standard resolution. Radarsat images were processed

by the Canada Center of Remote Sensing and JERS-1 images by the Japanese

Space Agency. The major characteristics of the dataset are outlined in table 1.

The SAR imagery was radiometrically and geometrically calibrated according to

the procedures described in Costa et al. (2002). The test of radiometric stability

showed that the Radarsat and JERS-1 imagery was stable both temporally and

within the scene. Figure 3 shows the multi-temporal variation of backscattering (so)

across the range of acquisition for upland forested areas. For the study area,

upland forest was considered to have the most stable so values. The small dif-

ferences across the image range were associated with small variations in the terrain

topography. The small differences in the average backscattering between imagery

were associated with the slightly different environmental conditions at the time of

imagery acquisition. For instance, for Radarsat imagery, the lowest backscattering

values occurred in November (dry month), when the precipitation and the relative

humidity were the lowest and the temperature was the highest (approximately

20 mm, 70%, and 30‡C, respectively) compared with May (300 mm, 87%, and 26‡C,

respectively). The same pattern was observed for JERS-1 imagery, i.e. the image

acquired in December (dry month) showed lower backscattering values than the

imagery acquired in wet months, such as March and May. It is known that for a

given surface, if the roughness remains constant the so decreases when the moisture

content of the material decreases (lower dielectric constant) (Dobson et al. 1996).

The calibrated SAR imagery was submitted to two distinct procedures. (1)

Extraction of so (dB) from intensity imagery of known sites of plant communities;

details of this procedure can be found in Costa et al. (2002). (2) Classification of the

floodplain according to a region-based algorithm.

2.3.1. Classification procedureThe classification procedure consisted of the following steps: (1) filtering of SAR

imagery, (2) scaling from 16 to 8 bit, (3) applying water and upland masks over the

imagery, (4) segmentation, and (5) imagery classification. Details of the first three

steps are found in Costa et al. (2002).

The seasonally paired imagery (Radarsat and JERS-1 for the same season –

total of five pairs) was submitted to an automatic segmentation procedure (region

growing algorithm), and a region-based classification. For the automatic segmen-

tation procedure, a threshold of similarity for each paired data was required. The

definition of the threshold of similarity was critical for the success of the classi-

fication, since it defined the rules for merging regions. Figure 4 is an enlargement of

an area of the image overlaid with the regions built by using three different

similarity thresholds (in digital numbers): 20, 30 and, 40. The threshold 30 was

defined by the Least Significance Difference method (LSD) at 95% confidence level

(Snedecor and Cochran 1980). Note that when the thresholds were larger, fewer

regions were built, i.e. less detail was separated. Conversely, at lower thresholds,

1820 M. P. F. Costa

Table 1. Characteristics of the remotely sensed data.

Satellite Acquisition Incidence angle (‡) Coverage (km) Swath mode* Band/Polarization Pixel spacing (m) Resolution (m) Number of looks

Radarsat 27 May 1996 y43 1006100 S6-D C/HH 12.5612.5 26627 1647 August 1996

11 November 19965 April 19978 June 1999

JERS-1 16 May 1996 y35 75675 D L/HH 12.5612.5 26627 312 August 1996

22 December 199620 March 1997

*D stands for descending.

Mappingveg

etatio

ncommunities

18

21

more regions were built. However, more regions were not synonymous with better

results for the classification. In SAR imagery, this sometimes is caused by over

segmentation due to the speckle (Dong et al. 1999). A visual inspection of several

sectors of segmented imagery generated by the different thresholds showed that,

indeed, a threshold of 30 (the calculated LSD value) gave better results in terms of

defining regions. Table 2 presents the calculated threshold of similarities for each

pair of images.

The segmented imagery was submitted to a region-based supervised classifica-

tion according to the Battacharrya distance algorithm (Richards 1986). Digital

mask files of training and testing sites were created and overlaid over the paired

imagery to ensure that the selected regions were approximately the same for all the

seasonally paired imagery. Table 3 presents the number of training and testing

regions for each pair of images.

Figure 3. Multi-temporal variation of backscattering of upland forest across the rangedirection (near to far range). (a) Radarsat S6 and (b) JERS-1.

1822 M. P. F. Costa

3. Analysis and results

3.1. Temporal backscattering variability of different vegetation communities

The multi-temporal mean, lower and upper bound of so values at 95% con-

fidence level for aquatic vegetation, floodplain flooded forest, upland forest, pas-

ture, and savanna are presented in table 4. Generally, for Radarsat imagery, the

mean so values of a specific ground cover did not change significantly (pv0.05)

between seasons, except for November (dry season) when the so average values

decreased by approximately 5 dB (aquatic plants), 1 dB (floodplain flooded forest),

2 dB (pasture), 2 dB (savanna), and 1 dB (upland forest). For JERS-1 imagery,

likewise, the mean so values were only significantly different (pv0.05) for

November when the so values on average decrease by approximately 5 dB (aquatic

Table 2. Calculated threshold of similarities for each pair of imagery.

Period Combined images Threshold of similarity

May (high water) Radarsat S6 and JERS-1 30June (high water) Radarsat S6 and JERS-1 30August (falling water) Radarsat S6 and JERS-1 25November (low water) Radarsat S6 and JERS-1 25April (rising water) Radarsat S6 and JERS-1 25

Figure 4. Example of segmentation results using the region-growing algorithm withdifferent threshold values. (a) Threshold of 20, (b) threshold of 30, (c) thresholdof 40, and JERS-1 image enlarged in the background, and (d ) enlarged section of theaerial photograph of the same location showing the different ground covers (originalphotograph at 1:20 000 scale). Red box in figure 6(a) shows the location of selectedsub-region.


Table 3. Number of training and test regions.

Classified images

Number of samples

Aquaticvegetation

Floodplainflooded forest

Floodplainunflooded forest

Training Testing Training Testing Training Testing

Radarsat S6zJERS, November 48 41 10 8 60 29Radarsat S6zJERS, April 57 49 100 67 20 8Radarsat S6zJERS, May 122 72 109 71 – –Radarsat S6zJERS, June 76 54 120 64 – –Radarsat S6zJERS, August 83 59 97 72 12 8

Table 4. Multitemporal mean, lower, and upper bound of backscattering coefficients (dB) at95% confidence interval.

Month Ground cover

Radarsat S6 JERS-1

Mean Lower Upper Mean Lower Upper

November/December

Floodplain forest n~41Upland forest n~41

28.229.9

29.0210.1

27.529.7

27.428.0

28.328.2

26.627.8

Pasture n~41 212.0 212.6 211.5 212.6 213.3 212.1Savanna n~33 213.9 214.4 213.5 211.0 211.5 210.5

Aquatic plants n~11 211.9 212.4 211.3 213.6 214.1 213.4

March/April Floodplain forest 26.9 27.7 26.3 25.5 26.3 24.8Upland forest 29.6 29.9 29.4 27.4 27.4 27.1

Pasture 29.9 210.5 29.5 210.8 211.4 210.1Savanna 212.4 212.9 212.0 29.6 210.1 29.2

Aquatic plants n~16 27.4 27.8 27.0 210.5 211.1 210.1

May Floodplain forest 26.6 27.4 26.0 24.4 25.2 23.7Upland forest 28.6 28.9 28.4 26.8 27.1 26.6

Pasture 29.7 210.2 29.2 210.2 210.9 29.7Savanna 211.8 212.3 211.4 29.5 210.0 29.1

Aquatic plants n~12 27.1 27.5 26.7 28.8 29.2 28.4

June Floodplain forest 26.6 27.4 25.9Upland forest 29.3 29.5 29.1

Pasture 29.7 210.2 29.2Savanna 211.9 212.4 211.5

Aquatic plants n~35 26.9 27.5 26.3

August Floodplain forest 26.9 27.7 26.3 24.9 25.8 24.2Upland forest 29.1 29.4 28.8 27.4 27.7 27.2

Pasture 210.0 210.6 29.5 211.8 212.5 211.2Savanna 212.5 213.3 212.0 210.2 210.7 29.7

Aquatic plants n~15 26.7 27.0 26.4 29.0 29.5 28.5

n is the number of polygons sampled for each ground cover. The same polygons wereused to estimate the backscattering values monthly, with the exception of aquatic vegetation.Mean so and its confidence interval were calculated assuming that the samples have a normaldistribution due to the large number of pixels in each sampled polygon. Laur et al. (1996)calculated that a minimum number of 100 pixels per sampled polygon are required to yield a95% confidence with an error boundary of the estimated so at ¡1 dB. The average numberof pixels per sampled polygon was 215, 132, 666, 245, and 159 for aquatic vegetation,floodplain forest, upland forest, pasture, and savanna, respectively.

1824 M. P. F. Costa

plants), 3 dB (floodplain flooded forest), 2.4 dB (pasture), 1.5 dB (savanna), and

1 dB (upland forest). Nonetheless, for both Radarsat and JERS-1 imagery, the

highest so dynamic range was found for aquatic plants and floodplain flooded

forest, i.e. biotopes that were highly dependent on the water level variation. The

results showed that at minimum water levels, the so values for vegetation com-

munities of the floodplain were the lowest at both C and L bands. As the water

level rose so did the so values, until they reached a maximum value.

The multi-temporal so values of the distinct ground covers were a result of the

different scattering mechanisms, which in turn were dependent on the temporal

variability of the ground cover. Table 5 summarizes some of the main structural

characteristics of the vegetation communities and the predominant scattering

mechanisms. The following sections describe these scattering mechanisms for

different vegetation communities.

3.1.1. Floodplain forest

Regions of floodplain forest that were seasonally flooded showed a large

temporal so variation. At the low water period, the so values were very similar to

those observed for upland forest, which were a result of the interaction of the

radiation with the canopy elements. At the high water period, L band back-

scattering was higher than C band backscattering (2 dB difference) because of

differences between incidence angle and radiation wavelength. At an incidence

angle of 35‡ (JERS-1), the long wavelength of L band (23 cm) penetrated deep into

the tree canopy and interacted with the trunk and the water underneath; the reverse

pathway could also happen (Hess et al. 1995, Wang et al. 1995, Proisy et al. 2000).

This interaction, called double bounce mechanism, caused an average so value of

24 dB. At a 45‡ of incidence angle (Radarsat S6), the short C band wavelength

(5.6 cm) interacted mostly with the upper canopy layer (volume scattering

mechanism). The result was an average so value of 27 dB. Figure 5(a) illustrates

the dominant scattering mechanisms of C and L bands with floodplain flooded

forest.

During flooded conditions, the so values also varied with the degree of defo-

liation of the floodplain trees. Pseudobombax munguba (30 m height above water)

and Courupita guianensis (2 m height above water) are examples of floodplain trees

that lose their leaves during flooding conditions. For these trees, the so values were

25 dB (C band) and 23 dB (L band). The radiation penetrated deeper into the

canopy (no interference of leaves) and therefore a pronounced double bounce effect

between the tree trunk and water surface occurred.

3.1.2. Semi-aquatic vegetation

The areas covered with grass-like aquatic vegetation exhibited the greatest

temporal variation of so values. At maximum growing stage, when the water was at

its highest level, stands of Hymenachene amplexicaulis, the most common species of

aquatic vegetation in the study area, showed the following characteristics: density of

approximately 111 stems m22, grass-like structure, one stem of 0.85 m high and

0.4 cm of diameter, five leaves of 25 cm long and 2 cm wide at an angle to the

structure of roughly 45‡, and a vertical span occupied by the canopy structures of

approximately 40 cm. The interaction of microwave radiation with these plants

resulted in specular scattering, volume scattering, and double bounce scattering


Table 5. General discription of some important vegetation found in the study site (adapted from Dobson et al. 1996).

Ground cover

Non woody vegetation Woody vegetation

Pasture land Aquatic vegetation Upland Seasonally flooded

Surface Soil Water Soil Soil/water

Growth form Blade-like Shrubs (savanna- sparsevegetation)

Decurrent(savanna denseand secondary

forest)

Decurrent Tree-likeaquatic vegetation

Structuralcharacteristics

Trunk None Many small trunkswith randomorientation

Cylindrical Cylindrical Cylindrical

Branches Non-woody stalks or stems Many small branches Branches forkedand randomly

oriented

Branches forkedand randomly oriented

None

Foliage Blade-like erectophile -short

Blade-likeerectophile - tall

Broad leaves Broad leaves Broadleaves

Defoliated Blade-likeclump at top

of trunkScatteringmechanism

C-band (45‡) Volume scattering Volume scattering Surface-scattering(quasi-specular)

Volume-scattering Volumescattering

Doublebounce

Double bounce

L-band (35‡) Surface scattering(quasi-specular)

Volume scattering andspecular reflection

Surface scattering(quasi-specular)

Surface andvolume scattering

Doublebounce

Doublebounce

Double bounce

18

26

M.P.F.Costa

mechanisms. Figure 5(b) illustrates these mechanisms, which were a function of both

the sensor characteristics and the plant’s biophysical properties.

The interaction of the fully developed canopy elements with radiation of 5.6 cm

of wavelength at incident angle of 45‡ (Radarsat S6) resulted in the volume

scattering mechanism. The total backscattering was on average 27 dB. At this

incidence angle, the path of the radiation through the canopy was increased. Thus,

more attenuation by the scattering elements and less penetration of the radiation

occurred (Ulaby et al. 1982). In comparison, data acquired over Amazonian aquatic

vegetation with C-HH band and steeper incidence angle (v33‡) showed higher so

(approximately 24 dB) (Hess et al. 1995, Costa 2000, Novo et al. 2002). The steeper

incidence angle facilitated the deeper penetration of radiation into the vegetation

canopy, which might have resulted in a water/vegetation double bounce mechanism.

Le Toan et al. (1997) published similar so (approximately 28 dB) compared to the

values of this study for data acquired with steep incidence angle, same wavelength,

but VV polarization (ERS-1 configuration) over Indonesian flooded rice fields. We

Figure 5. Schematic representation of the scattering mechanisms at C and L bands for (a)floodplain flooded forest and (b) aquatic vegetation. The thickness of the returningarrows (1, 2, and 3) represents relative magnitude of scattered radiation.


are speculating that the vertically polarized radiation was strongly attenuated by

vertically oriented rice plants, while horizontally polarized radiation, such is the

case of Radarsat, penetrated deeply into the canopy.

At a wavelength of 23 cm and 35‡ incidence angle (JERS-1), deeper penetration

of the radiation within the canopy of the aquatic vegetation occurred. This resulted

in low so values (29 dB). At this wavelength, the structures of the aquatic vege-

tation were mostly transparent to the radiation (quasi-specular reflection), except

when dense canopies occurred (volume scattering) (figure 5(b)). Among the aquatic

vegetation, isolated high so values (27 dB) were observed at this wavelength. The

higher values were related to areas colonized by Echinochloa polystachya and Paspalum

fasciculatum. The higher so values suggested that canopy–water interaction (double

bouncemechanism)mayhaveoccurredwhentheradiationwas interactingwithtallgrass-

like plants (1.5 m), with large leaves (80 cm long and 3 cm wide), thick stems (2 cm), and

larger canopy gaps (27 plants m22) compared with H. amplexicaulis. Hess et al. (1995)

published comparable so values at L band for Amazonian aquatic vegetation.

Montrichardia arborescens is a tree-like semi-aquatic vegetation that colonizes

areas where moisture is retained even during the low water period. The general

structural characteristics of this species are as follows: dense elongated or round

patches of plants, 3 m height above water or 7 m total height, cylindrical vertical

trunk of 4 cm diameter, and a clump of approximately four broad leaves of 50 cm

long and 40 cm wide at an angle to the trunk of roughly 45‡. The average so was

26 dB and 25 dB at Radarsat and JERS-1, respectively. These values were a result

of the same type of interaction (double bounce) observed between microwave

radiation and floodplain defoliated trees. At the low water period, the so from

M. arborescens remained high due to the high moisture content of the areas colonized

by this species. At this period, average so values were 28 dB (C band) and 27 dB

(L band).

3.1.3. PastureShort grass-like vegetation, a few random short non-woody shrubs, and bare

soil covers the pastureland of the study area. The grass-like vegetation was equi-

valent in size to the wavelength of C band when compared with L band. Con-

sequently, at L band, relatively higher transmissivity of the radiation through this

vegetation occurred and, therefore, the radiation interacted with the ground surface.

This interaction resulted in low backscattering values for the driest (November) and

wettest (May) periods, 213 dB and 210 dB, respectively. Conversely, C band radia-

tion was expected to interact with the volume of the vegetation due to the small size

of the scattering elements. However, the observed backscattering values (212 dB

for the driest and 210 dB for the wettest period) suggested that the interaction was

mostly a result of the roughness of ground surface. Note that at both C and L

bands the lowest values occurred during the dry season due to the decrease of

moisture content of the ground.

3.1.4. SavannaA similar interaction, i.e. interaction with the roughness of the ground surface,

was evident for savanna areas where the vegetation (3 m height) was sparse and the

ground surface was composed of bare sandy soil and short grass. During the dry

season, so values were on average 214 dB for C and 211 dB for L band. During the

wet season, savanna areas showed values of 212 dB for C and 210 dB for L band.

1828 M. P. F. Costa

Again, the slightly increased values were due to the higher moisture content of the

surface during the wet season. In general, for both seasons, the low values of so

suggested that radiation at both wavelengths interacted mostly with large patches of

bare soil and grass (surface scattering) than with the shrubs.

3.1.5. Upland forestUpland forest showed the lowest temporal variation of the backscattering

values. In the study area, upland forest is not a typical dense rain forest; it is mostly

composed of secondary forest and dense savanna (RADAMBRASIL 1976). The so

values were on average 29 dB for C and 27 dB for L band, and the lowest values

(less than 1 dB difference) were for the dry season imagery. For C and L bands,

regardless of the period of imagery acquisition, the so values were primarily a result

of the interaction of the radiation with the structures of the canopy (volume

scattering).

3.2. Zonation of the floodplain plant communities according to the time of

inundation

The previous section helped to understand the temporal scattering mechanisms

that occur due to the interaction of microwave radiation and vegetation communities

and ground surface. This knowledge was applied to define the classification scheme.

The classification procedure resulted in temporal maps of the floodplain with the

following classes: floodplain unflooded forest, floodplain flooded forest, aquatic

vegetation, upland vegetation (upland forest, savanna, and pasture), and water. The

upland vegetation class was a pre-defined mask applied over the imagery. The water

class was also a pre-defined seasonal mask applied over the imagery (details of these

procedures can be found in Costa et al. 2002).Overall, the classification accuracy of the multi-wavelength, C and L band

composite, exceeded 95%. The seasonal thematic maps are shown in figure 6, and

the matrix of error of the classification is shown in table 6. Aquatic vegetation areas

were accurately classified (approximately 95%) at any season. Floodplain flooded

forest showed an average classification accuracy of 90%. The main classification

error occurred between floodplain unflooded forest and flooded forest, for the low

water period imagery (November). When water levels rise, the different forested

areas within the floodplain present three distinct conditions of ground surface: dry,

moist, and flooded. It is speculated that the observed misclassification happened

because moist and flooded surfaces show similar strong backscattering due to the

high moisture content of the ground.

Clearly, the different backscattering values, which were a result of the inter-

action of C and L bands with the distinct ground covers, explain the high classi-

fication accuracies. That is, different scattering mechanisms occurred from the

interaction of short and long wavelengths with the canopy of the vegetation and

the undermeath surface. Other researchers have shown similar results in which

multi-wavelength SAR combinations improved the general classification of vege-

tated areas to values higher than 90% for forested areas (Pierce et al. 1994, Dobson

et al. 1996, Bergen et al. 1998, Kellndorfer and Pierce 1998), agriculture areas

(Lobo et al. 1996), and wetlands (Hess et al. 1995, Costa et al. 1998). Costa et al.

(1998) classified individual and combined sub-scenes from a set of original images

acquired in May for the same study area in the Amazon. The classified maps were

compared with a ground truth map that was generated from visual interpretation of


the aerial photography acquired for the same period. The comparison showed that

the classification of Radarsat images alone yielded confusion among the vegetated

ground cover. The JERS-1 images alone yielded confusion between floodplain

flooded forest and upland forest and pasture and aquatic vegetation and, in addition,

(a) (b)

(c)

(e)

(d )

Figure 6. Thematic classification maps. Cyan~aquatic vegetation; yellow~floodplainflooded forest; orange~floodplain unflooded forest; green~upland forest (forest,pasture and savanna); blue~water. Red box in 6 (a) represents the area of figure 4.(a) November – low water, (b) April – rising water, (c) May – high water, (d ) June –high/falling water, (e) August – falling water.

1830 M. P. F. Costa

did not separate narrow water channels as well as Radarsat. Generally, the multi-

wavelength combination provided better classification of the ground cover.In summary, the analyses of our multi-temporal results and the results of Costa

et al. (1998) show that a combination of different wavelengths yield better classi-

fication than individual C or L bands.

The seasonal calculated area of aquatic vegetation and floodplain flooded forest

and unflooded forest of the floodplain are shown in table 7. By November/

December, the Amazon River water level, as measured in Obidos, was already high

by approximately 2.5 m, and field observations showed that regions on the margins

of the central lake and nearest to the Amazon channel were starting to flood. In

these regions, large areas of aquatic vegetation at the beginning of the growth cycle

were observed. As a result, the measured total area covered by aquatic vegetation in

November was 342 km2. By April, the measured area decreased to 217 km2. At this

month, the Amazon River water level was 4 m higher than in November, causing a

2 m increase in the water depth of regions colonized by aquatic vegetation (Costa

et al. 2002). Possibly, the flood exterminated colonies of plants (established in

November) that could not keep pace with the water level change, as mentioned by

the local villagers. Further, during the rising water period, buffalo and cattle graze

large areas of freshly growing aquatic vegetation. This might also partially explain

the larger area occupied by aquatic vegetation in November than in April.

Maximum coverage of aquatic vegetation occurred in May (397 km2), during high

water. After May, the water started to recede and some aquatic vegetation com-

munities detached from the bottom and were carried away towards the Amazon

River. This was commonly observed during the field campaigns. The occupied area

Table 6. Matrices of confusion for the classifications.

% Classified as

True class

Aquaticvegetation

Floodplainflooded forest

Floodplainunflooded forest

Radarsat S6zJERS-1, November – overall accuracy~97.02%Aquatic vegetation 99.1 0 0.91Floodplain flooded forest 0 58.9 41.Floodplain unflooded forest 3.3 11.3 85.

Radarsat S6zJERS-1, April – overall accuracy~96.94%Aquatic vegetation 94.4 5.6 0Floodplain flooded forest 0.9 97.8 1.2Floodplain unflooded forest 0 2.9 97.1

Radarsat S6zJERS-1, May – overall accuracy~97.39%Aquatic vegetation 95.3 4.5 –Floodplain flooded forest 0 100 –Floodplain unflooded forest – – –

Radarsat S6zJERS-1, June – overall accuracy~95.47%Aquatic vegetation 97.4 2.6 –Floodplain flooded forest 0.6 99.36 –Floodplain unflooded forest – – –

Radarsat S6zJERS-1, August – overall accuracy~93.85%Aquatic vegetation 93.4 5.5 1.11Floodplain flooded forest 4.6 95.5 0Floodplain unflooded forest 0 8.1 91.9


decreased to 296 km2 (June) and 282 km2 (August) following the decrease of water

levels.

The area classified as floodplain flooded forest increased from November to

May, and decreased again from June to August, clearly reflecting the annual water

level variation. The areas classified as floodplain unflooded forest decreased from

November (333 km2) to April (87 km2) and May/June (0 km2), and started to

increase again by August (85 km2), once again following the water level. Large areas

of floodplain forest were not flooded in November when the water level was on

average 2.5 m high. This area increased considerably in April when the water level

variation was 6 m. By May/June, the areas of floodplain forest were completely

flooded. By August, when the water level receded, the areas of floodplain flooded

forest decreased.

Four distinct areas of zonation were characterized based on (i) the knowledge

that the time of inundation period is the primary force of zonation in the

Amazonian varzea (Junk and Piedade 1997, Worbes 1997), (ii) the understanding of

the interaction of microwave radiation with distinct ground covers, and (iii) the

seasonal thematic maps that were produced.

Two different ground covers were classified as flooded for at least 300 days

year21. The first ground cover contained aquatic vegetation continuously from the

beginning of rising water in November until August. In the Amazon, aquatic vege-

tation tolerates flood conditions for more than 300 days year21 as well as high rates

of sedimentation. These plants are one of the pioneer species that fix their roots

(Junk and Piedade 1997). In the study region, H. amplexicaulis is the dominant first

settler. In areas occupied by this species, both sedimentation and accumulation of

decaying organic material are greatly increased, which facilitate the formation of

habitats that are colonized by pioneer tree-like communities. These pioneer tree-like

communities comprise the second ground cover that was flooded continuously from

November to August. M. arborescens (aninga), a tall tree-like aquatic plant, Salix

sp. (oierana), and C. guianenses (castanha de macaco), both shrub-like trees,

tolerate flood conditions of 300 days year21 (Junk and Piedade 1997, Worbes 1997).

Early and late secondary settlers colonized the areas of floodplain forest that

were flooded at least from April to August (these areas were not flooded in

November), i.e. approximately 150 days year21. The most common species in the

floodplain forest in the study area are Cecropia latiloba (imbauba), Pseudobombax

munguba (munguba), and Astrocaryum jauari (jauarı). Unfortunately, the temporal

frequency in which the satellite imagery was acquired did not allow a separation of

areas of early and secondary colonizers. It is not known when these areas started to

flood due to the lack of image acquisition in January and February. Furthermore, it

Table 7. Total area (km2) per class.

MonthAquatic

vegetationFloodplain

flooded forestFloodplain

unflooded forest Upland*Open

water*

November 342 93 333 844 1045April 217 156 87 844 1247May 397 304 – 844 1112June 296 290 – 844 1227August 282 238 85 844 1208

*Area of upland represents the upland mask; areas of open water represent the openwater mask for each period.

1832 M. P. F. Costa

is not known when the flood receded from these areas due to the lack of imagery of

September and October. It is only known that these areas were flooded from April

to August, i.e. at least 150 days year21, and that they were not flooded in November,

which suggests that these colonizers do not tolerate flooded conditions all yeararound.

Finally, climax forest, composed of very tall dense-wood species (Worbes 1997),

colonized areas that were flooded only in May and June (possibly July, but the data

was not available), totalling approximately 60 days year21. These areas were not

flooded from August to November, i.e. these species were restricted to shorter flood

conditions. They were mostly seen in the northern areas of the classified maps

(figure 6(a) and (e)).

4. Summary and conclusions

This investigation examined the temporal variation of SAR backscattering and

the use of such data for mapping the spatial distribution and time of inundation

and zonation of vegetation communities in the Amazon floodplain.

In the first part of the analysis, so values indicated that at periods of minimum

water level the backscattering coefficients of both C and L bands were the lowest.

As the water level rose, so did the backscattering values. JERS-1 imagery exhibited

a larger dynamic range of backscattering in response to the ground cover for thetwo extreme periods of water level (10 dB) and within the scene (6 dB) compared

with Radarsat imagery. This suggests that the longer L band wavelength was more

sensitive to thickness and size of the vegetation scattering elements compared with

the C band shorter wavelength. The sensitivity of C and L bands to the scattering

elements provoked different scattering mechanisms. The possible scattering mechan-

isms for aquatic vegetation were as follows: at C band (45‡ of incidence angle),

volume scattering was dominant and at L band (35‡ of incidence angle), volume

scattering and specular reflection were dominant, however, for some species, doublebounce occurred. For defoliated flooded forest, double bounce occurred for C and

L bands. For foliated flooded forest, a combination of double bounce and volume

scattering prevailed at L band and volume scattering prevailed at C band. For

upland forest, volume scattering occurred for both C and L bands. For pasture and

savanna areas, surface scattering prevailed at both C and L bands.

In the second part of the analysis, the sensitivity of SAR imagery to temporal

changes and different ground covers was used for mapping the floodplain vege-

tation communities. The analysis of the backscattering differences between the flood-

plain and the surrounding upland areas suggested that a combination of Radarsatand JERS-1 was the optimal choice for mapping the seasonal habitats within the

floodplain according to a multi-wavelength region-based classification. The tem-

poral mapping achieved an accuracy of approximately 95%. These maps allowed

the determination of the spatial distribution and time of inundation and zonation of

different vegetated areas in the floodplain. Grass-like semi-aquatic vegetation, tree-

like semi-aquatic plants, and some floodplain shrub-like trees colonize regions flooded

for at least 300 days year21. Secondary settlers such as well-developed floodplain

forest, colonize regions flooded for approximately 150 days year21. Floodplainclimax forest colonizes regions flooded for approximately 60 days year21. The

zonation according to the time of flooding is a result of the different adaptations of

the vegetation for tolerating flood stress (Junk and Piedade 1997).

The method presented here, based on multi-temporal C and L band SAR

imagery, can provide quantitative information on spatial distribution and time of


inundation and vegetation cover of the Amazon floodplain. This method can be

further applied to the scale of the whole Amazon basin and used to further under-

stand vegetation adaptations to flood conditions, to map preferential habitats for

fish and human use of the floodplain, and to further elaborate the carbon budget of

the floodplain.

Acknowledgments

The Canadian International Space Agency and Fundacao de Amparo a

Pesquisa do Estado de Sao Paulo funded this research. Financial contributions for

field campaigns from Dr John Melack under the LBA project is also acknowledged.

CAPES (Coordenacao de Aperfeicoamento de Pessoal de Nıvel Superior), Brazil,

provided the graduate scholarship to the author. The National Space Agency of

Japan and the Canadian Space Agency provided JERS-1 and Radarsat imagery,

respectively. The support of several colleagues and friends during fieldwork is also

acknowledged, especially Dr Novo, Dr Mantovani, and Gilson.

ReferencesBERGEN, K. M., DOBSON, M. C., PIERCE, L. E., and ULABY, T., 1998, Characterizing carbon

in a northern forest by using SIR-C/X-SAR imagery. Remote Sensing of Environment,63, 24–39.

COSTA, M. P. F., 2000, Net Primary Productivity of Aquatic Vegetation of the AmazonFloodplain: A multi-SAR satellite approach. PhD thesis, University of Victoria,Victoria, Canada.

COSTA, M. P. F., NOVO, E. M. L. M., AHERN, F., MITSUO II, F., MANTOVANI, J. E.,BALLESTER, M. V., and PIETSCH, R. W., 1998, The Amazon floodplain throughradar eyes: Lago Grande de Monte Alegre case study. Canadian Journal of RemoteSensing, 24, 339–349.

COSTA, M., NIEMANN, O., NOVO, E., AHERN, F., and MANTOVANI, J., 2002, Biophysicalproperties and mapping of aquatic vegetation during the hydrological cycle of theAmazon floodplain using JERS-1 and RADARSAT. International Journal of RemoteSensing, 23, 1401–1426.

DOBSON, M. C., PIERCE, L. E., and ULABY, F. T., 1996, Knowledge-based land-coverclassification using ERS-1/JERS-1 SAR composites. IEEE Transactions on Geoscienceand Remote Sensing, 34, 83–99.

DONG, Y., FORESTER, B. C., and MILNE, A. K., 1999, Segmentation of radar imagery usingthe Gaussian Markov random field model. International Journal of Remote Sensing,20, 1617–1639.

FORSBERG, B. R., ROSENQVIST, A., PIMENTEL, T. P., and RICHEY, J. E., 2000, Modelingof flooding patterns and methane emissions in the Jau River floodplain (CentralAmazon) using JERS-1 imagery. Resume. In Wetlands 2000, Quebec. p. 131.

HESS, L. L., MELACK, J. M., FILOSO, S., and WANG, Y., 1995, Delineation of inundated areaand vegetation along the Amazon Floodplain with SIR-C synthetic aperture radar.IEEE Transactions on Geoscience and Remote Sensing, 33, 896–903.

JUNK, W. J., 1997, General aspects of floodplain ecology with special reference toAmazonian floodplains. In The Central Amazon Floodplain: Ecology of a pulsingsystem, edited by W. J. Junk (Berlin: Springer), pp. 3–17.

JUNK, W. J., and PIEDADE, M. T., 1997, Plant life in the floodplain with special reference toherbaceous plants. In The Central Amazon Floodplain: Ecology of a pulsing system,edited by W. J. Junk (Berlin: Springer), pp. 147–181.

KELLNDORFER, J. M., and PIERCE, L. L., 1998, Towards consistent regional-to-global-scalevegetation characterization using orbital SAR systems. IEEE Transactions onGeoscience and Remote Sensing, 36, 1396–1411.

LE TOAN, T., RIBBES, F., WANG, L., FLOURY, N., DING, K., KONG, J. A., and FUJITA, M.,1997, Rice crop mapping and monitoring using ERS-1 data based on experiment andmodeling results. IEEE Transactions on Geoscience and Remote Sensing, 35, 41–55.

LOBO, A., CHIC, O., and CASTERAD, A., 1996, Classification of Mediterranean crops with

1834 M. P. F. Costa

multisensor data: per-pixel versus per-object statistics and image segmentation.International Journal of Remote Sensing, 17, 2385–2400.

MELACK, J. M., and FORSBERG, B. R., 2000, Biogeochemistry of Amazon floodplain lakesand associated wetlands. In Biogeochemistry of the Amazon Basin, edited by M. E.McClain, R. L. Victoria and J. E. Richey (Oxford: Oxford University Press),pp. 210–260.

LAUR, H., BALLY, P., MEADOWS, P., SANCHEZ, J. B., and LOPINTO, E., 1996, Derivation ofthe backscattering coefficient so in ERS.SAR.PRI products. ESA/ESRIN, Issue 2,Rev. 2.

NOVO, E. M. L. M., LEITE, F. A., AVILA, J., BALLESTER, V., and MELACK, J., 1997,Assessment of Amazon floodplain habitats using TM/Landsat data. Ciencia eCultura, 49, 280–284.

NOVO, E. M., COSTA, M., MANTOVANI, J. E., and LIMA, I. T., 2002, Relationship betweenmacrophyte stand variables and Radar backscatter at L and C bands – Tucuruıreservoir – Brazil. International Journal of Remote Sensing, 23, 1241–1260.

PIERCE, L. E., ULABY, F. T., SARANDI, K., and DOBSON, M. C., 1994, Knowledge-basedclassification of polarimetric SAR images. IEEE Transactions on Geoscience andRemote Sensing, 32, 1081–1086.

PROISY, C., MOUGIN, E., FROMARD, F., and KARAN, M. A., 2000, Interpretation ofpolarimetric radar signatures of mangrove forest. Remote Sensing of Environment, 71,56–66.

RADAMBRASIL, 1976, Folha SA.21-Santarem. Levantamento de Recursos Naturais, Vol.10.Rio de Janeiro, Brazil.

RICHARDS, J. A., 1986, Remote Sensing Digital Analysis, 1st edn (Berlin: Springer-Verlag).SNEDECOR, G. W., and COCHRAN, W. G., 1980, Statistical Methods (Iowa: Iowa State

University Press).ULABY, F. T., MOORE, R. K., and FUNG, A. K., 1982, Microwave Remote Sensing: Radar

remote sensing and surface scattering and emission theory, Volume II (Norwood, MA:Artech House).

ULABY, F. T., MOORE, R. K., and FUNG, A. K., 1986, Microwave Remote Sensing fromTheory to Applications (Artech House).

WANG, Y., HESS, L. L., FILOSO, S., and MELACK, J. M., 1995, Understanding the radarbackscattering from flooded and non-flooded Amazonian forests: results fromcanopy backscatter modeling. Remote Sensing of Environment, 54, 324–332.

WASSMANN, R., and MARTIUS, C., 1997, Methane emissions from the Amazon floodplain.In The Central Amazon Floodplain, edited by W. J. Junk (Berlin: Springer-Verlag),pp. 137–142.

WORBES, M., 1997, The forest ecosystem of the floodplains. In The Central AmazonFloodplain, edited by W. J. Junk (Berlin: Springer-Verlag), pp. 223–260.


Use of sar satellites for mapping zonation of vegetation communities in the amazon floodplain

Technology