Data Cover

Springer ****

Landscape Ecology (****) **:*** ***

DOI 10.1007/s10980-005-4120-z

-1

Research article

Land use land cover conversion, regeneration and degradation in the high

elevation Bolivian Andes

Jodi S. Brandt1,2 and Philip A. Townsend1,3,*

1

Appalachian Laboratory, Center for Environmental Science, University of Maryland, 301 Braddock Road

Frostburg, MD 21532, USA; 2Instituto Interuniversitario Boliviano de Recursos Hidricos, Universidad

Autonomo Juan Misael Saracho, Tarija, Bolivia; 3Current address: Department of Forest Ecology and

Management, University of Wisconsin-Madison, 1630 Linden Drive, Russell Labs, Madison, WI 53706, USA;

*Author for correspondence (email: abo95r@r.postjobfree.com)

Received 2 March 2005; accepted in revised form 14 October 2005

Key words: Classi cation, Deserti cation, Remote sensing, South America, Spectral mixture analysis

(SMA)

Abstract

Regional land-cover change a ects biodiversity, hydrology, and biogeochemical cycles at local, watershed,

and landscape scales. Developing countries are experiencing rapid land cover change, but assessment is

often restricted by limited nancial resources, accessibility, and historical data. The assessment of regional

land cover patterns is often the rst step in developing conservation and management plans. This study

used remotely sensed land cover and topographic data (Landsat and Shuttle Radar Topography Mission),

supervised classi cation techniques, and spectral mixture analysis to characterize current landscape pat-

terns and quantify land cover change from 1985 to 2003 in the Altiplano (2535 4671 m) and Intermediate

Valley (Mountain) (1491 4623 m) physiographic zones in the Southeastern Bolivian Andes. Current land

cover was mapped into six classes with an overall accuracy of 88% using traditional classi cation tech-

niques and limited eld data. The land cover change analysis showed that extensive deforestation,

deserti cation, and agricultural expansion at a regional scale occurred in the last 20 years (17.3% of the

Mountain Zone and 7.2% of the Altiplano). Spectral mixture analysis (SMA) indicated that communal

rangeland degradation has also occurred, with increases in soil and non-photosynthetic vegetation fractions

in most cover classes. SMA also identi ed local areas with intensive management activities that are

changing di erently from the overall region (e.g., localized areas of increased green vegetation). This

indicates that actions of local communities, governments, and environmental managers can moderate the

potentially severe future changes implied by the results of this study.

regimes (Barry and Seimon 2000). In contrast to

Introduction

temperate mountain regions, the South American

South American Andean ecosystems are extremely highlands have a long history of human occupation

vulnerable to climatic factors and anthropogenic and landscape transformation driven by anthro-

activities (Brush 1982). Historically, global climate pogenic activity (Ellenberg 1979; Messerli et al.

change cycles have been shown to profoundly 1997). The majority of its current inhabitants are

in uence shifting vegetation zones and hydrologic subsistence farmers using traditional agricultural

608

and pastoral practices especially adapted to these assessment is di cult in rapidly changing and

sensitive mountain ecosystems. With recent mod- remote regions of developing countries because

ernization and increased population pressure, the limited eld data are available and detailed eld

balance between humans and the land has been analyses are not practical. Recently, the increased

threatened (Ellenberg 1979; Seibert 1983; Baied availability of remotely sensed (RS) data and the

and Wheeler 1993; Hamilton and Bruijnzeel 1997; standardization of satellite image analysis tech-

Roberts et al. 1998; Rundel and Palma 2000). niques has allowed the study of remote locations at

Population growth stimulates an increased need for scales not possible using traditional, eld-intensive

crop and pasture area, resulting in deforestation, methods (Roberts et al. 2003; Cingolani et al.

increased pressure on existing crop and grazing 2004; Townsend et al. 2004). A common applica-

lands, and inappropriate use of marginal land. tion is classi cation of the landscape into land

Modernization results in the abandonment of tra- cover classes, producing a categorical and quanti-

ditional agricultural practices, such as the hus- tative representation of a study area (Lillesand and

bandry of endemic grazing animals (e.g. llama) and Kiefer 1994). A regional classi cation, linked with

communal land management. Road construction, water quality and stream data, can provide infor-

mining, and urbanization are additional causes of mation on the impacts of observed landscape

serious, long-term degradation in the region. condition on water resources (Ballester et al. 2003).

Once degraded, vegetation and soil regeneration Classi cation of past images allows a quanti ca-

are restricted by climatic and topographic condi- tion of trends in land cover change over time,

tions. The result of landscape degradation in arid, which can be used to estimate future conditions

mountainous environments is two-fold: (1) desert- (Hall et al. 1995; Verburg et al. 2002b). A relatively

i cation, or a long-term reduction in the amount or new image analysis technique, spectral mixture

diversity of natural vegetation (UNEP 1992) and analysis (SMA), quanti es landscape components

(2) disturbance of the hydrologic cycle. Vegetation at a sub-pixel scale (Adams et al. 1995; Roberts

cover and land use are important determinants of et al. 1998). Unlike image classi cation, which as-

in ltration and erosive processes of precipitation, signs each pixel to a discrete land cover category

and therefore of biological integrity and stream based on its spectral re ectance, SMA identi es

water quality (Roth et al. 1996; Allan et al. 1997; materials of interest in the image (e.g. vegetation,

Johnson et al. 1997; Wang et al. 1997; Wang and soil, and water) and quanti es the proportion of

Yin 1997). Deforestation results in a decrease in these materials in each pixel, allowing an assess-

in ltration, evapotranspiration, and stream dry ment of changing landscape condition. For exam-

season base ows, and an increase in overland ow, ple, it has been used detect changes in vegetation

erosion, stream peak ow, and stream sediment abundance in semi-arid environments (Elmore

loads (Hamilton and King 1983). Mountain et al. 2000; Okin et al. 2001), pasture condition in

streams in regions with seasonal precipitation the humid tropics (Numata et al. 2003), and

regimes are particularly vulnerable to in uences of rangeland degradation in semi-arid Mediterranean

catchment vegetation cover and land use change as regions (Hostert et al. 2003).

they are smaller and relatively unable to bu er This study integrates a variety of geospatial data

seasonal variability in precipitation, water ow and and methods to assess patterns of land use and

material uxes (Flecker and Feifarek 1994; land cover conversion (LULCC) in a remote and

Monaghan et al. 2000). In a region where people extremely sensitive region of the southeastern

depend on untreated surface water for drinking, Bolivian Andes. The research was designed in

household use, and irrigation, the decrease in dry conjunction with local environmental managers

season stream ows and the contamination of and The Nature Conservancy. The primary

drinking water with harmful pathogens are devas- objectives of the study were to (1) accurately

tating consequences of landscape degradation. characterize past and current land cover patterns

The assessment of regional patterns of current and processes and (2) interpret the results to make

land use and past land cover conversion is the rst the research meaningful from a management

step in developing sound land management plans perspective. Topographic characterization of the

that could prevent broad scale, irreversible study area was performed with Shuttle Radar

watershed and stream degradation. However, this Topography Mission (SRTM) elevation data.

609

Satellite images from 1985 (Landsat TM) and 2003 contains the headwaters for downstream rivers that

(Landsat ETM) were used to determine current supply water for local communities and for Tarija, a

and past land cover patterns, to quantify types and city of 110,000 inhabitants. The arid climate and

rates of change through the past two decades, and mountainous terrain control and limit human land

to estimate future change based on observed rates use. Soils suitable for agriculture are restricted to

of past change (Verburg et al. 2002). the at, moist, fertile soils of the valleys. Forested

areas are scarce, and limited to very steep slopes

where cultivation and grazing are not possible. Low

seasonal rainfall and cold nighttime temperatures

Methods

limit agricultural production and forest regenera-

tion.

Study area

Historically, the region was densely populated

The study site is a 5367 km2 area of the Upper by subsistence farmers using traditional agricul-

Parana River basin in southeastern Bolivia sur- tural and pastoral practices. Recent population

rounding the Sama Mountain Range Biological growth and development driven by the exploita-

Reserve (Sama) (Figure 1). Sama encompasses a tion of natural gas reserves in Tarija is causing

unique area of transition between the Eastern rapid environmental and social change throughout

Andean Cordillera and the Altiplano (high plain). southeastern Bolivia, resulting in an increase in

Elevations climb dramatically from 1400 m at the immigration, infrastructure, and urban develop-

eastern edge of the study area to 4650 m at peaks ment. Due to the environmental constraints and

of the Sama Cordillera in the center of the study the present dense population, land available and

area. The mountainous terrain drives considerable appropriate for increased use and production are

climatic and topographic variability and is extremely limited, and potential for land cover

responsible for the formation of the two physio- conversion is very low. Yet, population and pres-

graphic zones. The Intermediate Mountain and sure on the land continues to grow, resulting in

Valley (Mountain) zone, east of the peaks of the intensi cation of current land use in addition to

Sama Cordillera, is at a lower elevation (2367 m), land use conversion. Throughout the study area,

and has a temperate climate with higher average substantial areas are devoid of vegetation and

annual temperature (18 C), compared to the topsoil. Information from local inhabitants, in

Altiplano zone (3619 m elevation, 11 C). High addition to historical Landsat images and topo-

elevations in the Altiplano are subject to extreme graphic maps developed in the 1970s, indicate that

temperature uctuations with intense solar radia- these areas were, in the recent past, productive

tion during the day and sub-freezing nighttime grasslands and forest, suggesting ongoing forest

conditions. The region has a seasonal precipitation loss, rangeland deserti cation, and agricultural

regime with greater than 85% of the annual pre- expansion. The main threats to land and water

cipitation falling between November and March resources in the region as identi ed by local

(Carpio et al. 2002). Local variability is driven by environmental managers are: (1) the advance of

orographic precipitation. Air systems moving from the agricultural and pastoral frontier caused by

east to west generate relatively high annual pre- population growth, (2) deterioration of the land-

cipitation (1318 mm/year in Calderillas, Figure 1) scape due to unsustainable agricultural practices,

in the Mountain zone (Carpio et al. 2002). As the overgrazing, and the presence of non-native

air crosses the Sama mountain range, it is depleted grazing species; and (3) deforestation due to

of most of its moisture, forming a rain shadow in logging for rewood (Ayala Bluske 1998).

the arid Altiplano, where average annual precipi-

tation ranges from 350 to 500 mm.

Ecosystems in the study area are unique and Data collection and image processing

extremely sensitive. The reserve contains four

distinct eco-regions and is home to several Field datasets for geocorrection, image classi ca-

plant and animal species endemic to the unique tion, and accuracy assessment were collected

combination of climate, altitude and geomorphol- throughout the study area in January March

ogy. In addition to biodiversity, the park region 2004 using a GPS. Easting, northing, altitude, and

610

Figure 1. The Sama Reserve and elevation contours in the study area.

descriptive information were recorded for 35 the eld as homogeneous areas representative of

ground control points (GCPs). To characterize the various land covers present in the study area.

vegetation cover and to test classi cation accuracies, FVPs were randomly chosen prior to eldwork,

85 training points (TPs) and 64 eld veri cation and subset to include only points within 2 km of a

points (FVPs) were collected. TPs were selected in road. When possible, each point in this subset was

611

visited in the eld. However, due to private prop- a DEM developed from 1:50,000 topographic

erty boundaries, rugged terrain, and time con- maps of the study area.

straints, many points could not visited. In these

cases, the point was collected as close to the

coordinates as possible. In addition to FVPs and Maximum likelihood classi cation

TPs, incidental points (IPs) were collected as

training data for image analysis and classi cation. Supervised classi cation of the 2003 and 1985

IPs included a description of the land cover and its images was performed using the maximum likeli-

distance and direction from a GPS or reference hood classi cation (MLC) parametric rule (Jensen

point while driving or walking. 1996). Input layers for classi cation included the

We recorded land use and land cover (LULC) Landsat bands 1, 2, 3, 4, 5, 7 and slope from the

SRTM DEM. The nal classi cation included (1)

observations and took photographs at each FVP

and TP to aid in image processing, classi cation, forest, (2) agriculture (crop and lowland pasture

and accuracy assessment. LULC observations plots), (3) pasture (upland grassland used as

were recorded for an approximate 100 100 m communal rangelands), (4) bare (land with

sampling area surrounding each GPS point extremely sparse or no vegetation), and (5) water.

(Justice and Townshend 1981). Land use and land cover conversion (LULCC)

Two rainy season satellite images, April 1, 1985 analysis was performed by comparing land cover

(Landsat-5 TM sensor) and April 29, 2003, classi cation in 1985 and 2003 for each pixel, and

was summarized for the entire study area, and the

(Landsat-7 ETM+) were used for the land cover

analyses. The 2003 image was geocorrected using Mountain and Altiplano zones separately.

the GCP s collected during eldwork. Root mean Accuracy assessment for the 2003 classi cation

square (RMS) error for the 2003 image was 5.7 m used 112 eld reference points, including the

(x=4.0 m and y=4.0 m) using 29 eld GCPs. The randomly selected FVPs and TPs not used as

1985 image was co-registered to the corrected 2003 training data during image classi cation. The GPS

image. RMS error was less than 1 pixel (21.4 m) coordinates for each reference point were located

(x=19.8 m and y=8.0 m) using 18 GCPs. on the classi cation to determine the land cover

assignment for an individual pixel. Mapped land

DNs were used for all image analyses. Spectral

normalization of the two images was necessary to cover and the actual ( eld reference) land cover at

account for di erences in atmospheric conditions, that point were compared and quantitatively

sensor variation, or other factors. Spectral summarized into a confusion matrix. Detailed eld

normalization was performed according to the data to test the accuracy of the 1985 classi cation

Relative Normalization method (Collins and were not available.

Woodcock 1996) by extracting cell values from

both images of temporally invariant areas of

extreme brightness (sand dunes) and darkness Spectral mixture analysis

(deep water) to encompass the entire re ectance

range. A linear regression model was generated Spectral mixture analysis (SMA), an image anal-

from the extracted pixel values and applied to the ysis technique that quanti es relative abundance

1985 image to calibrate it to the 2003 image. Cloud of speci c landscape components (called end-

and shadow masks for the 1985 and 2003 images members), was conducted to assess landscape

were developed using the software eCognition change and degradation within land cover classes

by determining changes in endmember abundance.

(Baatz et al. 2003), and were applied prior to

Endmembers were selected from the 2003 image

image analysis.

Topographic characterization of the study area using a combination of the pixel purity index (PPI)

was performed with a 90 m-resolution digital ele- (RSI 2000), eld data, and analysis of spectral

vation model (DEM) derived by the US Geological signatures (a pixel s spectral re ectance in each

Survey (USGS) from data collected in February image band). Endmember selection is an iterative

2000 on the Shuttle Radar Topography Mission process. Upon selection of an endmember set,

(SRTM). Gaps in the satellite data caused by SMA is performed to determine the proportion of

each endmember in each pixel. SMA produces a

incomplete SRTM sensor coverage were lled with

612

fraction image for each endmember and an image ridge top, represents the igneous and metamor-

depicting root mean square (RMS) error by pixel. phic formations of the ridges. The GV endmem-

RMS values indicate the ability of the SMA model ber was taken from an agricultural eld in the

to explain the composition of each pixel. To be Tarija Valley. The NPV endmember was taken

used, endmember data must satisfy the following from a at area of senesced grass near the run-

conditions: (1) patterns of fraction images coincide way of the Tarija airport. The shade endmember,

with actual eld conditions; (2) endmember frac- although not a physical material, was necessary

tions for the landscape components of interest are to account for illumination e ects, and was taken

between 0 and 1, indicating that the purest pixels from Laguna Tajzara, the deepest and clearest

were selected as endmembers; (3) endmember water body in the image.

fractions for a pixel sum to one, indicating that the A principal components (PC) transformation

endmember set adequately characterizes the was performed on the Landsat images to demon-

materials in the eld; and (4) the error band shows strate dimensionality of the spectral mixing space

low RMS error for the landscape components of as represented by orthogonal scatterplots of the

interest. rst three PC bands (Figure 3). Plotting the loca-

Target endmembers were: (1) green vegetation tion of endmembers in the spectral mixing space

(GV), (2) non-photosynthetic vegetation (NPV), indicates the endmembers relationship to the

(3) bare soil, and (4) shade. Due to the aridity image pixels and their ability to e ectively model

and geomorphologic variability within the study the image (Small 2004). Ideally, the mixing space is

area, there were three distinct bare soil materials bounded by the endmembers, indicating that the

in the images. Due to limits in the dimensionality endmembers are pure representations of the

of the data (six bands), one of these soil end- di erent materials present on the landscape. In this

members, sand, was excluded from the endmem- analysis, four of the ve endmembers (GV, dark

ber set since it was less of interest than the other soil, light soil, and shade) plot near the edge of

endmembers. The nal set, extracted from the spectral mixing space in at least one of the three

2003 image, included ve endmembers (Figure 2). projections of PC bands 1, 2, and 3 (Figure 3).

The light soil endmember, taken from the exten- However, NPV consistently plots within the

sive formations of severely eroded sedimentary mixing space, possibly introducing error into the

badlands north of Tarija, represents the light mixing model. Thus, the model has the potential to

and erodible lacustrine soils of the Tarija valley. mistakenly represent NPV-dominated pixels as a

The dark soil endmember, extracted from a bare mixture of the other four endmembers. The nal

Figure 2. Endmember spectra used in the spectral mixture analysis.

613

The central location of NPV in the mixing

space, as well as the spectral similarity between

some soils and NPV (Adams et al. 1995; Numata

et al. 2003; Souza et al. 2003), made the selection

of appropriate NPV and soil endmembers the

most challenging aspect of the endmember selec-

tion process. The nal NPV and dark soil end-

members are almost identical in Landsat bands

1 4, but diverge in bands 5 and 7 (Figure 2). In

addition to inspection of spectral signatures of

potential endmembers throughout the endmember

selection process, fraction images were closely

examined to determine separability between the

potential NPV and soil endmembers. For example,

large variability exists within communal range-

lands in the study area. The region surrounding

the city of Tarija is characterized by sparse vege-

tation and severe erosion, whereas the sparsely

populated, mountainous region of the Sama

Cordillera is composed of abundant grasslands.

Fraction images produced from the nal

endmember set closely corresponded to actual eld

conditions, showing high NPV and low soil

fractions in the abundant mountain grasslands,

and low NPV and high soil fractions in the eroded

and sparsely vegetated Tarija valley.

Average RMS for the 1985 image (6.1) was

relatively high compared to the 2003 image (0.81).

Lower RMS values for the 1985 image were ob-

tained with mixing models that used endmembers

extracted from the 1985 image. However, using a

di erent endmember set for each image resulted in

inconsistent fraction results between the 2 years

for temporally invariant areas, such as undis-

turbed forest, sand dunes, and urban features.

Similar to using reference endmembers from a

spectral library, using identical image endmembers

for di erent images allows a direct comparison of

the resulting endmember proportions. Using the

single endmember set extracted from the 2003

image produced consistent endmember propor-

tions for temporally invariant areas for both years,

Figure 3. Scatterplots of the rst three principal components of

which was an important criterion of endmember

the 2003 image showing the dimensionality of the mixing space.

selection. Despite its high average RMS relative to

Endmembers are included to show their relationships within the

that of the 2003 image, the 1985 fraction images

mixing space: S = Shade, GV = Green Vegetation, NPV

= Non-photosynthetic Vegetation, DS = Dark Soil, LS met all other criteria of an acceptable mixing

= Light Soil.

model. Scatter plots showed a slightly negative

relationship between GV proportion and RMS in

endmember sets in the spectral mixing spaces of 1985, and no signi cant relationship in 2003.

both the 1985 image and the 2003 image exhibited Not all image components can be e ectively

similar patterns. modeled using a simple endmember model (Adams

614

et al. 1995; Elmore et al. 2000; Souza et al. 2003). determine if endmember proportions varied

The endmember set was selected to maximize signi cantly according to land cover class. Aver-

model performance in the forest, rangelands and age endmember fraction change for each conver-

agricultural regions of the study. Greater than sion class was calculated to determine endmember

90% of each image was modeled within the changes over time and with land cover conversion.

physically meaningful range from 0 to 1 for all Areas of extreme change were identi ed by calcu-

endmembers. Fraction values that were greater lating the mean and standard deviation for each

than 1.0 or less than zero occurred predominantly fraction change image. Pixels with values between

in areas of high RMS that were not of interest in +1 and 1 standard deviation from the mean were

the study, such as clouds, sand dunes, the heavily considered to be areas of relatively little or no

urbanized city of Tarija, and turbid surface waters. change to account for potential error due to

The 1985 image also showed high RMS error in co-registration (Washington-Allen et al. 1998;

the badlands, an extensive area that had experi- Elmore et al. 2000).

enced severe degradation (conversion from pasture

to bare land) between 1985 and 2003. Inclusion of

an incorrect number of endmembers can increase Results

RMS error (Roberts et al. 1998), and possibly the

inclusion of the light soil endmember (extracted Land use and land cover classi cation and change

from the degraded sedimentary formations of the analysis

2003 image) to model the 1985 image (before

Overall classi cation accuracy of the 2003 classi-

serious vegetation loss and erosion had occurred)

cation was 88%, with a Kappa statistic (KHAT)

caused the relatively high average RMS error of

of 0.82 (Table 1). Although overall accuracy was

the 1985 image. As for the 2003 image, RMS

above the 85% target accuracy (Foody 2002),

errors in 1985 were relatively low for the regions of

forest cover was poorly predicted (40% accuracy)

interest in the SMA, including unconverted

because small forest plots and narrow tree

pasture, forest, and agricultural areas.

boundaries in the intensively used oodplain

All bands were truncated so that any negative

valleys were often misclassi ed as agriculture or

values were assigned to be zero and values greater

pasture. Historical information collected during

than 1.0 were assigned to be 1, and then rescaled

eldwork was considered along with close analysis

so that all endmembers for a pixel summed to 1.

of the image and showed that unchanged features,

For the change analysis, the light soil and dark soil

such as sand dunes, lakes, major roads, the airport

fractions were added together to form a single bare

runway, and the city of Tarija, classi ed accord-

soil band. The shade fractions from both image

ingly in 1985. In addition, land cover changes

dates were compared. An overall increase in shade

characterized in the classi ed image matched

fraction from 1985 to 2003 was evident, possibly

patterns of known deserti cation, agricultural

due to di erences in the sensor or atmospheric

development and forest regeneration (Table 2).

conditions on speci c image dates. Therefore,

Comparison of the classi cations indicates

shade was removed from the images and the other

substantial land cover conversion between 1985

endmembers, GV, NPV, and soil, were rescaled to

and 2003 (Table 2, Figure 4). Net land use con-

sum to 1 by apportioning the shade fraction to the

version as a proportion of each zone was higher in

remaining endmembers. Fraction bands from 1985

the Mountain zone (17.3% of Mountain zone

and 2003 were stacked to perform change detec-

experienced conversion) than the Altiplano (7.2%

tion in endmember fractions on a pixel-by-pixel

converted). The Mountain zone experienced a

basis.

4.7% decrease in forest, a 4.5% increase in bare

Average endmember fraction for each land

land, and a 4.0% increase in agricultural land.

cover class was calculated to determine the

Compared to the land cover in 1985, forest

coincidence of the land cover mapping with the

decreased by more than half, and bare ground

spectral mixture analysis. A non-parametric

almost doubled. The increase in surface water in

equivalent of a 1-way ANOVA (Kruskal Wallis

the Mountain zone (99.2%) is a result of the

test) and pairwise comparisons of the ranked data

construction of the San Jacinto reservoir. Net

(Tukey s studentized range test) were performed to

615

Table 1. Error matrix from comparison of the land cover classi cation and reference data.

Classi cation data Actual LULC reference data User s accuracy

F A P B Total

F 6 6 1.00

A 7 37 1 1 46 0.80

P 2 2 41 45 0.91

B 1 14 15 0.93

Total 15-39-43-15-112

Producer s accuracy 0.40 0.95 0.95 0.93

Overall 0.88

KHAT 0.82

F = Forest, A = Agriculture, P = Pasture, B = Bare.

Table 2. Net changes in land cover as a proportion of each physiographic zone and as a proportion of land cover in 1985 for each

zone.

Land use class Percent of physiographic zone Change in percent Percent change relative to 1985

area conditions 1985 2003 Mountain

4.7% 52%

Forest 9.1% 4.4%

Agriculture 12% 16% 4.0% 35%

4.0% 5%

Pasture 75% 71%

Bare 4.6% 9.1% 4.5% 96%

Water 0.01% 0.14% 0.13% 992%

Altiplano

0.08% 90%

Forest 0.09% 0.01%

2.0% 38%

Agriculture 5.4% 3.4%

Pasture 82% 84% 2.0% 1%

Bare 11% 13% 2.0% 19%

1.1% 80%

Water 1.4% 0.28%

changes in land cover for the Altiplano included a in the Mountain zone (Table 3). 1985 data show

2% loss in agricultural land, a 2% gain in bare that soil fraction was lowest in the forest class

land, and a 2% gain in pasture. Although small as (0.04) and reached a maximum in the bare land

a percentage of the entire Altiplano zone, the rate class (0.52). GV fraction complimented the soil

of change as a proportion of area in 1985 was very fraction well, with the highest value in the forest

high. For example, the reduction in forest was class (0.51), and the lowest value in the bare class

almost 100%. The shrinkage of the Altiplano lakes (0.13). NPV trends are variable because NPV rep-

accounts for the 80% reduction in water. resents many di erent components on the land-

scape, and therefore many di erent processes can

a ect its pattern (Adams et al. 1995; Roberts et al.

1998; Okin et al. 2001). In this study area, NPV on

Spectral mixture analysis

the landscape includes senesced pasture, mature

Spectral mixture analysis (SMA) was performed to annual crops, perennial crops (grape trees), woody

determine relative proportions of green vegetation material from living vegetation (tree branches,

(GV), bare soil, and non-photosynthetic vegetation trunks, etc), and plant material from dead vegeta-

(NPV) throughout the study area. In both 1985 and tion (debris from deforestation or from crop har-

2003, average endmember fractions were signi - vest). Its variability is highly dependent on

cantly di erent (p

Contact this candidate