Brown et al..fm

Multitemporal, Moderate-Spatial-Resolution Remote Sensing of Modern Agricultural Production and Land Modification in the Brazilian Amazon

J. C. Brown1

Department of Geography and the Environmental Studies Program, University of Kansas, 1475 Jayhawk Blvd., 223, Lawrence, Kansas 66045-7613

W. E. Jepson Department of Geography, Texas A&M University, MS 3147, College Station,Texas 77843-3147

J. H. Kastens Kansas Applied Remote Sensing Program, Kansas Biological Survey, University of Kansas, 2101 Constant Avenue, Lawrence, Kansas 66047-3759

B. D. WardlowNational Drought Mitigation Center, School of Natural Resources, University of Nebraska-Lincoln, 811 Hardin Hall, 3310 Holdredge Street, P.O. Box 830988, Lincoln, Nebraska 68583-0988

J. M. Lomas Kansas Applied Remote Sensing Program, Kansas Biological Survey, University of Kansas, 2101 Constant Avenue, Lawrence, Kansas 66047-3759

K. P. PriceDepartment of Geography, University of Kansas, 1475 Jayhawk Blvd., 223, Lawrence, Kansas 66045-7613

Abstract: We present an extensive review of the literature on remote sensing andland change in Amazonia as part of a call for new methods to study the recent expan-sion of mechanized annual cropping. Following the review is a presentation of theuse of multitemporal Moderate Resolution Imaging Spectroradiometer (MODIS)250-meter Vegetation Index (VI) data to study processes of intensification of mecha-nized agriculture in Vilhena, Rondônia, Brazil, an Amazonian soy-producing munic-ipality. The case study shows that the high temporal resolution and moderate spatialresolution of the MODIS VI data hold promise for acquiring information necessaryto answer important questions about mechanized agriculture and its relationship todeforestation.

[email protected]

117

GIScience & Remote Sensing, 2007, 44, No. 2, p. 117–148.Copyright © 2007 by V. H. Winston & Son, Inc. All rights reserved.

118 BROWN ET AL.

INTRODUCTION

Deforestation as a result of logging, cattle ranching, and smallholder colonizationhas been the central focus of land-change studies in the Brazilian Amazon. Signifi-cant theoretical and methodological advances have resulted from the study of land useand land cover change (LUCC) in colonist farmer communities along major rainforesthighways. Landsat-based remote sensing analysis linked to longitudinal, field-basedstudy of colonist communities has been the empirical basis informing most integratedland-change science and modeling efforts across the Brazilian Amazon. Research onendogenous causes of colonist land-change patterns, specifically household economicstatus and age-sex profiles, have been used to explain the proximate causes of landcover change, particularly shifts from subsistence farming to small-scale ranching,agro-forestry, changes in fallowing practices, forest burning strategies, and regenera-tion of secondary forest cover (Moran et al., 1994; Moran et al., 1996; Walker andHomma, 1996; McCracken et al., 1999, 2002; Sorrensen, 2000; Walker et al., 2000,2002; Perz, 2001; Perz and Skole, 2003a, 2003b; Walker, 2003).

New agents of environmental change arrived in the region during the 1990s:mechanized commercial farmers. These farmers, their impacts on the environment,and the driving forces behind their decision-making may have wide coverage in pop-ular media, but the process of land change due to mechanized agricultural productionhas yet to attract wide attention in peer-reviewed research (Soares et al., 2004;Morton et al., 2006). Emerging from the dynamic development of Central Brazil’smodern agricultural sector (Soskin, 1988; Leclercq, 1989; Warnken, 1999; Jepson,2006a, 2006b), farmers have begun to expand production in areas typically known asthe Amazon (Hecht, 2005). While it may be an exaggeration to argue that soybeansare “relentless foes of the Amazon forest” (Branford and Freris, 2000; Rohter, 2003),dry forest regions in Mato Grosso and Rondônia are undergoing regional transforma-tion based on modern agro-pastoral production and cultivation of soybeans.

The economy and geography of mechanized agriculture in the Brazilian Amazonis distinct from that of smallholder colonist farming in a number of ways. Mechanizedagriculture, predominantly in the southern region of the Amazon basin, has followedmultiple development trajectories that link to particular historical and economic geog-raphies of private colonization, technological development, and cattle ranching(Brown et al. 2005b; Jepson 2006a, 2006b) that differ from Amazonian colonist settle-ment along the major trunk roads through the tropical forest opened in the 1970s and1980s (Hecht and Cockburn, 1989). Mechanized agriculture in the region emergedfrom two decades of agricultural development in the Cerrado, a neotropical savanna,of Brazil’s Center-West region (Jepson 2003, 2005). Commodity production is depen-dent on high-cost machines and low-wage workers as the necessary labor input. Thisis important for two reasons. First, labor demands are reduced through mechaniza-tion, and second, wage laborers are accessed through the market, not through familialobligation. Therefore, endogenous household factors, such as age-sex structure andcomposition, do not constrain production as they do in subsistence or peasant agricul-ture. Mechanized production requires considerable capital to purchase the necessarymachines, farm implements, technologies, and chemical inputs, and thus farmers areforced to marshal a broad range of political, social, and economic resources frommultiple non-state sources. Finally, the institutional arrangements that underpin

MULTITEMPORAL REMOTE SENSING OF AGRICULTURAL PRODUCTION 119

mechanized land use are based on geographically flexible and highly variable land-leasing contracts similar to other arrangements in more developed countries (Jepson,2003). In contrast, colonist farm production often occurs under precarious land-tenureregimes, which cause the claimants to spend more time defending their title by culti-vating the land themselves than engaging in complex land-leasing contracts (Alston etal., 1999).

We propose a research agenda and methodology to form the empirical basis ofevaluating some basic questions about land changes related to commercial farming inthe Amazon. Are commercial farming land uses directly replacing forest cover, and ifnot, what is the progression of land use strategies in agriculture that lead toward themost intensive annual cropping strategies now found in Amazonia? Answering suchquestions demands that researchers find ways to go beyond the binary classificationof land cover as forest/agriculture to a remote sensing strategy that allows classifica-tion of various levels of intensity of agricultural land use. In effect, this means track-ing forest conversion and changes in land cover once areas are deforested, because itmay take anywhere from one to 25 years for commercial agriculture to be establishedin areas that are deforested (Brown et al., 2005b; Jepson, 2005).

For some years, researchers have proposed similar agendas at conferences (Brownet al., 2005a, DeFries, 2005), but published, peer-reviewed studies are still lacking.One important exception is Morton et al. (2006), who showed the usefulness of Mod-erate Resolution Imaging Spectroradiometer (MODIS) 250 m data to track generalchanges in land cover for Mato Grosso, but no published study to our knowledge hasapplied MODIS to the study of changes in agricultural land use intensity on landsalready deforested. Our study focuses on the task of tracking processes of agriculturalintensification, commonly defined as “higher levels of inputs and increased outputs ofcultivated or reared products per unit areas and time” (Lambin et al., 2003, p. 214).Land use intensification has been central to Amazonian forest conservation concernsas a means to redress extensive cattle ranching and short-fallow production practicesof early colonist farmers, which resulted in land abandonment and advancement ofthe frontier. Agro-forestry systems have been implemented as intensification/forestconservation strategies: agroforestry is supposed to increase on-farm incomes on landalready deforested, thus relieving conversion pressure on surrounding forests (Ander-son, 1990; Anderson and Ioris, 1992; Browder, 1992, 1996; Brondízio and Siqueira,1997; Browder and Pedlowski, 2000).

The question of land intensification under mechanized agriculture has generatednew interest. Two contrasting perspectives define current understanding of land inten-sification through mechanized farming in Amazonia. One view is referred to as the“Borlaug hypothesis.” Norman Borlaug, credited as the “inventor” of the GreenRevolution, has long argued that agricultural intensification saves natural vegetationfrom conversion to agriculture. Widely found in the literature is the theory thatincreasing land use intensification (and increasing crop yields) curbs deforestationand increases secondary forest regeneration because it reduces the demand for farm-land (Rudel and Horowitz, 1993; Southgate, 1998; Rudel, 1999; EMBRAPA, 2002;WWF et al., 2003). A contrasting view, firmly opposed to any mechanized, capital-intensive cultivation in Amazonia, is that the new land use is ultimately unsustainablebecause higher income from higher yields simply encourages expansion intonew lands for profit, and land consolidation under mechanized annual cultivation

120 BROWN ET AL.

may expel small farmers and produce a rural exodus further into the forest frontier(Carvalho, 1999; Branford and Freris, 2000; Fearnside, 2001; Kaimowitz and Smith,2001; Rohter, 2003).

This debate fails to move forward because few studies to date have collected andanalyzed systematic empirical data linking spatially explicit land cover change to landintensification under mechanized agriculture at a regional scale. In the absence of acomprehensive and empirically informed, spatially explicit land change study, oppos-ing sides rely exclusively on untested hypotheses of intensification and forest changeto forward their political agendas. Such deeply ideological positions argue fora normative position on whether mechanized farming should or should not be an“appropriate” land use in the Amazon, ignoring evidence of material processes ofchange.

This paper contributes to land change science in three ways. First, the projectoffers a method to map and analyze land intensification levels of mechanized farm-ing, thus providing an important component to any project that seeks to determineempirically the impact of land use intensification on forest conversion. Second, thepaper contributes to the important but largely neglected study of land modification, acritical process in land change studies (Lambin et al., 2003, pp. 213–216; Geist, 2005,pp. 426–427). Land change studies that focus on human drivers of change would ben-efit from greater consideration of land modification, which could yield importantinformation on the quality and status of land covers. Finally, the paper argues thatmethods to develop spatially explicit land intensification could be scaled up to theregional or continental scale and thus aid in understanding land surface processes instudies of regional and global climate change and biogeochemical cycling.

The importance of land modification (e.g., agricultural intensification) in thestudy of mechanized tropical agriculture therefore demands that the land change stud-ies in the Amazon shift from a near exclusive reliance on Landsat datasets (highspatial resolution) to include MODIS (moderate spatial resolution, high temporal res-olution) datasets. Unlike Landsat, MODIS allows for the detection of subtle changesin land cover using continuous surface attributes at seasonal and inter-annual scales ata moderate spatial resolution. Moreover, as will be shown below, the 250 m resolutionof MODIS pixels is sufficient to link farmers’ decisions to land modification at afield-scale because individual fields of mechanized farmers are most often larger than250 m × 250 m.

We begin our discussion with a comprehensive review of the current state of landchange research in the Brazilian Amazon. We review the extensive and well-devel-oped approaches taken to study logging, colonization, and cattle ranching and arguethat such research is highly dependent on high-spatialresolution datasets, limiting acomprehensive analysis of high-input commercial agriculture, a new driver of landchange in the region. The second section proposes that current high-temporal, moder-ate-spatial-resolution remote sensing, originally applied to analysis of crop-relatedland use change in the U.S. Great Plains using MODIS, can yield crucial informationon the process of land modification within these tropical agricultural landscapes. Wefocus on the application of this method in southern Rondônia, Brazil, a region inwhich highly capitalized farming systems are replacing old pasture lands and someforests with large tracts of soybean, maize, and rice. The areas of conversion there arerelatively small, especially in comparison to areas of conversion in the neighboring


state of Mato Grosso, but it is this reduced area of interest that allows us to generate insitu datasets (field based) and ground data (farmer based) that inform and validate thesynoptic analysis of land change using MODIS. We conclude with a discussion ofhow this shift in sensor, from high-spatial resolution to high-temporal resolution,offers a new research avenue more appropriate for examining a distinctly new driverof land change in the Amazon region.

LAND CHANGE RESEARCH IN THE BRAZILIAN AMAZON

Remote sensing studies of land use and land cover change address wide-rangingtechnological, analytical, and scientific themes generated from research conducted byindividual scientists and international research programs. Three major research linesrepresent the general structure of past investigations using Earth observing satellitesas primary data sources: (1) global and regional land-change mapping and monitor-ing; (2) integrated social science and remote sensing; and (3) spatial modeling of landchange. Scientific literature and findings on the Brazilian Amazon contribute signifi-cantly to each line of research.

Land Change Mapping and Monitoring

The systematic and continuous collection and organization of land observationsusing multiple Earth-observing satellites provides scientists with an internally consis-tent synoptic time series to identify and track land changes at multiple temporal andspatial scales (Geist, 2005, pp. 427–433). The increasing scale, pace, and intensity ofanthropogenic environmental transformations has required the scientific communityto improve monitoring efforts, such as land characterization, land cover, mapping andland use change, at global and regional scales. While high–spatial resolution satellitesystems (~20–60 m) such as Landsat (Multispectral Scanner [MSS], ThematicMapper [TM], and Enhanced Thematic Mapper [ETM+]) and Satellite Probatoired’Observation de la Terre (SPOT) are available, they have not been useful for large-area land monitoring due to incomplete spatial coverage, infrequent temporal cover-age due to cloud cover, data cost, limited computational infrastructure, and impracti-cal data volumes (Cihlar, 2000; DeFries and Belward, 2000, p. 1084). The U.S.National Oceanic and Atmospheric Administration’s (NOAA) Advanced Very HighResolution Radiometer (AVHRR) meteorological satellite (4 km and 8 km GlobalArea Coverage [GAC]) data and 1 km Local Area Coverage [LAC] data) has beenmore useful for large-area land monitoring due to its global, daily repeat coverage.Many scientists have produced global land cover maps using the coarser spatial reso-lution AVHRR data (DeFries and Townshend, 1994; DeFries et al., 1995; Mayaux etal., 1998; DeFries and Belward, 2000; Hansen et al., 2002) to advance the study ofglobal climate change and carbon stocks (DeFries and Belward, 2000; Lepers et al.,2005). However, the increasing availability and decreasing cost of finer spatial resolu-tion global datasets (e.g., MODIS) in addition to new sensors (e.g., Earth Observing-1[EO-1] Advanced Land Imager [ALI]) have provided new land-change and character-ization information to support large-area land monitoring programs (Cihlar, 2000) andnew global land cover maps (Friedl et al., 2002). Morton et al. (2006) have recentlydemonstrated the utility of MODIS 250 m data for characterizing the general land use

122 BROWN ET AL.

types (e.g., cropland, pasture, or not in agricultural production) on recently deforestedlands in the state of Mato Grosso in the Brazilian Amazon.

Concern over rapid deforestation in the late 20th and early 21st centuries has ledto increased investment in research programs that use current satellite datasets tomonitor pan-tropical regions (Mayaux et al., 1998; Achard et al., 2002; Lambin et al.,2003; Mayaux et al., 2005). Land-change monitoring of Africa (Tucker et al., 1985;Mayaux, 1999, 2000) and Southeast Asia (Achard and Estreguil, 1995; Stibig et al.,2001), however, have been comparatively overshadowed by the comprehensiveefforts to monitor South American forests (Eva et al., 2004), with primary attentionand resources applied to the Brazilian Amazon (Malingreau and Tucker, 1990; Skoleand Tucker, 1993; Lucas et al., 2000; Morton et al., 2006).

The most significant international effort to monitor the Brazilian Amazonundoubtedly has been the Large-Scale Biosphere-Atmosphere Experiment in Amazo-nia (LBA). Perhaps the largest regionally focused global change project, the multi-national, interdisciplinary research program led by Brazilian scientists andgovernment, seeks to explain how Amazonia functions as a regional entity, how trop-ical forest conversion, regrowth, and selective logging influence biogeochemicalinteractions with the climate system, and how current regional land cover changesinteract with the Earth system (Davidson and Artaxo, 2004; Geist, 2005). LBA is atransect-based research program designed to cover major climatic variability andextent of land use changes representative of the region. In addition, site-specificinvestigations are developed to target particular land use changes (Keller et al., 2004,pp. S5–S7). While an exhaustive review of LBA research is beyond the scope of thissection (Large-Scale Biosphere Atmosphere Project, 2006), it is important to pointout some of the program’s major findings that have appeared in special journal issuesin the diverse areas of earth-system science (see Jasinski et al., 2005), atmosphericscience and climatology (Avissar et al., 2002; Gash et al., 2004), biology and ecology(Davidson and Artaxo, 2004; Keller et al., 2004), and remote sensing (Roberts et al.,2003).

Across the board, Earth-observing satellites have a central role on data gatheringand land monitoring in the LBA science plan. Clearly the capacity of satellites to pro-vide synoptic time-series data at multiple temporal and spatial scales has contributedto analyses of regional biogeochemical processes. Diverse sensors and methods havebeen employed to study land cover mapping and land use change (Roberts et al.,2003). Indeed, land classification is “the most fundamental application of remotesensing” (Roberts et al., 2003, p. 379). LBA research has advanced land covermapping of wetland and floodplain (Hess et al., 2003) and secondary vegetation(Lucas et al., 2000; Lu et al., 2003) using an array of satellite assets, includingIKONOS and Landsat, and algorithms for classification. Land cover mappingresearch also has significantly contributed to the remote sensing of forest degradationdue to selective logging and fire (Nepstad et al., 1999) and land classification vegeta-tion phenologies of the tropical savanna (cerrado) (Ferreira et al., 2003; Ferreira andHuete, 2004; Ferreira et al., 2004). These studies have supported overarching goals ofdetermining forest dynamics and biomass estimates to allow for integration of thesedata into land-surface climate models. In comparison, LBA-related large-scale moni-toring of land use change has focused primarily on pasture degradation (Asner et al.,2004b) to improve carbon and nutrient cycling estimates (Houghton et al., 2000).


Only a few region-wide studies have mapped forest conversion to agriculture and pas-ture as a means to improve overall carbon budget estimates (Cardille and Foley, 2003;Sakai et al., 2004; Jasinski et al., 2005; Morton et al., 2006).

Integrated Land Change Research

Remote sensing studies that address the human dimensions of change in Amazo-nia draw from diverse social science traditions (anthropology, economics, politicalscience, sociology, and geography) to discern patterns and process of land change(Liverman et al., 1998; Rindfuss et al., 2004). Studies on Amazon small colonistfarming, cattle ranching (Malingreau and Tucker, 1990), logging (Stone and Lefeb-vre, 1998; Asner et al., 2002, 2004a), gold mining (Almeida and Shimabukuro, 2000,2002), extractive reserves (Peralta and Mather, 2000), and biomass burning (Sor-rensen, 2000) have made important contributions to integrated social science–remotesensing analyses for the Brazilian Amazon.

A common thread of integrated land change research on the Amazon is the nearexclusive reliance on high spatial resolution datasets (e.g., Landsat and SPOT). Theattempt to “pixelize the social,” to link human actions and consequences to satelliteimagery through models (Geoghegan et al., 1998), requires robust ground-based stud-ies of land manager decision making to build the basic assumptions and rules of landuse to be studied. In the case of Amazon land use/land cover change (LUCC) studies,the pixel has been explicitly linked to the dominant subjects of study, namely, colonistfarmers and their production system. Representing small colonist farmers as majoragents of change in the region is clearly justified by the large areas they impact. Stud-ies of small farmers, their decision-making and impacts are facilitated by the fact thatthe individual land manager is “fixed” to a property or geographically defined areaover a relatively long period of time. This basic geographic reality has allowedresearchers to link specific social and ecological characteristics with longitudinal landchange, thus allowing the identification of proximate causes and patterns of landchange. High-spatial-resolution data also allow for a more robust assessment of bio-physical factors on land change decisions, such as soil fertility on forest succession inanthropogenic landscapes (Moran et al., 2002). The move toward finer resolutiondata, such as geocoded digital videography, further confirms the commitment toincreasingly higher spatial resolution data (Hess et al., 2002).

Modeling Land Change

Integrated remote sensing/social science has allowed researchers to constructmodels predicting spatial outcomes. Remote sensing data are a major source of inputdata to the land change models, influencing the cell sizes of the models and their out-puts in simulation maps. Remote sensing data also are used to validate model results.Despite methodological difficulties that arise in an integrated social science–GIScience research agenda (Rindfuss et al., 2004), this approach to “pixelizing thesocial” (Geoghegan et al., 1998) or spatial modeling is an emerging area of researchfor the study of Amazonian land change (Parker et al., 2003).

Cellular automata (CA), Markov, and econometric models, where there are noexplicit actors, are typical models developed within the tropical land-change research

124 BROWN ET AL.

community. Studies on the Brazilian Amazon have primarily involved CA and alter-native scenario modeling related to land change and road/infrastructure construction(Soares et al., 2004; Arima et al., 2005), transition scenarios and probabilities (Dale etal., 1994; Soares et al., 2001; Soares et al., 2002), parcel-level land cover change(Evans et al., 2001), and behavior in Amazonian colonization landscapes (Walker etal., 2004).

Agent-based and process-based models are used to project both quantity and pat-tern of change in the Amazon region; however, there is a constant in that their spatialscales are restricted to landscape-level analysis and are exclusively within the contextof logging, road building, and small colonist farming. Research in agent-based mod-eling (ABM), which focuses on the rules of land use decision making, dominatesmodeling research and simulation studies in which scientists have developed numer-ous tools to explore the spatial complexity of coupled human-environment systems,specifically land use outcomes. For example, Deadman et al. (2004) used an ABM,LUCITA (Land Use Change in the Amazon) to look at colonist behavior in theAmazon near Altamira, taking into account burn quality, subsistence requirements,household composition, and soil quality. The rules reflect a particular frameworkof analysis and set of assumptions ranging from utility optimization (neoclassicaleconomics) to life-cycle (demography) to risk minimization (microeconomics).

Evaluation of MODIS 250 m Data for Agricultural Land Cover Characterization

We propose a research agenda in remote sensing that employs a high temporalresolution approach in order to accomplish our goal to measure inter-annual landchange dynamics and its human dimensions. This should allow us to distinguishcrops, pasture, savanna types, and forest types, in addition to measuring change fromone of these classes to another (particularly between pasture and crops) to determineland intensification changes. In our current study, we are interested in measuring theintensity of agricultural land use, gauged by detecting crop types and measuring howmany crops per year are planted on a given piece of land. In this section, we willbriefly review the current status of high-temporal-resolution research using MODIS250 m VI data. In particular we will review the application of MODIS to annual cropproduction (e.g., maize, soybeans, wheat) studies from the U.S. Great Plains, anapproach that supports the study of intensification of mechanized production in theBrazilian Amazon.

MODIS 250 m VI data offer new opportunities for monitoring and characterizingthe land modifications and intensification of mechanized agricultural in the BrazilianAmazon given their unique combination of resolutions and global coverage. Theuse of time-series VI information to estimate several biophysical parameters (e.g.,biomass and leaf area), classify land cover types (DeFries and Townshend, 1994;DeFries et al., 1998; Hansen et al., 2000; Loveland et al., 2000), study vegetationdynamics (e.g., phenology) (Reed et al., 1994), and monitor vegetation conditions(Reed et al., 1996; Jakubauskas et al., 2002) has been demonstrated with multi-temporal (e.g., 10- or 14-day composite) AVHRR data. Most AVHRR-based applica-tions, however, have been limited to characterizing broad-scale land cover patternsand general land cover types due to the data’s coarse spatial resolution. MODIS has a


comparable temporal resolution (16-day composite) to AVHRR, but a substantiallyhigher spatial resolution that should allow for more detailed land cover mapping andmonitoring.

The 250 m bands were specifically added to MODIS to detect human-inducedland cover changes that occur at or near this spatial scale (Townsend and Justice,1988). The value of the 250 m scale for detailed land cover characterization has beendemonstrated by several initial studies, which found promising results for land coverchange detection (Zhan et al., 2002; Morton et al., 2006), percent forest cover(Hansen et al., 2002) and percent crop (Lobell and Asner, 2004) mapping, generalland- cover mapping (Wessels et al., 2004), crop mapping (Wardlow, 2005; Wardlowand Egbert, 2005) and phenology characterization (Wardlow et al., 2006, 2007).

Two MODIS 250 m VI data sets are available, which include the NormalizedDifference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI). TheNDVI is a normalized difference of the red and near-infrared spectral bands,

NDVI = (ρNIR – ρRed) / (ρNIR + ρRed),

where ρRed and ρNIR are the bidirectional surface reflectances recorded for therespective MODIS bands. The NDVI is based on the inverse relationship between theabsorption of red energy by chlorophyll and the reflectance of near infrared energy bythe internal leaf structure (i.e., spongy mesophyll) of the vegetation. The NDVI hasbeen the most widely used of the VIs over the past 25 years and is produced fromMODIS primarily to serve as continuity index to existing 20+ year AVHRR NDVItime series (Justice et al., 1998). The EVI takes the form

EVI = G * (ρNIR – ρRed) / (ρNIR + C1 * ρRed – C2 * ρBlue + L),

where ρBlue, ρNIR, and ρRed are partially atmospherically corrected (ozone absorp-tion) surface reflectances, L is the canopy background adjustment (L = 1), C1 and C2are aerosol resistance term coefficients that use the blue band to correct the red band(C1 = 6 and C2 = 7.5), and G is a gain factor (G = 2.5) (Huete et al., 1994; Huete et al.,1997). The EVI is designed to minimize the effects of the atmosphere (e.g., aerosols)and canopy background (e.g., soil and plant litter) (Huete et al., 1997, 2002) that canlead to non-vegetation–related variations in the time-series data. It is also intended tohave improved sensitivity over high biomass areas, such as the tropical rainforest, andnot saturate like the NDVI in such areas (Huete et al., 1994).

Initial assessments of time-series MODIS EVI and NDVI data found that themultitemporal trajectories of both VIs well represented the seasonal response of gen-eral biome types (e.g., grasslands, shrublands, and forest) (Huete et al., 2002) andspecific crop types (e.g., maize, soybeans, and wheat) (Wardlow, 2005; Wardlow etal., 2007) and were capable of separating the different cover types. The EVI main-tained a lower value than the NDVI throughout the growing season for all cover types(Huete et al., 2002; Wardlow, 2005; Wardlow et al., 2007), which reflects the EVI’sdesign to avoid saturation over high biomass areas that can occur with the NDVI.Huete et al. (2002) found the EVI to be more sensitive than NDVI in certain applica-tions, exhibiting a strong contrast between cover types with different canopy struc-

126 BROWN ET AL.

tures (e.g., broadleaf versus needleleaf forest). Wardlow et al. (2007) and Wardlow(2005), however, found the two VIs to be highly correlated (r > 0.90) across the grow-ing season for all crop types and only had minimal differences in their responses,which were primarily limited to each crop’s senescence phase and outside the grow-ing season. Huete et al. (2002) also found the NDVI to become asymptotic at ~0.9over primary tropical forest in the Brazilian Amazon due to the saturated red bandsignal, while the EVI maintained sensitivity. Wardlow et al. (2007) and Wardlow(2005) found the NDVI to approach an asymptotic level for some crops in Kansas,but did not observe saturation over the 2,000+ field sites evaluated. These results indi-cate that the EVI and NDVI capture similar seasonal events, but their relationship canchange over different cover types and locations. Both VIs appear to be effective formonitoring most spatial and temporal changes in land cover conditions, but theirapplicability for a specific application should be further explored on a case-by-casebasis.

Wardlow (2005) conducted a detailed analysis of time-series MODIS 250 m EVIand NDVI data for regional-scale, crop-related land use/land cover mapping in theU.S. Central Great Plains. An initial pilot study in southwest Kansas found that accu-rate and detailed crop-related land use/land cover (LULC) patterns could be mappedusing either MODIS 250 m VI dataset. A hierarchical-structured series of four cropmaps were produced, in which crops were classified as: (1) crop/non-crop; (2) generalcrop types (alfalfa, summer crops, winter wheat, and fallow); (3) summer crop types(corn, sorghum, and soybeans); and (4) irrigated/non-irrigated crops. Overall andclass-specific accuracies generally ranged between 85% and 95% for these maps.Both VIs were found to produce comparable crop maps in terms of both their classifi-cation accuracy (1–2% difference in overall and class-specific accuracies) and thespatial cropping patterns that were mapped (>90% pixel-level thematic agreement).Based on these findings, this hierarchical crop mapping protocol was further testedacross the state of Kansas using only the MODIS NDVI (Wardlow, 2005; Wardlowand Egbert, 2005), and similar classification results were produced over this larger,more diverse area. The four crop maps at the state level had overall and class-specificaccuracies that ranged between 85% and 95%. The classified areas for each crop classhad a high level of areal agreement (<5% difference) with the United States Depart-ment of Agriculture’s (USDA’s) crop acreage statistics, and their classified patternsreflected the state’s general cropping pattern. In general, the protocol’s performancewas relatively consistent across the state’s range of climatic conditions (i.e., semi-aridto sub-humid), and the cropping patterns of many fields larger than 40 hectares wereretained at 250 m resolution in the maps.

Recent research by Morton et al. (2006) demonstrated the capability to detectdeforestation and classify general agricultural land use activities (e.g., crop produc-tion and cattle pasture) using information extracted from multitemporal MODIS 250m VI data. They noted, however, that the discrimination of single cropped and fallowlands from pasture was problematic due to their similar phenological behaviors. Toaddress this issue, a multi-year land use trajectory approach had to be adopted toimprove the classification accuracies of these land use classes.

The remainder of this paper will illustrate an application of multitemporalMODIS 250 m VI data analysis to the study of mechanized agricultural in the BrazilianAmazon, specifically to measure levels of agricultural land use intensity. We contend


that the application of current methods of MODIS analysis from the U.S. Great Plains(Wardlow, 2005; Wardlow and Egbert, 2005) comprises a potentially useful approachto assess landscape change dynamics of tropical mechanized agriculture.

MODIS-BASED LAND CHANGE RESEARCH IN SOUTHERN RONDÔNIA

Our goal in this section is to show how MODIS data can be used to studychanges in agricultural land use intensity in the Amazon. In Rondônia (Fig. 1), newagricultural development processes are unfolding as highly capitalized farms arereplacing old pasture, croplands, and some forest ecosystems with large tracts ofmechanized production in soybeans, maize, and rice. The establishment of a grainshipping port in Porto Velho on the Rio Madeira, construction of new regional storageand purchasing facilities, and development of new tropical soybean varieties has fur-ther supported agricultural expansion in southern Rondônia. Since the early 1990s,soybean production in Rondônia has grown considerably. First introduced to the stateexperimentally in the early 1980s, the area planted grew dramatically from 4,640 toalmost 30,000 ha between 1997 and 2002 (IBGE, 2003). Previous Landsat analyseshave demonstrated that primary forests are less likely, but certainly not immune, todeforestation for purposes of mechanized agriculture (Brown et al., 2005b).

Fig. 1. Location of the main soybean producing municipalities in southeastern Rondônia,Brazil. Vilhena, the site of our data collection, borders Mato Grosso along the BR-364 highway.

128 BROWN ET AL.

MODIS 250 m VI Data Description and Processing

A time series of MODIS 250 m EVI and NDVI data spanning from the Septem-ber 14, 2000 composite period to the August 29, 2005 composite period (spanning115 total 16-day periods) was acquired for MODIS tile h12v10 (Collection 4) fromthe Land Processes Distributed Active Archive Center (http://lpdaac.usgs.gov/main.asp). Data from both EVI and NDVI were assembled for the study to carry outpreliminary comparisons, but EVI was used for the bulk of our analyses. For eachcomposite period, the VI image data were converted from the Hierarchical DataFormat–Earth Observing System (HDF–EOS) to an ERDAS Imagine image format.The composited images were then stacked sequentially to produce the respective EVIand NDVI time series.

These processing steps were repeated for a time series of MODIS EVI and NDVIdata from both the Terra (MOD13Q1) and Aqua (MYD13Q1) platforms. For each VI,the data from the Terra and Aqua platforms were merged by composite period inorder to produce the most cloud-free time series possible. The rationale was thatcloud-covered areas during the morning overpass of Terra might be cloud free duringthe afternoon overpass of Aqua on some date within the 16-day composite window. Amaximum value compositing (MVC) technique was employed on a pixel-by-pixel,composite period basis to produce the merged Terra-Aqua EVI and NDVI time-seriesdata sets. For each pixel in a given composite period, the maximum VI value betweenthe Terra and Aqua observations was retained in the merged VI time series that wereanalyzed in this study. The MVC was used because cloud cover reduces VI values andthe selection of the maximum value favors the retention of cloud-free observations inthe time series. The cloud cover that remained in the EVI and NDVI datasets wasminimal for most composite periods over the study area analyzed in this research. Asa result, no further processing was performed to reduce cloud cover effects in the twoVI time series.

The stacked time series of EVI and NDVI images were then reprojected from thesinusoidal projection to the geographic projection (WGS84). The native 231 m spatialresolution of the MODIS VI data was retained during resampling in order to maxi-mize the spatial acuity of the imagery. These data, however, are still referred to asMODIS 250 m data throughout the paper to be consistent with the literature, whichuses “250 meters” as the common spatial reference for the higher resolution MODISdata products. Lastly, the Vilhena region was subset for analysis from the time seriesof EVI and NDVI images.

Farmer Interviews and Field Site Analysis

In situ data were collected in the municipality of Vilhena, Rondônia, Brazil inJuly and August 2005 to aid in interpretation of MODIS data. The timing and type ofcropping practices of farmers engaged in mechanized annual cropping were recordedfor individual field sites since 2003. The field-level information was integrated into aGIS as specific polygons with attribute data that were used to aid in the classificationand accuracy assessment of the MODIS time-series data.

Interviews proceeded after obtaining oral consent to participate in the researchproject, following the protocol outlined by Institutional Review Board Guidelines for


the respective universities. Thirteen farmers or farm managers were interviewed,whose area under cultivation in 2005 represented 17,300 ha, or 54%, of mechanizedagriculture in the Vilhena region for 2004. In addition to a basic socio-economicsurvey, each interviewee was asked to identify on a Landsat TM image their fieldswhere mechanized annual cropping had taken place. To obtain field data, fieldwork-ers presented the farmer with a Landsat TM image (229/69, 8/16/2002) of the studyarea at a scale of 1:350,000. Each field, or talhão, indicated by the farmer was a poly-gon in which cropping practices and management were homogenous throughout thearea during a given agricultural year (September through August). These fields wereoutlined on the paper image with a marker. Fields for the 2004–2005 agricultural yearwere marked first, followed by fields for 2003–2004 and earlier as time, farmermemory, or record keeping allowed. If farmers expressed doubt about their ability torecognize a field boundary or report accurate planting and cropping practice informa-tion, that information was discarded. Each field received a unique identifier, and foreach field the following data were collected: date(s) of planting; type(s) of crop (e.g.,soy, corn, millet, rice); brand name and variety of crop; yields; and type of fallow(e.g., allowing weeds to grow or planting a cover crop). Each field outline was thendigitized on-screen over a Landsat TM image using ERDAS Imagine image process-ing software (Leica Geosystems). After converting the digitized field boundaries toshape files, the resulting polygons were used to extract field site–specific VI datafrom the 16-day, 250 m MODIS time series EVI and NDVI composite imagery.

Quantifying Land Modification

We analyzed the data collected and described above to reveal the recent inter-annual dynamics of intensification in agricultural land use for the Vilhena region.Based on our interviews with farmers, we were able to assign any pixel from our insitu field data to an agricultural land use intensity class. For this exercise, each pixelwas assigned to one of two intensity classes to test, in the simplest way, whether itwas possible to classify agricultural intensity, and via change detection, quantify overtime and space any process of intensification or de-intensification. For our purposes,we refer to intensification as the transition of an agricultural field from managementthat is “not soybean-corn double crop” (or “other”) to “soybean-corn double crop.”De-intensification would be from “soybean-corn double crop” to “not soybean-corndouble crop.” Based on our field work and discussions with representatives ofBrazil’s Federal Agriculture Research Agency (EMBRAPA), we confirmed that themost intense use of land in Vilhena is the double-cropping of soybeans followed by aplanting of corn (safrinha) the same agricultural year. All other forms of cropping inVilhena (e.g., single soybean crop, single rice crop, single soybean crop followed by acover crop [most often millet]) represent a level of investment that is much lower thanthe soybean-corn double crop practice.

Through consideration of crop cycles typical of the area, in conjunction withvisual examination of MODIS EVI time series from the sampled field sites, it wasdetermined that the “natural” crop year break for the Vilhena region occurs betweenMODIS periods 16 (August 29–September 13) and 17 (September 14–September 29).Periods 16 and 17 fall in the middle of the dry season, when fields are idle. SinceMODIS composite EVI data are not available prior to February 2000, five “crop”

130 BROWN ET AL.

years of MODIS data (September 2000 to September 2005) were available for theanalysis. The first crop year will be identified as “00-01,” the second as “01-02,” etc.Field site data to be analyzed were collected that correspond to the final three of thesefive years (i.e., 02-03, 03-04, and 04-05).

As our focus is solely on cropland, VI values from the dry season when fields areidle are of marginal use for the research. Thus, we excluded from our analysis theMODIS data from dry-season periods. According to our data from farmer interviews,soy was planted most often from mid-October to mid-November. Moreover, by theend of July, farmers have typically finished harvesting any corn. Figure 2 shows thegrowing-season time periods used for the analysis; the first three and last threeMODIS periods consisting of data collected from July 28–October 31 were excluded.

Cropland Mask and Sampled Field Site Data Description. A cropland maskshapefile was manually extracted from a Landsat scene using ESRI Arc software (seeFig. 3). Attention was focused on the cropland in the immediate vicinity of Vilhena,in roughly a 0.4° × 0.45° box. The manual extraction was done referencing two Land-sat scenes, one from August 13, 2001, and another from July 6, 1996, to assure thatdesignated croplands existed prior to the initial crop year considered (i.e., 00-01).Most sampled field sites were within the area designated as cropland. Specifically,there is a 72% overlap of the sampled field site pixels (1205 of 1681 unique MODISpixels) with the cropland mask (which consists of 5718 pixels).

There is spatial redundancy in the sampled field sites, as 22 successive-year pairsfrom the 82 field-site samples exhibit substantial spatial overlap. In terms of pixels,3106 pixels are used across the three years of field site data: 395 in 02-03, 1133 in 03-04, and 1578 in 04-05. This progression demonstrates a recent-year bias of the sam-pling method, which was not intentional but instead a byproduct of farmers remem-bering and having the quickest access to data from the most recent years. Of the 1681unique pixels, 559 correspond to exactly one sample year, 819 for exactly two years,

Fig. 2. Mean pixel-level EVI for 82 field sites, split into two classes


and 303 are used in all three years. This spatial redundancy will impart an upwardbias to the field site classification accuracy results, but inter-annual EVI variation dueto variations in climate and field management is expected to dominate the spatialredundancy effect, so this bias is expected to be small.

Classification Methods and Accuracy Assessment. Three elementary classifi-cations are simultaneously used for this portion of the study. First, a simple “best fit”classifier is considered. For this classification, the sum of squared errors (SSE) iscalculated between the time series of interest and the two field-data derived, class-specific mean EVI time series (like those shown in Fig. 2). The time series is thenassigned a class according to which template curve provides the better fit as measuredby SSE. For simplicity, refer to this classification procedure as “SSE.” The secondclassifier is similar to the first, except that the correlation statistic is used for classassignment rather than SSE. The time series in question is correlated to the templatecurves for each class, and class is assigned according to the template curve offeringthe highest correlation. This classification procedure is referred to as “CORR.” Thethird classifier involves an elementary Fourier analysis. First- and second-orderharmonic components are calculated for the time series in question. If the second har-monic amplitude is greater than or equal to the first harmonic amplitude, then the time

Fig. 3. The manually delineated cropland mask is shown in the cross-hatched area, over-laid on the Landsat scene from which it was generated.

132 BROWN ET AL.

series is placed in the “soybean-corn double crop” class. We refer to this classificationprocedure as “HARM.”

The above three classification methods are extraordinarily simple compared toother classification techniques commonly used in remote sensing, such as decisiontrees, unmixing models, and ISODATA. The paucity of available field site data pre-cludes the use of more complex classification methods for the present study. All threeclassification methods used in this research are supervised. For SSE and CORR, atime series in question is compared to user-defined, class-specific template timeseries. Unlike SSE and CORR, the HARM classifier has no underlying dependenceon an external parametric “model.” Rather, HARM involves a simple statistical testreasoned to be appropriate for the classification problem at hand.

Although the three classification methods are elementary, they are somewhat dis-tinct with respect to the kind of time-series information that is used for classification.SSE provides a simple time-series proximity test. CORR provides a time-seriesshape- or phase-matching test, and does not computationally depend on proximity.HARM capitalizes on the difference in modality between the two classes considered,as the “soybean-corn double crop” class exhibits strong bimodality that is not foundin the “other” class (see Fig. 2). The final classified maps were based on the majorityvote from the three classification procedures, which is referred to as the “VOTE”method.

Classification accuracies from all four methods (SSE, CORR, HARM, andVOTE) are expected to be approximately 80%, based on an accuracy assessmentusing the available field site data. For this assessment, in-sample accuracies werecomputed, and a block cross validation (CV) analysis was performed to assess therobustness of the observed in-sample accuracies. The purpose of the CV exercise is toinvestigate the level of overfitting of the derived template curves in the in-sampleevaluation, which will produce optimistic in-sample accuracies (this is referred to as“selection bias,” which occurs when the same data are used for both model determi-nation and performance evaluation). For the HARM method, which is entirely objec-tive and whose performance should be free of selection bias, the main benefit of thecross validation exercise is to examine variability of the method across years.

Block cross validation involves regimented splitting of a data set into two sub-sets, one to be used for model training and the other to be used for model performanceevaluation. The field data at hand provide a somewhat natural splitting regimen, aspartitioning can be performed using crop year groupings. First, an approach is takenby which one of the three sampled field data crop years is used for training (i.e., com-puting the two template class curves), and the remaining two years of data are usedfor testing. This method is referred to as CV(2), with the “2” referring to the numberof years that are held out for testing. An alternative CV approach is examined bywhich a unique pair of years is used for template class curve estimation, and theremaining year is used for testing. This method is referred to as CV(1). Finally, usingthe template curves shown in Figure 2, the classification methods were applied to allpixels in the cropland mask. The number of class-specific hectares within the crop-land mask was calculated taking into account that there are ~5.5 ha per MODIS pixel.The final classified maps of agricultural intensification were obtained from maps inwhich the binary classification of agricultural intensity was determined using the“VOTE” method.


RESULTS

Results from the cross validation exercises, along with the results from the in-sample evaluation, are shown in Table 1. As reported earlier, and as indicated by theresults in the last four columns of Table 1, expected accuracy for any of the classifica-tion methods is approximately 80%. More specifically, the observed CV(2) andCV(1) pixel mean accuracies across the four methods are 83.5% (range of [80.4,85.5]) and 83.6% (range of [80.4, 85.8]), respectively. The observed in-sample meanaccuracy is 84.2% (range of [80.4, 87.0]). The “80.4%” lower bound result fromHARM is invariant due to the objectivity of the HARM method and the balanced res-ampling used during the CV exercises. Among the 24 CV iterations (6 data splits × 4methods), the lowest observed accuracy was 76.3%, which occurred when HARMwas applied to just 03-04 data. All other observed accuracies exceeded 79%, withmost (18 of 24) exceeding 82%. Using the template curves shown in Figure 2, thefour classification methods were applied to all pixels in the cropland mask across allyears, with the “VOTE” used to produce a series of agricultural intensity maps withthe binary classification of “soybean-corn double crop” and “not soybean-corn doublecrop.” Based on those maps, we created change detection maps between each agricul-tural year in order to visualize any spatial variability in intensification between years.

Class 1 (soybean-corn double crop) areas from the different classifications aredisplayed in Table 2. Looking across the five study years, results from all methodsdemonstrate a pattern of intensification (as it has been defined for this study, i.e.,progression from “not soybean-corn double crop” to “soybean-corn double crop”).Figure 4A shows the result of the change detection analysis between the 2000–2001and 2004–2005 agricultural years: 11847 ha (37.7% of the total cropland area) wereidentified as “intensified,” 2249.5 ha (7.2%) were “de-intensified,” and 17353 ha(55.2%) were found to have “no change” to their status one way or the other. Figure4B shows individual change detection maps revealing the intensification/de-intensifi-cation that occurred between each crop year during 2000–2005.

Upon examination of the data and the intensification maps, MODIS 250 m time-series data appear of sufficient spatial and temporal resolution to detect land modifi-cation within the study area from one year to the next using our binary classificationof “soybean-corn double crop” and “not soybean-corn double crop.” A high spatialvariability appears to exist in agricultural intensity from one year to the next (Fig.4B). In several fields, it is clear that farmers intensify production in successive yearsand soon after de-intensify, or vice versa. Despite the variability in intensity levelsacross space and time, there is a clear regional trend upward in intensification. By ourestimates, soybean-corn double cropping increased from 5038 ha in the 00-01 cropyear to 14636 ha in the 04-05 crop year within the area of the crop mask (VOTEmethod, Table 2).

CONCLUSION

High temporal resolution data from MODIS 250-m show promise in determiningcropping frequency in mechanized agriculture in the Brazilian Amazon. Such tempo-ral data allow researchers not only to track deforestation in near real time, something

134 BROWN ET AL.

Tabl

e 1.

Cla

ssifi

catio

n A

ccur

acy

Ass

essm

ent:

CV

(2),

CV

(1),

and

In-S

ampl

e R

esul

ts

Trai

ning

ye

ar(s

)Te

stin

gye

ar(s

)

Cla

ss 1

trai

ning

sa

mpl

es:#

fiel

d si

tes

(# p

ixel

s)a

Cla

ss 1

test

ing

sam

ples

:# fi

eld

site

s

(# p

ixel

s)a

Cla

ss 2

trai

ning

sam

ples

:# fi

eld

site

s

(# p

ixel

s)a

Cla

ss 2

test

ing

sam

ples

:# fi

eld

site

s

(# p

ixel

s)a

SSEb

CO

RR

bH

AR

Mb

VO

TEb

02–0

303

–04,

04–

055

(116

)21

(112

3)9

(279

)47

(158

8)84

.582

.879

.084

.1

03–0

402

–03,

04–

056

(403

)20

(836

)22

(730

)34

(113

7)84

.680

.582

.783

.7

04–0

502

–03,

03–

0415

(720

)11

(519

)25

(858

)31

(100

9)88

.386

.879

.887

.8

Ove

rall

CV

(2) a

ccur

acy

26 (1

239)

52 (2

478)

56 (1

867)

112

(373

4)85

.583

.080

.484

.9

02–0

3, 0

3–04

04–0

511

(519

)15

(720

)31

(100

9)25

(858

)83

.880

.280

.983

.0

02–0

3, 0

4–05

03–0

420

(836

)6

(403

)34

(113

7)22

(730

)87

.585

.476

.386

.7

03–0

4, 0

4–05

02–0

321

(112

3)5

(116

)47

(158

8)9

(279

)89

.187

.189

.989

.1

Ove

rall

CV

(1) a

ccur

acy

52 (2

478)

26 (1

239)

112

(373

4)56

(186

7)85

.883

.080

.485

.1

In-s

ampl

e re

sults

26 (1

239)

26 (1

239)

56 (1

867)

56 (1

867)

87.0

83.7

80.4

85.6

a Cla

ss 1

refe

rs to

the

“soy

bean

-cor

n do

uble

cro

p” c

lass

. Cla

ss 2

refe

rs to

the

“oth

er”

clas

s.b A

ccur

acie

s are

repo

rted

as p

erce

ntag

e of

test

ing

pixe

ls th

at w

ere

corr

ectly

cla

ssifi

ed.


INPE (Instituto Nacional de Pesquisas Espaciais) researchers are already doing withtheir Earth Observation programs, but also to determine the dynamic character ofhuman activity replacing those forests. The high-temporal resolution approach out-lined above holds promise for determining spatial and temporal variability of variouspasture and agricultural management practices. Such data, which are potentially scal-able to the entire Amazon Basin, can help add a finer resolution to current efforts in anumber of global and regional land-change mapping and monitoring projects. More-over, this approach will provide more precise information on the land transitionprocess, from land conversion (deforestation) to land modification. As an example,there is a frequently observed, multi-year progression of forestland to soybean field inthe Vilhena region. In particular, single crop rice is often used as a transitory covertype following deforestation, but prior to the planting of soybeans. That newly defor-ested land is not converted immediately to soybeans is due to the presence of post-deforestation ground clutter that interferes with the necessary near-to-ground harvest-ing of soybeans. Examining our field data, we had one field site that demonstratedthis progression in its most compact form: deforestation occurred during the first cropyear, rice was grown during the second year, and single crop soybeans were planted inthe third year. Figure 5 shows the EVI time series corresponding to this field site. It isworth noting that none of the field sites for which deforestation occurred during ourstudy period demonstrated a soybean crop immediately following deforestation.

The effort to offer 250 m resolution data with MODIS makes it especially appli-cable to regional integrated social science and remote sensing projects. The 250 mresolution is sufficiently fine to detect individual field sites in our study region. More-over, as operators in Vilhena and other pioneer regions of the Amazon consolidatetheir mechanized production systems, we expect to find fields that are larger thanthose found in Vilhena, thus facilitating the ability to study changes in these newlandscapes. What our study here did not test, however, is how small fields can bebefore they are no longer useful to develop classifications of various cropping prac-tices using the MODIS 250 m data. Nevertheless, Vilhena’s relatively small fields andour small sample size were sufficient to have achieved informative results. It isbeyond the scope of the current paper to present any analyses of the data we collected

Table 2. Cropland Hectares (and % of total) Designated as “Soybean-Corn Double Crop”a

Crop year

Method 00–01 01–02 02–03 03–04 04–05

SSE 2866 (9.1) 2800 (8.9) 9191 (29.2) 11330 (36.0) 14603 (46.4)

CORR 8019 (25.5) 5885 (18.7) 10571 (33.6) 12474 (39.7) 15010 (47.7)

HARM 8657 (27.5) 6100 (19.4) 10577 (33.6) 16231 (51.6) 15400 (49.0)

VOTE 5038 (16.0) 4257 (13.5) 9598 (30.5) 12111 (38.5) 14636 (46.5)

aPercentages are indicated in parentheses.

136 BROWN ET AL.

from farmers on their socio-economic characteristics. It is possible, however, to linkany of the socio-economic data to the geographic database of our field sites to tracethe relationship of any number of human variables to changes and patterns in agricul-tural intensity levels. For example, we could affirm whether agricultural intensifica-tion after deforestation occurs more quickly on land that is rented or owned byfarmers. Once the human variables important in explaining variation in land modifi-cation are determined, it should be possible, based on current Landsat-based land-change modeling results, to conduct spatial modeling of land change across largerareas according to various possible scenarios.

We referred earlier to the Norman Borlaug hypothesis, which states that agricul-tural intensification leads to conservation of natural forest cover. It is a hypothesisinvoked or refuted by proponents and opponents, respectively, in the current debateabout the impact of soybean agriculture in the Amazon. The methods and our findingsin this preliminary study open the possibility of testing relationships between intensi-fication and forest change in a spatially and temporally explicit approach across local,regional, or Basin-wide scales. With sufficient resources for field work to validateclassifications, it should be possible to amass time-series signatures for the majoragricultural cropping practices across the Brazilian Amazon so that we can produceyearly empirical data on agricultural intensity levels and track deforestation patterns,in a similar vein to research presented by Morton et al. (2006), in order to understand

Fig. 4. A. Intensification of agriculture in Vilhena, Rondônia, from 2000–2005. B (facingpage). Annual intensification of agriculture in Vilhena, Rondônia, from 2000–2005.


the relationship between the two at any scale and across any time period for whichdata are available. The contribution our research agenda brings to studies of agricul-tural intensification is that we distinguish two types of intensification. (1) Verticalintensification is the change in total agricultural production in areas where society isgenerally not concerned about effects of agricultural expansion: non-forest areas.Total production increases in such areas potentially relieve pressure to convert exist-ing forests. (2) Horizontal intensification is change in total production on recentlydeforested land, where mechanized agriculture has directly replaced forests. Thecomparison of total production increases between (1) and (2) gives the most realisticview of mechanized agriculture’s spatial-temporal relationship to forests.

Resolving several issues could aid in developing a Basin-wide monitoring pro-gram for mechanized agriculture and land change. One crucial issue, whether we will

Fig. 4. Continued.

138 BROWN ET AL.

continue to have high-temporal-resolution data at 250 m spatial resolution in thefuture, may or may not be resolved with the proposed ~400 m resolution VIIRS(Visible Infrared Imager/Radiometer Suite) sensor on the National Polar-OrbitingOperational Environmental Satellite System (NPOESS), which is designed to providesome continuity to MODIS. The 380+ m resolution of NPOESS leads to an approxi-mately 2.5-fold increase in pixel area, likely putting the data on the bad side of theagricultural landscape granularity threshold. As demonstrated by Townshend andJustice (1988), the 250 m spatial resolution is near the coarser spatial resolutionthreshold at which most general land cover changes can be resolved from remotelysensed imagery. Regardless of sensor, however, we know little about how the differ-ent vegetation indices vary with particular cropping strategies. There are apparent dif-ferences between EVI and NDVI with respect to the agricultural managementpractices we originally studied and their associated VI profiles. Choosing the bestone, or combination of profiles, to monitor agriculture will require further study. Fig-ure 6 shows mean NDVI and mean EVI for the soybean-corn double cropped pixels.Although the EVI and NDVI profiles from the first crop soybeans are largely similar,the second crop corn VI signatures are more distinct. It would be helpful to determinewhether this effect is due to the presence of particular atmospheric conditions duringthe corn growing season or corn plant physiological characteristics that amplify thedistinction between EVI and NDVI response. Moreover, until our methods can betested in other growing areas across the Basin, we are cautious about claiming themethods can transfer from one region to the next. For example, it remains to be seenwhether a soybean-corn double crop time-series signature from Vilhena can be usedto detect the same cropping practice in another region, such as in north-central MatoGrosso or in the soybean growing areas around Santarém and Paragominas, Pará.Crop timing, seed varieties, and atmospheric interference particular to regions are justsome of the many potential problems we may encounter as we attempt to scale-upour analyses. In addition, areas that have considerable cerrado (tropical savanna) and

Fig. 5. Land in transition: EVI from Field Site #73 across three crop years. This field sitedemonstrates that soybeans can occur on a deforested field site following just one year oftransitory cover (rice, in this case).


pasture will require further validation to distinguish between VIs of these land covers,which are similar in appearance to one another. Certainly, larger sample sizes areneeded to be able to classify particular crops and cropping practices beyond the verybasic binary classification we conducted above.

Mechanized annual cropping in Amazonia is likely to expand, considering therapid growth in world demand for vegetable protein. Thus, this new agriculturalsystem will play a major role in regional land change and land modification. High-temporal-resolution remote sensing appears to be a promising source of Basin-wide,empirical and systematic data to help resolve the intense debates that now existbetween those concerned about soybean expansion into natural areas and promotersof mechanized soybean and other annual cropping in Amazonia.

ACKNOWLEDGMENTS

J. C. Brown and Wendy Jepson would like to offer our thanks for the supportgiven by Donald Sawyer, Director of the Instituto Sociedade, População e Natureza inBrasília. The research affiliation enjoyed with the Instituto made our field research inVilhena, Rondônia possible. Thanks as well to Vicente Godinho of EMBRAPA inVilhena, Rondônia for facilitating our field work in the region. For making us feelwelcome and facilitating our contact with farmers in the region, we would especiallylike to thank Ildo Gubert, Roberto Carlos de Mattos (and their families and friends),and all those in the Prefeitura of Vilhena who took an interest in our work. Thanksespecially to the farmers who participated in this research.

Fig. 6. Mean EVI and NDVI time series obtained by averaging across 1239 soybean-corndouble crop pixels. Note the similarity between the soybean (first mode) and corn (secondmode) components of the NDVI profile, and the corresponding distinction in the EVI profile.The source of this pronounced behavior difference between the EVI and NDVI data remainsunknown.

140 BROWN ET AL.

REFERENCES

Achard, F. and C. Estreguil, 1995, “Forest Classification of Southeast Asia UsingNOAA AVHRR Data,” Remote Sensing of Environment, 54:198–208.

Achard, F., Eva, H. D., Stibig, H. J., Mayaux, P., Gallego, J., Richards, T., and J. P.Malingreau, 2002, “Determination of Deforestation Rates of the World’s HumidTropical Forests,” Science, 297:999–1002.

Almeida, R. A. and Y. E. Shimabukuro, 2000, “Detecting Areas Disturbed by GoldMining Activities through JERS-1 SAR Images, Roraima State, Brazilian Ama-zon,” International Journal of Remote Sensing, 21:3357–3362.

Almeida, R. A. and Y. E. Shimabukuro, 2002, “Digital Processing of a Landsat-TMTime Series for Mapping and Monitoring Degraded Areas Caused by Indepen-dent Gold Miners, Roraima State, Brazilian Amazon,” Remote Sensing of Envi-ronment, 79:42–50.

Alston, L. J., Libecap, G. D., and B. Mueller, 1999, Titles, Conflict, and Land Use:The Development of Property Rights and Land Reform on the Brazilian AmazonFrontier, Ann Arbor, MI: University of Michigan Press, 227 p.

Anderson, A. (Ed.), 1990, Alternatives to Deforestation: Steps toward SustainableUse of the Amazon Rain Forest., New York, NY: Columbia University Press, 281p.

Anderson, A. and E. M. Ioris, 1992, “Valuing The Rain Forest: Economic Strategiesby Small-Scale Forest Extractivists in the Amazon Estuary,” Human Ecology,20:337–370.

Arima, E. Y., Walker, R. T., Perz, S. G., and M. Caldas, 2005, “Loggers and ForestFragmentation: Behavioral Models of Road Building in the Amazon Basin,”Annals of the Association of American Geographers, 95:525–541.

Asner, G. P., Keller, M., Pereira, R., and J. C. Zweede, 2002, “Remote Sensing ofSelective Logging in Amazonia—Assessing Limitations Based on Detailed FieldObservations, Landsat ETM+, and Textural Analysis,” Remote Sensing of Envi-ronment, 80:483–496.

Asner, G. P., Keller, M., and J. N. M. Silva, 2004a, “Spatial and Temporal Dynamicsof Forest Canopy Gaps Following Selective Logging in the Eastern Amazon,”Global Change Biology, 10:765–783.

Asner, G. P., Townsend, A. R., Bustamante, M. M. C., Nardoto, G. B., and L. P. Olan-der, 2004b, “Pasture Degradation in the Central Amazon: Linking Changes InCarbon And Nutrient Cycling With Remote Sensing,” Global Change Biology,10:844–862.

Avissar, R., Silva Dias, P. L., Silva Dias, M. A. F., and C. A. Nobre, 2002, “TheLarge-Scale Biosphere-Atmosphere Experiment in Amazonia (LBA): Insightsand Future Research Needs,” Journal of Geophysical Research-Atmospheres,107 (D20), 8086 [doi: 10.1029/2002JD002704, 2002].

Branford, S. and N. Freris, 2000, “One Great Big Hill of Beans,” The Ecologist,30:46–47.

Brondízio, E. S. and A. D. Siqueira, 1997, “From Extractivists to Forest Farmers:Changing Concepts of Caboclo Agroforestry in the Amazon Estuary,” Researchin Economic Anthropology, 18:233–280.

Browder, J. O., 1992, “The Limits of Extractivism,” BioScience, 42:174–182.


Browder, J. O., 1996, Planting Seeds of Change in the Ashes of the Rainforest: TheRondonia Agroforesty Pilot Project, Blacksburg, VA: Virginia Polytechnic Insti-tute and State University, College of Architecture and Urban Studies.

Browder, J. O. and M. A. Pedlowski, 2000, “Agroforestry Performance On SmallFarms In Amazônia: Findings from the Rondônia Agroforestry Pilot Project,”Agroforestry Systems, 49:63–83.

Brown, J. C., Jepson, W., and K. P. Price, 2005a, “Modern Agriculture on TropicalFrontiers: Rethinking Methods for Land-Change Research,” 6th Open Meeting(OM) of the Human Dimensions of Global Environmental Change ResearchCommunity, Bonn, Germany, 9–13 October.

Brown, J. C., Koeppe, M., Coles, B., and K. Price, 2005b, “Soybean Production andConversion of Tropical Forest in the Brazilian Amazon: The Case of Vilhena,Rondonia,” Ambio, 34:456–463.

Cardille, J. A. and J. A. Foley, 2003, “Agricultural Land-Use Change in BrazilianAmazonia between 1980 and 1995: Evidence from Integrated Satellite and Cen-sus Data,” Remote Sensing of Environment, 87:551–562.

Carvalho, R., 1999, A Amazônia Rumo ao ‘Ciclo da Soja’ (Amazon Headed towardthe Soy Cycle), São Paulo, Brazil, Programa Amazônia, Amigos da Terra,Amazônia Papers, no. 2.

Cihlar, J., 2000, “Land Cover Mapping of Large Areas from Satellites: Status andResearch Priorities,” International Journal of Remote Sensing, 21:1093–1114.

Dale, V. H., Oneill, R. V., Southworth, F., and M. Pedlowski, 1994, “Modeling Effectsof Land Management in the Brazilian Amazonian Settlement of Rondonia,”Conservation Biology, 8:196–206.

Davidson, E. A. and P. Artaxo, 2004, “Globally Significant Changes in BiologicalProcesses of the Amazon Basin: Results of The Large-Scale Biosphere-Atmo-sphere Experiment,” Global Change Biology, 10:519–529.

Deadman, P., Robinson, D., Moran, E., and E. Brondízio, 2004, “Colonist HouseholdDecisionmaking and Land-Use Change in the Amazon Rainforest: An Agent-Based Simulation,” Environment and Planning B–Planning & Design, 31:693–709.

DeFries, R. S., 2005, “Deforestation Monitoring in the Amazon with MODIS Data,”6th Open Meeting (OM) of the Human Dimensions of Global EnvironmentalChange Research Community, Bonn, Germany, 9–13 October.

DeFries, R. S. and A. S. Belward, 2000, “Global and Regional Land Cover Character-ization from Satellite Data: An Introduction to the Special Issue,” InternationalJournal of Remote Sensing, 21:1083–1092.

DeFries, R. S., Hansen, M. C., and J. R. G. Townshend, 1995, “Global Discriminationof Land Cover Types from Metric Derived from AVHRR Pathfinder Data,”Remote Sensing of Environment, 24:459–479.

DeFries, R. S., Hansen, M. C., Townshend, J. R. G., and R. S. Sohlberg, 1998, “Glo-bal Land Cover Classifications at 8km Spatial Resolution: The Use of TrainingData Derived from Landsat Imagery in Decision Tree Classifiers,” InternationalJournal of Remote Sensing, 19:3141–3168.

DeFries, R. S. and J. R. G. Townshend, 1994, “NDVI-Derived Land Cover Classifica-tions at a Global Scale,” International Journal of Remote Sensing, 15:3567–3586.

142 BROWN ET AL.

EMBRAPA (Agricultural Brazilian Research Corporation), 2002, Sistemas deProdução 1: Tecnologias de Produção de Soja—Região Central do Brasil (Pro-duction Systems 1: Soybean Production tEchnologies—Central Region ofBrazil), Brasilia, Brazil: EMBRAPA.

Eva, H. D., Belward, A. S., De Miranda, E. E., Di Bella, C. M., Gond, V., Huber, O.,Jones, S., Sgrenzaroli, M., and S. Fritz, 2004, “A Land Cover Map of SouthAmerica,” Global Change Biology, 10:731–744.

Evans, T. P., Manire, A., de Castro, F., Brondízio, E., and S. McCracken, 2001, “ADynamic Model of Household Decision-Making and Parcel Level LandcoverChange in the Eastern Amazon,” Ecological Modelling, 143:95–113.

Fearnside, P. M., 2001, “Soybean Cultivation as a Threat to the Environment inBrazil,” Environmental Conservation, 28:23–38.

Ferreira, L. G. and A. R. Huete, 2004, “Assessing the Seasonal Dynamics of theBrazilian Cerrado Vegetation through the Use of Spectral Vegetation Indices,”International Journal of Remote Sensing, 25: 1837–1860.

Ferreira, L. G., Yoshioka, H., Huete, A., and E. E. Sano, 2003, “Seasonal Landscapeand Spectral Vegetation Index Dynamics in the Brazilian Cerrado: An Analysiswithin The Large-Scale Biosphere-Atmosphere Experiment in Amazonia(LBA),” Remote Sensing of Environment, 87, 534–550.

Ferreira, L. G., Yoshioka, H., Huete, Y. and E. E. Sano, 2004, “Optical Characteriza-tion of the Brazilian Savanna Physiognomies for Improved Land Cover Monitor-ing of the Cerrado Biome: Preliminary Assessments from An AirborneCampaign over an LBA Core Site,” Journal of Arid Environments, 56:425–447.

Friedl, M. A., Mciver, D. K., Hodges, J. C. F., Zhang, X. Y., Muchoney, D., Strahler,A. H., Woodcock, C. E., Gopal, S., Schneider, A., Cooper, A., Baccini, A., Gao,F., and C. Schaaf, 2002, “Global Land Cover Mapping from MODIS: Algorithmsand Early Results,” Remote Sensing of Environment, 83:287–302.

Gash, J. H. C., Huntingford, C., Marengo, J. A., Betts, R. A., Cox, P. M., Fisch, G.,Fu, R., Gandu, A. W., Harris, P. P., Machado, L. A. T., Von Randow, C.,and M. A. S. Dias, 2004, “Amazonian Climate: Results and Future Research,”Theoretical and Applied Climatology, 78:187–193.

Geist, H. (Ed.), 2005, Our Earth’s Changing Land: An Encyclopedia of Land-Use andLand-Cover Change, Westport, CT: Greenwood Press.

Geoghegan, J., Pritchard, L., Olneva-Himmelberger, Y., Chowdhury, R. R., Sander-son, S., and B. L. Turner II, 1998, “‘Socializing The Pixel’ and ‘Pixelizing TheSocial’ in Land-Use and Land-Cover Change,” in People And Pixels: LinkingRemote Sensing And Social Science. Washington DC: National Academy Press,51–69.

Hansen, M. C., DeFries, R. S., Townshend, J. R. G., and R. Sohlberg, 2000, “GlobalLand Cover Classification at 1 km Spatial Resolution Using a Classification TreeApproach,” International Journal of Remote Sensing, 21:1331–1364.

Hansen, M. C., DeFries, R. S., Townshend, J. R. G., Sohlberg, R., Dimiceli, C., andM. Carroll, 2002, “Towards an Operational MODIS Continuous Field of PercentTree Cover Algorithm: Using AVHRR and MODIS Data,” Remote Sensing ofEnvironment, 83:303–319.

Hecht, S. B., 2005, “Soybeans, Development, and Conservation on the Amazon Fron-tier,” Development and Change, 36:375–404.


Hecht, S. B. and A. Cockburn, 1989, Fate of the Forest: Developers, Destroyers, andDefenders of the Amazon, London, UK and New York, NY: Verso, 266 p.

Hess, L. L., Melack, J. M., Novo, E., Barbosa, C. C. F., and M. Gastil, 2003, “Dual-Season Mapping of Wetland Inundation and Vegetation for the Central AmazonBasin,” Remote Sensing of Environment, 87:404–428.

Hess, L. L., Novo, E., Slaymaker, D. M., Holt, J., Steffen, C., Valeriano, D. M.,Mertes, L. A. K., Krug, T., Melack, J. M., Gastil, M., Holmes, C., and C.Hayward, 2002, “Geocoded Digital Videography for Validation of Land CoverMapping in the Amazon Basin,” International Journal of Remote Sensing, 23:1527–1555.

Houghton, R. A., Skole, D. L., Nobre, C. A., Hackler, J. L., Lawrence, K. T., andW. H. Chomentowski, 2000, “Annual Fluxes of Carbon from Deforestation andRegrowth in the Brazilian Amazon,” Nature, 403:301–304.

Huete, A., Justice, C., and H. Liu, 1994, “Development of Vegetation and Soil IndicesFor MODIS-EOS,” Remote Sensing of Environment, 49:224–234.

Huete, A., Didan, K., Miura, T., Rodriguez, E. P., Gao, X., and L. G. Ferreira, 2002,“Overview of the Radiometric and Biophysical Performance of the MODISVegetation Indices,” Remote Sensing of Environment, 83:195–213.

Huete, A., Liu, H. G., Batchily, K., and W. J. D. V. Leeuwen, 1997, “A Comparison ofVegetation Indices over a Global Set of TM Images for EOS-MODIS,” RemoteSensing of Environment, 59:440–451.

IBGE (Brazilian Institute of Geography and Statistics), 2003, Sistema IBGE de Recu-peração Automática (Automated Database System. Municipal AgriculturalProduction), Brasilia, Brazil, Instituto Brasileiro de Geografia e Estatística [http://www.ibge.gov.br].

Jakubauskas, M. E., Peterson, D. L., Kastens, J. H., and D. R. Legates, 2002, “TimeSeries Remote Sensing of Landscape-Vegetation Interactions in the SouthernGreat Plains,” Photogrammetric Engineering and Remote Sensing, 68:1021–1030.

Jasinski, E., Morton, D., DeFries, R., Shimabukuro, Y., Anderson, L., and M. Hansen,2005, “Physical Landscape Correlates of the Expansion of Mechanized Agricul-ture in Mato Grosso, Brazil,” Earth Interactions, 9 (special issue).

Jepson, W., 2003, A Political Ecology of Land-Use and Land-Cover Change in theCentral Brazilian Savanna since 1970, Los Angeles, CA, University of Califor-nia, Los Angeles, Department of Geography.

Jepson, W., 2005, “A Disappearing Biome? Reconsidering Land-Cover Change in theBrazilian Savanna,” Geographical Journal, 171:99–111.

Jepson, W., 2006a, “Private Agricultural Colonization on the Brazilian Frontier,1970–1980,” Journal of Historical Geography, 32(4):839–863.

Jepson, W., 2006b, “Producing a Modern Agricultural Frontier: Firms and Coopera-tives in Eastern Mato Grosso,” Economic Geography, 82(3):289–316.

Justice, C. O., Vermote, E., Townshend, J. R. G., DeFries, R., Roy, D. P., Hall, D. K.,Salomonson, V. V., Privette, J., Riggs, G., Strahler, A., Lucht, W., Myneni, R.,Knjazihhin, Y., Running, S., Nemani, R., Wan, Z., Huete, A., Vanleeuwen, W.,Wolfe, R., Giglio, L., Muller, J.-P., Lewis, P., and M. Barnesley, 1998, “The Mod-erate Resolution Imaging Spectroradiometer (MODIS): Land Remote Sensing

144 BROWN ET AL.

for Global Change Research,” IEEE Transactions on Geoscience and RemoteSensing, 36:1228–1249.

Kaimowitz, D. and J. Smith, 2001, “Soybean Technology and the Loss of NaturalVegetation in Brazil and Bolivia,” in Agricultural Technologies and TropicalDeforestation. Wallingford, UK and New York, NY: CABI Publishing in associa-tion with Center for International Forestry Research, p. 195–211.

Keller, M., Alencar, A., Asner, G. P., Braswell, B., Bustamante, M., Davidson, E.,Feldpausch, T., Fernandes, E., Goulden, M., Kabat, P., Kruijt, B., Luizao, F.,Miller, S., Markewitz, D., Nobre, A. D., Nobre, C. A., Priante, N., Da Rocha, H.,Dias, P. S., Von Randow, C., and G. L.Vourlitis, 2004, “Ecological Research inThe Large-Scale Biosphere-Atmosphere Experiment in Amazonia: EarlyResults,” Ecological Applications, 14:S3–S16.

Lambin, E. F., Geist, H., and E. Lepers, 2003, “Dynamics of Land-Use and Land-Cover Change in Tropical Regions,” Annual Review of Environment andResources, 28:205–241.

Large-Scale Biosphere Atmosphere Project, 2006, Bibliography [http://lba.cptec.inpe.br/lba/].

Leclercq, V., 1989, “Aims and Constraints of the Brazilian Agro-Industrial Strategy:The Case of Soya,” in The International Farm Crisis, Goodman, D. And M. Red-clift (Eds.), London, UK: Macmillan, 275–291.

Lepers, E., Lambin, E. F., Janetos, A. C., DeFries, R. S., Achard, F., Ramankutty, N.,and R. J. Scholes, 2005, “A Synthesis of Information on Rapid Land-CoverChange for the Period 1981–2000,” BioScience, 55:115–124.

Liverman, D., Moran, E. F., Rindfuss, R. R., and P. C. Stern, 1998, People and Pixels:Linking Remote Sensing and Social Science, Washington DC: National AcademyPress, 244 p.

Lobell, D. B. and G. P. Asner, 2004, “Cropland Distributions from Temporal Unmix-ing of MODIS Data,” Remote Sensing of Environment, 93:412–422.

Loveland, T. R., Reed, B. C., Brown, J. F., Ohlen, D. O., Zhu, Z., Yang, L., and J. W.Merchant, 2000, “Development of a Global Land Cover Characteristics Databaseand IGBP DISCover from 1 km AVHRR Data,” International Journal of RemoteSensing, 21:1303–1330.

Lu, D., Mausel, P., Brondízio, E., and E. Moran, 2003, “Classification of SuccessionalForest Stages in the Brazilian Amazon Basin,” Forest Ecology and Management,181:301–312.

Lucas, R. M., Honzak, M., Curran, P. J., Foody, G. M., Milne, R., Brown, T., and S.Amaral, 2000, “Mapping the Regional Extent of Tropical Forest RegenerationStages in the Brazilian Legal Amazon Using NOAA AVHRR Data,” Interna-tional Journal of Remote Sensing, 21:2855–2882.

Malingreau, J. P. and C. J. Tucker, 1990, Ranching in the Amazon Basin—Large-Scale Changes Observed by AVHRR,” International Journal of Remote Sensing,11:187–189.

Mayaux, P., Achard, F., and J. P. Malingreau, 1998, “Global Tropical Forest AreaMeasurements Derived from Coarse Resolution Satellite Imagery: A Comparisonwith Other Approaches,” Environmental Conservation, 25:37–52.


Mayaux, P., De Grandi, G. and J. P. Malingreau, 2000, “Central African Forest CoverRevisited: A Multisatellite Analysis,” Remote Sensing of Environment, 71:183–196.

Mayaux, P., Holmgren, P., Achard, F., Eva, H., Stibig, H. J., and A. Branthomme,2005, “Tropical Forest Cover Change in the 1990s and Options for Future Moni-toring,” Philosophical Transactions: Biological Sciences, 360:373–384.

Mayaux, P., Richards, T., and E. Janodet, 1999, “A Vegetation Map of Central AfricaDerived from Satellite Imagery,” Journal of Biogeography, 26:353–366.

McCracken, S. D., Brondízio, E. S., Nelson, D., Moran, E. F., Siqueira, A. D., and C.Rodriguez-Pedraza, 1999, “Remote Sensing and GIS at Farm Property Level:Demography and Deforestation in The Brazilian Amazon,” PhotogrammetricEngineering and Remote Sensing, 65:1311–1320.

McCracken, S., Siqueira, A. D., Moran, E., and E. Brondízio, 2002, “Land UsePatterns on an Agricultural Frontier in Brazil,” in Deforestation and Land Use inthe Amazon, Wood, C. H., and R. Porro (Eds.), Gainsville, FL: University ofFlorida Press, 162–192.

Moran, E., Brondízio, E., and S. McCracken, 2002, “Trajectories Of Land Use: Soils,Succession, And Crop Choice,” in Deforestation and Land Use in the Amazon,Wood, C. H., and R. Porro (Eds.), Gainesville, FL: University of Florida, 193–217.

Moran, E., Packer, A., Brondízio, E., and J. Tucker, 1996, “Restoration of VegetationCover in the Eastern Amazon,” Ecological Economics, 18:41–54.

Moran, E. F., Brondízio, E., and P. Mausel, 1994, “Secondary Succession,” Research& Exploration, 10:458–476.

Morton, D. C., DeFries, R. S., Shimabukuro, Y. E., Anderson, L. O., Arai, E., del BonEspirito-Santo, F., Freitas, R., and Morisette, J., 2006, “Cropland ExpansionChanges Deforestation Dynamics in the Southern Brazilian Amazon,” Proceed-ings of the National Academy of Sciences of the USA, 103:39, 14,637–14,641.

Nepstad, D. C., Verissimo, A., Alencar, A., Nobre, C., Lima, E., Lefebvre, P.,Schlesinger, P., Potter, C., Moutinho, P., Mendoza, E., Cochrane, M., and V.Brooks, 1999, “Large-Scale Impoverishment of Amazonian Forests by Loggingand Fire,” Nature, 398:505–508.

Parker, D. C., Manson, S. M., Janssen, M. A., Hoffmann, M. J., and P. Deadman,2003, “Multi-Agent Systems for the Simulation of Land-Use and Land-CoverChange: A Review,” Annals of the Association of American Geographers,93:314–337.

Peralta, P. and P. Mather, 2000, “An Analysis of Deforestation Patterns in The Extrac-tive Reserves of Acre, Amazonia from Satellite Imagery: A Landscape Ecologi-cal Approach,” International Journal of Remote Sensing, 21:2555–2570.

Perz, S. 2001, “From Sustainable Development to ‘Productive Conservation’: ForestConservation Options and Agricultural Income and Assets in The Brazilian Ama-zon,” Rural Sociology, 66:93–112.

Perz, S. and D. L. Skole, 2003a, “Social Determinants of Secondary Forests in theBrazilian Amazon,” Social Science Research, 32:25–60.

Perz, S. G. and D. L. Skole, 2003b, “Secondary Forest Expansion in the BrazilianAmazon and the Refinement of Forest Transition Theory,” Society & NaturalResources, 16:277–294.

146 BROWN ET AL.

Reed, B. C., Brown, J. F., Vanderzee, D., Loveland, T. R., Merchant, J. W., and D. O.Ohlen, 1994, “Measuring Phenological Variability from Satellite Imagery,” Jour-nal of Vegetation Science, 5:703–714.

Reed, B. C., Loveland, T. R., and L. L. Tieszen, 1996, “An Approach for UsingAVHRR Data to Monitor U.S. Great Plains Grasslands,” GeoCarto International,11:13–22.

Rindfuss, R. R., Walsh, S. J., Turner, B. L., Fox, J., and V. Mishra, 2004, “Developinga Science of Land Change: Challenges and Methodological Issues,” Proceedingsof the National Academy of Sciences of the United States of America,101:13,976–13,981.

Roberts, D. A., Keller, M., and J. V. Soares, 2003, “Studies of Land-Cover, Land-Use,and Biophysical Properties of Vegetation in the Large Scale Biosphere Atmo-sphere Experiment in Amazonia,” Remote Sensing of Environment, 87:377–388.

Rohter, L., 2003, “Relentless Foe of The Amazon Jungle: Soybeans,” The New YorkTimes, September 17, 2003 [http://www.nytimes.com].

Rudel, T., 1999, “Critical Regions, Ecosystem Management, and HumanEcosystem Research,” Society & Natural Resources, 12:257–260.

Rudel, T. and B. Horowitz, 1993, Tropical Deforestation: Small Farmers and LandClearing in the Ecuadorian Amazon, New York, NY: Columbia University Press,252 p.

Sakai, R. K., Fitzjarrald, D. R., Moraes, O. L. L., Staebler, R. M., Acevedo, O. C.,Czikowsky, M. J., Da Silva, R., Brait, E., and V. Miranda, 2004, “Land-UseChange Effects on Local Energy, Water, and Carbon Balances in an AmazonianAgricultural Field,” Global Change Biology, 10:895–907.

Skole, D. L. and P. Tucker, 1993, “Tropical Deforestation and Habitat Fragmentationin the Amazon: Satellite Data from 1978 to 1988,” Science, 260:1905–1910.

Soares, B., Alencar, A., Nepstad, D., Cerqueira, G., Diaz, M. D. V., Rivero, S., Solor-zano, L., and E. Voll, 2004, “Simulating the Response of Land-Cover Changes toRoad Paving and Governance along a Major Amazon Highway: The Santarem-Cuiaba Corridor,” Global Change Biology, 10:745–764.

Soares, B. S., Assuncao, R. M., and A. E. Pantuzzo, 2001, “Modeling the SpatialTransition Probabilities of Landscape Dynamics in an Amazonian ColonizationFrontier,” Bioscience, 51:1059–1067.

Soares, B. S., Cerqueira, G. C., and C. L. Pennachin, 2002, “Dinamica—a StochasticCellular Automata Model Designed to Simulate the Landscape Dynamics in anAmazonian Colonization Frontier,” Ecological Modelling, 154:217–235.

Sorrensen, C. L., 2000, “Linking Smallholder Land Use and Fire Activity: ExaminingBiomass Burning in the Brazilian Lower Amazon,” Forest Ecology and Manage-ment, 128:11–25.

Soskin, A. B., 1988, Non-Traditional Agriculture and Economic Development: TheBrazilian Soybean Expansion, 1964–1982, Westport, CN and London, UK:Praeger, 172 p.

Southgate, D., 1998, Tropical Forest Conservation: An Economic Assessment of theAlternatives in Latin America, New York, NY: Oxford University Press, 192 p.

Stibig, H. J., Malingreau, J. P., and R. Beuchle, 2001, “New Possibilities of RegionalAssessment of Tropical Forest Cover in Insular Southeast Asia Using SPOT-


Vegetation Satellite Image Mosaics,” International Journal of Remote Sensing,22:503–505.

Stone, T. A. and P. Lefebvre, 1998, “Using Multi-Temporal Satellite Data to EvaluateSelective Logging in Para, Brazil,” International Journal of Remote Sensing,19:2517–2526.

Townshend, J. R. G. and C. O. Justice, 1988, “Selecting the Spatial Resolution ofSatellite Sensors Required for Global Monitoring of Land Transformations,”International Journal of Remote Sensing, 9:187–236.

Tucker, C. J., Townshend, J. R. G., and T. E. Goff, 1985, “African Land-Cover Classi-fication Using Satellite Data,” Science, 227:369–375.

Walker, R., 2003, “Mapping Process to Pattern in The Landscape Change of the Ama-zonian Frontier,” Annals of the Association of American Geographers, 93:376–398.

Walker, R., Drzyzga, S. A., Li, Y. L., Qi, J. G., Caldas, M., Arima, E., and D. Vergara,2004, “A Behavioral Model of Landscape Change in the Amazon Basin: TheColonist Case,” Ecological Applications, 14:S299–S312.

Walker, R., and A. K. O. Homma, 1996, “Land Use and Land Cover Dynamics in theBrazilian Amazon: An Overview,” Ecological Economics, 18:67–80.

Walker, R., Moran, E., and L. Anselin, 2000, “Deforestation and Cattle Ranching inthe Brazilian Amazon: External Capital and Household Processes,” World Devel-opment, 28:683–699.

Walker, R., Perz, S., Caldas, M., and L. G. T. Silva, 2002, “Land Use and Land CoverChange in Forest Frontiers: The Role of Household Life Cycles,” InternationalRegional Science Review, 25:169–199.

Wardlow, B. D. 2005, The Development Of A Crop-Related Land Use/Land Cover(LULC) Mapping Protocol in the U.S. Central Great Plains Using Time-SeriesMODIS 250-Meter Vegetation Index (VI) Data, Lawrence, KS, Ph.D. dissertation,University of Kansas, Department of Geography.

Wardlow, B. D. and S. L. Egbert, 2005, “State-Level Crop Mapping in the U.S.Central Great Plains Agroecosystem Using MODIS 250-Meter NDVI Data,”Pecora 16 Symposium, Sioux Falls, SD, October 25-27, unpaginated CD-ROM.

Wardlow, B. D., Egbert, S. L., and J. H. Kastens, 2007, “Analysis of Time-SeriesMODIS 250-m Vegetation Index Data for Crop Classification in the U.S. CentralGreat Plains,” Remote Sensing of Environment, in press.

Wardlow, B. D., Kastens, J. H., and S. L. Egbert, 2006, “Using USDA Crop ProgressData for the Evaluation of Greenup Onset Date Calculated from MODIS 250-Meter Data,” Photogrammetric Engineering and Remote Sensing, 72:1225–1234.

Warnken, P. F., 1999, The Development and Growth of the Soybean Industry in Brazil,Ames, IA: Iowa State University, 208 p.

Wessels, K. J., DeFries, R. S., Dempewolf, J., Anderson, L. O., Hansen, A. J., Powell,S. L., and E. F. Moran, 2004, “Mapping Regional Land Cover with MODIS Datafor Biological Conservation: Examples from the Greater Yellowstone Ecosystem,USA and Pará State, Brazil,” Remote Sensing of Environment, 92:67–83.

WWF, CIRAD, and UnB-CDS, 2003 “Prevenção dos Impactos da Produção e Comer-cialização de Grãos na Amazônia Legal (Prevention of the Impacts of Production

148 BROWN ET AL.

and Marketing of Grains in the Legal Amazon),” Paper presented at a conferenceat Hotel San Marco, Brasília, Brazil.

Zhan, X., Sohlberg, R. A., Townshend, J. R. G., Dimiceli, C., Carroll, M. L., Eastman,J. C., Hansen, M. C., and R. S. DeFries, 2002, “Detection of Land CoverChanges Using MODIS 250 m Data,” Remote Sensing of Environment, 83:336–350.

Documents

Brown et al..fm