LCLUC-SYPR Project Description

Contents:

• Challenges and Broader Goals
• SYPR
• Research Questions & General Structure
• Details of the Research Plan
• Project Linkages
• Beyond the Project
• Bibliography

Land-Cover and Land-Use Change (LCLUC) in the Southern Yucatán Peninsular Region (SYPR)

Spatially explicit probability approaches for modeling and projecting deforestation and land conversion linked to remotely sensed imagery

Challenges and Broader Goals

The international research community repeatedly identifies improved understanding of land-use/cover change as pivotal to most global environmental and sustainability research (Bioscience 1994; Meyer & Turner 1994). This same community identifies the conversion and modification of tropical forests and wetlands to be two of most important cover changes because of the implications to potential climate change, biodiversity loss, and sustainable use. The central research needs for these and all other land use/covers are improved data inventory and monitoring, and improved understanding of the processes of change in order to provide more robust models and projections. Such models and projections will not follow from research privileging the natural sciences or focusing on the social "impacts" (e.g., climate change/anomalies) alone because the majority of the change in question involves decisions by land managers (individuals and groups) about what land is to be used, how it is to be used, and for what purposes. These human dynamics and their interplay with biophysical processes are inadequately understood and require major improvements if the goals of the larger community are to be met.

The IGBP-IHDP Core Project on LUCC identifies three interlinked research foci designed to meet these critical needs (Turner et al. 1995). Two of these foci identify the necessity of linking remote sensing/GIS- and ground/field-based approaches in order to achieve the larger needs of the global change community and specifically, improved models and projections. Focus 2 emphasizes approaches anchored in remote sensing-GIS from which can be derived the spatio-temporal coverage of land-use/cover change in question as well as the monitoring of it. Focus 1 emphasizes approaches embedded in fieldwork on the use dynamics of the land manager in order to better understand the behavior of agents making the changes in the first place.

This LCLUC-SYPR project seeks to wed these two foci in a profoundly unique way by "socializing the pixel" and "pixelizing the social" in order to produce new kinds of land use/cover change models linked explicitly to remotely sensed approaches. It applies this approach to one of the most important tropical forests remaining in Mexico and Central America which is undergoing massive land-use/cover changes. As well, it helps to establish a regional Mexican research center with the capacity to undertake long-term monitoring of change in the region. The project thus strikes to the very heart of goals called for by NASA's NRA-96-MTPE-03 and by IGBP-IHDP LUCC.

SYPR

The Southern Yucatán Peninsular Region (SYPR) contains one of the largest expanses of extant humid tropical forests remaining in Mexico-Central America, encompassing all of the southern portions of the states of Quintana Roo and Campeche (Mexico), and the northern Petén (Guatemala) and Belize. The "natural" vegetation of SYPR is dominated by seasonally humid forests on well drained uplands and within large wetlands or bajos (Flores Guido 1987; Lundell 1934). The entire region is karstic and was once the heartland of the ancient Maya who denuded the forests until A.D. 900-1000.

This study focuses on a 10,000 km2 swath across the base of the peninsula, from the Caribbean Sea to the Gulf of Mexico, through the two Mexican states are an area that is now incurring major forest conversion. Save for seasonal extraction of chicle, the study area remained ephemerally used from the time of ancient Maya abandonment until the middle of this century, when low-scale logging of tropical hardwoods began. Major population growth and land-use change did not begin until the end of 1967 when a paved highway connecting the two coasts was built through the region (Rt. 186). This artery opened the forest lands to ejido-migrants (smallholder farmers on communal lands), private ranchers, and ultimately, large-scale development projects during the petro-dollar affected 1980s. Major deforestation ensued, including the clear-cutting of large wetland forests. The failure of these projects as well as concern about the scale of deforestation along the highway, led to the creation in the 1990s of the Calakmul Biosphere Reserve in a portion of the region as well as to NGO- and government-sponsored efforts to decrease deforestation in general. Today the region is central the "Mundo Maya" international development scheme to create an ecotourism-archaeological tourist economy integrated within an increasingly intensifying agrarian sector.

Virtually the entire period of modern deforestation and land intensification is captured on Landsat imagery, either MSS and TM.

Discussions with GCTE (IGBP) indicate that the proposed SYPR work will be added to an "extended" transect scheme for tropical forests in America, thus adding to the overall inputs and outcomes of global transect effort.

Research Questions & General Structure

The major research questions of the project are: What are the pace and scale of deforestation, other land conversion, and use intensification in SYPR since the 1960s more generally and the 1980s specifically: How have these changes affected biotic diversity, ecosystem productivity and biomass, and the structural and compositional complexity of natural communities? What are the major "forcing factors" to which the local land managers respond and how can they be inserted into LCLUC models? And, what improvements in projecting land use and cover can be gained by inserting socioeconomic components into spatially explicit LCLUC models?

To accomplish these tasks requires a complex research plan that integrates remote sensing with ecological and socioeconomic studies in highly innovative and exploratory ways. This project seeks to increase and expand understanding of the modeling and projecting capabilities of the LCLUC program and the LUCC community more broadly. The basic structure of our LCLUC SYPR plan is outlined in Figure 1. The plan is sufficiently complex that its general structure and connections are briefly reviewed before proceeding to foci descriptions.

This project (Fig. 1) uses two parallel and integrative approaches linked to ecological studies to achieve its goals. The Focus 2 approach seeks to create and test a spatially explicit, empirical diagnostic models (Markovian in type) of land-cover change (E) drawn primarily from remote sensing. It begins with the classification (A) of MSS and TM imagery tied to ecological fieldwork and GPS field checks (J) to create a suite of land covers (B) by type and attributes (J). Both the classification and land-cover outcomes account for the role of biophysical data (C, e.g., soils, elevation). Land-cover is tied to the imagery pixel by pixel and serves as one attribute of land-use (D). The probabilities of land-cover change based on three previous time periods (D, T1 3) will be derived and tested against recent imagery (T4) to determine the robustness of this approach.

The Focus 1 approach is developing empirically and theoretically informed behavioral models of different types of land managers/users in the region, coupling general census and archival data (G) by state and municipio with intensive household observations and surveys (F). This evidence is used to create a suite of land-use histories by management type (H). This history is linked to specific pixels (by GPS) identified in Focus 2 work to provide inputs for the ecological study (J). Together, this information (J & H) serves as the central component to assign land-uses by pixel (D) and as the basis from which a spatially explicit, behavioral model of land-use (I) can be generated and tested in a time series similar to that noted for Focus 2.

The two approaches are then be compared to determine the amount of explained variance gained in land-use/cover change by moving from the more common land-cover model A (E) to the novel land-use model B (I). Because the latter approach accounts for variability in land-uses through time, it is not encumbered by the stationarity issues in the Focus 1 Markovian approach and, we suspect it will provide more robust, near-term (â10 yrs) projections (K, L) of land-use and their impacts on land-cover. The amount of improvement gained, if any, is a central issue facing the LCLUC community. It also serves as a basis for assessing scenarios of LCLUC, and thus can serve as a planning tool.

Details of the Research Plan

Focus 2

This research activity will produce a Landsat data inventory from which land-cover maps for the SYPR can be constructed from 1975 to the present, but focusing on the period after 1984, for the purpose of spatially explicit analysis and sustained monitoring of land-conversion and its ecological impacts in the region.

Tasks A-D (Table 1; Fig. 1, A-B) produce a series of cover maps noted in Table 2 generated through image processing of Landsat Thematic mapper (TM) imagery. The previous experience at Clark with Landsat Multispectral Scanner (MSS) data of SYPR for 1975, 1986, and 1990 classified forest, secondary forest, grassland, cropland, bare soils/road, and water. Using TM data of higher spatial and spectral resolution should facilitate the identification of more detailed classes as well as higher overall accuracy in the previously identified level of classes. For example, it is anticipated that the general class of secondary forest will be superseded by additional successional stage classes and that classes of semi-evergreen and seasonal wetland forests can be identified. Such refinement would modify the transition probabilities associated with conversion of different types of forest. It is also anticipated that additional classes of agriculture can be refined, such as rice cultivation, maize-swiddens (milpas), and small-scale plantations, which have distinct transition probabilities from other agricultural practices.

In order to attain the highest degree of information extraction in the image classification, a variety of techniques will be employed. First, the Principal Components Analysis (PCA) utilized in previous MSS analysis yielded two components for visible (forest vs. non-forest) and infrared (water vs. cloud shadow). PCA applied to TM data, however, should provide more detailed distributional information on the components of brightness and greenness as well as an additional component tied to soil and canopy moisture (Bryant et al., 1993 [for MSS]; Lillesand & Kiefer, 1995; Crist & Cicone [1984 for TM]). Second, a two-pass approach to the imagery will be utilized to identify subtle differences in similar classes. This approach, also referred to as cluster-busting (Jensen 1995) masks out readily identified classes obtained in the first pass and applies the second pass only to the more similar pixels in order to tease out the more subtle differences. Evans and colleagues (1992) used this technique with TM data to refine the pine forest class into two subclasses associated with crown density. This approach can be combined with a hybridization of supervised and unsupervised classification procedures which we will be able to utilize because of our extensive field data. Moran and associates (1994) document such 2-pass hybridization in a TM analysis in Amazonia by adding to the first pass discrimination of forest and secondary growth, a second-pass level of information extraction which identified three successional stages of secondary growth. Third, ancillary data on historic land use and cover will be incorporated to weight confusion classes in classification procedures and particularly in converting patterned cover data (patches of forest within grassland) to the appropriate land use class (pasture). In a related manner, ancillary data, such as historical aerial photography and field data keyed to present and historic land use activity, will serve to improve the quality of current (1996-97) land use and cover classifications, and to a lesser extent, the classification of 1980's TM imagery.

The objective of the TM classification is to derive the best possible land cover/use maps for input to the "model A" approach (Fig. 1, E), not to develop image processing techniques to be extended to the rest of the tropics. Thus, reliance upon aerial photography, field data, and other forms of ground truth information will necessarily be more thorough than can be expected in operational projects utilizing remotely sensed data. Once the mechanics of the model are resolved, then the extension of the model to other regions having different exogenous variables and different quality of land cover/use data from which to operate can be evaluated.

Finally, the conceptual progression to the proposed work utilizing TM data will draw upon lessons learned at Clark in working with MSS data. The TM focus restricts the temporal window of Landsat data to be used to 1984 to 1996-97. The MSS data from 1975 to 1992 can be linked to the TM at the most generalized level, and will be investigated for purposes of model exploration only.

Both the classification and land-cover steps (A & B) will be established through the use of biophysical data, such as soil, elevation, and slope, that exist in map form. The resulting land cover categories will serve as input to the ecological study (J), which through GPS siting will verify the authenticity (and variability) of the cover categories. Finally, the land-cover categories provide a "first-order" set of attributions on which to assess land cover (D).

Task E develops a spatially explicit Markov approach to address use/cover change (Lambin 1994; Usher 1981). This information is then used to identify a suite of probabilities that a given pixel (land cover) will remain the same or change and into some other category, and to derive the annual probabilities of change over time based on various heuristics of adjacency. Cross classification of the 1984, 1988, and 1993 (T1-3, D) land-cover and land-use maps provides the transition for each cell in the imagery. From these maps and maps representing biophysical 

TABLE 1: RESEARCH DESIGN AND SCHEDULE

Task letters and color codes refer to Figure 1.

TABLE 2: SEQUENCE OF LAND-USE/COVER MAPS BY DATA SOURCE

variables, the values are extracted for each cell for use in multinomial logistic regression. If multicolinearity is present, a factor analysis will be employed and the first n factors that are meaningful will be extracted. The factor scores are saved as new independent variables and employed in multinomial logistic regressions against a particular cover or use transition. This procedure provides probability outputs for the Markov analysis (Trexler & Travis 1993). Map algebra functions create x number of transition probability maps which are normalized for the time period in question. Comparisons of the transition probability maps are made on a cell-by-cell basis. The highest transition probability is selected for each cell to determine the projected change in land cover and use, although experiments will be made using "joint" transition probabilities where two or more categories score high. The results are assessed against the actual cover-use of the cell for the last date in question. The model is calibrated and re-evaluated again.

Focus 1

This activity will produce a spatially explicit land-management and land-use history of SYPR from which the identified forces operating on the land manager will be developed into a LCLUC model linked to imagery of Focus 2.

Task F-H, D provide the base evidence for the activity. The general history of SYPR (H) will be developed from the archival and various census records (G), focusing on the period from 1967 (the construction of the highway 186) to the present. It will include information on all major national and international policies that directly affect land decisions in the region (e.g., NFTA), such as ejido (villages based on usufruct tenure) policy and resettlement strategies, NGO sponsored agricultural projects, establishment and enforcement of reserve lands, government subsidies, and so forth. The major changes in these "shaping forces" will be used as markers for assessments of land-use practices, such as the recent government policy permitting the purchase and ownership of former ejido lands. Clark's initial work in the area indicates that substantial headway can also be made with the archival evidence in creating land-use maps for several past periods.

The region is dominated by the following land uses/users: ejido smallholders (semi subsistence farmers), NGO-agricultural projects (indebted smallholders), incipient market orchards, small "ranches" (some with absentee owners), government forest (poorly enforced use), and timber concessions on forest land (selective hardwood logging). From these, a comparative sample of land-manager units (mostly households) will be surveyed and repeatedly visited to obtain detailed information on the their land-use histories (F): why they followed which practices; when they changed practices and why; and so forth. Those land managers with extended family tenure in the region will be favored in order to derive the information in question, backed by the archival and census data and the experts from ECOSUR-Chetumal. Their lands will be linked to imagery through the use of GPS, providing essential data for interpretations in the ecological study.

These data and interpretations serve as the "second order" evidence and primary basis for determining land use. Importantly, the data serves as the base for the LCLUC probability model B.  Task I. This model is based on a highly novel approach that attempts to "socialize the pixels" derived from the Focus 2 work. It not only incorporates the use-cover linkage from that focus (E) and the biophysical feedbacks on land use as ascertained from the ecological studies (J), but it also changes the transition probability of each pixel by linking it to the behavior of the land-manager and to socioeconomic conditions affecting the various managers (G-H).

The ultimate goal of this behavioral-based, empirical modeling approach (I) is to predict simultaneously when, why, and where human-induced land-use change occurs. Econometric models are proposed that have the potential to estimate and predict these changes spatially and temporally. These models, based in economic theory, will be more sophisticated than the base Markov analysis (E) and more difficult to developed.

The explanatory (exogenous) variables for the econometric models will be drawn from three analytical realms: the socioeconomic, political-institutional, and biophysical. While there exists a well established theoretical literature to understand land-use decisions, there is considerably less investigation into the biophysical feedback mechanisms on them. To the extent that the utility maximizing or satisfising opportunities of an agriculturist are diminished by the environmental degradation ensuing from deforestation, for example, these mechanisms will be explicitly incorporated into the models.

The estimated models will produce a set of nonstationary, conditional transition probabilities, conceptually similar to the initial Markov modeling, conditioned on features of the pixel of observation and its surroundings. Exogenous factors that affect the desirability of land in a given use include such factors as soil type, access to roads and villages, or international policy affecting input subsidies. The probability of a land-use change might also be affected by the surrounding land uses (e.g., the nearest recently converted pixel). Explanatory variables relating to the individual decision maker, such as wealth, household size, tenure status, are to be included as well as changes over time of the macroeconomic environment, such as interest rates, credit availability, population growth, and government projects.

Temporally dynamic considerations, such as the probability that conversion will depend on the initial state of the parcel or cell, are incorporated into the modeling approach. The cumulative history of the pixel will matter where there is accumulation and/or depreciation of natural, human, and structural capital (e.g., soil depletion). We might also expect a lagged adjustment to past valuations of alternative states. Given the near irreversibility of some land-use conversions, an individual might only respond to a persistent forcing function after a long period to - a threshold - as opposed to sudden responses to a one-time change in socioeconomic conditions. The information gathered during the proposed survey work will assist us in the development of the model specifications. The general models that follow permit a very rich family of specifications and set of hypothesis testing. These socioeconomic, political-institutional, and biophysical exogenous variables are in many cases operative across different spatial scales of analysis, therefore we propose also testing the sensitivity of the model to spatial scale.
A discrete choice probabilistic approach is the first method proposed, as the data on land use is categorical and the choice of land use is mutually exclusive. We propose first to adapt the general dynamic panel-data model of Heckman (1981). We consider that there is a continuous latent (unobserved) random variable reflecting utility or net returns (where "returns" are very generally defined and, for example, could contain issues concerning risk aversion and subsistence farming) from pixel i in land use m at time t. This index of utility or returns, (I_imt) may be represented simply by a function of exogenous variables (X_imt) and parameters (ß_mt) and an error structure (0_imt):

I_imt = f (X_imtt; ß_mt) + 0_imt for all m = 1,..., M

In the most general form, we hypothesize that a pixel i in the landscape that is currently in land use m is converted to land use k at time t if:

I_ikt > I_imt

for all land uses m = 1, ..., M. As we do not observe the latent variable (utility or net returns) but only observe the discrete choice of a particular land use, we can estimate the probability that the utility or net returns from a particular land use is greater under land use k than m

Prob (I_ikt > I_imt)

Let I_ikt = V_ikt(X_ikt) + 0_ikt and V_ikt = ß'_kt X_ikt, (where as above, the X's are the exogenous variables and the ß's are the coefficients to be estimated), then given assumptions on the distribution of the error terms, this probability can be estimated. For example, if we assume the errors are distributed as extreme value then the model is multinomial logit; if we assume the errors come from a normal distribution then the model is multinomial probit.

The application of this style of model is a recent innovation, following such works as that on the effects of roads on deforestation in northern Mexico (Nelson & Hellerstein, 1995), Belize (Chomitz & Gray, 1995), and the Brazilian Amazon (Pfaff, 1996), using spatial data and preliminary spatial econometric modeling. The proposed research would continue work by J. Geoghegan (this project) based on more advanced spatial econometric discrete choice modeling (discussed above) using GIS data to predict land-use change for a region of Maryland (Geoghegan et al., 1996a; 1996b). The choice of explanatory variables will be modified for this project, as the factors influencing land-use change will be different (e.g., the urban-rural wage differential is important in the SYPR but not in Maryland; and allowing for subsistence as well as cash crops in the SYPR).

The second econometric modeling approach that will be explored is hazard or duration models, which estimate the instantaneous probability of a transition between land-use states conditional on the duration of the initial land-use state of a pixel. The first step in this approach is to estimate the impact of a set of independent variables on a dependent variable termed a spell, which in this application is the length of time a pixel is in a given land use. Spells can either be measured in the temporal or spatial dimension, thus permitting an investigation of land-use dynamics from two angles. In the temporal sense, a spell is the length of time that elapses from the beginning of a state until a transition or until measurement is taken (Lancaster, 1990). The spatial counterpart is the Euclidean distance measured between events, where the event is some distinguishing characteristic of the pixel such as the market orientation of an associated farm (Pellegrini & Reader, 1996).

Both modeling approaches - discrete choice and duration, which under certain specifications are very similar in spirit - will be explored. The final choice in models to be used in this project will depend both on data availability, as each approach has different data requirements, as well as on theoretical econometric advances that must be made in order to do the proposed estimation. The time series analysis and test of the resulting models (of kind B) will follow the same logic as that employed for model A (T_1-3 moves to T₄).

The ecological study constitutes Task J. It identifies what ecological systems/land covers (e.g, forest or grassland types) have been and are being converted or modified, pixel-by-pixel, with emphasis on change-impacts on biodiversity, ecosystem productivity and biomass, and structural and compositional complexity. Based on the imagery classification (A-B), specific sets of land-cover pixels, will be examined to determine the vegetation complexes in question. Within these, a stratified sampling design - followed by Harvard Forest researchers at Luquillo LTER (Puerto Rico, tropical forest) - facilitates areal coverage while yielding community- and ecosystem level results (Foster & Bose 1994; Foster 1996; Zimmermann et al. 1996). Using the imagery analysis (B) and DEM and soil data, the vegetation will be stratified by site classes. Within each class randomly assigned plots will be analyzed for vegetation composition, structure and biomass distribution using standard methods (Motzkin et al. 1996; Foster et al. 1997). Multivariate analyses (CANOCO, DECORANA) will be used to evaluate vegetation-site relationships and vegetation composition. These samples will be linked to the specific land-use histories (H) developed for the pixel sets/land covers.

This task is essential for the project, not only verifying the land-cover classifications in way that is consistent with the ecological-user community (e.g., GCTE-IGBP), but determining the systems attributes noted. These attributes constitute the basis for projecting the ecological consequences of land-use/cover change (K-M) for biotic diversity, productivity and biomass, and structural and compositional complexity.

Tasks K-M involves validity and sensitivity tests of modeling approaches A and B (Fig. 1) in order to employ projections and scenario constructions to monitor and assess LCLUC in SYPR, a mission central to ECOSUR-Chetumal. Both models will be compared for their robustness to predict and project land-cover change (L) and their implications for ecological attributes of structure and composition, productivity and biomass, and biodiversity (J). As well, model B will be used to project land-use changes more generally, including scenario development under different assumptions. Owing to the socioeconomic underpinnings of model B, changes in such exogenous factors as NAFTA or government land policies can be inserted as they develop and their influences LCLUC addressed.

Extra funding will sought to hold a closing workshop in conjunction with the modeling efforts and units noted in the next two sections (i) to evaluate the output tasks of this project (K M), (ii) to develop post-project activities to improve the model and its use by ECOSUR-Chetumal, and (iii) to address issues of the model's applicability elsewhere in Latin America.

Project Linkages

LCLUC-SYPR is linked through the GPMI to Carnegie Mellon University's "Center of Excellence for Integrated Assessment for Global Environmental Change." This Center is a focal point of expertise on integrated modeling of environmental change. The GPMI serves as the Center's link to the land-use/cover change community. Thus the LCLUC SYPR modelling described in this proposal will be continually vetted through the Center for critique and improvement as a form of integrative assessment. The Center provides the project with a set of some of the most advanced integrated modelers bridging the biophysical and human dimensions of global change.

As well, LCLUC-SYPR will apply for formal status as an IGBP-IHDP LUCC project, networking the project with a large set of land-use/cover studies elsewhere, including IIASA's LUC modeling effort (Focus 3-LUCC) the steering committee on which the PI serves.

Beyond the Project.

This project ultimately leads to a LCLUC model for SYPR and SYPR, as part of an expanded IGBP transect of tropical American forests, will be linked to similar LCLUC studies in Amazonia and Central America through the auspices of the IGBP-IHDP LUCC program. The aim here is to determine the robustness of a pan-American model of tropical deforestation as part of IGBP-IHDP LUCC's activity to determine the suites of models required to build from regional to continental and global assessments.

Finally, the project enhances ECOSUR-Chetumal's capacity to undertake long-term monitoring of LCLUC in SYPR, a role that unit seeks to undertake as part of its larger mission.

Bibliography

Bioscience. 1994. Special Issue on Global Land use Change. 44 (5).

Bryant, E.S., R.W. Birnie, & K.D. Kimball. 1993. A practical method of mapping forest change over time using Landsat MSS data: a case study from central Maine. 25th International Symposium, Remote Sensing and Global Environmental Change, Graz, Austria, April, 1993.

Chomitz, K. M. and D. A. Gray. 1995. Roads, lands, markets, and deforestation: A spatial model of land use in Belize. World Bank Policy Research Working Paper # 1444.

Crist, E.P. & R.C. Cicone. 1984. Application of the tasseled cap concept to simulated Thematic mapper data. Photogrammetric Engineering and Remote Sensing 50(3):343-352.

Evans, D.L., L. Schoelerman, and T. Melvin. 1992. Integration of information derived from Landsat Thematic Mapper data into a national forest geographic information system. Proceedings, Resource Technology '92, 3rd International Symposium on Advanced Technology in Natural Resources Management, pp. 517-522, Bethesda, MD: American Society of Photogrammetry and Remote Sensing.

Flores Guido, J. S. 1987. Uso de los Recursos Vegetales en la Península de Yucatán: Pasado, Presente y Futuro. INIREB, Xalapa.

Foster, D. R. & E. R. Bose. 1995. Hurricane disturbance regimes in temperate and tropical forest ecosystems. In Wind and Trees. M. Coutts & J. Grace, eds. Cambridge: Cambridge University Press, pp. 305-399.

Foster, D. R., M. Fluuet and E. Boose. 1997. Human or natural disturbance: landscape-scale dynamics of the tropical forests of Puerto Rico. Ecology (in review).

Foster, D. R., B. Slater & G. Motzkin. 1997. Land-use history as long term broad scale disturbance: change in regional forest patterns in New England. Ecological Monographs (forthcoming).

Geoghegan, J., N. Bockstael, and L. Wainger. 1996a. Spatial landscape indices in a hedonic framework: an ecological economics analysis using GIS. Ecological Economics (forthcoming).

Geoghegan, J., N. Bockstael, and D. Lipton. 1996b. The economics of land-use change in the Patuxent watershed. In Land Use and Management in Maryland. eds. E. Brown and C. Buniski. The Aspen Institute and the Center for Agricultural and Natural Resource Policy, University of Maryland, College Park (forthcoming).

Heckman, J. J. 1981. Statistical models for discrete panel data. In Structural Analysis of Discrete Data with Econometric Applications. ed. C. F. Manski and D. McFadden. Cambridge: MIT Press.

Jensen, J.R. 1995. Introductory Digital Image Processing: A Remote Sensing Perspective. New Jersey: Prentice Hall, 2nd Edition.

Keifer, N. 1988. Economic duration data and hazard functions. J. Economic Literature 26: 646 679.
Lambin, E. F. 1994. Modelling Deforestation Processes: A Review. TREES SERIES B: Research Report no. 1. Luxembourg: European Commission.

Lancaster, T. 1990. The Econometric Analysis of Transition Data. Cambridge: Cambridge University Press.

Lillesand, T.M. & R.W. Kiefer. 1995. Remote Sensing and Image Interpretation. New York: John Wiley and Sons, 3rd Edition.

Lundell, C. 1934. Preliminary Sketch of the Phytogeography of the Yucatan Peninsula. Carnegie Institute of Washington, Pub. 436. Washington, D.C. Carnegie Institute.

Meyer, W. B. & B. L. Turner II, eds. 1994. Changes in Land Use and Land Cover: A Global Perspective. Cambridge: Cambridge University Press.

Moran, E. O. , E. Brondizo, P. Mausel, and W. You. 1994. Integrating Amazon vegetation, land-use, and satellite data. Bioscience 44: 329-339.

Motzkin, G., D. R. Foster, A. Allen & J. Harrod. 1996. History and vegetation patterns of sandplains in central New England: the role of land-use and implications for conservation biology. Ecological Monographs 66: 345-365.

Nelson, G.C. and D. Hellerstein. 1995. Do roads cause deforestation: Using satellite images in econometric analysis of land use. Staff Paper 95 E-488, Department of Agricultural Economics University of Illinois, Urbana-Champaign.

Pfaff, A. S. P. 1996. "What drives deforestation in the Brazilian Amazon: evidence from satellite and socioeconomic data. Draft Manuscript, Department of Economics, Columbia University, June 1996.

Pellegrini, P. and S. Reader. 1996. Duration modeling of spatial point patterns. Geographic Analysis, 28, 219-243.

Trexler, J. C. & J. Travis. 1993. Nontraditional regression analyses. Ecology 74: 1629-1637.

Turner II, B. L., D. Skole, S. Sanderson, G. Fischer, L. Fresco, R. Leemans, et al. 1995. LandUse and Land-Cover Change. Science/Research Plan. IGBP Report No. 35/HDP Report No. 7. Stockholm.

Usher, M. B. 1981. Modelling ecological succession, with particular reference to Markovian models.Vegetatio 46: 11-18.

Zimmermann, J., T. Aide, M. Rosario, M. Serrano and J. Herrara. 1996. Effects of land-use history and a recent hurricane on forest structure and composition in the Luquillo Experminent Forest, Puerto Rico. Ecological Applications (in press).

Copyright 2000 Graduate School of Geography, Clark University