Land change, which includes changes in land use, land cover, and environmental functions related to human-driven processes, can be described and projected through land change models (LCMs). Models of land change are applied from the level of individual parcels within urban areas to vast expanses of global forests and are used to explain, forecast, and project past, present, and future land and landscape conditions important for decision and policy making at many different scales. Due in part to an explosion in observational and monitoring data on land cover and spatially explicit environmental and socioeconomic data, as well as advances in analytical and technological infrastructure, LCM is now entering a phase with new possibilities for development that could help address a large range of decisions that affect human-environment systems. These advances permit problems to be addressed in greater detail and with better representation of the underlying processes. Researchers continue to push LCMs to treat increasingly complex problems and to support robust measures for addressing them.
These developments are taking place in the context of emerging national and international attention to global climate change and sustainability, be it America’s Climate Choices (NRC, 2011a,b), the U.S. Global Change Research Program (2012), or the new sustainability initiative, Future Earth, of the International Council of Science (Reid et al., 2010). These and other programs seek a more integrative understanding of human-environment systems and cooperation between the science and decision-making communities to tackle critical problems associated with human-environment systems. Having their basis in models that emphasize changes in land use or land cover, LCMs are confronted with new demands as a result of these integrated research and problem-solving goals.
In this context, the time is appropriate to examine LCMs and to determine their applicability for a myriad of scientific and decision-making applications; their fit with current themes, concepts, and data; and the improvements needed to provide the quality of output increasingly expected of them. To this end, the authoring committee was asked by the U.S. Geological Survey and the National Aeronautics and Space Administration to review the present status of spatially explicit LCM approaches and describe future data and research needs so that model outputs can better assist the science, policy, and decision-support communities. They were also asked to provide guidance on the verification strategies and data and research requirements needed to enhance the next generation of models. The committee was asked specifically to (1) assess the analytical capabilities and science and/or policy applications of existing and emerging modeling approaches; (2) describe the theoretical and empirical basis and the major technical, research, and data development challenges associated with each modeling approach; and (3) describe opportunities for improved integration of observation strategies (including ground-based survey, satellite, and remote sensing data) with LCM to improve LCM outputs to better fulfill scientific and decision-making requirements.
In addressing these tasks, the committee necessarily had to place bounds on the scope of this assessment. Rather than reviewing specific models, the committee focused on the need to understand differences among modeling approaches which are implemented in various ways and for different purposes. While this means that many specific models, of which there are likely thousands of instances, are not specifically mentioned in the report, the categories of modeling approaches addressed are used in the vast majority of these models. Additionally, because of the emphasis of the sponsoring agencies on Earth observations, we place greater emphasis on modeling approaches for which these observations are relevant inputs, though not to the exclusion of considering other important data requirements.
Despite existence of a broad literature on modeling coupled with land-transportation systems, we chose to focus on modeling approaches that can use transportation patterns as inputs to land change processes. Approaches to modeling transportation systems themselves were outside the scope of our assessment. Additionally, while significant efforts have been made to develop spatial optimization approaches to be used for developing land use and cover patterns that optimize some specific objectives (what are referred to as normative models), these have not been well integrated with models focused on understanding and forecasting land changes (what we refer to later as positive models), the third chapter of the report identifies opportunities for doing so. The report primarily focuses its assessment on positive models.
Land systems, from cityscapes to landscapes, have long been examined to understand the causes and consequences of their spatial organization (e.g., Beckman, 1972), and various models have been developed to guide in the explicit design of these systems to deliver desirable societal and environmental outcomes (e.g., Waddell et al., 2003). These traditions notwithstanding, attention to changes in land systems and to their modeling has been elevated in importance over the past quarter century as awareness of the role of land systems in environmental change and sustainability has increased (Reid et al., 2010; Rounsevell et al., 2012; Watson et al., 2000). Changes in land systems have significant consequences for local to global climate and environmental change (Foley et al., 2005; Pielke, 2005). For this reason, decisions and policies related to land systems de facto will serve as strategies for mitigating and adapting to these changes and to reaching a more sustainable world (NRC, 2010a,b). Various scientific and practitioner communities seek to address new types of questions and problems with LCMs, such as configuring land systems to ameliorate climate change and developing scalable models of land change with improved capacity to be coupled with other environmental and socioeconomic models addressing specific topics. These efforts have been facilitated by major improvements in the amount and quality of data, methods, and technologies relevant for observing and monitoring, analyzing, and modeling land system change (NRC, 2003, 2008). The resulting mosaic of LCMs is large, tackles various parts of land system change differently, and includes models with different strengths for various science or practitioner communities.
In general, scientists build LCMs to test theories and concepts of land change associated with human and environment dynamics and to explore the implications of these dynamics for future land changes under scenarios that elude real-world observation. The policy and practitioner communities are concerned with guiding land use decision making, for which LCMs provide value by enabling exploration of the possible outcomes of those decisions. These distinctions notwithstanding, LCMs inform and are used by many research and practitioner communities to address topics related to the processes of and outcomes from land change across a wide range of domains of relevance to environmental change and sustainability, including:
1. Land-climate interactions;
2. Water quantity and quality;
3. Biotic diversity, ecosystem function, and trade-offs among ecosystem services;
4. Food and fiber production;
5. Energy and carbon (sequestration); and
6. Urbanization, infrastructure, and the built environment.
LCMs are especially relevant for these and related topics because land systems are expressed spatially as land uses and land covers; these and related attributes result from dynamics in land systems and from a series of human-environment interactions (Turner et al., 2007). LCMs are used to describe, project, and explain the changes in and dynamics of land use and cover, but they can also represent the dynamics in these broader land system interactions. They consider social and biophysical conditions, processes, and variables to address the land system at large, or to target specific social (e.g., vulnerability to hazards) or biophysical (e.g., water quality) outcomes.
Dynamics of land use and land cover are complex, involving multiple social and biophysical processes and outcomes. To account for this complexity, LCMs may be linked or coupled with climatic, ecological, biogeochemical, biogeophysical, and socioeconomic models (Polasky et al., 2008; Robinson et al., 2007), such that other models are an input to the LCM, the LCM is an input to other models, or the models are coupled bidirectionally.
Complexity is present in land system dynamics because of social and biophysical heterogeneity, spatial and social interactions, natural and human adaptation, and feedbacks among system components. This leads to variation in outcomes by geographical location, social group, or ecosystem type, and to nonlinear dynamics that can complicate attempts to validate and predict models. Virtually all LCMs produce outcomes that are spatially explicit, either in terms of land use and cover or specific biophysical (e.g., NPP, leaf area index, roughness) or socioeconomic (e.g., income levels and distributions, age) variables. Complexity typically enlarges the sensitivity of model outcomes to boundary conditions.
Improvement and Challenges
The past two decades have witnessed an expansion and improvement of our understanding of land change dynamics and our ability to project changes into the near-term future through many types of LCMs, especially those drawing on remote sensing data of land cover (as opposed to land use) and directed to changes in the biophysical dynamics of land systems (Agarwal et al., 2002; Lambin, 1997; Parker et al., 2003). Models have improved in their ability to treat spatial, temporal and decision-making complexity as described by Agarwal (2002) and render detailed outputs, from spatial scales of 1 m to 500 km. Model performance is linked to both the quality and resolution of the data employed and the degree of fidelity in representing the processes of land change. Machine learning, data mining, and statistical methods have advanced to improve our ability to identify patterns in the changes we observe. Economic modelers have taken advantage of spatially explicit data sets to build and improve models with varying levels of detail on economic decision making. Agent-based models have increased our capacity to address different types of agents (e.g., households, land managers) and their behaviors, especially when backed by empirical data about
that behavior (Manson and Evans, 2007). Creative approaches have been developed for integrating LCMs across scales, across different approaches, and to other types of models, including biophysical and socioeconomic types.
These advances notwithstanding, LCMs confront a number of limitations that stem from both data constraints and limits to our understanding of underlying processes. Data constraints can be characterized in several ways, including limitations due to the sensor or source (e.g., spatial, spectral, and temporal resolutions), limitations due to development of model inputs from the raw data (e.g., lack of a single ontology for land use, land cover, or other land variables that can be used for classification across all applications), and limitations due to poor coordination of or restricted access to a variety of public and proprietary primary data about the land systems. Process representations are confounded by their complexity, and by temporal nonstationarity in land change processes (e.g., changes in zoning, policy, or environmental conditions), prompting an emphasis on near-term projections and highlighting the possibility that there are very real limits to the level of prediction we can expect LCMs to exhibit (Batty and Torrens 2005). In particular, the further into the future that model outputs are projected or forecast, the greater is the uncertainty of those outputs. In addition, feedback mechanisms within land systems are commonly not represented well, a shortcoming of increasing significance as LCMs address trade-offs of ecosystem services and their socioeconomic consequences. Perhaps most importantly, models are only beginning to account for spatial and social interactions among different land units, land users, and the environmental processes linked to them, especially as affected by the shape and pattern of land units and the network structures of social interactions. Finally, only a few models attempt to treat cross-scale dynamics—ascending or descending spatiotemporal scales of land use and cover and land change processes—rather than treating adjacent scales as boundary conditions.
This report relies on and refers to a number of key concepts that underlie our understanding of both LCMs and the problems LCMs are built to address. Here, we address two major categories of topics, organized around the ideas of pattern and process, and of projection, forecast, and scenario, and define a number of key terms (Box 1.1).
Pattern and Process
Data on land change provide information about patterns that can be described over space and/or time. These patterns of composition and configuration are based on observations of various state variables in the land system (e.g., land use, land cover, land value, and land management). Spatial patterns can be described in the form of maps, or in quantitative measures derived from maps that charac-
Terminology and Definitions
Boundary conditions – attributes and processes affecting the dynamics within a model that are set from outside the model and are not affected by dynamics within the model.
Calibration – parameters are set in a way that a model reproduces outcomes similar to those observed for the specific time and place of a case study.
Diagnosis – developing a degree of trust in the model through verification, calibration, and validation.
Drivers – variables that influence a land change variable (outcome).
Endogenous and exogenous variables – factors that are generated or determined from within a system (endogenous) or outside a system (exogenous) and can be developed within or outside a model, though they often change over time and/or space.
Equifinality – the principle that an observed pattern can be generated by multiple different processes.
Land change – change in land surface characteristics that are usually instigated by human action and that have consequences for environmental system functions.
Land cover – the biophysical qualities of the land surface (e.g., impervious surfaces, vegetation, water, bare soil).
Land system – a set of biophysical processes and human actors and organizations, together with the interactions among them, expressed spatially in the form of a mosaic of land units with different kinds and degrees of land uses and land covers.
Land use – the human intent given to or activity carried out on the land surface (e.g., housing, parks, and cultivation).
Linked, nested, and coupled models – linked model refers to one model result (process) affecting another; nested models are linked models arranged hierarchically by scale; and coupled models allow models to interact dynamically and often in two directions.
terize the organization or configuration of objects or values in the map. In land change contexts, this often involves characterizing the size, shape, distribution, and connectivity or continuity of land cover or land use, attributes that can have significant impacts on human-environment systems (e.g., Chan et al., 2006; Laurence and Williamson, 2001; McGarigal et al., 2012). Temporal patterns can describe changes in the composition and configuration of land over time and can be described graphically as time trends or with derived statistics that characterize the changes, trends, or variability over time. For example, forest transitions, involving loss and then regrowth of forest area within countries, exhibit a regular
Outcome validation – comparison of model outcomes to data from a specific real-world case on those same outcomes.
Path dependence – ultimate or later outcomes of dynamic processes are dependent on the outcomes of earlier iterations of the process. Early outcomes can constrain the choices later on in the process.
Pattern versus process – patterns are descriptions of observed phenomena over some time interval or spatial area, whereas processes are the mechanisms that generate observed patterns.
Projection, forecast, prediction, and scenario – projection refers to a description of a future land system and pathway leading to it; forecast or prediction to the most likely projection; and scenario to a plausible description of a possible future state of a land system.
Scale – the extent and resolution of a variable or model in time, space, thematic categories, or organizations.
Scale dynamics – the process interactions across spatiotemporal-organizational scale.
Spatial interactions – the relationships between variables and processes across geographical locations, often described as pairwise relationships or spatial patterns.
Spatially explicit model – a model in which the data and outputs are specified for some set of geographical locations (e.g., from pixels to regions or continents).
Stationarity – processes during one time period or in one location are the same as those in the subsequent time period or other locations (antonym: nonstationarity).
Structural validation – evidence that the processes in a model accurately reflect those operating in a specific real-world situation.
Verification (debugging) – demonstration that the model’s code expresses the intention of the modeler.
temporal pattern that some in the land change community have sought to explain (e.g., Rudel, 1998, 2005).
To explain observed land change patterns or trends, land change science seeks to understand process. This understanding can be represented with degrees of formality varying from informal conceptual models to formal mathematical or computational models. Stochastic aspects of this understanding might be included with otherwise deterministic processes to represent uncertainty and statistical variability in system behavior.
These distinctions are important because modeling approaches have been developed to support science and decision making in land change for a variety of
reasons. Two broad purposes of models are to predict/project (PP) and to explain/learn (EL). While the two categories of purposes need not be mutually exclusive, PP can be carried out without the goal of gaining insight into fundamental system behavior that is the aim of EL.
Likewise, it is possible to use models for EL in ways that do not provide particularly strong PP capabilities. Models that use different structures or parameter values may fit observed data equally well, a circumstance referred to as equifinality (Beven and Freer, 2001). A model can provide a good fit to data for the wrong or unknown reasons, however, and not capture the internal form and processes of the real world. The ability to distinguish between model forms with similar degrees of fit is termed identifiability, which is constrained by the adequacy of observations, in terms of both accuracy and range of system properties. Where processes are represented dynamically, the same model configuration can produce potentially very different outcomes (a circumstance referred to as multifinality) depending on variations in boundary conditions or stochasticity in the parameters or processes, further challenging attempts at identification.
Modeling approaches range from inductive (or pattern based) to deductive (process based) (Overmars et al., 2007), though in practice these approaches are commonly used together in an iterative manner. Inductive approaches seek to observe relationships between outcomes and drivers in order to infer some process-level understanding from the patterns in data about specific cases. Deductive approaches, on the other hand, present models of general processes to specific cases or experiments that can be used to test these general explanations. These approaches can be used jointly in a form of abduction (Paavalo, 2004), in which results from analyses of pattern suggest a set of possible explanations that can be tested, the findings of which are fed back into the pattern analysis, often creating a chain of linked explanations for the phenomenon in question. In fact, most operational models reside somewhere between induction (i.e., pattern based) and deduction (i.e., process based). Where models fall on this continuum affects their characteristics and how their results can be interpreted by users. For example, a completely inductive approach uses no theory or thesis in developing the model structure and might rely instead on machine learning to establish relationships between driving variables and outcomes. (Even these approaches, however, require selection of the variables, which necessarily draws on some level of theory and process understanding.) This might limit the types of policies the model can be used to evaluate to those involving only changes in the input patterns, restricting their use for policies involving changes in process (e.g., incentives to the users of land to make different choices). On the other hand, a more strictly process-based model developed without recourse to data on (a) specific case(s) might have rich structural detail about the mechanisms but might perform poorly when it comes to reproduction of the observed patterns in data. Trading off these various strengths is an important component of model choice.
Projection, Forecast, and Scenario
Our statement of task highlights “future needs for higher-resolution and more accurate projections.” Various research communities use the term projection—and affiliated terms, forecast and scenario—in different ways. In this report, we follow the definitions set out by the Intergovernmental Panel on Climate Change (IPCC).
The IPCC Third Assessment Report (2001) identifies a projection as “any description of the future and the pathway leading to it.” For our purposes, projections can be derived from a LCM (e.g., future land use), although other input sources for such projections could exist (e.g., assumptions about land policies and economic activity). The IPCC distinguishes projections from forecasts or predictions (terms that are used interchangeably) by defining the latter two as projections that are branded as “most likely,” often based on models to which some level of confidence can be attached.
In contrast, the IPCC defines a scenario as “a coherent, internally consistent, and plausible description of a possible future state of the world” (IPCC, 1994). Scenarios are used to make projections, but under a specific set of assumptions that can include projections from other models (e.g., economic forecasts). The uncertainties about these projections may be unknown, but they may represent plausible directions for the initial conditions and external driving forces for a set of processes (e.g., land change) (IPCC, 2000). Usually any set of scenarios will include a baseline against which others can be compared, often related to projections based on currently known conditions and processes. Scenarios can be exploratory, and describe how the future might unfold under a set of known processes or current trends, or normative, in that they describe a future that might be achieved (or avoided) if a specific set of actions is taken. The IPCC SRES scenarios (IPCC, 2000) follow the exploratory approach, based on a set of demographic and economic assumptions, whereas the representative concentration pathways follow a more normative approach, wherein they are defined with respect to the outcomes and used to evaluate ways those outcomes might be achieved (Moss et al., 2010).
LCMs are used widely in the overlapping arenas of science and practice (decision making, policy, or real-world application in the private and public domains), and a clear understanding of the different requirements of each of these communities as well as their close connection provide context for the later sections of this report. In general, science is concerned with generating and organizing knowledge, and the field of land change science and related disciplines build LCMs to formalize and test land change and associated theory and explore scenarios where real-world experimentation is not possible. For the
most part, science-driven modeling efforts tend to become more process based as understanding of a problem and the processes evolve, beginning with data-based exploratory models and proceeding toward more explicit representations of process in the models. The value of different approaches to LCM in science varies according to both the nature of the modeling approaches and the nature and specificity of the questions.
The communities engaged in practice are concerned with setting sound and defensible guidelines for policy and action, and LCMs provide value by revealing current and exploring possible future outcomes from policy or action. Given the number and range of different decision-making units charged with different facets of human-environment systems (e.g., climate or water regulation, housing or zoning, and corporations), LCMs often focus on or link to the missions of the individual practitioner groups. In consideration of the roles of models in decision making, Figure 1.1 provides a useful conceptual model of how decision making proceeds (Verdung, 1997) and serves as a touchstone for later discussions about the relative value of different modeling approaches at different stages in the decision process.
Problems (e.g., flooding) about which decisions and actions are needed are identified in a variety of ways and various interventions are designed to mitigate or reduce the problem. Upon evaluation of alternative interventions, some decision is reached and an intervention implemented. Following implementation, the
FIGURE 1.1 Decision-making process. SOURCE: Verdung, 1997.
intervention is evaluated and the results fed back into identifications of similar problems in the future. Models can contribute at each stage of this idealized process, but the information requirements are different at each stage.
Viewing LCMs in the context of the policy cycle will allow managers and policy makers to make more informed decisions regarding the type of model output needed during each phase of policy making. Framing LCM’s in terms of the user needs rather than modeler interests represents a departure from previous LCM reviews.
The Chesapeake Bay study provides an example of a context within which LCMs contribute to formulation of policy and management decisions through coupling with other environmental models (Box. 1.2). Land cover changes affect runoff and nutrient loadings, and can be adjusted in scenarios to evaluate the effects of alternative land-related patterns on achieving water quality goals.
A variety of science and policy communities, including climate change modelers, integrated assessment modelers, and the IPCC, are asking for better information about land change at global scales. Spatial LCMs would provide platforms for exploring future scenarios of rural-urban form and structure to support decisions about these mitigation strategies. Similar effects of spatial patterns on ecological and biogeochemical processes have demonstrated a need for spatially explicit characterizations of land change (Debinski and Holt, 2000; Irwin and Bockstael, 2007; Robinson et al., 2007).
Despite the differences noted among the uses of LCMs, the science and practice communities are connected bidirectionally in the same way that humans and the environment are connected. Science informs policy and practice, and the outcomes or needs of the latter often prompt additional scientific research. Such feedbacks are instrumental to the advancement of LCMs.
We note above that temporal non-stationarity in land change processes creates uncertainty in LCM projections, especially the further into the future that the model is applied. In addition, an important property of all models is that they contain some irreducible level of uncertainty that can be inherent in the model structure (i.e., the basic equations or algorithms), the parameter values, input data, or all of these. Uncertainty is endemic in our understanding of land use decision making and of the physical state and function of the land system.
The future state of land systems will be determined by a combination of individual and societal constraints and opportunities about which decisions have yet to be made. LCMs implicitly or explicitly address decision making in the face of opportunities and constraints and biophysical processes not subject to human decision making, but which may be fundamentally affected by it. For example, a model can simulate where humans will change the landscape, a decision-making component that is inherently uncertain. Other parts of the model follow physical
Chesapeake Bay Land Change Model as Input to the Chesapeake Bay Watershed Model
The Chesapeake Bay Program (CBP) is a consortium of federal agencies, states, and Washington, D.C., that aims to develop strategies and implement methods to restore the health of the Chesapeake Bay. A key component of the CBP is the use of an LCM, based on SLEUTH, to provide scenarios of land use and land cover that can be input to the Chesapeake Bay Watershed Model. Current and future scenarios of land change are entered as input to the CBP Scenario Builder to develop parameter sets to parameterize the watershed model for simulation of stream flow, nitrogen, phosphorus, and sediment loading to the Bay. Different scenarios for land use change are developed to assess the long-term potential of achieving water quality goals and to help set policies on development and watershed to local-level restoration methods.
SOURCE: Chesapeake Bay Progream, Modeling Workgroup: http://www.chesapeakebay.net/groups/group/modeling_team.
principles that may not change but are subject to data and parameter uncertainty. For instance, models that connect land change with the storage and movement of water are guided by physical laws of hydraulics. Incomplete knowledge in the current state of water infrastructure, surface topography, and subsurface structure, however, introduces substantial uncertainty into understanding and prediction of fundamental physical processes. In addition, decision-making processes and physical processes interact in a model, for example, as humans change landscapes by installing or retrofitting drainage infrastructure toward specific water management goals, which can change over time (Huang et al., 2013). This provides
strong feedbacks to land, to the biological and geophysical properties of the landscape, and to human decision making, which may need to be incorporated into a LCM. In some instances, human use of the land can drive the environment to thresholds that, once crossed, change environmental (ecosystem) function (DeFries et al., 2004) and amplify uncertainty.
Model diagnosis, a fundamental part of any modeling exercise, seeks to learn about the behavior of the model relative to the real system in order to interpret model output appropriately. This diagnosis accounts for the essential modeling steps of verification, calibration, and validation. When model parameters are fit by calibration to historical data, additional uncertainty is introduced due to the inherent temporal nonstationarity of processes. Model diagnosis must, therefore, also account for nonstationarity in the data and processes, and stationarity assumptions in the model. Model diagnosis is critical to evaluating the representation of interacting processes, the manner in which the landscape is represented, and the uncertainty in each of these forms. In Chapter 3 we outline a framework for best practices in model diagnosis in the context of LCMs.
Land change modeling confronts a rapidly changing data infrastructure that affords large opportunities for improved models and their application, but it also entails a number of important issues about data quality and assessment that require consideration during model development and testing. These include issues related to remote sensing of land change as well as other data on the various social and biophysical characteristics that define the land system and its drivers within LCMs.
Remote Sensing of Land Change
The science of characterizing and measuring land change from remote sensing data has evolved significantly over the past 40 years. The earliest methods—and still widely used—identify land change as the difference between classified land cover maps from two or more points in time or through direct detection of changes themselves based on one or more images (Coppin et al., 2004). Although this remains the most common approach to characterize land change due to its simplicity and flexibility, it is more appropriate for identifying coarse-scale quantitative changes based on specific points in time (e.g., deforestation, urban expansion) than it is for identifying nuanced qualitative changes (e.g., changes in urban density, vegetation health) or dynamic processes that evolve over time, like crop rotation and change.
Advances in change detection analysis provide new opportunities for LCM. Most notably, new algorithms that use the entire set (or a subset) of the long archival record of many sensors, such as Landsat or MODIS, take advantage of
the temporal signal to identify land changes not possible through the traditional differencing approach (Zhu et al., 2012). New methods developed from allied fields such as geostatistics (Kaliraj et al., 2012), artificial intelligence (Ghosh et al., 2008), and time-series analysis (Verbesselt et al., 2010) further improve and expand the type of information that can be extracted from remote sensing data. Moreover, change detection methods are moving away from individual pixel analysis and incorporating spatial neighbors and shape. Further, a whole community has emerged around the use of objects as the unit of analysis in object-based image analysis and change detection (Chen et al., 2012). Many of these new algorithms and methods have yet to be fully utilized in LCMs.
Here we briefly take stock of some of the key advances in remote sensing that significantly contribute to the development of LCMs. In Chapter 3, we identify new challenges for remote sensing for LCMs and emerging developments in algorithms and image processing that could meet these challenges.
1. The growing depth of the Landsat archive, coupled with the no-cost policy governing its access, has driven a lot of efforts to model observed changes. This data set can be seen as an important impetus and enabler of the current state-of-the-science in LCM.
2. The growing constellation of private and government sensors offer a wide range of spatial resolutions, from submeter (Pleides, Quickbird, WorldView, GeoEye) to 250-500 m (MODIS). Revisit frequency from daily to weekly observations makes it possible to examine land change with high temporal frequency. With archival records from Landsat, long time series are being coupled with high-temporal-frequency and high-spatial-resolution analyses. This increased temporal frequency improves our ability to characterize dynamics in the land system, diagnose temporal nonstationarity, and develop empirical parameter sets for LCMs that are better tuned to historical changes.
3. Data from land imaging sensors cover a wide spectral range, from the visible to the long-wavelength infrared (also known as thermal imaging) and with different spectral bandwidths. Combined with the trend toward higher radiometric resolution with dynamic ranges up to 16 bit (65,536 unique values per pixel per spectral band), the land change community now has available a large range of high spectral detail in a single pixel. This helps to refine the detail within which land attributes and types can be represented, and it opens the possibility of incorporating more thematic detail into LCMs.
4. There are growing numbers of algorithms for spectral mixture analysis to estimate the fractions of different materials or cover-types within mixed pixels. Rather than assuming a single land cover type per pixel, the resulting maps provide continuous fields that represent fractional cover of different categories and that could be the targeted outcome represented in future LCMs.
5. Algorithm developments that move away from pixel-based analysis provide new looks at landscapes using shapes and objects as the unit of analysis.
Outputs from these analyses are particularly useful for identifying human activity or management of the landscape (e.g., agricultural plots) and hold the potential to interact more directly with models of human decision making.
6. Multiple LiDAR remote sensing platforms, from spaceborne to airborne systems, provide information in the third dimension. For example, the Geoscience Laser Altimeter System on ICESat collected data that can be used to map vegetation canopy parameters including tree height, biomass volume, and stand density. Sensors of this sort further expand the land variables that can be sensed and included in LCMs, either as inputs or outputs.
Whereas for many years, the primary land imaging sensors were Landsat and the Advanced Very High Resolution Radiometer, the LCM community now has at its disposal a large range of additional sensors and types of information for LCM inputs. As discussed in Chapter 3, however, the community is only beginning to harness the synergies possible from algorithm advances in both the LCM and remote sensing communities.
Other Data for Land Change Models
In addition to data from remote sensing, data from a wide variety of other sources are central to development of the ability in LCMs to both characterize aspects of the land system that cannot be directly observed with Earth observation technology (e.g., property ownership) and incorporate information about the social, historical, ecological, and other drivers of land change (e.g., population growth, economic activity).
A common challenge that arises with using these data for LCMs is that, with the exception of customized surveys, none were designed explicitly for use in LCMs but rather for different audiences and different purposes. For example, census data are collected over various administrative units and usually at decadal intervals. In addition, they can usually be used only at some aggregate level (e.g., census blocks), not at their finest spatial resolution. Other data may be recorded or collected more intermittently. Land values, for example, may be updated only when a property changes ownership. Four measurement and characterization issues common to all data for LCMs are briefly reviewed below and opportunities for future advancement are outlined in Chapter 3.
Format of variables
The subjective decision to format variables in LCMs as continuous or discrete influences the type of questions that can be answered and the type of analyses that can be employed (Southworth et al., 2004). With discrete variables, each spatial unit is represented by a single categorical value, and these data can detect wholesale changes in either the land use or cover (e.g., agriculture to urban) or the input drivers (e.g., change in land ownership). Continuous variables allow
measurement and characterization of finer details, such as qualitative changes in land use or cover (e.g., agricultural intensification or variations in net primary productivity) or modification in driver variables (e.g., changes in agricultural inputs such as fertilizers). The use of crisp categories simplifies analysis, but it requires that the number of categories to use must be determined and categories must be applied consistently over time. More categories add more detail, but they also make the analysis more complicated. It can be desirable to aggregate categories to reduce the data to a small number of important categories; the manner in which aggregation is performed influences the signals of land change.
Data accuracy and reliability
Understanding the accuracy and reliability of the data used in LCMs is essential. With remotely sensed data, accuracy is affected by the data source (i.e., sensor) and by the processing steps involved in creating the final map. The lineages of other LCM data are often more difficult to identify. For example, the specific instruments, field methods, original source material, or analyses used may not be recorded or evident. Consequently, it is often difficult to assess the accuracy or reliability of these other data sources. Additionally, the increasing availability of “volunteered geographic information” (Goodchild, 2007), for which individual citizens provide geographic information, presents a new trove of information for observing and analyzing change (NRC, 2010c). While the accuracy of these data is nearly impossible to assess, especially given the magnitude of the information available and the lack of strict protocols in its collection, the volume of data can often be used in creative ways to develop robust characterizations of the phenomenon of interest.
Multiple time points
Inherent to all LCMs is information over time. New methods are needed to consider the use of data with high temporal frequency in LCMs (de Beurs and Henebry, 2005). While the availability of temporal data from coarse-resolution imagery (i.e., > 100 m resolution) has been common for a couple decades, the availability of high-frequency data with finer resolutions (i.e., <50 m resolution) is a recent occurrence. A challenge for other data sources, such as field surveys and censuses, is how to sustain consistent collection efforts over the long term to create time-series data that could be more useful for LCMs.
An Abundance of New Data Sources
The growth of new data sources from satellites, aircraft, ground sensors, and “citizen science” presents new opportunities to measure and characterize land systems. The massive increase in data-gathering methods and data sets has not been matched by parallel increases in approaches for turning raw geographic data into more meaningful information about land systems or inputs to LCMs. The
increase in information that describes and measures land systems requires expertise in inference and knowledge of the local systems in order to make effective use of it. In other words, new skills are required to synthesize and qualify data from multiple sources and disciplines. Furthermore, measurements from different agencies, sensors, and researchers may share the same or similar categories but under significantly different conditions and assumptions (i.e., the semantics are different)—a common problem in integrating land classes undertaken by different research programs. Therefore, LCMs cannot simply integrate these data into a single database; land change scientists must identify robust approaches to translate the raw observations into meaningful information (Di Gregorio, 2005).
Land change models have been and continue to be critical to a large range of uses and users in science and practice. Indeed, the demands on LCMs continue to increase in terms of product outcomes and uses. These demands confront a series of problems, the broad outline of which has been noted. Input to inform the opportunities identified in this report was gathered at committee meetings and a workshop, through the committee’s own expertise, and through a questionnaire distributed electronically to a variety of individuals and groups working on LCMs (see Appendix A for a full list of contributors). The workshop included national and international experts in land-climate interactions, water quantity and quality, food-fiber, energy, ecosystem services, and urbanization. In Chapter 2 we describe and compare approaches to LCM and suggest guidance for their appropriate application. In Chapter 3 we suggest ways to improve LCMs and outline several forward-looking issues.