Some key challenges in modeling coupled human-natural systems, illustrated through agent-based land-use modeling research

Dawn Cassandra Parker Associate Professor School of Planning, University of Waterloo

With contributions also from many others!

26 Oct, 2010

Motivation: Concern regarding human impacts on the natural landscape

• Increasing human appropriation of land • Degradation of land-based natural capital • Decline in land-based ecosystem services: • Carbon sequestration • Water purification • Nutrient cycling • Temperature and climate regulation (http://www.globallandproject.org/)

Increased demand for “coupled human natural systems” models

• Need to understand likely future levels of ecosystem services and future status-quo trajectories of change • Need to develop management strategies and potential policy interventions • Requires an understanding of how human incentives shape land-use decisions, and how these decisions alter ecosystem services (uni-directional linkages) • Also requires understanding how changes in ecosystem services alter (or fail to alter!) human incentives

(A) (B) (C Figure 2: Three approaches to linked systems: natural to social (A), natural, social, natural (B), and fully linked (C) (Parker, Hessl, Davis, 2008)

Carbon Sequestration Atmospheric CO 2 Fossil Fuel Consumption Forest Productivity Forest landowner decisions Figure 1: Indirect Linkages between landowner decisions, forest productivity, C sequestration, and atmospheric C (Parker, Hessl, Davis 2008)

From here forward:

• What are agent-based land-use models?

• What are some practical, current challenges in developing models like these?

• As illustrated through some ongoing work in the deciduous US East Coast and Midwest: • Modeling joint effects of LUCC and land manager behavior (water quality and carbon sequestration) • Modeling the carbon implications of timber harvest behavior • Which challenges should be highest priorities for investment and research (maybe yours!)?

H/E Interactions ABM/LUCC Land-Use Modeling Complexity Theory Agent-based models of land use consist of: • An electronic representation of a landscape • An “agent-based” simulation of decision-makers whose choices alter the landscape (usually a computer program) (Parker et. al 2003; Parker, Berger, Manson 2002)

Cellular/Spatial Landscape Model

• May or may not be based on real-world maps via geographic information systems layers • May contain a variety of geographic and socioeconomic features such as: • Slope, elevation, vegetative cover, soil types, zoning restrictions • Road and rail networks, information on social networks (who knows who) • Models of “spatial diffusion,” such as how air pollution spreads and disperses across a region

Figure 7, LUCC Report #6

Agent-Based Model

• Autonomous decision-making agents • Interaction environment • Interdependencies among agents, their environment, or both • Rules governing sequencing of actions and information flows

What is an “agent”?

• Goal-oriented entity • Model of cognition that links goals and behavior: • Capable of autonomous action • Capable of responding to changes in its environment • Generally represents a land manager in land use change models

Agent-Based Model of Decision Making

• Each individual decision maker is represented through a set of rules that link information about his/her environment to a decision • Decisions often depend on the agent’s physical environment (the landscape) • Decisions may also depend on what other agents do as well - can lead to “path dependence”

H/E Interactions ABM/LUCC Land-Use Modeling Complexity Theory Advantages of ABM/LUCC for land-use modeling: • Your model can have a realistic (and appropriate) geographic representation • Potential links with geographic information systems for data input and output visualization • Modeling of structures that are nested in time and space (cross-scale)

H/E Interactions ABM/LUCC Land-Use Modeling Complexity Theory Modeling human/environment interactions: • Socioeconomic and biophysical models can be linked spatially • Simulation approach allows for feedbacks between dynamic social and environmental processes • Applications include crop yields, hydrology, forest growth, pest species modeling, endangered species populations

H/E Interactions ABM/LUCC Land-Use Modeling Complexity Theory Human/environmental landscapes are

complex:

• Characterized by: • Interdependencies (one agent’s action depend on what another has done previously) • Heterogeneity (diverse variation in the same type of object) • nested hierarchies (overlapping structures in time and space)

Soccer example: Nested hierarchies

goalkeeper defense midfield offense

Complex properties of soccer …

• Local actions of players lead to reoccurring, recognizable patterns that are semi-stable • A small change in strategic play can lead to large changes in the state of the system • In general, successful strategic play is general the result of interdependent team effort (“whole greater than the sum of its part”). In short, the relationship between players is of key importance.

Properties of complex systems:

• Analytical intractability • Path dependency • Linkages across hierarchies • Emergence: “Organization into recognizable macroscopic patterns” (Epstein and Axtell, 1999)

Quilting examples of emergence

Start with two very simple objects that are heterogeneous with respect to shape, size, and color …

Quilts, scaling up …

By defining the relationship between other similar objects at a very local level, we obtain some structure and pattern …

Quilts, scaling up again …

And by combining these elements, we create structure at a coarser spatial scale …

One emergent form …

And another …

Key sources of agent heterogeneity:

• Pecuniary and non-pecuniary motivations: profits, preservation of family farm, environmental ethic • Experience and knowledge • Financial, physical, and human capital • Access to credit • Cultural preferences • Expectation formation mechanisms • Decision strategies

Types of Interactions Agent-Agent

• Information transfer • Technology diffusion • Land markets • Local labor exchange • Community-based resource management

Agent-Environment

• Hydrology (ground and surface) • Erosion • Deforestation • Transport of pollutants • Species migration • Soil fertility

How might ABM be used to study H-E interactions?

• To link socioeconomic drivers of resource use to their biophysical impacts • To explore the effects of feedbacks between humans and their environment • To examine whether current systems of resource use are sustainable • To design policies to encourage more sustainable resource use

Why the interest in ABM?

• “Change” of modelling approach?

• ABM seen as a tool for building: • finer-scale process-based models • with more flexible representation compared to analytical or systems dynamic models • with the ability to incorporate theories and drivers from many social science perspectives • facilitating spatially explicit, fine scale, coupled models of human-environment interaction in the land system

Can ABM address these challenges?

• Bridge the gap between theoretical/process based and empirical/pattern-based modeling • Model 3-way feedbacks between land use/management, land cover, and landscape function • Provide useful information for scientific and policy analysis regarding the effects of human incentives on the natural world

Some practical challenges

• Challenges include: • Social and natural models have different key driving variables • Scale issues (conceptual and empirical) • Data constraints • Common challenges in many coupled human natural models • Which challenges are most important? (binding constraints) • Which should be highest priorities for investment?

Big issues:

• Desire to link socioeconomic drivers of LULCC to LULCC to effects on key ecosystem services • Relevant processes, and therefore models of them, often operate on different variables of interest at different spatial and temporal scales • Data often available/collected for independent process models -> matching problem • Data availability constraints

Big questions:

To what extent can coupled-HE models be improved by: • Better process models vs...

• Better data?

In short • Are limitation practical or conceptual?

• What investments should we target and why?

Focusing the question: 3 research applications

• Residential household-level best management practices, storm water runoff, and water quality in the Potomac Gorge, USA • Residential landscaping and carbon sequestration in ex-urban Michigan, USA • Timber management decisions and carbon sequestration in West Virginia, USA

A preview of the main issues

• Process-based limitation: the biophysical dimensions of the ecosystem service of interest are not the central drivers of the land management decisions (dependent variable, not scale mismatch) • Data-based limitations: • collection costs • confidentiality and other institutional constraints • lack of field studies to link social behavior to inputs of biophysical models

Exploring changes in residential land use and land management via ABM land market models

• Longer-term goal is to develop agent-based models of residential suburban and ex-urban land markets that: • Connect land market and land management behavior of residential agents • Allow exploration of the relative contributions of land-use vs.. land management change (categorical change and change in intensity) to environmental changes • Preliminary work on water quality done in the Potomac Gorge watershed (DC metro area); full-fledged model under development in Southeastern Michigan; application in for development of a Waterloo Region model • Explore the value added of the land market component through comparison to comparable models without a land market component (*Can talk more about this if wanted .. )

Environmental impacts of residential development

• Suburban and ex-urban development bring about environmental change • Evaluation of impacts requires understanding of: • Location and timing of land-use change • Characteristics of new or modified development • Land management behavior of new land managers • Three factors jointly determined via land market interactions

Residential development and water quality (Potomac Gorge)

• Residential land use contributes to decreased water quality through increased nutrient loadings and changes in hydrology (flow) • Main research question: What linkages exist between residential land use in the Potomac Gorge watershed (DC area) and the degradation of water quality in tributary streams (a primary threat to rare and endangered species in the Gorge) • Both social science (resident survey) and natural science (water quality) models were planned; social science component had to be dropped.

Residential development and carbon sequestration

• Landscaping choices in existing and new developments may have dramatically different carbon profiles • Main research question: Will ex-urban development in Southeastern Michigan produce a landscape-scale source or sink of carbon, given observed landscaping strategies of developers and landscaping preferences of residential agents? • Collaboration with UM (Dan Brown et al.) to extend Project SLUCE • See recent WICI talk for many more details

Sources of environmental impacts: Behavior of resident land managers

Land managers are not homogeneous. Water quality and carbon profiles also depend on: • Landscaping preferences and practices • Tree, turf, and horticultural choices also • Fertilizer and pesticide use • Management of organic matter and debris • Willingness to adopt BMPs: • Green roofs • Rain gardens • Pervious pavers • Evidence from human ecology/environmental sociology/economics that these two factors vary with agent resources, information, attitudes, beliefs, and values & neighborhood influences

PoGo WQ model: Dep. Var mismatch issue

• What are the dependent variables?

• Household model: residential landscaping, building, and storm water management decisions; • Water quality model: nutrient runoff • Potential link between models: Decisions about best management practices • Households care about economic cost, social implications, transaction costs, and flood risk reduction when thinking about WQ BMPS • They may not know about, and rarely consider, local or off-site nutrient loadings and other WQ impacts

PoGo WQ model: Scale issues

• WQ models: coarse spatial scale (watershed); calibrated with fine temporal scale data for few locations • Social science models (adoption of rain barrels, rain gardens, pervious pavers): fine spatial scale (parcel level), coarse temporal scale (10 years for RS land cover data, annual for parcel data, often 1 point in time for a survey) • Institutional/jurisdictional boundaries cross watersheds: zoning, parcel sizes, setbacks, WQ regulations, etc.

PoGo WQ model: Data issues

• *Coefficients to translate BMP assumptions into nutrient reductions for WQ models are sparse or missing. Most come from experimental, not field studies • Lack of time series of land cover data • Lack of access to/expense of obtaining parcel-level land use data • Institutional constraints on data collection • Concern about time and attention demands of surveys

SLUCE II model: Dep. Var mismatch issue

• What are the dependent variables?

• Household model: choice of house to buy, residential landscaping modifications and maintenance; • Carbon model: carbon sequestration based on biomass growth/removal • Potential link between models: Landscaping decisions • Households stick with initial landscape design, care about aesthetics, imitate neighbors, economic cost, social standing, local regulations, when thinking about residential landscaping • They may not know about, and rarely consider, the carbon sequestration impacts of their landscaping

SLUCE II model: Scale issues

• Landscaping decisions are more fine-grained in time and space (at a parcel level) • Existing carbon model designed/calibrated to run on a coarser scale • Problems could be solved with time, money, and computing power —but it would require recalibrating models

SLUCE II model: model: Data issues

• *Coefficients to translate landscaping decisions into biomass changes for carbon model are sparse or missing; have to be estimated from field work.

• Other data constraints less binding due to long term nature and resources of the project • Cost limitations limit number of in-depth interviews/field surveys • Complicated sample stratification due to desire to stratify by development type, LULCC history, and biophysical conditions • Concern about time and attention demands of surveys

Timber management decisions and carbon sequestration in US Eastern deciduous forests

• Carbon dynamics in mixed-stand deciduous forests are poorly understood; stands are reaching maturity, land manager motivations and strategies are diverse • Main research question: How might the ability of central hardwood forests to store C change in the future under current conditions? Under alternative policy and economic regimes? • Collaboration with West Virginia University (Amy Hessl et al.) with PhD student Sean Donahoe

Experimental field work for model calibration

Timber model: Dep. Var mismatch issue

• What are the dependent variables?

• Timber harvester: timber harvest (event, species type, harvest strategy) • Carbon model: carbon sequestration based on biomass growth/removal • Potential link between models: Timber harvested

Timber model: Dep. Var mismatch issue

• Harvesters make timber harvest considering only 2 3 high-value species, using various harvest heuristics (diameter limit cuts, clear cuts, etc.) They may or may not consider future biomass, but do not consider carbon • Timber rights may be sold off; separation between land owner motivation and timber harvest; may not be forward-looking behavior • Carbon model wants consistent and uniform data about species composition and corresponding canopy cover.

Timber model: Scale issues

• Like SLUCE II, scale issues are not conceptually high hurdles--both models could run at a parcel scale • In this case, the carbon model was modified and recalibrated at a parcel scale using local data inputs —so there is less mismatched between the scales at which models are designed to operate.

• Problem seem to be solved with time, money, and computing power • However, Sean tried to apply the new model to the landscape scale, using state level data, and it didn’t perform well.

• Instead he developed a simpler, alternative growth and carbon model

Timber II model: model: Data issues

• *Compromises needed to be made on dependent variable (total biomass removal vs. timber harvest strategy) due to data quality issues • Design of FIA database was good (gathers data on both management type and intent and biophysical measurements of growth and harvest • But, sampling strategy was too sparse and data quality too poor to meet all of our goals for parcel level analysis (data collection was designed for aggregation). Probably influenced failure of the carbon model. • Confidentiality constraint on ownership type and location; could be solved with more time and money • Land market dynamics mean that highest valued timber is not available for harvest —need data to model land markets

Conclusions and recommendations: Dependent/state variables

• More conceptual work needed on the problem of mismatch of dependent variables between models: • socioeconomic and biophysical models generally are built around different state variables • This reflects missing feedbacks/incentives in the system.

• The practical result is that disciplinary conceptual and empirical models tend not to match.

• We also need to consider whether models and developed under mismatched incentives will still be valid under a regime of corrected incentives (i.e., carbon markets)

Conclusions and recommendations: Scale issues

• I see these as less important than I did in 2007!

• Many can be solved by time, money and computing power • However, scale-related process and data matching issues should be considered when designing data collection protocols • Socioeconomic models tend to have better spatial coverage and poor temporal coverage; biophysical models, the reverse.

Conclusions and recommendations: Data issues

• Improved field study and survey data to develop technical linkages between land management behavior and biophysical model inputs is

the highest priority

• Coupled modeling needs should be considered when designing public data collection protocols • Issues of privacy and time cost for survey respondents will continue to be important • We face new problems in Canada due to the compromised census sampling protocol.

Final Thoughts

• Lack of empirical studies that link social and biophysical dependent variables not a surprise: • Practically each system has different driving variables; • Disciplinary studies designed in isolation focus efforts on the driving variable of interest • The gap results • These deficits have been discovered (and in some cases resolved through interdisciplinary research teams funded by new, targeted funding initiatives.

• Hope/predict to see similar funding initiatives in Canada, too.

Acknowledgements

• ALMA-v1.0: Tatiana Filatova and Anne Van der Veen, University of Twente; funding from NWO-ALW (LOICZ-NL) project 014.27.012 and NSF 041406 • Potomac Gorge project: Ryan Albert, Robin A. Brake, Susan A. Crate, R. Christian Jones, Brandy Holstein, Atesmachew Hailegiorgis, and Charles Nguyen; Departments of Computational Social Science, Environmental Science and Policy & Sociology, George Mason University; Giselle Mora-Bourgeois, Urban Ecology Research Learning Alliance, National Park Service; Funding from Chesapeake Watershed CEUS. • U. Michigan: Dan Brown, Bill Currie, Joan Nassauer, Scott Page, Rick Riolo, Derek Robinson, etc. funding for grant development from NSF BCS-0119804 , new funding from NSF CNH-0813799 • WVU: Amy Hessl, Sarah Davis, Bill Peterjohn, Richard Thomas, Maction Komwa, Sean Donahoe, NSF grant 0414565