LandCover and LandUse Change

1
Land-Cover and Land-Use Change in the Southern
Yucatán Peninsular Region Project
2
Major Participants

• George Perkins Marsh Institute Graduate School
  of Geography - Clark University
• Harvard Forest - Harvard University
• El Colegio de la Fronter Sur (ECOSUR)
• with cooperation of
• Center for Integrated Studies-Carnegie Mellon
  University
• For a list of individual researchers and their
  contributions to the project see the SYPR web
  page (earth.clarku.edu/lcluc)

3
Funding Sources

• Overall Project
• NASA - LCLUC Program
• ECOSUR
• Center for Integrated Studies, Carnegie Mellon
  University
• Ecological Research
• Harvard Forest
• Conservation Research Foundation
• National Science Foundation
• Modeling Research
• Center for Integrated Studies, Carnegie Mellon
  University
• NASA New Investigator Fellowship
• Specific Dissertation Research
• Fulbright Fellowship (Garcia Robles)
• NASA Earth Systems Science Fellowships (2)
• NSF Geography/Regional Science Decision, Risk,
  and Management Science/INT-Americas Doctoral
  Dissertation Improvement Grants (3)
• InterAmerican Foundation Fellowship

4
Problem Overview

• LCLUC-SYPR seeks
• to identify the trajectories and pace of various
  land-use/cover changes in the region
• to develop detailed understanding and
  explanations of causes and consequences of these
  changes
• to develop a suite of models capable of
  explaining and projecting spatially explicit use
  and cover changes
• to do all the above through integrative of
  studies employing historical-empirical
  narratives, behavioral, structural and ecological
  theory, and remote sensing and GIS analysis.

5
Specific Goals and Components of LCLUC-SYPR

• Understand the dynamics of land-use/cover changes
  in SYPR, especially deforestation and agriculture
  since the 1960s
• Via the following questions
• How does the historical narrative of the
  land-use/cover changes and the causes and
  consequences of these changes affect
  understanding, explanation, and modeling the
  current conditions of SYPR?
• What is the explanatory role of agent-based and
  structural-based theories, and can they be merged
  to improve understanding of land change?
• What is relation between market and non-market
  production among households and does getting
  this relationship right improve explanations of
  land change?

6

• Why is chili production apparently succeeding,
  and what is its longer term economic viability
  and land impacts?
• Does the presence of different resources
  institutions (rules of access) matter in terms of
  land-use/cover conditions?
• What is the structure and function of the three
  main forest types present in SYPR?
• What is the role of nutrient cycling in forest
  succession and land-use decisions?
• What is the role of hurricanes in land-use/cover
  change?
• Can TM imagery analysis be pushed to the level of
  specification needed for detailed land-change
  studies?
• Does spatial explicitness improve or hinder
  explanation.

7
Specific Goals of LCLUC-SYPR

• Develop and test the applicability of three types
  of spatially explicit models that explain and
  project land-use/cover changes, especially for
  forest and agriculture.
• Disaggregate household ejido models using
  survey, census, and environmental information
  linked to remotely sensed imagery (pixelizing the
  social Focus 1 research)
• Aggregate imagery-based models built from TM
  imagery but incorporating biophysical and social
  information (socializing the pixel Focus 2
  research)
• Dynamic spatial simulation models used to
  project land-use/cover changes under different
  scenarios (Focus 4 research)

8
(No Transcript)
9
22,005.6 km2
10
Environmental Schema of Region
The entire SYPR is karst with large, seasonally
inundated poljes (bajos) dispersed throughout.
These features increase in number and size on the
low-lying eastern and western flanks of the
region and support bajo forest. The center of
region consists of rolling hills with moderate
topography (100-300 m
Increasing precipitation decreasing dry season
asml) dominated by mediana or upland forest and
secondary forest. Poljes and uplands constitute
the major soil distinction, with minor
variation on uplands linked to depth of the
limestone bedrock. A strong precipitation
gradient exists diminishing east to west and
increasing north to south.
Decreasing precipitation increasing dry season
11
Human Impacts in SYPR (1000 BC - Present)
Maya entry - deforestation begins
Maya occupation - landscape change
Maya collapse - return of forest
Minor occupancy - sustained forest
European Conquest
1000 BC
AD 900
AD 1880
AD 1500
1880
1894
1901
1934
1955
1967
1989
1982
1993
Hurricane Janet
Rt. 186 Paved
Debt crisis
Calakmul biosphere reserve
Oil boom
Initial findings suggest that large structural
and policy changes link strongly to major shifts
in rates of deforestation.
12
Focus 3 Environmental Assessment and
Reconstruction leading to the ecological
consequence of and impacts on land change

• The aims of this work are
• 1 to identify and measure the principal
  biophysical perturbations on SYPR forests,
• 2 identify the basic structures of the
  different forest types and successional growth,
• 3 determine the role nutrient cycling on use
  and forest, and 4 incorporate these dynamics
  into the various SYPR land-change models.
• ATTENTION Slides marked by are not be used or
  cited without permission of the project.

13
Hurricane Simulation a major historical
disturbance
Simulated hurricanes
Gilbert 1988
Janet 1955
Fujita damage
Hurricane frequency (1886-1996) Modeled from
wind speed, direction, and duration, producing
estimates of potential damage.
14
Area Damaged by Hurricane Janet, 1955 Timber
Concession Records
15
Constructed Precipitation for SYPR
1

• An. Ave. Precipitation
• by station (mm)
• Yohaltun 877
• Silvituc 1286
• Xbonil 1147
• Conhuas 993
• Zoh Laguna 907
• N. Bravo 1228
• Blanca 1324
• !950-70s data to be
• added

2
5
3
4
mm
6
7
16
Composition, Structure, Dynamics, and
Regeneration of Forests SYPR study sites
17
Composition, Structure, Dynamics, and
Regeneration of Vegetation

• Six vegetation sites established across the
  region representing gradient of precipitation,
  soil depth, topography, and hurricane exposure.
• Minimally 10 vegetation plots (circular 500 m2
  each) per forest type at each site.
• Tree (gt10 cm dbh) and liana species composition
  and structural characteristics of forest record
  for each plot (as well as epiphyte load).
• In nested 100 m2, stems gt5 cm dbh or gt2 m in
  height are recorded.
• Seedling layer to be investigated in future.

18
First Two Axes of a Non-metric Multidimensional
Scaling of 500 m2 Forest Plots in SYPR.
Note The ordanation was based on abundance of
tree species. A similar result was obtained using
tree species prescence, and tree species basal
area. Triangles bajo forest squares upland
forest circles plots of successional ages
indicated by numbers. All successional forest
plots were originally upland forest. Plots on
bajo and upland forest are clearly
distinguishible along these two axis, as
supported by a MRPP test (plt0.001). As abandoned
agricultural fields reach 20-30 years of
succession, they aquire a tree species
composition undistinguishible from that of mature
upland forests (Pérez-Salicrup and Foster nd)
19
Changes in Nutrient Cycling During Succession
Three sites serve studies of nutrient stocks
(biomass soil) and cycling (production, turn
over constraints) as function of forest age
under the dominant shifting or swidden
cultivation in the region.
20
Litter production in mature forests
.
Increasing precipitation Decreasing hurricane
frequency
21
Linking Biomass and Species Composition to
Nutrient Cycling Processes
Litter production as an index of productivity
with links to soil nutrient status.
22
Variation in Forest Floor Mass vs. AgeAcross
Study Area
23
Variation in Monthly Litter Fall vs. Age
Across Study Area
24
Focus 1 Historical Narrative and Socio-Economic
Conditionleading to econometric analysis and
theory-based models

• Critical aims are to determine
• 1 the influence of previous political economic
  conditions on those now present,
• 2 the rationality of semi-subsistence land
  managers and its land management implications,
  and
• 3 the institutional and infrastructural
  conditions in which managers operate, and
• 4 to create models that move from the
  magnitude of change to the specific locations
  of the change.
• ATTENTION Slides marked by () are not be used
  or cited without permission of the project.

25
Population Centers for the SYPR, 1930-1960
26
Example Historical Data SYPR Population (1910 -
2000)
Population Density (Persons/km2)
Population (thousands)
27
Forest Concessions for the Southern Yucatan
Peninsula, circa 1920 (Estado de Campeche)
28
Example Historical Data SYPR in Transition
Laguna Corporation controls 760,460 ha in
Campeche 1915-1935
1915
Eco-archeological tourism in the Calakmul zone
2000
29
Likely Areas of Hardwood Extraction, 1930-1960
30
Smallholders, Infrastructures, and Institutions

• rationality of land users
• conditions in which they operate
• land management practices

31
Land Tenure and Ejidos in SYPR Sample ejido area
18,703 km2
19 00
19 00
N
25 KM
LEGEND
Ejido
18 30
Forest amplifications
to outside SYPR
ejidos
Private
National
Ejidos in sample
18 00
89 30
90 00
90 30
89 00
32
LCLUC-SYPR Example Question Guiding Theory
Construction and Household Surveys

• To what extent is the expansion of agriculture
• (i.e., forest to open land) propelled by
• 1 On-site demands for subsistence crops
• or
• 2 External markets for commercial crops?

33
Predominant Land Uses in SYPR Tentative Survey
Results
Mean plot size (ha)
S.D. plot size (ha)
Median plot size (ha)
Land Use
sample
100 52 28 23
4.65 1.34 25.9 12.9
4.34 1.36 28.9 20.8
4 1 15 5.75
Maize Chili Pasture used Pasture unused
Subsidies facilitate land cleared for pasture.
Presumably landholders speculate that livestock
programs (subsidies) will follow.
34
Indicators of Economic Links to Subsistence
Market Production
Selling maize 41 Purchasing maize 27 Self
sufficient in maize 31
of households
Sell chili No sell chili Sell maize 27
15 No sell maize 24 34
Sell labor No sell labor Hire labor 57
14 No hire labor 23 6
35
Spatially Explicit Land Models

• Previous spatially-explicit econometric models
• Aggregate socioeconomic data
• Profit maximizing behavior
• LCLUC-SYPR Project
• Individual land manager data linked to the pixel
• Test hypotheses concerning hh behavior linked to
  structural and infrastructural changes

36
Signature Development Sketch-Maps of Surveyed
Household Fields
Parcels linked to imagery by GPS
Sketch maps and GPS link the actions of the land
managers with land-use/cover change. The maps
not only aid in classification, they facilitate
spatially explicit analysis of change and
ultimately permit regional assessment.
Land Use History (gt20 yrs) via sketch-map
37
Conceptual and Empirical Framework for
Identifying Determinants of Land Manager
Behavior
Determinants of to sell or not to sell
Determinants of area cultivated
Supply-side on-farm labor units farm capital
land endowment soil quality
education Demand-side family size percent
children off-farm labor units Market
transaction costs distance to market distance
to plot transportation to plot (horse or
foot) vehicle ownership
Sellers
distance to market labor units distance to
plot farm capital land endowment soil
quality inverse mills
Area Cultivated in Semi-subsistence Crop
OLS model 1
Probit model
Non Sellers
consumption units labor units distance to
plot farm capital land endowment soil
quality inverse mills
OLS model 2
38
Separability of Production and Consumption by
Farmstead Type sell maize or notNon-linear
Specification of Demographic Variables
Endogenous variable ha in maize, estimated with
OLS. Control variables not presented. pgt.10 p
gt.05 pgt.01
39
Focus 2 TM Imagery Analysis, GIS and
theRegional Conditionleading to empirical
diagnostic models
A critical aims of this work are 1 to create
imagery classifications sufficient in detail for
SYPR-type studies, 2 to rework various data
for micro-spatial scale GIS use, and 3 to
create region-wide models that estimate correctly
in terms of the specific location of change.
2 ATTENTION Slides marked by () are not be
used or cited without permission of the project
40
Target Classification Scheme
Clouds Cloud Shadows Urban, Roads and
Quarries Water Savanna Herbaceous Wetland
Vegetation Seasonally Inundated Lowland
Forest Well-drained Upland Forest Cropland P
asture Successional Forests -
herbaceous - shrub-dominated -
arboreal Successional Pteridium (bracken fern)
Another invasive, tajonal (Viguiera dentata),
also exists in the area. Further work may permit
this invasive to separated from the general
category of herbaceous succession.
41
Image Processing Methodology Overview
Geometric Correction
Noise Removal I Haze Removal
3 Bands
Noise Removal II PCA
3 Bands
Texture Analysis
1 Band
NDVI
Training Site Signature Development
Signature Evaluation
Supervised Classification
Change Detection and Modeling
42
Noise Removal of Landsat TM DEHAZING
ERDAS Imagine Dehazing algorithm, using Tasselled
Cap Transformation RGB False Color
Composite TM Bands 4,3,2
43
Noise Removal of Landsat TM Principal
Components Analysis

• Perform PCA on 6 of 7 Landsat TM bands (no
  thermal band)
• Results striping and noise eliminated, data
  redundancy reduced

PC1, Visual bands
PC2, Visual bands
PC3, Visual bands
44
Incorporating Spatial ContextTexture Analysis

• Image texture -- distinctive spatial and spectral
  relationships among neighboring pixels
• Focus on overall pattern of variation in each
  category, rather than spectral average
• Texture Analysis on PCA results (1) improved
  separability of upland and lowland forests,
  cropland and pasture, transitional edges, (2)
  allowed the detection and separation of 4 classes
  of secondary (successional) vegetation

3x3 variance
RGB Composite 3 PCA bands
RGB Composite 3 Texture bands
45
Improving Signature Separability Using Texture
46
Compilation of PCA, Texture and NDVI Bands

• Normalized Difference Vegetation Index (NDVI) is
  a measure of relative biomass
• NDVI (IR-R)/(IRR)
• 3 PCA bands, 3 Texture Bands, 1 NDVI band
  layer-stacked to produce final 7-band image for
  signature development and classification

47
Improving Classification Contingency Using
Texture NDVI
48
Signature Development Sketch-Maps of Surveyed
Household Fields
Parcels linked to imagery by GPS
Sketch maps and GPS link the actions of the land
managers with land-use/cover change. The maps
not only aid in classification, they facilitate
spatially explicit analysis of change and
ultimately permit regional assessment.
Land Use History (gt20 yrs) via sketch-map
49
Training Sites for Signatures for Multiple Years
1995
1996
1994
1992
1988
1987
50
Improvements in Signature Separability Using
Sketch Maps
Lowland/Upland Forest
Pasture/Cropland
YoungSec/Pteridium
YoungSec/OldSec
51
Characterizing Uncertainty Soft Classifiers
Bayesian Maximum Likelihood Classifier
PCA Composite
Land-cover Categories
Bayesian Classifier w/ uncertain area unclassified
Classification Uncertainty
Average Uncertainty 0.37 Secondary Vegetation
Trees 0.29 Secondary Vegetation Shrubs 0.19
Secondary Vegetation Herbaceous 0.41 Upland
Forest 0.13 Cropland 0.23 Pasture
52
Example Classification for Modeling
Land-Use/Cover Change
53
Modeling Approaches

• Disaggregate, household and ejido-based
• Aggregate, imagery-based
• Dynamic spatial simulation

54
Enhanced Survey Regression Model of
Deforestation A Trial

• Aim to model the amount and location of
  deforestation associated with individual farmers
  (ejidatarios).
• In this trial, market behavior assumed and
  choices explained as a function of
  socio-demographic, market, environmental, and
  geographic variables obtained from household
  (farmer) surveys 2 southern ejidos omitted.
• The model tests whether the same variables that
  increase the probability of deforestation in the
  aggregate imagery-basaed model (see Focus 2)
  have the same general effect in explaining the
  amount of deforestation that an individual land
  manager undertakes
• The results show that location variables in the
  aggregate model are not statistically
  significant however, variables consistent with
  behavioral themes, such as off-farm income and
  livestock sold are.

55
Predicted Residuals from Enhanced Survey
Regression Model of Deforestation
Underprediction (-0.048 to -23.127) Overprediciton
(0.08 to 25.215)
56
Trial Model of Land-Use/Cover Change

• Deforestation Land-cover changes from Forest
  (mature and successional) to Cropland/Pasture
• Succession Land-cover changes from
  Cropland/Pasture to young secondary vegetation
• Model 2 time periods 1988-1992, 1992-1995

57
Explanatory Variables

• Biophysical soils, topography, initial cover
  class
• Locational and infrastructure distances to
  roads, village centers, markets, nearest other
  land cover
• Landscape pattern indices fragmentation,
  diversity, richness, number of different classes
  (NDC)
• Socioeconomic demographic, wealth indicators

58
Biophysical Variables
Low
High
High
Low
59
Locational and Infrastructure Variables
Distance to Nearest Road
Distance to Nearest Cropland
High
Low
Low
High
60
Landscape Pattern Indices
Low
High
Low
High
61
Demographic and Socioeconomic Variables
62
Model Estimation

• Binomial LOGIT model
• Endogenous Variable Forest Persistence vs.
  Deforestation
• Exogenous Variables GIS layers
• Base Case Forest Persistence
• Estimate Parameters
• evaluate significance of exogenous variables in
  model explanation
• enable prediction of deforestation probability

63
Model Output Deforestation in 1988-92

• Logit Estimates
• Number of obs 805898 chi2(21) 147541.4 gt 5
  
• Prob gt chi2 0.0000 Log Likelihood
  -219751.11 Pseudo R2 0.2513
• depen Coef. Std. Err. z
  Pgtz 95 Conf. Interval
• elev -.021297 .0002431 -87.615
  0.000 -.0217734 -.0208206
• slop .0366753 .0024783 14.799
  0.000 .0318179 .0415327
• sl2 -.6190486 .0247161 -25.046
  0.000 -.6674912 -.5706059
• sl3 -.5856845 .0091304 -64.146
  0.000 -.6035798 -.5677892
• baj -1.043225 .0263979 -39.519
  0.000 -1.094964 -.9914857
• med -.2162192 .0251993 -8.580
  0.000 -.265609 -.1668295
• sec2 .3272907 .0245467 13.333
  0.000 .27918 .3754013
• sec3 -.7267448 .0513751 -14.146
  0.000 -.8274381 -.6260514
• disr -.0205669 .0002015 -102.072
  0.000 -.0209619 -.020172
• dism .0006189 .0001467 4.218
  0.000 .0003313 .0009065
• disv .3046899 .0184065 16.553
  0.000 .2686137 .340766
• disc -.0008068 .0000101 -79.685
  0.000 -.0008267 -.000787
• div .1112383 .0127281 8.740
  0.000 .0862917 .1361849
• rich -.000519 .0015288 -0.339
  0.734 -.0035155 .0024775

64
Spatially Explicit Probability Maps
DEFORESTATION PROBABILITY 1988-1992
1
0
Deforestation Probability
65
Hardening Predictions
FOREST PERSISTENCE / DEFORESTATION 1988-1992
DEFORESTATION PROBABILITY 1988-1992
Threshold at p0.3198?

• The critical threshold
• is estimated using the
• total amount of actual
• pixels deforested.

Forest Persistence
Deforestation Probability
Deforestation
66
Comparing Predicted Actual Change
Crosstabulate Predicted Change with Actual Change
Predicted Change (map to be replaced to match
current threshold)
Sampling Prediction Region
Forest Persistence
Deforestation
Prediction Region
67
Results of Binomial LOGIT Model

• The Pseudo-R2 for the 1988-1992 model is 0.25
  (n805,898) and for the 1992-1995 model is 0.37
  (n744,530).
• Most of the variables included in the model are
  statistically significant.
• The signs of the coefficients of variables, such
  as slope, distance from markets and villages,
  distance to near cropland, remain constant
  through time.
• Deforestation pixels are most likely to occur on
  early secondary growth (where deforestation cut
  of successional growth as well mature forest).
• Coefficient of variables such as cattle density,
  population, slope are counter intuitive (e.g.,
  the higher the slope the higher the probability
  of deforestation).

68
Kappa Index of Agreement Statistics
MODEL FOREST PERSISTENCE/DEFORESTATION 1988-1992
Traditional Kappa (-1 ? K ? 1)
Modified Kappa (R. Pontius)
Category KIA
(p0.5) (p0.4) Forest Persistence
0.9873 0.9595 Deforestation 0.1224 0.2427 Overall
0.9310 0.9280
69
Focus 2 Summary

• High level of detail extracted from Landsat TM
• experimental image processing techniques
• exhaustive multi-date training sites from
  land-use histories
• detailed classification of cover and succession
• Model Estimation
• use of spatially explicit endogenous exogenous
  variables
• probabilistic prediction maps
• Model Validation and Future Directions
• Optimization of threshold for probability of
  change
• Kappa statistics to Relative Operating
  Characteristic ROC
• Dynamic transition probabilities and projections
• Incorporate spatially explicit decision-making
  rules derived from Focus 1 household model as
  additional exogenous variables
• Incorporate constraints on land use (change)
  e.g., forest reserves
• Incorporate spatially explicit ecological
  feedbacks derived from Focus 3 on future land
  use/cover

70
A conceptual actor-institution-environment
framework is mapped onto a computer model to form
an agent-based dynamics spatial simulation
Focus 4 Dynamic Spatial Simulation (DSS).

• The conceptual framework joins
• Actors agrarian decision making interpreted by
  bounded rationality resource profiles
• Institutions land tenure subsidies
• Environment simple ecological relationships

71
Conceptual Model Relationships
Actors
Actor D-M via decision variables resource
profiles
Actor D-M via decision variables resource
profiles
Active participation
LUCC
Through actors
Institutions
Environment
72
Assertions and Model Configurations

• Different forms of decision making actor
  behavior as simple rules vs. group-based rules
• Actor heterogeneity a single kind of
  smallholder-agent vs. different classes of agents
• Institutions no institutions vs. an array of
  institutions
• Reactive environment unchanging landscape vs.
  an adaptive, reactive environment with endogenous
  transition rules

73
ADSS Implementation
Agent Based Model
The conceptual framework is mapped onto an
agent-based model and generalized cellular
automata within an operating shell.
Institutions Land tenure Subsidies
Actors D-M model
Resources
Shell Monte Carlo Model parameters User
interface Calibration Validation
Environment Land-use/cover Env
hydro/soils/slope/aspect Distance to
market/transport
Generalized Cellular Automata
74
Challenges

• Dimensionality time, space scale
• Construction calibration (esp. agents),
  validation, visualization software integration
• Conceptual barriers adequacy of conceptual
  model, uncertainty, qualitative knowledge ethics

75
Added Elements of Study

• Forest reconstruction from aerial photography
• Household fertility and education study
• Resource institution typology
• Landscape fragmentation assessment
• Chile management strategies
• Origins and structure of chile production
• Forest succession and management practices