From Pixels to Processes: Detecting the Evolution of Agents in a Landscape

1
From Pixels to Processes Detecting the Evolution
of Agents in a Landscape
Department of Geography, SUNY Bufallo, February
2007

• Gilberto Câmara
• Director
• National Institute for Space Research
• Brazil

2
Knowledge gap for spatial data
source John McDonald (MDA)
3
The way remote sensing data is used

• Exctracting information from remote sensing
  imagery
• Most applications use the snapshot paradigm
• Recipe analogy
• Take 1 image (raw)
• Cook the image (correction interpretation)
• All salt (i.e., ancillary data)
• Serve while hot (on a GIS plate)
• But we have lots of images!
• Immense data archives (Terabytes of historical
  images)

4
The challenge of remote sensing data mining

• How many cutting-edge applications exist for
  extracting information in large image databases?
• How much RD is being invested in spatial data
  mining in large repositories of EO data?
• How do we put our image databases to more
  effective use?

5
Land remote sensing data mining A GIScience view

• A large remote sensing image database is a
  collection of snapshots of landscapes, which
  provide us with a unique opportunity for
  understanding how, when, and where changes take
  place in our world.
• We should search for changes, not search for
  content
• Research challenge How do model land change for
  data extracted from a land remote sensing
  database?

6
MSS Landsat 2 Manaus(1977)
7
TM Landsat 5 Manaus (1987)
8
Can we avoid that this.
Source Carlos Nobre (INPE)
9
Fire...
.becomes this?
Source Carlos Nobre (INPE)
10
Dynamic areas (current and future)
New Frontiers
INPE 2003/2004
Intense Pressure
Deforestation
Forest
Future expansion
Non-forest
Clouds/no data
11
Modelling Land Change in Amazonia

• How much deforestation is caused by
• Soybeans?
• Cattle ranching?
• Small-scale setllers?
• Wood loggers?
• Land speculators?
• A mixture of the above?

12
Agent-based models

• Recent emphasis on agent-based modeling for
  simulation of social processes.
• Simulations can generate patterns similar to
  real-life situations
• How about real-life modelling?
• We need to be able to describe the types of
  agents that operate in a given landscape.

13
Extracting Land Change Agents from Images

• Land change agents can be inferred from land
  change segments extracted from remote sensing
  imagery.
• Different agents can be distinguished by their
  different spatial patterns of land use.
• This presentation
• Description of methodology
• Case studies in Amazonia

14
Research Questions

• What are the different land use agents present in
  the database?
• When did a certain land use agent emerge?
• What are the dominant land use agents for each
  region?
• How do agents emerge and change in time?

15
Challenge How do people use space?
Loggers
Competition for Space
Source Dan Nepstad (Woods Hole)
16
What Drives Tropical Deforestation?
of the cases
? 5 10 50
Underlying Factors driving proximate causes
Causative interlinkages at proximate/underlying
levels
Internal drivers
If less than 5of cases, not depicted here.
sourceGeist Lambin
17
Different agents, different motivations

• Intensive agriculture (soybeans)
• export-based
• responsive to commodity prices, productivity and
  transportation logistics
• Extensive cattle-ranching
• local export
• responsive to land prices, sanitary controls and
  commodity prices

18
photo source Edson Sano (EMBRAPA)
Large-Scale Agriculture
Agricultural Areas (ha) Agricultural Areas (ha) Agricultural Areas (ha) Agricultural Areas (ha)
  1970 1995/1996
Legal Amazonia 5,375,165 32,932,158 513
Brazil 33,038,027 99,485,580 203
Source IBGE - Agrarian Census Source IBGE - Agrarian Census Source IBGE - Agrarian Census
19
photo source Edson Sano (EMBRAPA)
Cattle in Amazonia and Brazil Cattle in Amazonia and Brazil Cattle in Amazonia and Brazil Cattle in Amazonia and Brazil
Unidade 1992 2001
Amazônia Legal 29,915,799 51,689,061 72,78
Brasil 154,229,303 176,388,726 14,36
20
Different agents, different motivations

• Small-scale settlers
• Associated to social movements
• Responsive to capital availability, land
  ownership, and land productivity
• Can small-scale economy be sustainable?
• Wood loggers
• Primarily local market
• Responsive to prime wood availability, official
  permits, transportation logistics
• Land speculators
• Appropriation of public lands
• Responsive to land registry controls, law
  enforcement

21
Landscape Analysis Land units associated to
agents
Space Partitions in Rondônia
linking human activities to the landscape
22
Agent Typology A simple example
Tropical Deforestation Spatial Patterns
Corridor, Diffuse, Fishbone, Geometric (Lambin,
1997)
23
Landscape Ecology Metrics

• Patterns and differences are immediately
  recognized by the eye brain
• Landscape Ecology Metrics allow these patterns in
  space to be described quantitatively

Source Phil Hurvitz
24
Fragstats (patch metrics)
25
Some patch metrics

• PARA perimeter/area ratio
• SHAPE perimeter/ (perimeter for a compact
  region)
• FRAC fractal dimension index
• CIRCLE circle index (0 for circular, 1 for
  elongated)
• CONTIG average contiguity value
• GYRATE radius of gyration

26
Increased fragmentation on Rondonia, Brazil
27
Region-growing segmentation
28
Remote sensing image mining
29
Patterns of tropical deforestation (example 1)
30
Patch metrics for example 1
31
Decision tree classifier

• C4.5 decision tree classifier (Quinlan 1993).
• Each node matches a non-categorical attribute and
  each arc to a possible value of that attribute.
• Each node is associated the numerical attribute
  which is most informative among the attributes
  not yet considered in the path from the root.

32
Decision tree for patterns
metrics are perimeter/area ratio (PARA) and
fractal dimension (FRAC)
33
Validation set for decision tree (ex 1)
Validation showed 81 correctness
34
Case Study 1Rondônia
Objective To capture patterns and to
characterize and model land use change processes
TM/Landsat, 5, 4, 3 (2000)
Prodes (INPE, 2000)
Escada, 2003.
35
Spatial patterns in the Vale do Anari
irregular, linear, regular
36
Land use patterns Spatial distribution Clearing size Actors Main land use Description
Linear (LIN) Roadside Variable Small households Subsistence agriculture Settlement parcels less than 50 ha. Deforestation uses linear patterns following government planning.
Irregular (IRR) Near main Settlement main roads Small (lt 50 ha) Small farmers Cattle ranching and subsistence agriculture Settlement parcels less than 50 ha. Irregular clearings near roads following settlement parcels.
Regular (REG) Near main Settlement main roads Medium- large (gt 50 ha) Midsized and large farms Cattle ranching Patterns produced by land concentration.
37
Decision tree for Vale do Anari
38

• Changes in Incra parcels configuration by (Coy,
  1987 Pedlowski e Dale, 1992 Escada 2003)
• Fragmentation
• Transference
• Land concentration

39
Vale do Anari 1982 -1985
REG
Patterns/Typology IRR Irregular Colonist
parcels LIN Linear roadside parcels REG
Regular agregation parcels
Pereira et al, 2005 Escada, 2003
40
Vale do Anari 1985 - 1988
REG
Pereira et al, 2005 Escada, 2003
41
Vale do Anari 1988 - 1991
REG
Pereira et al, 2005 Escada, 2003
42
Vale do Anari 1991 - 1994
Pereira et al, 2005 Escada, 2003
43
Vale do Anari 1994 - 1997
REG
Pereira et al, 2005 Escada, 2003
44
Vale do Anari 1997 - 2000
REG
Pereira et al, 2005 Escada, 2003
45
Vale do Anari 1985 - 2000
REG
REG
Pereira et al, 2005 Escada, 2003
46
Marked land concentration Government plan for
settling many colonists in the area has failed.
Large farmers have bought the parcels in an
illicit way
47
Case study 2 Xingi-Iriri watershed in the state
of Pará
48
Spatial patterns in the Xingu-Iriri region
linear, small irregular, irregular, medium
regular, large regular
49
Land use patterns Spatial distribution Clearing size Actors Main land use Description
Linear (LIN) Roadside Variable Small households Subsistence agriculture Roadside clearings, following main roads
Small irregular (SMALL) Near main settlements and main roads Small (lt 35 ha) Small farmers Family labour and cattle ranching Near main roads and settlements up to 10 Km.
Irregular (IRR) Near main settlements and main roads Small (35 190 ha) Small farmers Cattle ranching Associated to small family households
Medium Regular (MED) Isolated or near secondary roads 190 900 ha Medium farmers Cattle ranching Associated to medium to large farms
Large Regular (LARGE) Isolated or at the end of secondary roads Large (gt 900 ha) Large farmers Cattle ranching Isolated, may have airstrips
50
Decision tree for Terra do Meio spatial patterns
51
Trend towards land concentration where large
farms dominate over small settlements.
52
Conclusions

• Pattern classification in maps extracted from
  images of distinct dates enables associating land
  change objects to causative agent
• Pattern classification techniques associated to
  remote sensing image interpretation are a step
  forward in understanding and modelling land use
  change.
• Next step develop agent-based models for
  deforestation in Amazonia

53
References

• Mining Patterns of Change in Remote Sensing Image
  Databases.Marcelino Silva, Gilberto Camara,
  Ricardo Souza, Dalton Valeriano, Isabel
  Escada.Fifth IEEE International Conference on
  Data Mining. Houston,TX, USA, November 2005.
• "Remote Sensing Image Mining Detecting Agents of
  Land Use Change in Tropical Forest Areas
• Marcelino Silva, Gilberto Câmara, Ricardo Souza,
  Dalton Valeriano, Isabel Escada.
• International Journal of Remote Sensing,
  under review (manuscript available from the
  author).