A Predictive Model of Archaeological Potential

1
A Predictive Model of Archaeological Potential

• Locating Maya Ruins with GIS and Satellite Imagery

2
Presentation Outline

• Goals of this Project
• Study Area
• Physical Setting
• Cultural Setting
• The Maya Culture and Civilization
• Central Place Theory and Ancient Maya Settlement
  Patterns
• Data
• Archaeological Site Database
• Independent Variables
• Predictive Modeling Methods
• Model estimation through Binary Logistic
  Regression
• Results
• Interpreting Predictive Relationships
• Assessing Classification Accuracy
• Employing the Predictive Model
• Biases and Limitations
• Conclusions

3
Goals of this Project

• The purpose of this project is to generate an
  effective and useful archaeological predictive
  model of ancient Maya settlements within an
  inadequately investigated region of northwestern
  Belize

4
Study Area
5
Study Area Physical Setting

• Extreme southern portion of the Yucatán Peninsula
• Limestone bedrock characterized by shallow soils,
  rough relief, little surface water, solutional
  depressions. and sink holes.
• Elevation ranges from 11 to 253 meters.
• Landscape is crossed by several escarpments, each
  about 60 meters high.

6
Study Area Cultural Setting

• Nearly 100 archaeological sites have been
  identified in this region.
• 69 archaeological sites are used in this study
  and are indicated here.
• Most known sites are located near the main road.
• Culturally, the Maya who inhabited this region
  are related to those who lived in the Petén of
  Guatemala.

69 Archaeological sites and methods by which some
were found.
7
The Maya Geographic and Temporal Contexts
The Maya inhabited present day southeastern Mexico
, Belize, Guatemala, and portions of Honduras and
El Salvador.
The first people we recognize as Maya first
settled this region 10,000 to 6,000 years ago and
still inhabit this area today.
After Coe, 200512 and Demarest, 20043
8
The Maya Geographic and Temporal Contexts
9
The Maya Culture and Civilization
Physical Hallmarks of Maya Civilization include
massive architecture, large stucco masks, stelae
and altars, and ball courts.
Tourtellot, Belli, Rose, and Hammond, 2003.
10
The Maya Culture and CivilizationMonumental
Architecture
Range Structure, La Milpa
Plazas are large open areas situated between
buildings which typically face the cardinal
directions
11
The Maya Culture and CivilizationLarge Stucco
masks
Large Residential Structure, Altun Ha Belize
Beginning in the Late Preclassic, main stairways
of structures were adorned with giant plaster
masks
12
The Maya Culture and CivilizationStelae and
Altars
Stelae and altars are typically found together
and were used to commemorate activities of the
ruling class.
13
The Maya Culture and CivilizationBall Courts
Hammond and Tourtellot III, 1993.
Ball courts are found in plazas of the largest
Maya settlements and served ceremonial and ritual
functions.
14
The Maya Culture and CivilizationCommon People

• Common people lived in much smaller communities
  with smaller architecture and a lack of many of
  the hallmarks discussed here.

Main residential structure, Medicinal Trail,
Belize
Stone tool workshop, Medicinal Trail, Belize
15
The Maya Culture and CivilizationCommon People
Stone tool workshop, Medicinal Trail, Belize
Ceremonial plaster bench within a residential
structure at Medicinal Trail
16
Central Place Theory and Ancient Maya Settlement
Patterns

• Central Place Theory, seeks to explain the
  size, number, and distribution of towns based on
  trade and marketing of different types of goods
  (Wheeler et al., 1998153-154).
• Settlement size and market function dictate the
  spatial and hierarchical relationships between
  settlements in a region.
• Four ranks have been identified for the study
  area.

17
Central Place Theory and Ancient Maya Settlement
Patterns

• Large, high ranking (tier 1) sites are located
  distant from each other.

• As the size and rank of sites
• decreases, so does the
• distance between them.

• The smallest of sites
• (tier 4) are often
• located very close
• to one another.

Settlement distribution by rank within the study
area
18
Predictive Modeling

• Predictive modeling is, the production of one
  or more outputs from one or more inputs through
  the use of a specified rule (Wheatley and
  Gillings, 2002166).

19
Predictive Modeling

• Binary Logistic Regression is the preferred
  method for predictive modeling.
• Probability(y) ______________1______________
• (1 Exp -(a0 b1X1 b2X2 bnXn))
• Where Exp is equivalent to the base of the
  natural log system, e (2.718), raised to the
  contents of the parenthesis. A is a constant, b
  represents the coefficient of variable X. The
  probability of presence (1) or absence (0) of
  dependent variable y runs from zero representing
  no probability, to one representing the highest
  probability of site presence.
• Binary Logistic Regression is a powerful modeling
  tool with relaxed statistical assumptions
  including
• No colinearity between variables
• No missing values
• Variables do not have to be normally distributed
• All variables important to the model must be
  included
• Dependent variable must be binary (site presence
  or absence)

20
Data Dependent Variable

• A database of 69 archaeological sites was used
  to generate the dependent variable.
• A subset of 50 sites was chosen and used for
  prediction
• The other 19 will be used to test the model once
  a model has been estimated.

Programme for Belize/ University of Texas Camp
Library
All data used to create the site database was
hand copied from literature in this library.
21
Data Explanatory Variables

• Based on the three models presented here and
  other archaeological predictive models, the
  following variables were chosen

22
Data Explanatory Variables

• Greenness, NDVI, and wetness are all derived
  using weighted linear combinations and basic map
  algebra to multiple spectral bands of Landsat
  satellite imagery
• In other words, areas which are ripe for the
  growth of healthy vegetation.
• It is believed that the Maya would have preferred
  to live in areas which can sustain vegetation and
  crops.

23
Data Explanatory Variables
Vegetation is green where it is dense. Areas
that are dry, have little green
vegetation. Notice areas that stand out in all
three images.
24
The discovery of San Bartolo
Bill Saturno of Hew Hampshire University used
high resolution multi-spectral imagery from
IKONOS to examine the spectral reflectance of the
landscape.
He and Tom Sever from NASA identified a
correlation between the reflectance of specific
bands with large ancient Maya structures.
http//www.sanbartolo.org/technology.htm
25
Remote Sensing in Belize
Last Spring, I spent the semester abroad in the
study area. Upon my return to the U.S. I
attended a conference where Saturno and Sever
presented their findings.
26
Predictive Modeling Methods
Binary Logistic Regression
Archaeological Predictive Model
Backward Wald Stepwise Variable Removal Method
27
Results Predictive Model
Tasseled Cap Greenness, eastern facing aspects,
and the interaction between the sum of flat areas
and distance to those areas.
28
Results

• The best predictive model is estimated as
• Probability(y) 1 / (1 (Exp (-4.1265
  (aspectEast 0.7548)
• (flatInteractive -0.0005) (greenness
  0.0664)))).

29
Results

• Interpreting the coefficients

Greenness is the most important variable to the
predictive model followed by the flatInteractive
variable and aspectEast.
30
Results
Notice the similarities between Tasseled Cap
Greenness and the model
31
Results Assessing Classification Accuracy

• Is the model more effective at correctly
  predicting archaeological sites than random
  chance?
• Proportional chance accuracy ((numPresence /
  N)² (numAbsence/N) ²) 125

32
Results Model Validation

• Kvammes Gain statistic is used to assess the
  predictive accuracy of a probability surface.
• The predictive model is input to the GIS to
  generate a probability surface.
• Probabilities at each of the 19 training sites is
  assessed

33
Results Model Validation

• Gain 1 ( area with high probability /
  sites with high probability).
• Gain 1 (42.82 / 57.89) 0.26

34
Results Employing the Predictive Model
Areas indicated in orange should be
investigated for tier 1 archaeological sites.
35
Results Employing the Predictive Model
Areas indicated in orange should be
investigated for tier 2 and 3 archaeological
sites.
36
Biases and Limitations

• Possible exclusion of variables pertinent to the
  model (requirement of Binary Logistic
  Regression).
• Measurement error of explanatory variables.
• Multi-spectral imagery was not corrected for
  atmospheric conditions.
• Lack of information about locations of known
  non-sites.
• The flatInteractive variable does not accurately
  measure the amount of arable land located within
  a close distance to a settlement.

37
Biases and Limitations

• Remember distFlat flatSum flatInteractive
• Where distFlat measures distance to the nearest
  flat area for agriculture.
• Low values are preferred such that the Maya can
  be close to their fields
• Where flatSum measures the amount of flat areas
  within a 5-10 minute walking radius of a
  settlement.
• High values indicate larger areas of flat land.
• Therefore, when these variables are multiplied,
  neither low nor high values are preferred.
• The variable would be better off measured as a
  ratio
• distFlat / flatSum

38
Conclusions

• This project has accomplished a variety of
  things
• A comprehensive database of archaeological sites
  in the Rio Bravo Conservation and Management Area
  has been compiled.
• The utility of GIS and remote sensing for the
  prediction of ancient Maya archaeological sites
  has been established.
• A predictive model has been generated which
  improves the ability of the archaeologist to
  locate new sites by 25.
• This project also opens up a variety of future
  field work opportunities.

39
Acknowledgements
Questions?

• Dr. Tom Crawford, Advisor
  Dr. Ron Mitchelson

Dr. Rebecca Torres Dr.
Charlie Ewen
Dr. Fred Valdez, Jr. and the University of Texas