University of Southern California Dissertations and Theses

Predicting archaeological site locations in northeastern California’s High Desert using the Maxent model

(USC Thesis Other)

Predicting archaeological site locations in northeastern California’s High Desert using the Maxent model

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Predicting Archaeological Site Locations in Northeastern California’s High Desert
using the Maxent Model

by

Megan Christine Oyarzun

A Thesis Presented to the
Faculty of the USC Graduate School
University of Southern California
In Partial Fulfillment of the
Requirements for the Degree
Master of Science
(Geographic Information Science and Technology)

August 2016

Copyright © 2016 by Megan Christine Oyarzun

To my parents, Roger and Janet Farschon, for your emotional, physical, financial and academic
support. Without you none of this would have been possible.

iv

Table of Contents
List of Figures ................................................................................................................................ vi
List of Tables ................................................................................................................................ vii
Acknowledgements ...................................................................................................................... viii
List of Abbreviations ..................................................................................................................... ix
Abstract ........................................................................................................................................... x
Chapter 1 Introduction .................................................................................................................... 1
1.1 Motivation ............................................................................................................................2
1.2 Project Purpose and Scope ...................................................................................................4
1.3 Structure of this document ...................................................................................................6
Chapter 2 Background .................................................................................................................... 7
2.1 Prehistoric Archaeology of Northeastern California ...........................................................7
2.2 Archaeological Site Prediction Models................................................................................9
2.2.1. Far Western prehistoric site sensitivity model .........................................................11
2.2.2. BLM Distance to water model .................................................................................13
2.2.3. Review of existing models .......................................................................................14
2.3 Maxent for Predicting Prehistoric Archaeology ................................................................15
Chapter 3 Methodology ................................................................................................................ 17
3.1 Study Area .........................................................................................................................17
3.2 Software .............................................................................................................................17
3.3 Archaeology Site Location Data ........................................................................................18
3.3.1. Prehistoric Data Preparation for Maxent .................................................................19
3.4 Environmental Evidence Layers ........................................................................................21
3.4.1. Terrain Features – Slope and Aspect .......................................................................22
3.4.2. Tool Stone Sources ..................................................................................................23
3.4.3. Geologic Units .........................................................................................................24
3.4.4. Large Game Corridors .............................................................................................25
3.4.5. Water Sources ..........................................................................................................26
3.5 Other Data ..........................................................................................................................27
3.6 Maxent Modeling ...............................................................................................................28

v

Chapter 4 Results .......................................................................................................................... 31
4.1 “Kitchen Sink” Results ......................................................................................................31
4.2 Ecological Region Results .................................................................................................33
4.3 Archaeological Site Type Results ......................................................................................34
4.4 Probability Distribution .....................................................................................................36
4.5 Evaluation of Maxent Models ............................................................................................39
4.5.1. Study Area Evaluation .............................................................................................39
4.5.2. Project Scale Evaluation ..........................................................................................42
4.5.3. Discussion ................................................................................................................47
Chapter 5 Conclusions .................................................................................................................. 49
5.1 Discussion ..........................................................................................................................49
5.2 Limitations .........................................................................................................................50
5.3 Future Work .......................................................................................................................50
5.4 Conclusion .........................................................................................................................51
References ..................................................................................................................................... 52

vi

List of Figures
Figure 1 Study Area ........................................................................................................................ 2
Figure 2 Prehistoric features within the Study Area ....................................................................... 4
Figure 3 Kniffen’s map of the traditional Pit River Tribal Boundary ............................................ 8
Figure 4 Pit River Tribal Boundary and Study Area ...................................................................... 8
Figure 5 Archaeological Site Prediction Models Comparison ..................................................... 10
Figure 6 Far Western Study Area and Ecological Zones .............................................................. 12
Figure 7 Archaeological Site Locations Map ............................................................................... 19
Figure 8 Terrain Features – Aspect and Slope .............................................................................. 23
Figure 9 Tool Stone Sources ......................................................................................................... 24
Figure 10 Geologic Units .............................................................................................................. 25
Figure 11 Large Game Corridors .................................................................................................. 26
Figure 12 Water Sources ............................................................................................................... 27
Figure 13 Ecological Regions ....................................................................................................... 28
Figure 14 “Kitchen Sink” AUC .................................................................................................... 32
Figure 15 “Kitchen Sink” Probability Distribution Map .............................................................. 36
Figure 16 Ecological Region Probability Distribution Maps ....................................................... 37
Figure 17 Site Type Probability Distribution Maps ...................................................................... 38
Figure 18 Ecological Regions and Distance from Tool Stone ...................................................... 42
Figure 19 Evaluation of “kitchen sink” model within survey area ............................................... 45
Figure 20 Evaluation of ecological region model within survey area .......................................... 46
Figure 21 Evaluation of site type, lithic scatter model within survey area ................................... 47
Figure 22 Evaluation of site type, rock features model within survey area .................................. 47
Figure 23 Model Success Curve ................................................................................................... 48

vii

List of Tables
Table 1 Existing Model Performance within Study Area ............................................................. 14
Table 2 Archaeological Site Location Data .................................................................................. 21
Table 3 Environmental Evidence Layers Source and Resolution ................................................. 22
Table 4 Replicates chosen for each Maxent run ........................................................................... 29
Table 5 Model Parameters ............................................................................................................ 30
Table 6 “Kitchen Sink” environmental variables ......................................................................... 32
Table 7 Ecological region environmental variables...................................................................... 33
Table 8 Ecological region AUC .................................................................................................... 34
Table 9 Site type environmental variable ..................................................................................... 35
Table 10 Archaeological Site Type AUC ..................................................................................... 35
Table 11 Model Percent Contribution Comparison ...................................................................... 40
Table 12 Survey Area Model Performance .................................................................................. 43

viii

Acknowledgements
I am grateful to BLM NECA Archaeologists for their assistance and support, especially to
Jenifer Rovanpera for the tireless work she did to input and track down data for me. I would like
to thank Bureau of Land Management, Applegate Field Office for providing me the subject
matter and the time to complete this thesis. I am grateful for the data provided to me by the
archaeologists at the Modoc and Lassen National Forests. A special thank you to Karen Kemp,
without your support, advice and pushing me to get back on track, I may never have completed
this thesis.

ix

List of Abbreviations
ASCII American Standard Code for Information Interchange
AUC Area Under the Receiver Operating Characteristic Curve
BLM Bureau of Land Management
DEM Digital Elevation Model
GIS Geographic Information Systems
GPS Global Positioning Systems
NAD83 North American Datum 1983
NECA Northeastern California Archaeologists
ROC Receiver Operating Characteristic
SHPO State Historic Preservation Officer
USDA United States Department of Agriculture
USDI United States Department of Interior
USFS United States Forest Service
USGS United States Geological Survey
UTM Universal Transverse Mercator

x

Abstract
Prehistoric sites and artifacts are common across the country side in the high elevation desert of
California’s northeastern corner. For decades archaeologists have been researching, surveying
and cataloging archeological sites on lands managed by the Bureau of Land Management
(BLM). While thousands of sites have been recorded, it is hard to say how many remain
undiscovered. Multiple archaeological site prediction models have been completed covering the
area to assist archaeologists in locating and recording sites. This project tests the hypothesis that
the site type Maxent model can be as good or a better predictor of archaeological site probability
than the Maxent models that do not categorize by site type. The site type Maxent model will also
be as good or a better predictor of archaeological sites than the previous models at a project
scale. To test this hypothesis three models were run (1) the “kitchen sink”, all 3,729 sites within
the study area, (2) ecological region, using all sites categorized by the ecological region in which
they fall, and (3) site type, a subset of 1,332 sites, categorized by the prehistoric people use at
that site. Maxent uses the spatial location of individual archaeological sites and environmental
variable rasters to produce a probability of distribution raster. At the study area scale the Maxent
software’s built-in validation tools, environmental variable performance and Area Under the
Receiver Operator Curve (AOC) the three Maxent models were compared and to test the
hypothesis. At a project scale a 5,800 km
2
archaeological survey area was used to compare how
well the Maxent models and the previous models were able to predict recorded site locations.
This project was unable to definitively prove the hypothesis; however the results show that the
site type Maxent method of modeling provides a successful method for predicting archaeology
site locations at the study area and project scales, with some additional work being needed.

1

Chapter 1 Introduction
Archaeological survey records date back to the late 1960s on public lands in California’s far
northeastern corner. These records, including more recently documented sites, are how federal
land management agencies preserve what was left behind by prehistoric people. This area is
exciting in an archaeological context, due to the density and types of sites found and the
proximity to other important areas, including Paisley Caves, a nearby site with the oldest radio
carbon dated artifacts in North America (Gilbert 2008).
For decades archaeologists have been researching, surveying and cataloging
archeological sites on lands managed by the BLM and may not have scratched the surface of
what still exists on the landscape. Prehistoric lakes, lava flows, large game populations,
grasslands, and woodlands provided a diverse landscape where Native Americans established
dwellings, gathered and hunted for food, and constructed tools and weapons for survival. The
sites and artifacts that were left behind tell the story of how Native Americans lived and
recording and preserving this cultural history is the only way to ensure that the story can be told.
The best way to preserve archaeological sites is to know where they are located, catalog
the artifacts on those sites, and monitor to see that they are conserved. It is excessively expensive
to do intensive archaeological survey over hundreds of thousands of acres, so focusing on areas
that provided food, water, shelter or other resources to prehistoric residents will aid archeologists
in finding additional archaeological resources. The California State Historic Preservation Officer
(SHPO) has requested that a predictive model be developed and continually updated by the BLM
to assist northeastern California archeologists in their inventory efforts, specifically to direct
field surveys, by conducting more intensive survey where prehistoric archaeological sites are
most likely to occur.

2

The study area is the western portion of the BLM Applegate Field Office, composed of
portions of Modoc, Lassen and Siskiyou counties in California. There are approximately 501,000
acres of BLM managed lands within the study area. Figure 1 shows the BLM lands within the
study area. The purpose of this project is to use the Maximum Entropy software (Maxent)
method of modeling to predict the probability of prehistoric archaeological sites occurring on
BLM managed lands within the study area.

Figure 1 Study Area
1.1 Motivation
The BLM currently has two archaeological site prediction models that cover the study
area. Jerome King and Kim Carpenter of Far Western Anthropological Research Group, Inc.
completed a model in 2004 for the BLM using Weights of Evidence prediction method. A few
years later BLM archaeologists completed a much simpler model internally, using only distance

3

to water. This is the model BLM archaeologists are currently using. Both models are discussed in
further detail in Chapter 2. Both models have around a 70% success rate (70% of recorded sites
fall within areas mapped as having a high probability for archaeological sites). It is hoped that
with additional data and a different modeling approach the success rate can be improved. Since
these models were developed over 10 years ago, the BLM has located and collected information
on hundreds of archaeological sites. Additionally, existing paper records associated with
hundreds of sites have been entered into tabular and spatial databases that increase BLM’s ability
to use predictive models.
The 501,000 acres of BLM managed lands are managed for multiple uses, ranging from
wilderness and recreation to cattle grazing and mining. Any proposed project or use requires a
determination by BLM Archaeologists as to whether it will have detrimental effect to prehistoric
cultural resources. In order to make this determination, a field survey must be conducted to
locate and record the resources within the proposed project area. Projects can range in size from
fractions of an acre to 100s of thousands of acres. With the use of a predictive model, BLM
Archaeologists have the ability to do intensive survey in areas that have a high probability to
contain sites and less intensive survey in areas less likely to contain sites. Using modeling to
guide field survey instead of doing intensive survey on entire project areas can result in savings
of considerable amounts of time and money. The more efficiently the model is able to predict
prehistoric site locations, the more efficient and cost effective field surveys can become.
The Maxent software was chosen to create a new archaeological site prediction model
because it limits human biases and requires presence only locations. The software was developed
in 2004 by Phillips, Dudík and Schapire and has proven to effectively model species distribution
(Merow, Smith and Silander 2013). Within a defined study area, Maxent extracts environmental

4

indicators at species presence locations and uses that information to generate the probability that
a species will occur across the study area (Phillips, Dudík and Schapire 2004). For the purpose of
this project the ‘species’ are recorded prehistoric archaeological sites, which include but are not
limited to lithic scatters, habitation sites, rock features (hunting blinds and rock alignments) and
rock art (Figure 2). Environmental variables provide evidence about the landscape’s suitability
for habitat; thus in this project, “habitat suitability” is the suitability for prehistoric human use.
The environmental evidence was chosen based on the knowledge of BLM Archaeologists as well
as basic human necessity, terrain (slope and aspect), distance to water (springs, waterways and
water bodies), geologic mapping, tool stone sources and large game corridors.

1

1.2 Project Purpose and Scope
The purpose of this project is to produce a reliable archaeological site prediction model
with the Maxent software. In the last few years the BLM has made an attempt to computerize
paper site records, part of this effort is to add attribute information to spatial site data. The
attribute information for each site contains varying information on site type, artifacts and features
present within the site and brief terrain description. This new data provides an opportunity to

1
Photos by Jennifer Rovanpera (2014)
Figure 2 Prehistoric features within the Study area
Habitation site on the left and rock art panel on the left
1

5

create a model based on how a site was used at a specific location, as opposed to just the
location. Environmental factors may vary depending on the use of a site, a habitation site may
need to be closer to water sources than a site used for hunting large game. To test this concept
three different approaches were assessed. The first approach is the “kitchen sink” approach,
using a large amount of presence only data, with no site type categorization, to predict the
presence of archaeological sites. The second approach is a site type approach and categorizes
sites into two types (habitation and rock features). Based on the available attribute information,
site location probabilities are predicted for each site type. The third approach categorizes sites by
ecological region, assuming that within an ecological area the environmental variables would be
more closely related, this is based on the Far Western model methods (King, et al. 2004).
The hypothesis is that the site type Maxent model can be as good or a better predictor of
archaeological site probability than the Maxent models that do not categorize by site type. The
site type Maxent model will also be as good or a better predictor of archaeological sites than the
previous models at a project scale. Thus the initial expectation is that the second approach, site
type, will do the best at predicting the presence of archaeological sites within the study area and
the kitchen sink approach should be the least successful. If correct, this hypothesis might explain
why the two previous models used by the BLM have similar success at predicting archaeological
sites even though they are were produced with greatly different approaches and complexity. To
test the hypothesis at the study area scale, the performance of each model run was evaluated
from tools built into Maxent, at the project scale the models were evaluated by looking at a
survey area to see if the highest probability areas captured recorded sites.

6

1.3 Structure of this document
The goal of this project is to use currently available data to determine if the site type
approach for archaeological site prediction modeling produces meaningful results. This project
focuses on the usefulness of the Maxent tool and Geographic Information Systems (GIS) to
inform BLM Archaeologists on how future data collection and input can assist in improving
model success in the future. Chapter 2 gives background on the archaeology of the study area as
well as, previous archaeological site prediction models and the use of Maxent for predicting
archaeology sites. Chapter 3 outlines the data and software used to model prehistoric
archaeology site locations using Maxent software. Chapter 4 presents the results of each of the
Maxent model runs, “kitchen sink”, ecological region and site type and compare those models at
the study area and project scales. Chapter 5 summarizes the conclusions made after comparing
the results.

7

Chapter 2 Background
There are two main topics to address when discussing the predictive modeling of archaeological
site predictions. The first topic is the prehistoric people themselves and the associated
archaeological sites that were left behind. Second is existing methods for predicting
archaeological site locations of those people.
2.1 Prehistoric Archaeology of Northeastern California
Without first understanding how prehistoric people used the landscape and what
environmental variables were desirable or undesirable, it is not possible to model where the
remains of their existence will occur. According to US Forest Service and BLM documentation
the earliest humans occupied the area during the Early Holocene, roughly 12,000 years ago. The
earliest people did not settle in one place, but moved around gathering food. It was not until
roughly 7,000 to 5,000 years ago that the first settlements were established. The most well
documented era of prehistoric occupation is the Terminal Prehistoric period, 600 years ago to the
first contact with western European settlers, this is the era that is described below (USDA Forest
Service, USDI Bureau of Land Management 2007).
In 1920s both Kniffen (1928) and Merriam (1926) published articles on the geography of
the Pit River Tribe in California, based on interviews with tribal members. Kniffen’s map, shown
in Figure 3, depicting tribal boundaries is still in use by the Pit River Tribe today. Figure 4 shows
the Kniffen tribal boundaries overlaid with the study area. The Pit River Tribe were the
predominant inhabitants of the study area for this project. A smaller area in the northern portion
of the study area is part of the traditional homeland of the Modoc Tribe (Merriam 1926). The
boundaries of the Modoc people are not as well defined as that of the Pit River, but the southern

8

boundary is similar to the northern border of the Pit River as described by Kniffen (King, et al.
2004).

Figure 3 Kniffen’s map of the traditional Pit River Tribal Boundary (Kniffen 1928)

Figure 4 Pit River Tribal Boundary and Study Area

9

The Pit River Tribe is named after the Pit River which flows from the east side of the
study area to the southwest corner. The far northeastern and southwestern portions of the study
area are mountain pine forest and the northeastern area is dominated by high elevation lava flows
and marshes. These high elevation areas were only utilized by the Pit River Tribe in the summer
months when the snow pack had melted and travel on foot would be possible (Kniffen 1928).
Kniffen describes the main habitation sites of the Pit River Tribe as being along the river itself as
well as in the lower elevation valleys. These areas provided protection from snow in the winter
and a large selection of wild edible plants in the summer months. Although not desirable for
habitation, the lava flows in the north were visited frequently because of their abundance of raw
materials for making tools and weapons (Merriam 1926). The Modoc Tribe, having a similar
range of ecologically diverse territory, chose habitation sites much like those of the Pit River.
Within the study area, they chose lower elevation sites near lakes and marshes for the availability
of food sources (King, et al. 2004).
Both the Pit River and the Modoc gathered edible vegetation, fished and hunted large and
small game: mule deer, antelope, sage hen, and numerous small mammals (Kniffen 1928; King,
et al. 2004). The foothills of the Warner mountain range on the eastern side of the study area
were habitat to large numbers of deer and antelope (Kniffen 1928). In the lava flows in the north,
the Modoc hunted mountain sheep (King, et al. 2004). The marshes in the north and the low
desert plains in the south were gathering places for root vegetables (Kniffen 1928; King, et al.
2004).
2.2 Archaeological Site Prediction Models
As discussed in Chapter 1, there are two previously developed models that cover the
study area. The models are vastly different in their approach and complexity but similar in their

10

success at predicting prehistoric archaeological sites within the study area. Figure 5 compares the
two models discussed below and shows each model’s success for an area survey area during
2014 field season by BLM Archaeologists, using 20 meter transects.

Figure 5 Archaeological Site Prediction Models Comparison
Top Far Western model, Bottom BLM model

11

2.2.1. Far Western prehistoric site sensitivity model
Far Western Anthropological Research Group, Inc. was contracted by the BLM to
complete a report on the cultural resources of northeastern California and in 2004 they published
that report (King, et al. 2004). As part of that report they developed a Prehistoric Site Sensitivity
Model using the Weights of Evidence modeling technique. Weights of Evidence is a Bayesian
prediction method initially applied in medical diagnosis (Lusted 1968). This method was later
adapted to work with spatial data for use in geologic studies, treating raster cells as an ensemble
of independent models (Bonham-Carter 1994). Similar to logistic regression, Weights of
Evidence relies on the logistic transformation to deal with a continuous range of outcomes,
ranging from highly likely to highly unlikely (Bolstad 2010).
A Weights of Evidence model is trained on a set of specific points, the "training points",
in this case known archaeological sites, in combination with the corresponding evidential rasters
(King, et al. 2004). The map from Far Westerns report is displayed in Figure 6, the study area for
this project is the displayed as the BLM Field Office boundary in the north western corner. By
observing the presence and absence of training points in raster cells, weights are developed for
each cell of the evidential layer, the presence of training-points within a particular raster value
constitutes a positive weight, the absence of training-points a negative weight. Training points
will be associated with some values (positive weights) and not associated with other values
(negative weights) in an evidence layer. The “contrast” between the positive and negative
weights for an evidential layer is a strong measure of how predictive that layer is (King, et al.
2004). Although a proven modeling technique, Weights of Evidence was not chosen for this
project because the software is out of date and no longer compatible with the latest Esri software
which it needs to run.

12

Figure 6 Far Western Study Area and Ecological Zones (King, et al. 2004)
Far Western separated their study area into four ecological zones to better represent the
variability in ecological settings across the study area. The ecological zones were determined
based mainly watershed boundaries and vegetation communities, to account for environmental
differences across the study area (King, et al. 2004). They ran the Weights of Evidence model for
each of those ecological zones using slope, aspect, landform type, hydrologic features and

13

vegetation as the evidential layers. The archaeological site data was provided by the BLM, USFS
and the Northeast Information Center at California State University Chico. The resulting
sensitivity model was categorized as low (<0.5 times average site density), moderate (0.5 - 1.25
time average), high (1.25 - 3 times average) and very high (> 3 times average) (King, et al. 2004).
2.2.2. BLM Distance to water model
In 2007, the United States Forest Service (USFS) and the Bureau of Land Management
completed the Sage Steppe Ecosystem Restoration Strategy Draft Environmental Impact
Statement. This document was the first step in an effort by the USFS and BLM to restore
declining habitats on 6.5 million acres of Federal Land. In order to ensure that prehistoric
archaeology was preserved, while being able to complete restoration work on a large and diverse
area, the California SHPO requested the BLM create a predictive model to guide field work.
During a meeting of the BLM Northeastern California Archaeologists (NECA) group (Jenifer
Rovanpera, David Scott, Sharron-Marie Blood and Marilla Martin) in June 10, 2013, the
development of the model was discussed.
In 2010, the NECA group working with a BLM GIS specialist created a distance to water
model. The model initially had two parameters: (1) Distance from water source parameter, 200
meters from either side of a stream and surrounding a spring or natural water body; and (2) Slope
parameter, omitting any area with slope of 25 degrees or greater. During field surveys and testing
of the model it was decided that a large enough percentage of sites were falling outside the
model and the decision was made to remove the slope parameter from the model. The model
became purely a 200 meter buffer of water sources.

14

2.2.3. Review of existing models
The two models discussed in this section are very different and both have limitations in
predicting archaeology site locations. However, even with their great differences in approach,
they produce similar results at the project area scale, Table 1 shows the similarities with the
assumption that the BLM distance to water is comparable to the High and Very High sensitivity
categories of the Far Western model. This could be merely coincidence or an indicator that they
have a similar design flaw, not factoring in site type limits the ability of the model to produce
meaningful results.
Table 1 Existing Model Performance within Study Area
Model
Total Sites
within study
area
Sites within
high
probability
Percent
Found
BLM distance to water 1,467 1,050 72%
Far Western prehistoric
site sensitivity
1,467 1,045 71%

The BLM distance to water model makes assumptions about the importance of water to
prehistoric people. Assuming that all activities and necessities occur within a certain distance to
water sources is problematic. Water is necessary to sustain life, so being close to water is
important when selecting habitation sites. However, other activities that are also necessary to
sustain life, such as collecting or hunting for food, increase the likelihood of prehistoric people
moving away from water sources. Additionally, the availability of water on the landscape
changes seasonally and over longer periods of time due to variability in weather patterns and
climate.
Far Western’s model used vegetation as one of the evidential layers, vegetation has
changed drastically since the first prehistoric people inhabited the area 12,000 years ago. Large

15

changes in climate would have greatly affect the amount of rainfall, increased the size and
amount of lakes and meadows, which would have huge effects on vegetation communities. Far
Western also chose to omit tool stone sources as an evidence layer, even though the data was
available and the Weights of Evidence method used could report the success of the layer at
predicting sites (King, et al. 2004).
At a project scale the two models appear very similar and have varying success at
predicting archaeological site locations. For the project scale analysis shown in Figure 5,
prediction similarities are apparent between the two models: in the Far Western Weights of
Evidence model, water courses were a strong predictor and springs were not (King, et al. 2004).
Thus, the very high and high sensitivity areas are similar to the BLM water proximity model.
2.3 Maxent for Predicting Prehistoric Archaeology
The maximum entropy technique is what the Maxent software uses to make predictions.
Using a sample of locations within a defined area and a set of variables the Maxent technique
calculates a range of environmental values that are predictors of the sample locations, from that
range the distribution of maximum entropy is selected (Phillips, Dudík and Schapire 2004). A
presence only species data set, with spatial coordinates, multiple environmental variables and a
defined study area boundary are all that are need to run the Maxent software. What it predicts is
the environmental suitability across the study area by using the environmental conditions found
at each of the occurrence points (Phillips, Anderson and Schapire 2006). Maxent does multiple
iterations within a “black box” modeling technique, to optimize the suitability distribution (Kern-
Isberner, Wilhelm and Beierle 2014).
Maxent software was developed in 2004 by Phillips, Dudik and Schapire for use in
conservation of animal and plant species. Animal and plant species distribution is driven by

16

environmental variables. While human behavior is slightly less prone to environmental variables,
prehistoric people’s distribution is much more influenced by environmental variables than that of
modern people. This makes Maxent a good tool for predicting the environmental suitability of
locations for use by prehistoric people across the study area.
For this study the presence data is archaeological sites with locations recorded during
field survey. The environmental variables were selected based on King et al. 2009 and personal
communication with the BLM NECA group. Slope, aspect, distance to water sources, distance to
tool stone sources, distance to large game corridors and geologic units were selected to predict
the occurrence of prehistoric people across the study area.
Each of these environmental layers as well as the application of Maxent to these data are
next discussed in greater detail in Chapter 3.

17

Chapter 3 Methodology
In order to test the hypothesis, that using a site type model will be as good or a better predictor of
archaeological site location than an uncategorized model, the presence of site data was
categorized three different ways and three runs of the Maxent software were conducted. For each
run the environmental data remained the same. The following is a discussion of the geographic
context of the study area, the sources of each of the presence and environmental data layers, the
basis for the model set up and the tools that were used to assess and compare the models that
were produced.
3.1 Study Area
The study area, shown above in Figure 1, is the western portion of the Applegate Field
Office, BLM. It is located in northeastern California, containing 501,000 acres of public lands
managed by BLM and 1.9 million acres of USFS managed public lands. Ranging greatly in
ecological diversity, the study area contains pine forest, high desert plateau, wetlands,
grasslands, basalt lava flows and river basins. It ranges in elevation from approximately 3,000 to
7,500 ft. This is a rural area with no large cities. The largest disturbance to prehistoric sites since
the arrival of European settlers to the area has been from the clearing of land for agriculture as
Kniffen described in 1928.
3.2 Software
This project utilized Esri® ArcGIS™ version 10.3.1, including ArcMap and ArcCatalog
with the ArcGIS Spatial Analyst license. The XTools Pro version 11.1 toolbar for ArcGIS
desktop and Microsoft Excel® 2010 was also used in the preparation of data. The modeling was

18

done using Maxent version 3.3k. Maxent is a free software program available online for
download from Princeton University
2
.
3.3 Archaeology Site Location Data
Archaeology site location data has been collected within the study area in the form of site
records since the 1960s. Although this data is considered sensitive and is not provided to the
general public, it was graciously provided for use in this project by the BLM Applegate Field
Office, and the USFS (Modoc National Forest and Lassen National Forest) in the form of
ArcGIS geodatabases. Each of the three data sources is a combination of legacy data, data
digitized from 24k topographic maps and data collected with professional grade Global
Positioning Systems (GPS) devices. The data digitized from 24k topographic maps has an
accuracy of approximately 14 meters; the data collected with GPS devices has an accuracy of 10
meters or better. The majority of the data was collected during field survey of specific project
areas. These project specific surveys cause the data to have small clusters within the distributed
data as a whole. These clusters may influence the final Maxent output, this sampling bias may
cause the model to be weighted towards areas that have a higher number of samples (Phillips,
Dudík and Schapire 2004). Areas such as privately owned lands that have not been sampled, tend
to be areas around large water sources, lakes and rivers, as well as the most fertile lands for
agricultural production. Figure 7 shows the distribution of the archaeological site locations
across the study area.

2
https://www.cs.princeton.edu/~schapire/maxent/

19

3.3.1. Prehistoric Data Preparation for Maxent
Archaeological site presence data must be in a comma delineated text file with three
required fields, ‘species’, X-coordinate and Y-coordinate, to be compatible with the Maxent
software. The ‘species’ field allows for the categorization of site types: if all sites have the same
‘species’, this translates to the undifferentiated model, and if various ‘species’ (i.e. site types) are
given, this translates to multiple site types. The USFS data was in three formats, point, line and
polygon feature classes and the BLM data was in a polygon feature class. Both data sources
contained both historical as well as prehistoric data; for this study the historic data was removed.
The USFS data contained very little attribute information and could not be categorized into site
Figure 7 Archaeological Site Locations Map

20

types (e.g. ‘species’), while the BLM data had a large amount of attribute information which was
used to make site type categorizations. All data was projected into North American Datum 1983,
Universal Transverse Mercator (UTM) zone 10.
The polygon and line features were converted into point features using the Feature to
Point tool, in the Data Management toolbox within ArcMap. This tool converts the center point
of the feature into a point and exports the resulting data into a shapefile. To organize the site
location data for each of the three runs, the data was processed as follows. Table 2 summarizes
the resulting data prepared for Maxent.
1. The site type approach used only BLM site data because it was the only dataset that
contained attribute information about the artifacts and features at each site, providing
the basis to categorize by the type of site. There was sufficient attribute information
to create four categories, however due to the low number of sites in two of the
categories only the two with the highest number of sites were used. The categorized
point shapefile, was used as the ‘species’ input for the Maxent model.
2. The “kitchen sink” approach used all of the BLM and USFS site data and, using the
Merge tool from the Data Management toolbox, combined the individual layers into
one shapefile. The ‘species’ type distinction was not used for this run.
3. For the ecological approach, an Ecological Region layer was created (discussed in
further detail in Section 3.5). The ecological regions were intersected with the site
location point shapefile created for the kitchen sink approach, adding an ecological
region ‘species’ type to each site record.
The remaining steps were done for each of the three shapefiles created for the three
different approaches. The X and Y coordinates were calculated for each point within the attribute

21

table. The value at each point for each of the environmental evidence layers was also extracted
and added to the attribute table for each layer. The addition of this data helps Maxent run more
efficiently and save time. Each of the site datasets was then exported and converted into a
comma delineated text file.
Table 2 Archaeological Site Location Data
Maxent Run Data Source ‘Sp e cies’ Number of sites
Site Type BLM
Lithic Scatter 1,195
Rock Feature 137
Habitation* 90
Rock Art*
22
Eco Region BLM and USFS
Fall River 426
South Fork Pit River 1,029
Tule Lake 1,554
Warm Springs
720
Kitchen Sink BLM and USFS Archaeological Site
3,729
* These categories were not used because of the small amount of data
3.4 Environmental Evidence Layers
The environmental evidence layers used in this project were chosen because of the effect
they would have had on influencing the behavior of prehistoric people across the landscape. This
section discusses why each data category was chosen and the resulting layers. Table 3
summarizes the environmental variables and their data sources. Each of the environmental
variable layers must be in the form of an ASCII grid, with matching raster cell size and grid
placement to be compatible with the Maxent software. Esri ArcMap software allows for
geoprocessing environments to be set for all data processed within an ArcMap session and the
following environments were set: 1) Project all data into North American Datum 1983, UTM
Zone 10; 2) Clip all layers to the study area; 3) Raster analysis cell size of 30 meters; 4) Snap to
raster (aligned all raster grids to the aspect raster as this was the first raster created). This insured

22

that as all of the environmental evidence layers were identical, in shape, cell size, orientation,
and projection.
Table 3 Environmental Evidence Layers Source and Resolution
Environmental
Evidence Layer
Original Data
Format
Source Resolution
Slope Raster - GRID
USGS Digital Elevation Model 30
meter
+/-30 Meters
Aspect Raster - GRID
USGS Digital Elevation Model 30
meter
+/-30 Meters
Tool stone Source Vector - Point BLM Obsidian Source Layer +/-12 Meters
Geologic Units Raster - Tiff
USGS 1 arc Second Digital Elevation
Model
30 Meters
Large Game
Corridor
Vector - Polygon
California Department of Fish and
Wildlife
+/- 30 Meters
Streams and
Water bodies
Vector - Line
and Polygon
USGS National Hydrographic
Dataset
+/- 30 Meters
Springs Vector - Point BLM, digitized from USGS 24k Topo +/-12 Meters
3.4.1. Terrain Features – Slope and Aspect
Terrain features have a large impact on the movement of people across a landscape, as
well as the habitability of one area over the other. The steep slopes are difficult to traverse and
would be avoided for hunting and gathering activities as well as locations for habitation sites.
Areas of little to no slope would be ideal for habitation sites. On the steepest slopes (near 90°)
rock art can be found. Aspect greatly affects the temperature, hours of sunlight and the
vegetation on a slope.
The slope and elevation were created from United States Geologic Survey (USGS)
Digital Elevation Model (DEM) with a 30 meter cell size, in geographic coordinate system North
American Datum 1983 (NAD83). Using the Slope and Aspect tools available in the Spatial
Analyst within ArcMap, the output raster type of ASCII (American Standard Code for
Information Interchange) grid to be compatible with the Maxent software was chosen. As noted

23

above, the data was projected into North American Datum 1983, UTM Zone 10 during the
processing. Figure 7 shows the resulting aspect and slope rasters.

3.4.2. Tool Stone Sources
The tool stone sources are obsidian flows which form as slow moving lava cools and
forms glass flows (Weldon 2010). This important resource for prehistoric people is used to create
projectile points and other tools. It was also used for trade and prehistoric people would travel to
these sources (Merriam 1926). The tool stone source environmental layer is a vector point file, in
the NAD83, UTM zone 10 projection. The layer was provided for this research by BLM
Archaeologists, who compiled and maintain this layer. Data for the layer is gathered from several
sources, the Northwest Research Obsidian Studies Laboratory
3
, Geochemical Research
Laboratory
4
and BLM archaeology site record data. The points are center points or within a flow,
there is no way of knowing the size and shape of the obsidian flow that each point represents.

3
http://www.obsidianlab.com
4
http://www.geochemicalresearch.com
Figure 8 Terrain Features – Aspect and Slope

24

The tool stone sources layer had to be converted to a raster to use in Maxent. This was
done using the Euclidian Distance tool in Spatial Analyst within ArcMap. The Euclidian
Distance tool creates a continuous distance raster, where each cell’s value is the distance to the
nearest source. Using the environmental settings discussed earlier in this chapter, the raster was
created with a 30 meter cell size and clipped to the study area. It was then converted into the
ASCII grid format for use in Maxent. Figure 8 shows the resulting tool stone source raster.

3.4.3. Geologic Units
Geologic units were selected for this project because of the large amount of information
that can be inferred from the underlying geologic features. The geologic map unit gives
information on the age of geologic features. Basalt lava flows from the Pleistocene and Holocene
eras would mean active volcanic activity that would have been avoided by prehistoric people of
that period and they would have been free of vegetation for the period following. In later
prehistoric times geologic features are an indication of the possible soil depth and fertility.
Figure 9 Tool Stone Sources

25

The geologic unit layer is a vector polygon layer digitized from a 1:100,000 scale USGS
Geologic map of northeastern California, in geographic coordinate system NAD 1927. The layer
was converted from vector polygon to GRID raster using the Polygon to Raster tool in the
Conversion toolbox in ArcMap. The resulting raster has a 30 meter cell size, was clipped to the
study area and projected in NAD83, UTM zone 10. The raster was then converted into an ASCII
grid for use in Maxent, shown in Figure 9.

3.4.4. Large Game Corridors
The large game corridors are deer and pronghorn antelope migration corridors and
seasonal use areas. Large game provides an important food source that could feed many people
and for the Pit River tribe, large game drives involved multiple groups (Kniffen 1928).
Archaeological site records detail evidence of large game hunting within and near these
corridors, projectile points and a game drive (Scott and Oyarzun 2012). The data for large game
corridors used in this project were developed in the 1970s and then digitized and reviewed in
Figure 10 Geologic Units

26

2001 by California Department of Fish and Wildlife. The data are in vector polygon format and
in NAD83, UTM zone 10. Using the same methods as the geologic data, the polygon data were
transformed into a 30 meter, distance from source raster layer, before being converted into an
ASCII grid for use in Maxent (Figure 10).

3.4.5. Water Sources
Water sources were split into two categories, a spring layer and a natural watercourses
and water bodies layer. The reason water sources were separated into these categories was
described by King et al. (2004) who noted that as the number of watercourses and water bodies
increase in an area, the importance of springs as a factor for survival decreases. The two were
separated for this study to see if one had more of an impact on the model than the other.
The source of the water courses and water bodies is the USGS National Hydrographic
Dataset, medium resolution data, at the 1:100,000 scale. All man made features were removed
from both the vector line data for water courses and the polygon data for water bodies. No data
Figure 11 Large Game Corridors

27

was available about the width of the water courses so the water course lines were buffered by
one meter to convert the data into polygons and then merged with the polygon water body data.
The resulting layer was then converted into a 30 meter distance to water raster, using the same
methods as described earlier and then converted into the ASCII format for use in Maxent.
The spring data for this project was assembled from two sources, the BLM water source
improvements layer (collected with a professional grade GPS unit) and by digitizing from a
1:24,000 USGS topographic map. The two data sources were merged and the resulting distance
to springs layer was prepared in the same manner as the tool stone data layer described earlier in
this chapter. The resulting ASCII rasters for watercourses and water bodies, as well as for
springs is shown in Figure 11.

3.5 Other Data
The ecological regions were based off of the Far Western model ecological zones, shown
in Figure 6. However, for this project the ecological regions were adjusted to better represent a
Figure 12 Water Sources
Watercourses and Water Bodies on the left, Springs on the right

28

smaller study area than Far Western used. The USGS Watershed Boundary dataset, subbasins
were used for the basis of the layer. On the edges of the study area slivers of subbasins outside of
the study area boundary were combined to with subbasins within the study area. On the eastern
portion of the study area subbasins were divided based on the fifth level watershed boundaries, to
better represent the more cohesive environmental variables on the eastern side of the study area.
The resulting ecological regions are displayed in Figure 13.

Figure 13 Ecological Regions
3.6 Maxent Modeling
While the Maxent software is easy to use, with the input data in the correct format the
user must then set the parameters of the model to produce the best result for the data being
modeled. The following section outlines the model parameters selected and why those choices

29

were made. Within the software the user selects the ‘species’ to model (if any), the
Environmental layers, and output format and file type. In addition, the user can adjust settings for
each model run. The output for this project is Logistic, this output uses post processing to create
the probability that the ‘species’ will occur in each modeled location (Phillips, Dudík and
Schapire 2004).
For this project, in each of the three runs—site type, ecological regions and the “kitchen
sink”—the parameters were set the same, with the exception of the number of replicates run.
This decision was made based upon the time it would take for the model to run since replicates
are run for each ‘species’. Table 4 shows the number of replicates used for each of the three runs.
In order to produce the best results, over 20 test runs were made to evaluate different parameter
settings; only the parameters selected for the final runs are discussed here. Maxent also has many
available settings; only the selected settings or settings changed from the default settings are
discussed. Table 5 summarizes the selected parameters and the rationalization for each of those
selections.
Table 4 Replicates chosen for each Maxent run
Maxent Run Number of replicates Site Type (Species) Number of Sites
Site Type 25
Lithic Scatter
1,195
Rock Feature
137
Eco Region 10
Fall River
426
South Fork Pit River
1,029
Tule Lake
1,554
Warm Springs
720
Kitchen Sink 25 Archaeological Site
3,729

30

Table 5 Model Parameters
Parameter Selection/entry Rationalization
Create Response
Curves
Selected Response curves display how each of the environmental
variables performed for each ‘species ’ run
Default
Prevalence
0.8 Probability that a ‘species ’ will occur at any occurrence
point. Based on archaeology survey data, the probability
is high that there will be an occurrence within an
occurrence raster cell. Default is 0.5
Jackknife Selected Test determines the importance of each environmental
variable
Maximum
Iterations
500 Iterations of optimization algorithm, the more iterations
the more the model is trained
Random Seed Selected Different set of random points are selected for test and
training samples
Random Test
Percentage
20 Percent of random points set aside for testing the model
Regularization
Multiplier
5 More evenly distributed probability as this number
increases (default is 1)
Replicated Run
Type
Bootstrap Uses 20% of randomly selected points for each of the
replicates
Replicates See Table 4 Numbers chosen to be high enough to create average
and median outputs, while remaining small enough for
the Maxent to run in a reasonable amount of time

Although the number of replicates as well as the number of sites varies for each of the
runs, the built-in model validation tools provide enough information that the runs can be
compared. The Receiver Operating Characteristic (ROC) curve, Area Under the ROC Curve
(AUC), response curves and jack-knife testing, assess the models overall performance as well as
that of each of the environmental evidence layers (Phillips n.d.). All of these results are discussed
in the next chapter.

31

Chapter 4 Results
Three runs were conducted using the Maxent software program, each using the same Maxent
parameter settings as well as the same environmental evidence rasters and varying “species” or
site type presence point locations. This chapter discusses the output of each of the three runs and
assesses the fit of each model using the results of the built-in validation tools. The final product
is a probability distribution map that Maxent produces for each ‘species’ model run.
4.1 “Kitchen Sink” Results
The “kitchen sink” approach ran 25 replicates and only one ‘species’ type, archaeological
site, of which there were 3,729 sites. Using the bootstrap method, 20% of the total sites were
held back during each replicate run for testing.
Maxent provides some very important information in the output of the model run for
assessing for each environmental factor and the model as a whole. The percent contribution of
each environmental variable, how much that variable contributed each of the presence point
locations is summarized in Table 6. It also gives the permutation importance, which tests how
the model reacts if the values of that variable were altered (Phillips n.d.). Given this information,
the stability of each environmental factor can be assessed, an unstable variable has high percent
contribution and a high permutation value. A stable variable has high percent contribution and a
low permutation importance. Geologic unit had the largest percent contribution and a moderately
high permutation importance. Distance from game corridor had a moderately high percent
contribution and very high permutation importance value. Over all the environmental variables
are fairly unstable.

32

Table 6 “Kitchen Sink” environmental variables
Variable Archaeological Site

Percent
contribution
Permutation
importance
Geologic Unit 39% 12.8
Distance from watercourses and water bodies 23% 15.1
Distance from large game corridors 16% 31.0
Distance from tool stone sources 9% 15.6
Slope 8% 14.6
Distance from springs 3% 8.5
Aspect 2% 2.2

Another important test of the overall model performance is the ROC and AUC. The AUC
tells how well the model is able to predict the difference between the presences and random. The
model fit can be determined based on how close the AUC is to 1. Maxent averages the AUC
from each of the 25 replicates runs to come up with the AUC for the model. This model has an
AUC of 0.793 with a standard deviation of 0.003, shown in Figure 12. This model performed
well.
Figure 14 “Kitchen Sink” AUC

33

4.2 Ecological Region Results
The ecological region approach ran 10 replicates of the four ecological regions, Fall
River (426 sites), South Fork Pit River (1,029 sites), Tule Lake (1,554 sites) and Warm Springs
(720 sites). Using the bootstrap method 20% of the total sites for each ecological region where
held back during each replicate run for testing. The following is the results of the Maxent
assessment of the ecological region variables and the fit of the model AUC.
The percent contribution of each environmental variable, how much that variable
contributed each of the presence point locations is summarized in Table 7. Distance to tool stone
contributes the most to the model for each of the four ecological region models. For the Fall
River and South Fork Pit River Models distance to tool stone is very unstable, but in the Tule
Lake and Warm Springs models it is very stable.
Table 7 Ecological region environmental variables
Variable Fall River
South Fork
Pit River
Tule Lake
Warm
Springs

Percent
contribution
Permutation
importance
Percent
contribution
Permutation
importance
Percent
contribution
Permutation
importance
Percent
contribution
Permutation
importance
Geologic Unit 5% 0.7 0.4% 0.7 2% 18.1 3% 2.3
Distance from
watercourses and
water bodies
15% 22.8 12% 10.0 12% 10.3 11% 5.1
Distance from large
game corridors
5% 2.1 8% 13.7 12% 28.8 16% 49.5
Distance from tool
stone sources
72% 64.1 71% 63.3 63% 10.5 54% 1.1
Slope 0.9% 1.0 2% 6.4 6% 25.3 9% 14.3
Distance from springs 0.5% 9.3 0.4% 5.8 6% 4.7 7% 26.8
Aspect 0.1% 0.1 0.1% 0 0.3% 2.4 0.4% 0.9

34

The AUC for each of the four ecological region models show that each model performed
very well, with the South Fork model having the best fit. The AUC and standard deviations are
displayed in Table 8.
Table 8 Ecological region AUC

Fall River
South Fork
Pit River
Tule Lake
Warm
Springs
Mean
AUC
0.882 0.903 0.852 0.823
Standard
Deviation
0.007 0.004 0.007 0.007

4.3 Archaeological Site Type Results
The site type approach ran 25 replicates of the two site types, lithic scatter (1,195 sites)
and rock features (137 sites). Using the bootstrap method, 20% of the total sites for each of the
site types where held back during each replicate run for testing. The following is the results of
the Maxent assessment of the environmental variables and the fit of the model AUC.
The percent contribution of each environmental variable, how much that variable
contributed each of the presence point locations is summarized in Table 9 for the archaeological
site types. Distance to large game corridors contributes the most to the model for each of the site
type models. Distance from tool stone sources is also a high contribution to the models of both
site type model and is a much more stable indicator.

35

Table 9 Site type environmental variable
Variable
Lithic
Scatter
Rock
Features

Percent
contribution
Permutation
importance
Percent
contribution
Permutation
importance
Geologic Unit 5% 1.4 5% 3.7
Distance from watercourses and
water bodies
16% 26.5 12% 13.0
Distance from large game
corridors
48% 17.4 47% 41.8
Distance from tool stone sources 21% 32.2 15% 5.3
Slope 5% 2.1 7% 20.9
Distance from springs 5% 19.9 11% 14.1
Aspect 1% 0.5 3% 1.1

The AUC for each of the site type models show that each model performed very well,
with the rock features having the best fit of the two models. The AUC and standard deviations
are displayed in Table 10.
Table 10 Archaeological Site Type AUC

Lithic
Scatter
Rock
Features
Mean AUC 0.86 0.905
Standard
Deviation
0.003 0.011

36

4.4 Probability Distribution
Maxent produces an ASCII raster of the probability distribution. It is an average of the
replicates for each of the ‘species’ run. Each map displays a continuous probability distribution
raster where the probability of a site occurring is calculated for each 30 meter cell. Figures 13, 14
and 15 display the resulting rasters for the “Kitchen Sink”, ecological region and archaeological
site type respectively.
Figure 15 “Kitchen Sink” Probability Distribution Map
The probability distribution for the “kitchen sink” model, displays how the high percent
contribution from geologic unit, distance from watercourses and water bodies and distance from

37

large game corridors contributed to the distribution, with the highest probabilities falling within
one geologic type and within areas close to water and game corridors. Also, visible in the map is
the areas with the lowest and highest percent slope are lower probability.

Figure 16 Ecological Region Probability Distribution Maps
Each of the four ecological region probability distributions maps display the high percent
contribution of distance from tool stone sources. Tule Lake, Warm Springs and South Fork Pit
River highest probability areas correlate to areas close to tool stone sources. For the Fall River
model the highest probability area correlates to the area farthest away from tool stone sources.

38

Figure 17 Site Type Probability Distribution Maps

39

The site type probability distribution maps, lithic scatter and rock features, both show a
strong correlation with large game corridors, which had the highest percent contribution to both
models. The lithic scatter probability map also shows a strong connection to tool stone sources,
the highest probability areas are close to sources. The rock feature probability distribution also
shows high probability areas that are moderate slopes, while very steep and flat areas are low
probability.
4.5 Evaluation of Maxent Models
Each of the three Maxent runs had their own set of successes and challenges. In this
section the models are evaluated for the whole study using the tools built into the Maxent
software. The models are also evaluated for the project scale, using an area surveyed at 20 meter
transects and all sites within the survey area recorded. Each evaluation provides important
information on the reliability of Maxent for predicting archaeological site locations. The study
area scale evaluation gives an idea of how statistically sound each model is and the influence of
each environmental variable for predicting archaeological site locations. The project scale
evaluation gives an idea of how successful each model is at the project scale and gives the
opportunity to compare the Maxent model against the Far Western and BLM archaeological site
prediction models.
4.5.1. Study Area Evaluation
At the study area scale, all three models performed very well, when only taking into
account the AUC. The closer the AUC is to 1 the better the fit of the model is. As hypothesized,
the “kitchen sink” model run had the lowest AUC of 0.793 with a standard deviation of 0.003.
The ecological site and site type models preformed similarly with the AUC ranging from 0.823
to 0.905 on all the models. The percent contribution and permutation importance give a better

40

idea of how successful these models are from an archaeological context. Table 11 summarizes
the environmental variables with the highest percent contribution for each of the model runs. It is
important to keep in mind that the lower the permutation importance values the more stable the
environmental variable.
Table 11 Model Percent Contribution Comparison

Percent
contribution
Permutation
importance
"Kitchen
Sink"
Variable Archaeological Site
Geologic Unit 39% 12.8
Distance from watercourses and water bodies 23% 15.1
Ecological Region
Fall River
Distance from tool stone sources 72% 64.1
South Fork Pit River
Distance from tool stone sources 71% 63.3
Tule Lake
Distance from tool stone sources 63% 10.5
Warm Springs
Distance from tool stone sources 54% 1.1
Site Type
Lithic Scatter
Distance from large game corridors 48% 17.4
Distance from tool stone sources 21% 32.2
Rock Features
Distance from large game corridors 47% 41.8
Distance from tool stone sources 15% 5.3

Geologic unit and distance from watercourses and water bodies had the highest
contribution to the “kitchen sink” model. This makes sense when considering what would be
important factors for any type of site use. At a landscape level, prehistoric people would be more
inclined to select sites that are close to water sources and have less volcanic rock, making them
more easy to traverse and more likely to have fertile soils for food sources as well as being
habitat for game. The percent contribution was much more distributed over all of the

41

environmental variables for the “kitchen sink” model, this is most likely because this model only
took into account that an archaeological site existed at each location but not what the use was at
that site.
While the ecological region models were very successful from the perspective of the
statistical tools within Maxent, this is very misleading. Sites were categorized based on an
ecological region, so the sites used for each model were grouped in one portion of the study area.
The models should only be considered valid for the ecological region that they represent, as
shown in Figure 16. The highest percent contribution for each of the models was distance from
tool stone. However, the Fall River and the South Fork Pit River model have a very high
permutation importance, so tool stone is a very unstable predictor of archaeological site
probability. Figure 18 shows the distribution of tool stone sources and the distribution of
archaeological sites within each ecological site.

Figure 18 Ecological Regions and Distance from Tool Stone

42

The site type models are the most successful of the three runs. The highest percent
contribution for the lithic scatter model is distance from large game corridor which is a fairly
stable predictor. The second highest percent contribution is distance from tool stone sources,
however this is much less stable predictor than the distance from large game corridors. The rock
features model had the same two environmental variables with the highest percent contribution,
with distance to tool stone sources being the more stable of the two.
The highest percent contribution from distance from large game corridors and distance
from tool stone sources shows the success of the site type method. Lithic scatters are the remains
of creating tools and projectile points from tool stone sources, so these two variables being the
highest percent contribution are archeologically sound. Rock features are any rock placement,
rock stack, rock alignment, hunting blinds or other rock feature. These features could be
associated with hunting, either directly hunting blinds and rock alignments that were used for
large game drives, or indirectly, rock stacks used for navigation.
4.5.2. Project Scale Evaluation
The project scale survey example can be used to evaluate how the Maxent models
compare to the previous models, Far Western and the BLM distance to water model. Each of the
previous models captured over 70% of the sites within the high probability area of the models for
the whole study area. The project scale performance of the Far Western and BLM models were
discussed in Chapter 2 where Figure 5 shows the previous models’ performance at the project
scale within the survey area. This section looks at the performance of each of the Maxent models
for the same survey area, because the survey area was within the South Fork Pit River ecological
region, that was the only ecological region model that was used for comparison.

43

The survey area is approximately 5,800 km
2
(square kilometers) and after a being
surveyed using 20 meter transects, approximately 203 km
2
of sites were recorded. Table 12
compares the area, in km
2
, of recorded sites and how they fell within each of the models. It is not
possible to compare the Maxent models directly to the Far Western and BLM models because
the categories are different. However, in Chapter 2 the assumption was made that the BLM
modeled area was comparable to the High and Very High categories from the Far Western
model. For this section the assumption is made, for comparison purposes, that <50% probability
is Very Low, 50-70% is Moderate, 70-90% is High and 90-100% is Very High. Using this
assumption, the South Fork Pit River, lithic scatter and “kitchen sink” were better predictors at
the project scale than the BLM, Far Western and rock features models. Each of these models is
discussed in further detail below.
Table 12 Survey Area Model Performance
Model Square kilometers of area containing sites

Very Low Moderate High Very High
Percent Probability
< 50% 50-60% 60-70% 70-80% 80-90% 90-100%
"Kitchen Sink" 0 0 2 134 46 21

Lithic Scatter 1 0 3 1 130 68
Rock Features 165 32 2 4

South Fork Pit River 0 0 1 1 194 7

Sensitivity
Very Low Moderate High Very High
Far Western 0 90 94 19

Modeled Area Outside Within

BLM 33 170

44

The survey area probability distribution maps for each of the models are discussed
below. Each map shows the recorded sites outlined and overlaid on the probability distribution.
The warmer colors indicate higher probability of an archaeology site and the cooler colors lower
probability. The “kitchen sink’ performed well with 201 km
2
of the 203 km
2
of areas containing
recorded sites falling in the High and Very High probability areas and only 2 km
2
falling within
the moderate range. The majority of the survey area is categorized as above 70% probability
area; this is displayed in Figure 19. As the probability distribution map shows, even the 2 km
2
is
only a small portion of two sites and portions of the same sites also fall within the high
probability area.

Figure 19 Evaluation of “kitchen sink” model within survey area
For the ecological models, the only model evaluated is the South Fork Pit River. All but
two of the sites are within 80% probability and above, this is 201 km
2
out of the total 203 km
2

45

surveyed. The remaining 2 km
2
fall within the 60 to 80% range. For the South Fork Pit River
model the majority of the survey area is categorized as 80% probability and above, this is
displayed in Figure 18. The probability distribution map shows that the 2 km
2
is one small site on
the eastern edge of the study area, all other recorded sites are completely within high probability.

The site type models, lithic scatters and rock features must be evaluated by taking into
consideration the type of site that was located during the survey. In Figures 21 and 22 the site
types are symbolized differently to show which sites were labeled as lithic scatters and what
were labeled as rock feature. The lithic scatter model performed very well with 199 km
2
of the
surveyed sites within High and Very High probability. The rock features model performance was
the least successful of all the models with 156 km
2
within the Very Low and 43 km
2
within
Moderate probability. Rock features had the least amount of archaeological site locations put in
to train the model, only 137 rock features were labeled within the entire project study area. Only
Figure 20 Evaluation of ecological region model within survey area

46

4 sites labeled as rock features fell within the survey area and are indicated on the probability
distribution map by the blue outline, the pink outlines are the lithic scatter type labeled sites.

Figure 21 Evaluation of site type, lithic scatter model within survey area

47

4.5.3. Discussion
The study area and project scale evaluations produced different results. At the study area
scale, the Maxent rock feature site type model performed the best, at the project scale the same
model performed the worst. The “kitchen sink’ was the least successful at the study area scale
and was very successful at the project scale. The lithic scatter and South Fork Pit River models
performed well at both scales. With the exception of the rock features model, all the other
Maxent models outperformed the Far Western and BLM models at the project scale. These
results indicate that Maxent can be used as a tool for predictive modeling of prehistoric
archaeology, however some improvements should be made. The variability of success for the
rock feature model may be attributed to this model having the lowest number of archaeological
site locations to train the model, 137 sites compared to over 1,000 for the lithic scatter and South
Fork Pit River and over 3,700 for the “kitchen sink” model. It is possible that this behavior
Figure 22 Evaluation of site type, rock features model within survey area

48

suggests that too few sites do not provide sufficient data to train the model, while too many
uncategorized sites may confuse the model. This possible explanation is depicted in Figure 23.

Figure 23 Model Success Curve

49

Chapter 5 Conclusions
The purpose of this project is to achieve reliable archaeological site predictions with the Maxent
software. To do this the following hypothesis was tested, that the site type Maxent model is as
good or a better predictor of archaeological site probability than Maxent models that do not
categorize by site type. The site type Maxent model will also be as good or a better predictor of
archaeological sites than the previous models at a project scale. This Chapter discusses the
results of the Maxent models produced and how archaeological site prediction models may be
improved in the future for the study area.
5.1 Discussion
An archaeological site prediction model is one important tool that can be used to help
preserve and tell the story of the prehistoric people that lived on this land. Predicting human
behavior is a daunting task and predicting the behavior of 11,000 years of prehistoric people
within this study area will never be completely accurate. As discussed in Chapter 4, this project
has shown that the Maxent method of modeling provides a successful method for predictive
modeling of archaeology site locations for the study area, with some additional work being
needed. The Maxent models, for the most part, improved on the Far Western and BLM models
and were shown to be reliable using the tests built into the Maxent software and within the
project level survey example.
The evaluation of the models were not enough to prove or disprove the hypothesis,
further work is needed to make a definitive argument. Although, site type approach did perform
well with high AUC numbers for each of the four site type models, there are not enough
categorized sites to say definitively that the site type model will perform better at the project

50

scale than the two previous models. This project was able to prove the value of the commitment
by the BLM to input archaeological site attribute information into a spatial database.
5.2 Limitations
This project was limited by the completeness of the attribute information for the
archaeology site location data, because the data was provided by two different government
agencies data formats and the level of attribute information varied. However, because the BLM
and the USFS are both federal government agencies, the data collected in the field for each site
and entered into site records is similar and could be entered into a spatial database producing a
consistent data layer for the entire study area. This would make it possible to run the site type
modeling approach using all the known archaeology site locations.
Another data limitation of this project is the lack of archaeology site location data on
privately owned lands. As discussed in Chapter 3, no data is available for archaeology site
locations on private lands and because areas of fertile agricultural lands and surrounding large
water bodies and rivers are more likely to be privately owned, this lack of data may cause a
sampling bias. If archaeology surveys were conducted randomly across the landscape, including
all ownership and environment settings, this would limit the survey bias and could improve
modeling success.
5.3 Future Work
Further work with archaeologists is needed in order to get the required classifications of
prehistoric archaeology sites and to further improve the modeling results. Enough data exists
within the study area, with over 3,000 sites, however only around 1,400 sites had sufficient
attribute information to be categorized with the site type. The site type information does exist
within site records for each of these sites, it would be a time intensive, but important task to input

51

this data into a useable format for modeling. With more attribute information more site type
categories could be used for modeling. If that work were to be completed, the hypothesis could
be tested and a definitive conclusion could be made.
There may be other environmental variables that could improve or be better predictors of
archaeology sites within the study areas. As more research and field survey is conducted by
archaeologists, better information on environmental factors may be available. Also, the quality of
environmental variable information could improve in the future. As these advancements are
made predictive archaeological site models should be continuously updated and approved upon.
5.4 Conclusion
The California SHPO requested a predictive model be developed by the BLM to assist
northern California archaeologists in their inventory efforts, specifically to direct field surveys.
The study area has a large amount of known prehistoric archaeology sites, many of these known
sites are eligible for listing on the National Register of Historic Places because of their value to
provide important information on how the people of the past lived. Only a small fraction of the
study area has been surveyed, so thousands of additional sites remain undiscovered. These sites
could contain important and never before discovered information on the life of prehistoric people
and are degrading over time because of natural elements, erosion and human activities. The cost
of surveying and finding important archaeology sites on millions of acres of land is huge. This
project demonstrated that using known site locations and environmental variables can successful
predict where to inventory for important sites. Implementing this modeling approach and doing
the future work described in the previous section could save federal land management agencies
and taxpayers millions of dollars. This approach will also lead to more important sites being
located and the sites and artifacts studied to increase understanding of prehistoric people.

52

References
Bolstad, William M., 1943. Understanding computational bayesian statistics. 1st ed. Hoboken,
N.J: Wiley.
Bonham-Carter, Graeme. 1994. Geographic information systems for geoscientists: Modelling
with GIS. 1st ed. Vol. 13. Oxford ;New York: Pergamon.
Carpenter, Kimberly L. 2002. "Reversing the Trend: Late Holocene Subsistence Change in
Northeastern California." in Boundary Lands: Archaeological Investigations Along the
California-Great Basin Interface, edited by Kelly R. McGuire. 49-59. (Nevada State
Museum Anthropological Papers, 2002).
Elith, Jane, Steven J. Phillips, Trevor Hastie, Miroslav Dudik, Yung En Chee, and Colin J. Yates.
2011. A statistical explanation of MaxEnt for ecologists. Diversity and Distributions 17
(1): 43.
M. Thomas P. Gilbert, Dennis L. Jenkins, Anders Götherstrom, Nuria Naveran, Juan J. Sanchez,
Michael Hofreiter, Philip Francis Thomsen, et al. 2008. DNA from pre-clovis human
coprolites in oregon, north america. Science 320 (5877): 786-9. Green, Dee F., and
Robert J. Laacke. Predictive Models for Locating Prehistoric Travel Pathways and
Campsites. Warner Mountain Rager District: USDA Forest Service, 1998.
Kailihiwa, Solomon Ha'aheo, III. 2015. “Using Maxent to Model the Distribution of Prehistoric
Agricultural Features in a Portion of the Hōkūli‘a Subdivision in Kona, Hawai‘i.”
Master's thesis, University of Southern California.
Kemp, Karen K. 2012. The Hawai'i Island Crop Probability Map: An Update of the Crop
Growth Parameters for the Hawai'i County Crop Model: Final Report Prepared for the
County of Hawai'i Researcy and Development Department. GIS Research Laboratory.
Kern-Isberner,Gabriele, Marco Wilhelm, and Christoph Beierle. 2014. “A Novel Methodology
for Processing Probabilistic Knowledge Bases Under Maximum Entropy.” Florida
Artificial Intelligence Research Society Conference. Key Larg, Florida, May 16-18.
King, Jerome, Kelly McGuire, Kimberley Carpenter, Mary Maniery, Cindy Baker, Hellen
McCarthy, Heather Scotten. 2004. Class I Cultural Resources Overview and Research
Design for the Alturas, Eagle Lake, and Surprise Resource Areas. Class I Cultural
Resources Overview, Davis, California: Far Western Anthropological Research Group,
Inc.
Kniffen, Fred B. "Achomawi Geography." University of California Publication in American
Archaeology and Ethnology 23, no.5 (1928): 279-332.
52

53

Merow, Cory, Matthew J. Smith, and John A. Silander. 2013. A practical guide to MaxEnt for
modeling species’ distributions: What it does, and why inputs and settings matter.
Ecography 36 (10): 1058-69.
Merriam, Dr. C. Heart. 1926."Classification and Distribution of the Pit River Indian Tribes of
California." Smithsonian miscellaneous collections. Vol. 78, No. 3.
Phillips, Steven J. n.d."A Brief Tutorial on Maxent." Unpublished documentation of the Maxent
software.
Phillips, Steven J., Miroslov Dudík, and Robert E. Schapire. 2004. "A maximum entropy
approach to species distribution modeling." Proceedings of the Twenty-First
International Conference on Machine Learning. 655-662.
Phillips, Steven J., Robert P. Anderson, and Robert E. Schapire. 2006. Maximum entropy
modeling of species geographic distributions. Ecological Modelling 190 (3): 231-59.
Scott, David J., and Megan C. Oyarzun. 2012."A big game procurement site, western toe of the
Warner Mountains, Rock Spring, Modoc County, California." 33rd Biennial Great Basin
Anthropological Conference, Poster Presentation. Stateline, Nevada, October 18-21.
Survey, US Geologiv. 2005. Preliminary Integrated Geologic Map Databases for the United
States: Western States: California, Nevada, Arizona, Washington, Oregon, Idaho, and
Utah. U.S. GEOLOGICAL SURVEY OPEN-FILE REPORT 2005-1305.
USDA Forest Service, USDI Bureau of Land Management. 2007.Sage Steppe Ecosystem
Restoration Strategy Draft Environmental Impact Statement. Environmental Impact
Statement, Alturas, CA: Modoc National Forest.
Weldon, K. 2010. Obsidian: Hotter Side of Obsidian.
http://volcano.oregonstate.edu/book/export/html/205 (accessed April 2, 2016).

Abstract (if available)

Abstract Prehistoric sites and artifacts are common across the country side in the high elevation desert of California’s northeastern corner. For decades archaeologists have been researching, surveying and cataloging archeological sites on lands managed by the Bureau of Land Management (BLM). While thousands of sites have been recorded, it is hard to say how many remain undiscovered. Multiple archaeological site prediction models have been completed covering the area to assist archaeologists in locating and recording sites. This project tests the hypothesis that the site type Maxent model can be as good or a better predictor of archaeological site probability than the Maxent models that do not categorize by site type. The site type Maxent model will also be as good or a better predictor of archaeological sites than the previous models at a project scale. To test this hypothesis three models were run (1) the “kitchen sink”, all 3,729 sites within the study area, (2) ecological region, using all sites categorized by the ecological region in which they fall, and (3) site type, a subset of 1,332 sites, categorized by the prehistoric people use at that site. Maxent uses the spatial location of individual archaeological sites and environmental variable rasters to produce a probability of distribution raster. At the study area scale the Maxent software’s built-in validation tools, environmental variable performance and Area Under the Receiver Operator Curve (AOC) the three Maxent models were compared and to test the hypothesis. At a project scale a 5,800 km² archaeological survey area was used to compare how well the Maxent models and the previous models were able to predict recorded site locations. This project was unable to definitively prove the hypothesis

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Assessing the transferability of a species distribution model for predicting the distribution of invasive cogongrass in Alabama

PDF

Using Maxent to model the distribution of prehistoric agricultural features in a portion of the Hōkūli‘a subdivision in Kona, Hawai‘i

PDF

Building better species distribution models with machine learning: assessing the role of covariate scale and tuning in Maxent models

PDF

Predicting the presence of historic and prehistoric campsites in Virginia’s Chesapeake Bay counties

PDF

A Maxent-based model for identifying local-scale tree species richness patch boundaries in the Lake Tahoe Basin of California and Nevada

PDF

Testing LANDIS-II to stochastically model spatially abstract vegetation trends in the contiguous United States

PDF

Modeling geopolitics in Tikal through least cost paths

PDF

Modeling historic structure preservation candidacy on Fort Ord

PDF

Selection of bridge location over the Merrimack River in southern New Hampshire: a comparison of site suitability assessments

PDF

Predicting Hydromantes shastae occurrences in Shasta County, California

PDF

Spatial analysis of human activities and wildfires in the Willamette National Forest

PDF

Preparing for earthquakes in Dallas-Fort Worth: applying HAZUS and network analysis to assess shelter accessibility

PDF

Archaeological least cost path modeling: a behavioral study of Middle Bronze Age merchant travel routes across the Amanus Mountains, Turkey

PDF

Spatial distribution of the greater sage-grouse in the Powder River Basin in northeastern Wyoming

PDF

Modeling prehistoric paths in Bronze Age Northeast England

PDF

Using Maxent modeling to predict habitat of mountain pine beetle in response to climate change

PDF

Installing public electric vehicle charging stations: a site suitability analysis in Los Angeles County, California

PDF

A comparison of GLM, GAM, and GWR modeling of fish distribution and abundance in Lake Ontario

PDF

Crowdsourced maritime data: examining the feasibility of using under keel clearance data from AIS to identify hydrographic survey priorities

PDF

Stone tool raw material distribution network and predictability study in southern Illinois

Asset Metadata

Creator Oyarzun, Megan Christine (author)

Core Title Predicting archaeological site locations in northeastern California’s High Desert using the Maxent model

School College of Letters, Arts and Sciences

Degree Master of Science

Degree Program Geographic Information Science and Technology

Publication Date 06/30/2016

Defense Date 05/11/2016

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag archaeology,BLM,Maxent,OAI-PMH Harvest,predictive model

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Kemp, Karen (committee chair), Oda, Katsuhiko (committee member), Yang, Wei (committee member)

Creator Email farschon@usc.edu,mfarschon@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c40-259530

Unique identifier UC11281160

Identifier etd-OyarzunMeg-4486.pdf (filename),usctheses-c40-259530 (legacy record id)

Legacy Identifier etd-OyarzunMeg-4486.pdf

Dmrecord 259530

Document Type Thesis

Format application/pdf (imt)

Rights Oyarzun, Megan Christine

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA