University of Southern California Dissertations and Theses

Deep convolutional neural networks for remote sensing investigation of looting of the archeological site of Al-Lisht, Egypt

(USC Thesis Other)

Deep convolutional neural networks for remote sensing investigation of looting of the archeological site of Al-Lisht, Egypt

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Deep Convolutional Neural Networks for Remote Sensing Investigation of Looting of the
Archeological Site of Al-Lisht, Egypt

by

Timberlynn Woolf

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 2018

ii

For Alice

Copyright © 2018 Timberlynn Woolf
iii
Table of Contents
List of Figures ................................................................................................................................ iv
List of Tables .................................................................................................................................. v
Acknowledgements ........................................................................................................................ vi
List of Abbreviations .................................................................................................................... vii
Abstract…………………………………………………………………………………………viii

Chapter 1 Introduction…………………………………………………………………………….1

Chapter 2 Related Work…………………………………………………………………………..5

Chapter 3 Methodology………………………………………………………………………….18

Chapter 4 Results………………………………………………………………………………...27

Chapter 5 Conclusion…………………………………………………………………………….36

References………………………………………………………………………………………..39

Appendix A: GEP Code………………………………………………………………………….43

Appendix B: GEP Console Output………………………………………………………………47

Appendix C: DGF Code…………………………………………………………………………54

Appendix D: DGF Console Output……………………………………………………………...58

iv
List of Figures
Figure 1 Study Area ........................................................................................................................ 2
Figure 2 CaffeNet architecture ..................................................................................................... 14
Figure 3 VGG16 architecture…………………………………………………………………….15

Figure 4 Labeling of Pits..……………..…………………………………………………………21

Figure 5 Sample pit images…...…………………………………………….…………………....22
Figure 6 DGF WV3 Image and pit samples……………………………………………………...23
Figure 7 GEP Plotted Val and Train Loss……………………………………………………….29
Figure 8 GEP Plotted Val and Train Acc………………………………………………………...30
Figure 9 DGF Plotted Val and Train loss…………………………………………………..........33
Figure 10 DGF Plotted Val and Training Loss…………………………………………………..34

v
List of Tables
Table 1 AlexNet Architecture Layers .......................................................................................... 13
Table 2 GEPTraining, Testing, and Validation ............................................................................ 28
Table 3 GEP Measure of Error .................................................................................................... 31
Table 4 DGF Training, Testing, and Validation..………………………………………………..32

Table 5 DGF Measure of Error……………………………………………………………..........32

vi
Acknowledgements
Thank you to the DigitalGlobe Foundation for the satellite imagery grant and images courtesy of
the DigitalGlobe Foundation. Thank you to my brother, John, for all of your love and support
and to my faculty, family, and friends who have supported and encouraged me through this
process.

vii
List of Abbreviations
AlexNet CNN architecture developed by Alex Krizhevsky

BOVW Bag of visual words

Caffe Convolutional Architecture for Fast Feature Embedding

CaffeNet CNN architecture provided by Caffe

CNN Convolutional neural network

DGF DigitalGlobe Foundation

DNN Deep convolutional neural network

FC layer Fully-connected layer

GEP Google Earth Pro

GIS Geographic information system
GISci Geographic information science
HRRS High-Resolution Remote Sensing

IFK Improved Fisher kernel

ReLU Rectified linear units

SVM Support vector machines

SSI Spatial Sciences Institute
UFL Unsupervised feature learning

USC University of Southern California
VGG-F Fast CNN architecture developed by Chatfield

VGG16 16 layer Deep CNN

viii
Abstract
Looting of archaeological sites is a global problem. To quantify looting on a nationwide scale
and to assess the validity and scope of the looting reports and modern encroachment, satellite
archaeologist have turned to mapping looting from space. High-resolution satellite imagery has
become a powerful tool and resource for monitoring looting and site destruction remotely and
proves to be an independent way to cross check and analyze against varied and unreliable reports
from media and government agencies. It is estimated that over a quarter of Egypt’s 1100 known
archaeological areas have sustained major damage and site destruction directly linked to looting.
The organized looting and illicit trafficking of art and antiques, known as cultural racketeering, is
a multi-billion dollar worldwide criminal industry that thrives in Egypt during times of political
and economic turmoil and potentially funds drug cartels, armed insurgents, and even terrorist
networks. This study analyzes methods used to monitor site looting at the archaeological site of
al-Lisht which is located in the Egyptian governorate of Giza south of Cairo. Monitoring damage
and looting over time has been largely dependent upon direct human interpretation of images.
The manual image comparison method is laborious, time consuming, and prone to human-
induced error. Recently, partially-supervised methods using deep convolutional neural networks
(CNNs) have shown astounding performance in object recognition and detection. This study
seeks to demonstrate the viability of using deep convolutional neural networks (CNNs) within
the field of archaeology and cultural heritage preservation for the purpose of augmenting or
replacing the manual detection of looting. It brings recent advancements from the field of
Artificial Intelligence to an applied GIS challenge at the intersection of remote sensing and
archaeology. The objective is to show that CNNs are a more accurate and expedient method for
the detecting of looting with wide-ranging application beyond this specific research.
1
Chapter 1 Introduction
Cultural heritage theft and small-scale looting has been a part of the history of Egypt for
thousands of years. In recent years the looting has increased dramatically which has been shown
to be primarily a result of economic factors. Looting in Egypt has escalated dramatically from
economic and political instability seeing a dramatic spike with the onset of the economic crisis of
2009 and the Egyptian Revolution of 2011. In the recession of 2009, the consumer price index
rose and tourism fell with record numbers of unemployment reaching 36% (Parcak 2016). The
economic crisis was then shortly followed by the 2011 Egyptian revolution known as the Arab
Spring. During this timeframe Egypt experienced the highest overall values of both looting pits
and encroachment (Parcak 2016). High unemployment and financial incentives have increased
looting dramatically. An increase in looted goods flooded the underground antiquities market
during these timeframes and were thought to be potentially used to fuel international crime from
drug trafficking, illegal arms trafficking, and even terrorism. With archaeological site destruction
comes the loss of the material culture from the site and the loss of the piece of history that was
taken. Satellite Archaeologist are taking dramatic steps using satellite remote sensed imagery and
GIS for manual and partially-supervised processing and assessment in order to gain a better
understanding of the damage and loss and to aid the future prevention.
This project focuses on one previously studied archaeological site of al-Lisht, Egypt. Al-
Lisht has seen a dramatic spike in looting following the 2009 economic downturn and 2011 Arab
Spring Revolution. This study analyzes the looted site using the manual method of detection of
direct human interpretation of remotely sensed images monitoring/ detecting change over time.
In an effort to improve upon the manual comparison method which is time consuming and prone
to human-induced error, this study investigates the use of a partially-supervised method of deep

2
convolutional neural networks (CNN) for looting pit classification using high-resolution satellite
imagery.
1.1 Study Area
The study area is located south of Cairo archaeological site of al-Lisht in the Egyptian
governorate of Giza south of Cairo. Al-Lisht is the site of Twelfth and Thirteenth Dynasties
Middle Kingdom royals and elite burials, including two pyramids built by Amenemhat I and
Senuseret II.

Figure 1: Study Area al-Lisht
1.2 Motivation

Time is running out for many cultural heritage and archaeological sites. Once a site has been
looted there is no way to determine what exactly has been lost. It is a part of history that may
never be known to the world. Monitoring damage and looting over time has been largely
dependent upon direct human interpretation of images. The manual image comparison method is

3
laborious, time consuming, and prone to human-induced error (Lauricella et al., 2017) Current
advances in deep convolutional neural networks have achieved breakthrough performance in
object recognition and detection achieving up to 96% accuracy. This study seeks to show the
potential to expedite the looting detection process using Deep Convolutional Neural Networks
(CNNs).
Monitoring of looting is complicated in that it is an illicit activity, subject to legal
sanction (Contreras 2010). Poor or unreliable information about damage from looting has an
impact on policy making. It enables claims that the extent of looting damage is being
exaggerated. This allows artifacts that reach the market to be claimed as chance finds (objects
discovered not related to the activity of illicit digging), or pre-existing collections which in turn
do not call for strong policy making or response (Contreras et al., 2010). Additionally, it is
difficult to monitor the effectiveness of any ameliorating policies whether direct at demand (the
illicit trafficking of antiquities) or supply (illicit digging itself) (Contreras 2010). In an effort for
site preservation, and for stopping the illicit trafficking of antiquities, this study hopes to identify
the best method for monitoring, detecting, and for the prevention of looting of archaeological
sites over time. Additionally, this study hopes to bring awareness and understanding to local
government agencies who can use this information for counteractive measures and preventative
policies as well as proactive measures such as community outreach programs.
1.3 Project Purpose and Scope
The purpose of this project is to expand and build upon CNNs methods that have been
used for other forms of scene and object recognition and to train and apply it to the specific task
of monitoring looting. The intent of this effort is for it to provide a more accurate and expedient
method for looting pit detection. The scope of this project covers the previous manual methods

4
of looting detection, the use of partially-supervised methods, the use of CNNs for other forms of
scene classification and object recognition, and lastly how CNNs were used in this study for
looting pit detection.
The benefits of monitoring looting are that it provides a basis for quantification estimates
of damage from looting. It also can allow for temporal estimates to be made of the time periods
in which the looting occurred. Furthermore, quantifying looting in an expedient manor using the
CNN method can potentially provide a means of assessing damage still being done to
archaeological sites with the chance of linking the damage to the trade in illicit activities.
Possible uses of this study and its results can aid in predictive policing. Knowing what dates and
times the sites were looted and the economic factors that motivate them can aid in predicting of
looting and the predicting of possible attempts in the future. Actions can be taken to protect these
sites and provide possible monitoring of artifacts that could be illicitly trafficked making their
way to the antiquities black market from these sites. It can also be used to hypothesize what to
look for in the illegal international antiquities trade market and aids in the creation of an
international watch list. Monitoring looting helps us know what types of sites have been looted
and from what time period. This knowledge can aid in notifying international agencies such as
US Immigration and Customs Enforcement and INTERPOL (Parcak 2016). Additional
motivations are archaeological site discovery, looting detection, and ultimately, preservation of
cultural heritage sites and artifacts.

5
Chapter 2 Related Work
In satellite archaeology and remote sensing there are many methods employed for the
investigation of looting. However, there is still a lack of investigation using CNNs with high-
resolution remotely sensed imagery for object recognition and detection that could be potentially
used for looting detection. This section first discusses the manual direct human interpretation of
satellite images for looting detection. Second, it discusses other partially supervised methods and
other CNN methods for scene classification. Third it discusses CNNs and the potential for the
supervised deep convolutional neural networks to be trained for looting detection.
2.1 Manual direct human interpretation of satellite images for looting detection
In the article, Satellite Evidence of Archaeological Site Looting in Egypt: 2002-2013,
Sarah Parcak and her team turned to the use of satellite imagery and GIS to map the looting in
Egypt from 2002-2013. They used Google Earth Pro (GEP) to obtain satellite data from 2002-
2013. The imagery quality that was available varied over time with the changes that came in the
initial development phases in high-resolution satellite imagery. Initially image availability was in
the range of 20% in 2003 but changed dramatically over time increasing to around 50% in in
2005 and 70% for 2009 to 2011 (Parcak 2016). In order to reach 100% coverage they
extrapolated the data for incomplete coverage years (Parcak 2016).
Looting pit encroachment on sites was then assessed between the years of 2002-2013 to
determine the extent of the looting and the damage. They assessed the 1100 sites they surveyed
for damage and determined that 24.3% displayed evidence of damage and looting (Parcak 2016).
267 sites were georeferenced within ArcGIS and individual polygons were drawn over each
looting pit and a series of larger polygons were placed over areas that had been affected by
encroachment. Natural boundaries such as roads and rivers were used to determine the extent of

6
the perimeter for undefined sites. In the past satellite imagery did not have the resolution to do
this, but now it is possible with GEP. Polygons were drawn around the location of the sites
perimeter. By determining the perimeter, they were able to calculate the area of the site and the
percentage of the site that had been looted. From the data in the research gathered by Parcak and
her team quantifying site looting in Egypt, they were able to determine that looting in Egypt
since 2002 did not increase steadily over time, but fluctuated dramatically with recent political
and economic instability. By discerning and detecting the political and economic indicators that
seem to come before the dramatic increase in looting, efforts can be made to protect
archaeological sites that are known to be at risk.
A previous study of the Viru Valley, Peru by Contreras in 2010, used Google Earth
imagery as a useful tool for addressing the scale of looting damage to archaeological sites
(Contreras 2010). As with present day concerns when using remote-sensing, they encountered
problems with coverage, appropriate resolution, and surface visibility (Beck 2006; Ur 2006;
Scollar et al. 2008; Parcak 2009). The Viru Valley Study utilized remotely sensed imagery in
conjunction with historical site surveys and then analyzed their findings using GIS analysis.
2.1.2 Looting Pit Quantification Methods
The method Parcak used to calculate this was inverse distance weighting (IDW). Inverse
distance weighting is a special interpolation algorithm. Weights are proportional to the inverse of
the distance between points and the predicted location. The IDW weighs the values from specific
known points and uses the inverse of the distance of those points to assign and predict
approximate weights for the unknown area within the extent of the known points (Parcak 2016).
One of the drawbacks of this study and the IDW method is that it only indicates looting attempts

7
and does not predict if it was a successful attempt or not. There is no specific way to determine
what has been taken and to what extent, only that a site has been looted.
The Viru Valley, Peru 2010 study inspected the valley sites utilizing Google Earth for
signs of obvious and extensive looting and visible pitting on aerial and satellite images.
Correlation was made between looting pits identified in images and looting that was established
from areas known to be badly damaged by looting utilizing survey information. Areas noted for
looting were referenced with a polygon for later evaluation. 263 areas were identified using a
combination of historical images, ground survey data, known looted sites, and remotely sensed
imagery for further analysis. Once looted areas were identified, control points were selected and
a .jpg image was downloaded from Google Earth at highest resolution possible (4800x 3229
pixels) adjusted (contrast, brightness and color balance) for improvement of visibility of features,
and georeferenced in ArcGIS 9.2. Areas identified as damaged by looting were used to create
boundary polygons in ArcGIS. Resolution quality did not allow for accurate or adequate
counting of individual pits and thus did not allow for the direct calculation of the total number of
pits and density of pits in each site. Instead, the total looted area was approximated by bounding
the visibly disturbed areas. This lent to multiple polygons being defined for one select site
location. As a result of this effort, polygon shapefiles were used to calculate the looted area.
Published literature was then consulted to identify previously mentioned looting damage and
when the site was most likely dated to be looted. By doing so, they were able to assess possible
types of artifacts a particular site might yield or have yielded to the illicit antiquities market
(Contreras 2010). The research and methodology allowed for quantifying the scale of looting in
the Viru Valley. It also allowed for the identification of recent looting.

8
2.1.3 Remote sensing for looting detection
Previous use of satellite remote sensing in Syria dates back to 2012. Using satellite
imagery in Syria for site looting detection and destruction. Unfortunately, most reports of war-
related destruction in Syria come from journalists, and photos and videos posted on social media
by potentially biased actors in the conflict. The Syrian government’s Director General of
Antiquities and Museums releases periodic reports regarding looting and site destruction but has
been harshly criticized for being selective in reporting with political motivations. (Casana 2014).
Previous work to document and analyze looting and damage to archaeological sites in Syria as a
direct consequence of the ongoing civil war relied on-high resolution satellite imagery. The
Syrian site assessment effort was based in the years 2012 and 2013 using GeoEye and
WorldView imagery of 30 key sites. Additional free imagery was obtained from Google Earth
and Bing Maps. Anaysis was further expanded through an imagery grant from the DigitalGlobe
Foundation. In an attempt to safely and accurately quantify the true scale of the damage and
looting, archaeologist and analyst turned to high-quality satellite imagery. In april of 2012,
Quickbird imagery of the Roman city of Apamea was posted on Google Earth showing the full
extent of the looting damage to the site from the previous 8 months. Imagery revealed that
between July 2011 and April 2012, Apamea was intensely looted showing a pockmarked
landscape (Casana 2014). The image was picked up and distributed around the world by media.
Methodology for the analysis by Casana and Panhipour used freely available Google Earth and
Bing Maps imagery, alongside GeoEye-1, and Worldview-1 and 2 imagery provided by the
DigitalGlobe Foundation. Comparison was made between pre-war images and the most recent
images available high resolution satellite imagery has become a powerful tool and resource for
monitoring looting and site destruction remotely and proves to be an independent way to cross

9
check and analyze against varied and unreliable reports from media and government agencies
(Casana 2014).
2.1.4 Earlier Years of Remote Sensing
Earlier years of remote sensing using satellite imagery for looting detection in Iraq date
back to before and after 2003 war. 1900 sites were investigated for signs of looting. Limited data
was developed from the comparison of imagery from before and after the 2003 invasion of Iraq.
Areas of focus included archaeological sites of the Nippur area near Eridu and around Uruk.
These sites were chosen based on the reports of significant looting in the area and that the area
had been imaged at 60cm resolution by the DigitalGlobe Corporation (Stone 2008). The study of
DigitalGlobe Imagery spans from 2002 to 2006. 9728km
2
of imagery was examined, 0.87
percent was occupied by archaeological sites. The total number of sites studied in the end totaled
1949 sites. Intense looting was found to be most common close to the boundaries between settled
areas and the desert. Conclusions determined that site selection for looting is close enough to
draw a workforce yet far enough away to go undisturbed or noticed (Stone2008). Of the 1949
sites examined only 743 can be seen in more than one image, of those 213 were looted
representing 26 percent of all looting sites (Stone 2008). However, there are 348 pairs of images
taken at different times at sites allowing for progress in looting to be determined from previous
older activity. Of the 114 images that were imaged multiple times and imaged in 2003, 85
percent showed evidence of fresh looting (Stone 2008). The total area of intensive looting for
this time period adds up to 15.75km
2
which is larger than all archaeological investigations ever
conducted it Iraq at the time (Stone 2008). Spatial analysis of looting distribution suggested that
many interesting sites at the time, were intact which were in areas that had not been as badly hit.

10
2.2 Machine Learning
Lack of manual performance motivates research into computer assisted methods, mostly
from machine learning and homegrown algorithms. Common methods consist of Support Vector
Machines (SVMs), a supervised non-parametric statistical learning technique and Principal
Component Analysis (PCA)- PCA is a statistical method for analysis of multivariate datasets.
Benefits of SVMs are their ability to successfully handle small training datasets often producing
greater accuracy than traditional methods, it allows for more rapid identification of pits.
However, there are drawbacks with the kernel function and choices resulting in overfitting and
oversmoothing. SVMs can show poor performance with noisy data. Benefits of PCA with
multispectral satellite imagery is its ability to recombine data collected on the reflectance of
visible and non-visible spectra of light, highlighting patterns in the landscape allowing for the
display of looting pits. PCA with Interactive supervised classification tool in ArcGIS can define
training polygons on pits and isolates pixels corresponding to pits. PCA allows for a more
expedient quantification of looting patterns than manual visual inspection. A drawback of PCA
method is the cost and availability of multispectral imagery.
Looting pit identification history parallels general object recognition and detection history in
manual and machine learning. Methods like SVM and custom expert crafted methods
(SIFT/SURF) achieved reasonable but unsatisfactory performance, (see ImageNet challenge).
Recent findings favor usage of CNNs to SIFT/SURF, SVMs and PCAs. CNN fine-tuning yields
competitive accuracy on various retrieval tasks and has proven to have advantages in efficiency
over SIFT/SURF, SVMs and PCAs.
Advent of deep convolutional neural networks and ongoing refinements have achieved an
average of 96% and are now the dominant algorithmic approach for object recognition and

11
detection tasks (Krizhevsky et al., 2012) (Russakovsky, 2015) (Hu et al., 2015), see ImageNet
large scale visual recognition challenge for performance data.
2.2.1 Convolutional neural networks
Deep learning is a new take on specific sub-field of machine learning. It consists of
learning representations from data which puts an emphasis on learning successive layers of
increasingly meaningful representations (Chollet 2017). The deep in deep learning refers to the
layers of the model standing for the idea of successive layers of representations and how many
layers there are that contribute to the model referring to the models depth. Deep neural networks
map inputs (images) to targets (labels) via a deep sequence of data transformations (layers) and
the transformation is learned by exposure to the examples (Chollet 2017). Deep convolutional
neural networks (CNNs) architectures are typically comprised of several layers of various types
that can be summarized into categories. (1) Convolutional layers compute the convolution of the
input image with the weights of the network. Neurons in the first hidden layer only view a small
image window and learn low-level features. Deeper layers view larger portions of the image and
learn more expressive features by combining low-level features. Hyper-parameters characterize
each layer with the number of filters to learn, spatial support, stride between different windows
and zero padding which controls output layer size (Castelluccio et al., 2015). (2) Pooling layers
are inserted in-between successive convolutional layers and progressively reduce the size of the
input layer through local non-linear operations. They reduce the amount of parameters and
computation in the network and also controlling overfitting (Castelluccio et al., 2015). (3)
Normalization layers have the intentions of implementing inhibition schemes observed in the
biological brain (Russakovsky et al., 2015). (4) Fully connected layers are used in the last few
layers of the network. Neurons in a fully connected layer have full connections to all activations

12
in the previous layer. Their activations can hence be computed with a matrix multiplication
followed by bias offset ((Russakovsky et al., 2015). By removing the constraints, they can better
summarize the information that is conveyed by the lower-level layers in view of final decisions
(Castelluccio et al., 2015).
2.2.2 Deep CNN Architecture
AlexNet, developed by Alex Krizhevsky is a groundbreaking deep CNN architecture that
consists of five convolutional layers with the first, second, and fifth of which are followed with
pooling layers, and three fully-connected layers as displayed in Figure 2. AlexNet success is
attributed to its use of Rectified Linear Units (ReLU) non-linearity, data augmentation, and
dropouts (Hu et al., 2015) (Krizhevsky et al. 2014) table 1 below. ReLU is the half-wave rectifier
function f (x) = max (x,0) which can significantly accelerate the training phase. Data
augmentation effectively reduces overfitting (the error in the training set is driven to a very small
value, but when new data is presented to the network error is large). The network has memorized
the training examples but it has not learned to generalize to new situations (Hu et al., 2015)).
Augmentation reduces overfitting when training large CNNs, which generates more training
image samples. Training image samples are created by cropping small-size patches and
horizontally flipping these patches from original images (Hu et al., 2015).

13
Table 1: AlexNet architecture layers (Pedraza 2017)

The dropout technique reduces the co-adaptation of neurons by randomly setting zeros to
the output of each hidden neuron and is used in fully-connected layers to reduce substantial
overfitting. AlexNet has become the baseline architecture for modern CNNs by popularizing the
application of large CNNs in visual recognition tasks (Hu et al., 2015).
2.2.3 Caffe
CaffeNet is aPre-trained CNN Convolutional Architecture for Fast Feature Embedding
(CaffeNet) also called Caffe. Caffe is a fully open-source deep learning framework that allows
clear and easy implementations of deep architectures (Penatti 2015). For convolutional neural
networks in particular, it is one of the most popular libraries for deep learning. It is developed by
the Berkeley Vision and learning Center (BVLC) and community contributors. It provides

14
possibly the fastest available implementations for effectively training and deploying general-
purpose CNNs and other deep models (Hu et al., 2015). Caffe is implemented using C++ with
support to CUDA a NVidia parallel programming based on graphics processing units (GPU).
Caffe uses Protocol Buffer language, which makes it easier to create new architectures. Other
functionalities of Caffe include; fine-tuning strategizing, layer visualization, and feature
extraction (Penatti 2015). Caffe is very similar the CNN architecture AlexNet with the exception
of a few small modifications. CaffeNet allows for training without data augmentation and it
allows for the exchanging of the order of pooling and normalization layers (Hu et al., 2015).

Figure 2: Image Credit Hu et al., 2015. Transferring Deep Convolutional Neural Networks for
the Scene Classification of High-Resolution Remote Sensing Imagery. The Boxes show the size
of each feature layer and for the fully connected layers, the size of the output.

2.2.4 VGG16
VGG16 was developed and trained by Oxford’s Visual Geometry Group (VGG). It is a
deep convolutional neural network for object recognition and has achieved very good

15
performance on the ImageNet dataset securing first and second place in the ImageNet Challenge
in 2014. Its convolutional layers use 3x3 size its Max pooling layers use 2x2 size and it has fully
connected end layers with a total of 16 layers (VGG16). It has increased depth using architecture
with very small 3x3 convolutional filters and has shown that a significant improvement on the
prior-art configurations can be achieved by pushing the depth to 16 weight layers (Simonyan et.
al 2015).

Figure 3: VGG116 Architecture
2.2.5 Feature Extraction in Deep Learning
The motivation behind the use of feature extraction for this study was based in how
highly effective leveraging a pre-trained network has proven to be in deep learning while
working with a small image dataset. Training deep learning networks from scratch takes large
data and computational resources in order to obtain useful generalizable models (Rosebrock
2017). Deep convolutional neural networks applied to computer vision problems generate low-

16
level object recognition features (e.g. edge detection, line and curve detection) that are common
across image classification domains. By using a CNN previously trained on a large scale image
recognition dataset, it’s possible to bootstrap the training of a new CNN classifier using layers of
the previously trained network architecture and weights and with a new classifier and smaller
dataset. This process is called transfer learning. In transfer learning the general features for
image recognition (edges, etc) are transferred to the new network which is then optimized
against the specific image data that this study attempts to classify (looting pits).
For looting pit image classification task feature extraction was used. The VGG16
architecture convolutional base of the network which is trained on ImageNet was used to extract
interesting features from this study’s own pits versus non-pit images, and then trained a pits
verses non-pit densely-connected classifier (with only two classes) on top of these features.
VGG16 CNN was chosen as the starting point for transfer learning. VGG16 won ImageNet
Large Scale Image Recognition Challenge (2012). Publically available weights from training
against ImageNet dataset of 1.4 million images were used. The VGG16 architecture was
modified by removing the final fully-connected classification layers and applying the ImageNet
trained weights to the remaining lower layers. Note that the weights in these layers are not
updated as part of the training for pit classification as they already generalize for low-level object
recognition features. This study’s own fully-connected classifier was added fit to the recognition
task (number and size of inputs, two classes in the classification problem). The weights in this
network are the ones optimized by the gradient descent training process as described previously.
During training, images pass through the VGG16 ImageNet weighted base network which finds
general image features. The outputs of the VGG16 network feed into the fully-connected layers

17
(created specifically for classifying the looting pits) which become optimized for looting pit
image specific features over the course of training.
Feature extraction in deep learning differ from other methods. Features extracted by deep
neural networks are the network weights learned over training that minimize the loss function. In
other words, a feature in deep learning is a parameterization of the weights of the network. These
weights correspond to the patterns in the image data that lead to correct classification of the
image. While there are techniques for visualizing what the network is “seeing” in these features,
the outputs are contrived in the sense that the feature is the parameterization of the network (the
set of weights) while the visualization is a given data sample filtered through those weights
which is perhaps representative, but not the same thing.
In contrast to deep learning methods, feature extraction in GIS, and across many domains in
which statistical or machine learning methods are applied differ. Feature extraction is
traditionally performed by experts that ‘hand craft’ what they (or the experts) believe to be most
important to making a correct determination. These features tend to be human recognizable (e.g.
nose and two eyes, or a circular sand berm). Part of what makes deep learning exciting and yet
sometimes difficult for fields to adopt, is that the features that experts have long considered to be
discriminatory in classification tasks are less so compared with the ‘alien’ features found in the
data by neural networks as evidenced by the massive successes deep neural networks have
achieved over traditional methods in recent years.

18
Chapter 3 Methodology

The method used for this study uses a supervised Deep Convolutional Neural Network
(CNN) for object recognition and detection. CNNs are hierarchical architectures trained on large-
scale data sets, which have recently shown astounding performance in object recognition and
detection (Hu et al., 2015). The purpose of this analysis was to find a more expedient and
accurate method with lower false-positive rates for looting detection using CNNs.
3.1 Deep convolutional neural network analysis
This section focuses on the processes that were used to create a dataset of looting pits that
are representatives of the problem. These datasets were then applied to CNNs after for object
recognition and classification. In order to perform the first task, looting pit characteristics needed
to be defined. Looting pit selection and image acquisition choice is based on detailed
descriptions by previous archaeological studies, descriptions, and findings. The main stages
necessary to create a dataset of this kind are; image acquisition, looting pit selection, data
labeling, image processing, and dataset building.
3.1.1 Data
Two datasets were used fir this study. The first dataset was gathered from Google Earth
Pro (GEP). The image datasets consisted of the images of two classes; looting pits, and not pits,
that were used to teach the machine learning classifier what the different categories looked like.
The data set utilized satellite imagery from Google Earth Pro with a spatial resolution or ground
sample distance (GSD) of 2.5 meters. Sets of publically-available satellite imagery from Google
Earth Pro spanning from 2008 to 2018 were obtained of the sites al-Lisht, Dahshur, and Saqqara
pyramid field regions of Egypt. Building upon a previous study from 2017, Algorithmic
Identification of Looted Archaeological Sites from Space (Bowen et al, 2017), similar site

19
locations and satellite images were selected. Sites were previously selected from regions with
multiple instances of looted burial sites, substantial expanse of open desert, and multiple
distractors such as graveyards, modern buildings, and farmlands (Bowen et al, 2017).
The second dataset consists of a set of high-resolution satellite imagery of the pyramid
fields region in Egypt granted by the DigitalGlobe Foundation. Images were captured by the
WorldView-3 satellite at a panchromatic mean ground sample distance (GSD) of 0.30 m per
pixel. Collection Start = 2017-01-03T09:07:51.758306Z; Collection Stop = 2017-01-
03T09:08:17.814179Z; Country Code = ""; Number Of Looks = "1"; Cloud Cover = 0.000;
NW Lat = 30.01341060; NW Long = 31.11287091; SE Lat = 28.45609182; SE Long =
31.26501260. Once again, site selection was based on similar site locations. Sites were
previously selected from regions with multiple instances of looted burial sites, substantial
expanse of open desert, and multiple distractors.
3.1.2 Data Labeling
Once the GEP looting pits were selected and the images were resized to 500 x 500
pixels, over 300 images of looting pits were obtained. To use these images for training, they
must be labeled. In this stage the pits are labeled with their grid number, their site location, and
their latitude and longitude information for location. Additionally, this task consisted of
indicating which class they belonged to; positive or negative and pits or non-pits.
Once the DigitalGlobe Foundation looting pits were selected and the images were resized
to 112 x 112 pixels, over 300 images of looting pits were obtained. To use these images for
training, they must be labeled. In this stage the pits are labeled with their grid location, site
location, and their latitude and longitude information for location. Additionally, this task
consisted of indicating which class they belonged to; positive or negative and pits or non-pits.

20
3.1.3 Dataset Building
After gathering and labeling, two primary datasets were available with 500 samples each.
The image datasets consisted of the images of two classes; looting pits, and not pits, that were
used to teach the machine learning classifier what the different categories looked like. The first
dataset utilized satellite imagery from Google Earth Pro with a spatial resolution or ground
sample distance (GSD) of 2.5 meters. Sets of publically-available satellite imagery from Google
Earth Pro spanning from 2008 to 2018 were obtained of the site al-Lisht, Dahshur, and Saqqara
pyramid field regions of Egypt. The second dataset consisted of DigitalGlobe Foundation
WorldView-3 imagery from 2017 with a (GSD) of 0.30m (Figure 6).
3.2 Site Selection
Sites selected were al-Lisht, Dahshur, and Saqqara pyramid field regions of Egypt. Site
selection of these regions were based on and compared with previously obtained and analyzed
images of the selected archaeological site of these regions for algorithmic identification of
looting in these same areas. These regions were selected for this study and by previous studies
based on their multiple instances of looted burial sites, substantial expanses of open desert, and
for multiple distractors such as farmland, non-archaeological structures, modern graveyards, and
military bases with bomb craters (Bowen et al., 2017). For GEP imagery, a grid system was
used, control points were selected and a jpeg image was downloaded at the highest resolution
possible (4800x 3229 pixels), adjusted (contrast, brightness and color balance) for improvement
of visibility of features, with the ability to be georeferenced in ArcGIS Pro (Figure 4). Provider
date, catalog ID # of each image and images incorporated into GEP was identified and noted for
metadata. Images of these sites were obtained and classified to create a dataset of a collection of
images with each individual image labeled as a data point representing a looting pit, or not a

21
looting pit. A uniform number of images of each category were selected. 300 Individual looting
pit images were all labeled with site grid location (letter), site location name, location within that
grid (upper left etc.), data point count number and data point longitude and latitude. 300 non-
looting pit images were selected from the same regions (Figure 5). Both pit and non-pit images
were then loaded and cropped to 500 x 500 pixels in the online photo editor Pixlr and filed in an
image data tracker.
Figure 4: Labeling of pits Al-Lisht Google Earth Pro

22

Figure 5: Sample of images from dataset pits and non-pits (sand) Al-Lisht

Factors of variation that were considered through this process included viewpoint
variation, scale, variation, deformation, occlusion, illumination, background clutter and intra-
class variation. Looting pit data sets were curated with these in mind and based on previously
selected looting locations. Any looting pit location of question was excluded from this study
based on previously selected looting pit location. Looting pit selection, characteristics, and
definition as well as image acquisition choices were based on detailed descriptions by previous
archaeological studies and findings. Dataset samples were taken from each of the three pyramid
field sites. Primary visual references for looting pit site selection include; Algorithmic
Identification of Looted Archaeological Sites from Space (Bowen, et al. 2017) and Satellite
Evidence of Archaeological Site Looting in Egypt 2002-2013 (Parcak et al. 2016).

23

Figure 6: Image courtesy of DigitalGlobe Foundation WV3 2017 R10C3 and R11C3 al- Lisht

24
3.3 Machine Training
A common goal in machine learning is to achieve generalization. Models that generalize
perform well on data that has never been seen before. The goal when training a machine learning
model is to reduce the training loss as much as possible while ensuring that the gap between the
training and testing loss is reasonably small (Challot 2017). In order for the neural network to
learn, we have to find the combination of model parameters that minimize the loss function for
the given set of training data samples (pits vs non-pits) and their corresponding targets
(Rosebrock 2017). The loss function is used to guide the training process of a neural network. It
informs us what is happening internally inside the network. Ideally, training loss and validation
loss should be close and should get smaller and smaller. However, a loss of zero leads to
overfitting and is not desired (Ng 2016). Overfitting is the error in the training set is driven to a
very small value, but when new data is presented to the network error is large. The network has
memorized the training examples but it has not learned to generalize to new situations (Hu et al.,
2015).
After building the dataset and considering the image classification problem that needs to
be solved, selection of the loss function, the network architecture, and optimization method used
to train the model needs to be made. For this study the use of the neural network is for image
classification. Keeping the loss function selection in line with the nature of the task at hand, and
following nearly all deep learning image classification problems, cross-entropy loss is selected.
As this project is a two class project binary cross-entropy is used. Architecture selection for this
study was done on deep convolutional neural network VGG16 architecture developed by Karen
Simonyan and Andrew Zisserman for the ImageNet large-scale object recognition challenge
which achieved new benchmarks in 2014 outperforming other methods and neural networks

25
(Simonyan et al., 2015). For feature extraction, VGG16 architecture was used and used publicly
available weights from training against the Imagenet dataset. Optimization method selection is
RMSprop. Additionally, the dropout technique was used and activation consisted of using
Rectified Linear Unit (ReLU). The dropout technique reduces the co-adaptation of neurons by
randomly setting zeros to the output of each hidden neuron and is used in fully-connected layers
to reduce substantial overfitting. Rectified Linear Unit (ReLU) is the half-wave rectifier function
f (x) = max (x,0) which can significantly accelerate the training phase. Activation is thresholded
at zero.
3.3.1 Binary Cross-Entropy and RMSprop Optomizer for Gradient Descent
For this neural network, binary cross-entropy was used for a two-class classification
problem; pits or non-pits. The loss function quantified the quality of any particular set of weight
values for the weights in the network. It is differentiable which is what allowed for taking its
gradient. The gradient showed how steep the function was at that time or how steeply the loss
function was moving. The desire at that point, is to update the weights in the opposite direction
in order to flatten the gradient.
The optimization strategy gradient decent is used for taking the gradient from steep to
flat. It is the rule used to determine how much the weights (usually on an individual basis) are
updated relative to the current gradient. RMSprop was used as the gradient decent optimizing
algorithm. RMSprop uses the magnitude of the recent gradient to normalize the gradient
calculations. It attempts to iteratively find weights that are slightly better at reducing the loss
(given by the binary-cross-entropy loss function) than in the previous iteration. It is a process of
updating parameters of the function such that the gradient gets shallower and shallower until it is
basically flat. In turn, the loss function became smaller and smaller.

26
3.3.2 Backpropagation
Backpropagation starts after the forward pass where inputs are passed through the
network and output predictions are obtained. Once final loss value is obtained, it works
backwards from the top layers to the bottom layers, applying the chain rule to compute the
contribution that each parameter had in the loss value (Chollet 2017). This is the backward pass
where we compute the gradient of the loss function (which includes the loss function and
optimizer) at the final layer of the network and use the gradient to recursively apply the calculus
‘chain rule’ to update the weights layer by layer in our network (Chollet 2017). This is also
referred to as the weight update phase (Rosebrock 2017). Weights are adjusted layer by layer
because changing weight values in the reverse direction has a cascading effect on weights in
layers closer to the beginning of the network.
Feature extraction consisted of taking the convolutional base of the previously-trained
network, running the new data through it, and training a new data classifier on top of the output.
The convolutional base was run over the dataset, outputting a network feature array. It then used
that data as an input to a standalone densely-connected classifier (Challot 2017). Once features
were extracted from these images they were then flattened to be fed to the densely-connected
classifier. The densely-connected classifier was defined and used the dropout method for
regularization and was trained on the data and labels that were just recorded. Training happens
through a forward pass, calculating loss. Once loss is calculated the weights are updated through
backpropagation.

27
Chapter 4 Results
Looting of ancient archaeological sites continues to rise globally. Artifacts are stolen, lost
to the black-market, or sometimes destroyed. These pieces of history are often lost to mankind
forever. The goal of this study was to demonstrate the viability of using deep convolutional
neural networks (CNNs) within the field of archaeology and cultural heritage preservation with
the purpose of augmenting or replacing the manual detection of looting. It brings recent
advancements from the field of Artificial Intelligence to an applied GIS challenge at the
intersection of remote sensing and archaeology. Motivation is based in the need for
advancements in the field of cultural heritage preservation with the purpose of expediting and
aiding in the protection for ancient archaeological sites. Training is a transfer learning technique
that employs fine-tuning of previously trained deep convolutional neural networks that have
learned general features such as CaffeNet, AlexNet, or VGG16. The fine-tuned CNN is applied
to our dataset and adapts the network weights for the problem at hand. Using VGG16 it was
possible to achieve good results with 90-96% accuracy with only a few iterations.
4.1 GEP Findings
Ran on two GPU: NVIDIA GeForce GTX 1080 Ti, feature extraction was done on
VGG16 architecture using publicly available weights for training against the ImageNet dataset.
Feature extraction consisted of taking the convolutional base of the previously-trained network,
running the new data through it, and training a new data classifier on top of the output with a
batch size of 20 for 30 epochs.

28
Table 2: GEP Training, Testing, and Validation Loss and Accuracy Results
Training Validation Test
Training loss: 0.3871

Validation loss: 0.4158

Test loss is 0.4187
Training accuracy: 0.9150 Validation accuracy: 0.9600

Test accuracy is 0.9600

4.1.1 Loss
The loss function quantifies how well, or how bad the given predictor is at classifying the
input data points in the dataset. As stated earlier, smaller the loss the better the classifier is doing
at modeling the relationship between the input data and output class labels (Rosebrock 2017).
Loss is cumulative per epoch. At the beginning of each epoch, loss is zero. For each calculation
of the loss (given by the loss function) the loss is added to the loss metric on the graph. What is
seen over time in the plotted results in Figure: 7 below, is the total loss of training and validation
decreasing (generally) which means the weights of the network (in this case, the fully connected
classifier head on the convnet) are getting more accurate. However, when reading such a low
loss of near zero, it is evident that because the weights are near perfectly tuned to the training
and validation data, it has crossed the point where it has overfit the model which leads our model
to have problems generalizing.

29

Figure 7: Plotted GEP Training Loss and Validation Loss
4.1.2 Accuracy
In this gradient decent optimization, training accuracy increases with every epoch (Figure 8),
however a model that performs well on training data may not necessarily do well on data it has
never seen before (generalization). Training accuracy reaches 92%, validation accuracy reaches
96%, and test accuracy reaches 96%.

30

Figure 8: Plotted GEP Training and Validation Accuracy

4.1.3 Evaluating Error, Bias, and Variance
Bias relates to the ability of your model function to approximate the data, so high bias is
related to under-fitting (Ng 2011). Low bias in turn relates to your models ability to differentiate
well. A high bias percentile relates to misclassification. Variance is about the stability of your
models response to new training examples. High variance relates to overfitting (Ng 2011).
Measuring the error rate human level performance on looting pit classification has yet to be
quantified. For the sake of analysis, a conservative line was taken, assuming a baseline of 1%
error rate by a team of experts.

31
Table 3: GEP Measure of Training Error and Val Error for Calculating Bias and Variance
Human Error 1% Human Error – Train
Error = 7%

High Bias
Training Error 8% Training Error –
Validation Error = 4%

High Variance
Validation Error 4% Validation Error –
Test Error = 0
Overfitting of
Validation
Test Error 4% Test Error –
Validation Error = 0
Overfitting of
Validation

By looking at the algorithm error on the test set and the algorithm error on the validation
set, high bias, or high variance can be diagnosed. In this case, there is both high bias and high
variance. Some of the high bias relates directly to the ImageNet weights because they were not
optimized for the classes of images in this projects dataset. There is high bias because it is a
mostly linear classifier, but it also has high variance as well because of too much flexibility.
4.2 DGF Findings
Ran on two GPU: NVIDIA GeForce GTX 1080 Ti, feature extraction was done on
VGG16 architecture using publicly available weights for training against the ImageNet dataset.
Feature extraction consisted of taking the convolutional base of the previously-trained network,
running the new data through it, and training a new data classifier on top of the output with a
batch size of 20 for 30 epochs. One epoch is equal to one forward pass and one backward pass of
all the training examples. Batch size is equal to the number of training examples in one forward
/backward pass. Results:

32

Table 4: DGF Training, Testing, and Validation Loss and Accuracy Results
Training Validation Test
Training loss: 0.3228

Validation loss: 0.2895

Test loss is 0.3200
Training accuracy: 0.9275 Validation accuracy: 0.9200

Test accuracy is 0.9000

Table 5: DGF Measure of Training Error and Val Error for Calculating Bias and Variance
Human Error 1% Human Error – Train
Error = 6%

High Bias
Training Error 7% Training Error –
Validation Error = 1%

Low Variance
Validation Error 8% Validation Error –
Test Error = 2%
Some Overfitting of
Validation
Test Error 10% Test Error –
Validation Error = 2%
Some Overfitting of
Validation

33
4.2.1 Evaluating Error, Bias, and Variance
By looking at the algorithm error on the test set and the algorithm error on the validation
set, high bias, or high variance can be diagnosed. In this case, there is high bias and low
variance. Some of the high bias relates directly to the ImageNet weights because they were not
optimized for the classes of images in this projects dataset. As bias is a measure of how well
your model fits your data and how well it differentiates your data set, the lower the bias the
better job it is doing at differentiating. As seen with DGF data, with high bias there is
misclassification.
Figure 9: Plotted DGF Training Loss and Validation Loss

34

Figure 10: Plotted DGF Training and Validation Accuracy

4.2.2 DGF Accuracy
In this gradient decent optimization (Figure 10), training accuracy increases with every
epoch, however a model that performs well on training data may not necessarily do well on data
it has never seen before (generalization). Training accuracy reaches 93%, validation accuracy
reaches 92%, and test accuracy reaches 90%.
4.2.3 DGF Loss
The loss function quantifies how well, or how bad the given predictor is at classifying the
input data points in the dataset. The smaller the loss the better the classifier is doing at modeling
the relationship between the input data and output class labels (Rosebrock 2017). What is seen is

35
the total loss of training and validation decreasing epoch by epoch (figure 9) which means the
weights of the network (in this case, the fully connected classifier head on the convnet) are
getting more accurate. However, when reading such a low loss of near zero, it is evident that
because the weights are tuned to the training and validation data, it has crossed the point where it
has overfit the model which leads our model to have problems generalizing.
This study’s solution is to assist in looting prevention and detection by showing that
CNNs are a highly accurate and expedient method for machine learning and detecting of looting
with wide-ranging application beyond this specific research. This study applied machine learning
and deep learning algorithms to discover underlying patterns to its looting pit dataset which
enabled the correct classification of data points (pits) that the algorithm had yet to encounter. The
image datasets consisted of the images of two classes; looting pits, and not-pits, that were used to
teach the machine learning classifier what the different categories looked like with near human
accuracy but with a fraction of the time expenditure.

36
Chapter 5: Conclusion
This study utilized a pre-trained deep convolutional neural network VGG16 architecture
for feature extraction using publicly available weights for training against the ImageNet dataset.
Feature extraction consisted of taking the convolutional base of the previously-trained network,
running the new data through it, and training a new data classifier on top of the output. The
convolutional base was run over the dataset, recording its output to a numpy array. It then used
that data as an input to a standalone densely-connected classifier. The workflow of object
recognition and image classification process with a deep convolutional neural network for
looting pit recognition was covered and applied and yielded positive results. The applied
methodology has shown that CNNs can be highly efficient in expediting image recognition of
looting pits obtaining an overall accuracy of 96% with the GEP dataset and 90% accuracy with
the DGF dataset. Thoughts behind accuracy variation of the two datasets accuracy rating are
possibly due to image quality. The non-pit images in the DGF imagery are of a much higher
spatial resolution (0.30m) than that of the GEP images (2.5m) leading to higher detail quality and
a more varied non-pit dataset leading to a more complicated differentiation. GEP datasets
training and validation errors show high bias and high variance, while DGF dataset shows high
bias with low variance.
Convolutional neural networks are proving to be the best type of machine learning
models when faced with an image recognition task yielding higher and more accurate results
than other previously used methods. They are particularly favorable when working with small
datasets and are able to still yield decent to positive results. However, one of the drawbacks with
small datasets consist of overfitting, but this can be mitigated with data augmentation.

37
Future work to improve the shortcomings in performance of this CNN on this dataset
should start with following Ng’s recommendation for deep learning workflow as outlined in his
Nuts and Bolts of Applying Deep Learning talk from 2016, fine-tuning, and lastly visualizing
what the convnet learned.
Ng suggests starting by addressing high training error with high bias with the solution of
possibly training on a bigger model, training longer, ensuring optimization is working, or moving
to new model architecture. Once bias and training error is low, the next step is to address high
validation error with high variance. Ng suggests possible solutions for this include obtaining
more data, working on regularization, data augmentation or more data similar to test data or new
model architecture. Once high validation error with high variance is low, if test set data error is
high get more data, or possible new model architecture (Ng 2016).
After addressing variance and bias, the next recommendation would be fine-tuning. Fine-
tuning unfreezes a few of the top frozen layers of a frozen model base used for feature extraction
and together trains the newly added fully-connected classifier part of the model and those top
layers allowing for slight adjustments to the more abstract representations of the model that are
being reused which makes them more relevant to the task at hand. Fine-tuning adapts to a new
problem some of the representations previously learned by an existing model which helps the
model to perform better (Chollet 2017).
For a final recommendation, visualizing the intermediate activation outputs that the
convolutional neural network learned is useful in understanding how successive layers transform
their input in order to have a better understanding of the meaning of individual filters. Deep
learning models are often referred to as difficult to extract and present in a human readable form
but it is possible to do so with convolutional neural networks. Visualizing the networks filters

38
helps in understanding exactly what visual pattern concept each filter in the network is receptive
to. Heatmaps of class activation in an image can also be visualized and aid in the understanding
of which part of an image were identified as belonging to an individual class and in doing so
allows to localize objects in images.
Revolutionary recent advancements in the machine learning subfield of deep learning
have shown remarkable results in near-human level image classifications as well as in all
perceptual problems. Deep learning and AI’s long-term picture is looking bright as it is starting
to be applied to many problems in which it has the possible capability to be transformative to
many fields. Archaeology is one of them. The use of deep convolutional neural networks is
applicable and will prove to be a valuable tool for archaeological research and site preservation.
Within the field of archaeology efforts are currently being made to use GIS, remote sensing, and
algorithmic identification techniques to mitigate against looted archaeological sites. Building
upon these efforts, this study acquired high-resolution satellite imagery, built its datasets, and
was able to achieve its goal of using a deep convolutional neural network to obtain near-human
level accuracy (90% and 96%) of looting pit detection expediting the time and the process of
detection with the potential of broader future applications. The work done for this study using
deep convolutional neural networks for looting pit detection is an early example of what will one
day be dominant research modality.

39
References

Antiquities Coalition , 2016. Maps. https://theantiquitiescoalition.org/resources/maps/
(Accessed November 23, 2016).
Atwood, R. 2004. Stealing History: Tomb Raiders, Smugglers, and the Looting of the ancient
world. New York: St. Martin’s.
Brodie, N. & D. Contreras 2012. Lazrus & A.W. Barker (ed.), “All the Kings Horses: Looting,
antiquities trafficking and the integrity of the archaeological record”: 9-24. Washington,
D.C.: Society for American Archaeology.
Brodie, N. & C. Renfrew, 2005. “Looting the Worlds Archaeological Heritage: the inadequate
response”. Annual review of Anthropology 34:343-361
http://dx.doi.org.libproxy1.usc.edu/10.1146/annurev.anthro34.081.084.120551
Brodie, N. & K. Tubb, 2001. “Illicit Antiquities: The Theft of Culture and the Extinction of
Archaeology”. London: Routledge.

Bowen, E., et al., 2017. “Algorithmic Identification of Looted Archaeological Sites from Space”
https://www.frontiersin.org/articles/10.3389/fict.2017.00004/full (accessed 15 January
2018).

CAPMAS/CAPMASTAT 1.0: DevInfo adaptation released by Population Statistics and
Censuses Sector affiliated to the Central Agency for Public Mobilization & Statistics
(CAPMAS). http://www.devinfo.org/capmas/libraries/aspx/Catalog.aspx (accessed 26
October 2016).

Casana, J. & N. Panahipour 2014. “Notes on a Disappearing Past: Satellite- based monitoring of
looting and damaged to archaeological sites in Syria”.
Castelluccio,M. G. Poggi, C. Sansone, L. Verdoliva, 2015. “Land Use Classification in Remote
Sensing Images by Convolutional Neural Networks” https://arxiv.org/pdf/1508.00092.pdf
(accessed 9 November 2017).
Central Bank of Egypt, (2014). Economic research. Time series. Available at:
http://www.cbe.org.eg/english/economic+research/time+series/ (accessed 31 October,
2016).
Chollet, F. 2017. “Deep Learning with Python” Manning Publications. www. Manning.com
(accessed 28 April 2018)

40
Coluzzi, R. et al., 2010. “Satellite Imagery Time Series for the Detection of Looting Activities at
Archaeological Sites”. CNR-IMAA. EGU General Assembly 2010: 10569.

Contreras, D.& Brodie, N. 2010. “The Utility of Publicly Available Satellite Imagery for
Investigating Looting of Archaeological Sites in Jordan”. Journal of Field Archaeology
35: 101-14. http://dx.doi.org.libproxy1.usc.edu/10.1179/009346910X12707320296838

Crossona, A. 2012. “Egypt Looters Ransack Archaeological Sites”. The World, TI March 2012.
Available at: http://www.theworld.org/2012/03/egypt-looters-ransack-sites/(accessed 30
October 2016).
El Dorry, M.-A. 2011. “Why Do People Loot? The Case of the Egyptian Revolution”. al-Rawi
Egypt's Heritage Review 2011: 20-28.
Reske Henry, J. 2012. “Egypt's Poverty, Unemployment, Push Youths to Breaking Point".
Retrieved 10 February 2012. http://www.newsmax.com/Newsfront/Egypt-poverty-
unemployment-unrest/2011/01/31/id/384555/.
Hanna, M. 2013. “What has Happened to Egyptian Heritage After the 2011 Unfinished
Revolution?” Journal of Eastern Mediterranean Archaeology & Heritage Studies 1: 371-
75. http://dx.doi.org.libproxy1.usc.edu/10.5325/jeasmedarcherstu.1.4.0371
Hu, Fan, Xia, Gui-Song; Hu, Jingwen Zhang, Liangpei.2015. “Transferring Deep Convolutional
Neural Networks for the Scene Classification of High-Resolution Remote Sensing
Imagery”Remote Sensing; Basel Vol. 7, Iss. 11, (2015): 14680-14707. https://search-
proquest-com.libproxy1.usc.edu/docview/1748566503?pq-
origsite=summon&accountid=14749(accessed 9 November 2017)
Jemahs 2013. “Cultural Heritage in Times of Crisis”. Journal of Eastern Mediterranean
Archaeology Heritage Studies 1: 366-71.
Krizhevsky A. et al., 2014. “ImageNet Classification with Deep Convolutional Neural
Networks” https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-
convolutional-neural-networks.pdf (accessed 20 February 2018)
Ikram, S. et al., 2013. “Looting and land grabbing: the current situation in Egypt”. Bulletin of
the American Research Center in Egypt 202: 34-39.
Lauricella, A. et al., 2017 “Semi-automated detection of looting in Afghanistan using
multispectral imagery and principal component analysis” Antiquity Volume 91, Issue
359. https://www-cambridge-
org.libproxy1.usc.edu/core/journals/antiquity/article/semiautomated-detection-of-looting-

41
in-afghanistan-using-multispectral-imagery-and-principal-component-
analysis/A649663703443DBAB767D434587BBB1C (Accesed 10 November 2017).
Muller,T. 2016. National Geographic: “Plundering the past”. (accessed 25 October 2016)
http://www.nationalgeographic.com/magazine/2016/06/looting-ancient-blood-antiquities/
Ng, A., 2016, “Nuts and Bolts of Applying Deep Learning”
https://www.youtube.com/watch?v=F1ka6a13S9I (accessed 20 April 2018).

Parcak, S. 2007. “Going, Going, Gone: towards a satellite remote-sensing methodology for
monitoring archaeological tell sites under threat in the Middle East”. Journal of Field
Archaeology 42: 61-83.

Parcak, S. 2009. “Satellite Remote Sensing for Archaeology”. New York: Routledge.

Parcak, S. 2016. “Satellite evidence of Archaeological site looting in Egypt: 2002-2013”.
Antiquity Volume 90 issue 394, pp. 188-205. (accessed 30 October 2016).
https://www.cambridge.org/core/journals/antiquity/article/satellite-evidence-of-
archaeological-site-looting-in-egypt-
20022013/D23EA939FC4767D8BF50CAC6DE96D005

Pedraza A., Gloria Bueno, Oscar Deniz, Gabriel Cristóbal , Saúl Blanco,and María Borrego-
Ramos 3 2017. “Automated Diatom Classification (Part B): A Deep Learning
Approach”VISILAB Research Group, University Castilla La Mancha, Av. Camilo José
Cela s/n,Ciudad Real 13071, Spain. http://digital.csic.es/handle/10261/149382 (accessed
11 November 2017).

Penatti , A. 2015. “Do Deep Features Generalize from Everyday Objects to Remote Sensing
and Aerial Scenes Domains?”Computer Vision Foundation https://www.cv-
foundation.org/openaccess/content_cvpr_workshops_2015/W13/papers/Penatti_Do_Dee
p_Features_2015_CVPR_paper.pdf (accessed 9 November 2017).
Rosebrock, A. 2017. “Deep Learning for Computer Vision with Python ”Pyimagesearch
https://www.pyimagesearch.com/deep-learning-computer-vision-python-book/ ( last
acceded 20 May 2018).
Russakovsky et al., 2015 “ImageNet Large Scale Visual Recognition Challenge ”
https://link.springer.com/article/10.1007/s11263-015-0816-y (accessed 12 February
2018)

42
Simonyan, K. & Zisserman 2015. “Very Deep Convolutional networks for Large-Scale Image
Recognition ” https://arxiv.org/abs/1409.1556 (accessed 2 January 2018)
Stone, Elizabeth C. 2008. “Patterns of Looting in Iraq”, Antiquity 82, 125-138. (accessed 17
September 2017)
http://www.academia.edu/4764494/Elizabeth_C._Stone_2008_Patterns_of_Looting_in_Ir
aq_Antiquity_82_125-138
United Nations Statistics Division: UN Data: Egypt. (accessed 28 October 2016).
http://data.un.org/CountryProfile.aspx?crName=egypt
World Bank Database. 2014. http://www.worldbank.org/en/country/egypt (accessed 31 October
2016).
The World Bank, (2013) Egyptian
Overview. .http://www.worldbank.org/en/country/egypt/overview

43
Appendix A: GEP Code
import os
import matplotlib.pyplot as plt
import numpy as np
from keras.preprocessing.image import ImageDataGenerator
from keras.applications import VGG16
from keras import models
from keras import layers
from keras import optimizers

print("Starting image classification process...")
conv_base = VGG16(weights='imagenet', include_top=False, input_shape=(500, 500, 3))
print("VGG16 base summary: " + str(conv_base.summary()))
print("Starting training data augmentation")
print("Starting feature extration process...")

base_dir = '/home/jw/Timber_Thesis/pits_v_sand4'
print("Using base directory for training, validation and test images: " + base_dir)

train_dir = os.path.join(base_dir, 'train')
validation_dir = os.path.join(base_dir, 'validation')
test_dir = os.path.join(base_dir, 'test')
datagen = ImageDataGenerator(rescale=1./255)
batch_size = 20

def extract_features(directory, sample_count):
features = np.zeros(shape=(sample_count, 15, 15, 512))
labels = np.zeros(shape=(sample_count))

44
generator = datagen.flow_from_directory(
directory,
target_size=(500, 500),
batch_size=batch_size,
class_mode='binary')
i = 0
for inputs_batch, labels_batch in generator:
features_batch = conv_base.predict(inputs_batch)
features[i * batch_size : (i + 1) * batch_size] = features_batch
labels[i * batch_size : (i + 1) * batch_size] = labels_batch
i += 1
if i * batch_size >= sample_count:
break
return features, labels
print("Starting feature extration process on training images...")
train_features, train_labels = extract_features(train_dir, 400)
print("Ending feature extration process on training images...")
print("Starting feature extration process on validation images...")
validation_features, validation_labels = extract_features(validation_dir, 50)
print("Ending feature extration process on validation images...")
print("Starting feature extration process on test images...")
test_features, test_labels = extract_features(test_dir, 50)
print("Ending feature extration process on test images...")
print("Ending feature extration process...")
train_features = np.reshape(train_features, (400, 15 * 15 * 512))
validation_features = np.reshape(validation_features, (50, 15 * 15 * 512))
test_features = np.reshape(test_features, (50, 15 * 15 * 512))

45
model = models.Sequential()
model.add(layers.Dense(256, activation='relu', input_dim=15 * 15 * 512))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(1, activation='sigmoid'))
model.compile(optimizer=optimizers.RMSprop(lr=2e-5),
loss='binary_crossentropy',
metrics=['acc'])

print("Starting training of NN on training, validation features...")
history = model.fit(train_features, train_labels,
epochs=30,
batch_size=20,
validation_data=(validation_features,
validation_labels))

print("Ending training of NN on training, validation features...")

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(1, len(acc) + 1)

plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.legend()
plt.figure()

46
plt.plot(epochs, loss, 'bo', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()
plt.show()

print("Starting models evaluation of NN on test dataset...")
test_datagen = ImageDataGenerator(rescale=1./255)
test_generator = test_datagen.flow_from_directory(test_dir,
target_size=(500, 500),
batch_size=20,
class_mode='binary')

test_loss_and_metrics = model.evaluate(test_features, test_labels)
print("Ending evaluation of NN on test dataset...")
print('Available test metrics: ', str(model.metrics_names))
print('CNN model loss and accuracy on test dataset: ', str(test_loss_and_metrics))

print("Starting prediction on test dataset...")
predictions = model.predict(test_features)
print(predictions)
print("Ending prediction on test dataset...")
print("Ending image classification process...")
print("Done")

47
Appendix B: GEP Console Output
(deeplearning) jw@Orion:~/Timber_Thesis/pits_v_sand4$ python listing_529.py
Using TensorFlow backend.
/usr/lib/python3.6/importlib/_bootstrap.py:219: RuntimeWarning: compiletime version 3.5
of module 'tensorflow.python.framework.fast_tensor_util' does not match runtime version
3.6
return f(*args, **kwds)
Starting image classification process...
2018-04-05 07:15:55.763735: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU
supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2
AVX AVX2 FMA
2018-04-05 07:15:55.930950: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:892]
successful NUMA node read from SysFS had negative value (-1), but there must be at least
one NUMA node, so returning NUMA node zero
2018-04-05 07:15:55.931228: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030]
Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.721
pciBusID: 0000:01:00.0
totalMemory: 10.91GiB freeMemory: 10.08GiB
2018-04-05 07:15:56.048337: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:892]
successful NUMA node read from SysFS had negative value (-1), but there must be at least
one NUMA node, so returning NUMA node zero
2018-04-05 07:15:56.048597: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030]
Found device 1 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.721
pciBusID: 0000:02:00.0
totalMemory: 10.91GiB freeMemory: 10.75GiB
2018-04-05 07:15:56.049326: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1045]
Device peer to peer matrix
2018-04-05 07:15:56.049347: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1051] DMA:
0 1
2018-04-05 07:15:56.049352: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1061] 0:
Y Y
2018-04-05 07:15:56.049354: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1061] 1:
Y Y

48
2018-04-05 07:15:56.049378: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1120]
Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: GeForce GTX 1080 Ti, pci
bus id: 0000:01:00.0, compute capability: 6.1)
2018-04-05 07:15:56.049385: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1120]
Creating TensorFlow device (/device:GPU:1) -> (device: 1, name: GeForce GTX 1080 Ti, pci
bus id: 0000:02:00.0, compute capability: 6.1)
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 500, 500, 3) 0
_________________________________________________________________
block1_conv1 (Conv2D) (None, 500, 500, 64) 1792
_________________________________________________________________
block1_conv2 (Conv2D) (None, 500, 500, 64) 36928
_________________________________________________________________
block1_pool (MaxPooling2D) (None, 250, 250, 64) 0
_________________________________________________________________
block2_conv1 (Conv2D) (None, 250, 250, 128) 73856
_________________________________________________________________
block2_conv2 (Conv2D) (None, 250, 250, 128) 147584
_________________________________________________________________
block2_pool (MaxPooling2D) (None, 125, 125, 128) 0
_________________________________________________________________
block3_conv1 (Conv2D) (None, 125, 125, 256) 295168
_________________________________________________________________
block3_conv2 (Conv2D) (None, 125, 125, 256) 590080
_________________________________________________________________
block3_conv3 (Conv2D) (None, 125, 125, 256) 590080
_________________________________________________________________
block3_pool (MaxPooling2D) (None, 62, 62, 256) 0

49
_________________________________________________________________
block4_conv1 (Conv2D) (None, 62, 62, 512) 1180160
_________________________________________________________________
block4_conv2 (Conv2D) (None, 62, 62, 512) 2359808
_________________________________________________________________
block4_conv3 (Conv2D) (None, 62, 62, 512) 2359808
_________________________________________________________________
block4_pool (MaxPooling2D) (None, 31, 31, 512) 0
_________________________________________________________________
block5_conv1 (Conv2D) (None, 31, 31, 512) 2359808
_________________________________________________________________
block5_conv2 (Conv2D) (None, 31, 31, 512) 2359808
_________________________________________________________________
block5_conv3 (Conv2D) (None, 31, 31, 512) 2359808
_________________________________________________________________
block5_pool (MaxPooling2D) (None, 15, 15, 512) 0
=================================================================
Total params: 14,714,688
Trainable params: 14,714,688
Non-trainable params: 0
_________________________________________________________________
VGG16 base summary: None
Starting training data augmentation
Starting feature extration process...
Using base directory for training, validation and test images:
/home/jw/Timber_Thesis/pits_v_sand4
Starting feature extration process on training images...
Found 400 images belonging to 2 classes.

50
Ending feature extration process on training images...
Starting feature extration process on validation images...
Found 50 images belonging to 2 classes.
Ending feature extration process on validation images...
Starting feature extration process on test images...
Found 50 images belonging to 2 classes.
Ending feature extration process on test images...
Ending feature extration process...
Starting training of NN on training, validation features...
Train on 400 samples, validate on 50 samples
Epoch 1/30
400/400 [==============================] - 1s 2ms/step - loss: 0.8846 - acc: 0.5300 -
val_loss: 0.6629 - val_acc: 0.5000
Epoch 2/30
400/400 [==============================] - 0s 1ms/step - loss: 0.7143 - acc: 0.5375 -
val_loss: 0.6518 - val_acc: 0.5000
Epoch 3/30
400/400 [==============================] - 0s 1ms/step - loss: 0.6826 - acc: 0.5300 -
val_loss: 0.6872 - val_acc: 0.5000
Epoch 4/30
400/400 [==============================] - 0s 1ms/step - loss: 0.6697 - acc: 0.5950 -
val_loss: 0.6309 - val_acc: 0.8800
Epoch 5/30
400/400 [==============================] - 0s 935us/step - loss: 0.6180 - acc: 0.6675 -
val_loss: 0.6514 - val_acc: 0.5000
Epoch 6/30
400/400 [==============================] - 0s 1ms/step - loss: 0.6420 - acc: 0.6425 -
val_loss: 0.6210 - val_acc: 0.9600
Epoch 7/30
400/400 [==============================] - 0s 982us/step - loss: 0.6174 - acc: 0.6900 -
val_loss: 0.6079 - val_acc: 0.9800
Epoch 8/30

51
400/400 [==============================] - 0s 895us/step - loss: 0.6095 - acc: 0.6750 -
val_loss: 0.6036 - val_acc: 0.7000
Epoch 9/30
400/400 [==============================] - 0s 900us/step - loss: 0.5953 - acc: 0.7150 -
val_loss: 0.6217 - val_acc: 0.5000
Epoch 10/30
400/400 [==============================] - 0s 946us/step - loss: 0.5781 - acc: 0.7475 -
val_loss: 0.5890 - val_acc: 0.7000
Epoch 11/30
400/400 [==============================] - 0s 919us/step - loss: 0.5663 - acc: 0.7850 -
val_loss: 0.5690 - val_acc: 0.9800
Epoch 12/30
400/400 [==============================] - 0s 945us/step - loss: 0.5521 - acc: 0.8125 -
val_loss: 0.5701 - val_acc: 0.7000
Epoch 13/30
400/400 [==============================] - 0s 1ms/step - loss: 0.5605 - acc: 0.7900 -
val_loss: 0.5643 - val_acc: 0.6200
Epoch 14/30
400/400 [==============================] - 0s 1ms/step - loss: 0.5513 - acc: 0.8150 -
val_loss: 0.5392 - val_acc: 0.9800
Epoch 15/30
400/400 [==============================] - 0s 941us/step - loss: 0.5195 - acc: 0.8650 -
val_loss: 0.5746 - val_acc: 0.5800
Epoch 16/30
400/400 [==============================] - 0s 1ms/step - loss: 0.5164 - acc: 0.8475 -
val_loss: 0.5212 - val_acc: 1.0000
Epoch 17/30
400/400 [==============================] - 0s 1ms/step - loss: 0.4981 - acc: 0.8775 -
val_loss: 0.5104 - val_acc: 1.0000
Epoch 18/30
400/400 [==============================] - 0s 1ms/step - loss: 0.4797 - acc: 0.8650 -
val_loss: 0.5166 - val_acc: 0.8000
Epoch 19/30
400/400 [==============================] - 1s 1ms/step - loss: 0.5026 - acc: 0.8625 -
val_loss: 0.5001 - val_acc: 0.9600

52
Epoch 20/30
400/400 [==============================] - 0s 988us/step - loss: 0.4702 - acc: 0.8975 -
val_loss: 0.4904 - val_acc: 0.9600
Epoch 21/30
400/400 [==============================] - 0s 1ms/step - loss: 0.4657 - acc: 0.8900 -
val_loss: 0.4930 - val_acc: 0.8200
Epoch 22/30
400/400 [==============================] - 1s 1ms/step - loss: 0.4444 - acc: 0.8950 -
val_loss: 0.4916 - val_acc: 0.8800
Epoch 23/30
400/400 [==============================] - 0s 918us/step - loss: 0.4465 - acc: 0.9000 -
val_loss: 0.4685 - val_acc: 0.9400
Epoch 24/30
400/400 [==============================] - 0s 1ms/step - loss: 0.4276 - acc: 0.9150 -
val_loss: 0.4724 - val_acc: 0.9200
Epoch 25/30
400/400 [==============================] - 0s 1ms/step - loss: 0.4222 - acc: 0.9100 -
val_loss: 0.4505 - val_acc: 0.9600
Epoch 26/30
400/400 [==============================] - 0s 1ms/step - loss: 0.4007 - acc: 0.9300 -
val_loss: 0.4414 - val_acc: 0.9600
Epoch 27/30
400/400 [==============================] - 0s 935us/step - loss: 0.4027 - acc: 0.9150 -
val_loss: 0.4329 - val_acc: 0.9600
Epoch 28/30
400/400 [==============================] - 0s 1ms/step - loss: 0.4009 - acc: 0.9275 -
val_loss: 0.4312 - val_acc: 0.9600
Epoch 29/30
400/400 [==============================] - 0s 1ms/step - loss: 0.3884 - acc: 0.9400 -
val_loss: 0.4071 - val_acc: 0.9800
Epoch 30/30
400/400 [==============================] - 0s 1ms/step - loss: 0.3871 - acc: 0.9150 -
val_loss: 0.4158 - val_acc: 0.9600
Ending training of NN on training, validation features...
Starting evaluation of NN on test dataset...

53
Found 50 images belonging to 2 classes.
50/50 [==============================] - 0s 484us/step
Ending evaluation of NN on test dataset...
Available test metrics: ['loss', 'acc']
CNN accuracy on test dataset: [0.41870951652526855, 0.95999999046325679]
Ending image classification process...
Done

54
Appendix C: DGF Code

import os
import matplotlib.pyplot as plt
import numpy as np
from keras.preprocessing.image import ImageDataGenerator
from keras.applications import VGG16
from keras import models
from keras import layers
from keras import optimizers
from pathlib import Path

print("Starting image classification process...")
conv_base = VGG16(weights='imagenet', include_top=False, input_shape=(112, 112, 3))
print("VGG16 base summary: " + str(conv_base.summary()))
print("Starting feature extration process...")
base_dir = str(Path.cwd())
print("Using base directory for training, validation and test images: " + base_dir)

train_dir = os.path.join(base_dir, 'train')
validation_dir = os.path.join(base_dir, 'validation')
test_dir = os.path.join(base_dir, 'test')

datagen = ImageDataGenerator(rescale=1./255)
batch_size = 20

def extract_features(directory, sample_count):
features = np.zeros(shape=(sample_count, 3, 3, 512))

55
labels = np.zeros(shape=(sample_count))
generator = datagen.flow_from_directory(
directory,
target_size=(112, 112),
batch_size=batch_size,
class_mode='binary')
i = 0
for inputs_batch, labels_batch in generator:
features_batch = conv_base.predict(inputs_batch)
features[i * batch_size : (i + 1) * batch_size] = features_batch
labels[i * batch_size : (i + 1) * batch_size] = labels_batch
i += 1
if i * batch_size >= sample_count:
break
return features, labels

print("Starting feature extration process on training images...")
train_features, train_labels = extract_features(train_dir, 400)
print("Ending feature extration process on training images...")
print("Starting feature extration process on validation images...")
validation_features, validation_labels = extract_features(validation_dir, 100)
print("Ending feature extration process on validation images...")
print("Starting feature extration process on test images...")
test_features, test_labels = extract_features(test_dir, 100)
print("Ending feature extration process on test images...")
print("Ending feature extration process...")

56
train_features = np.reshape(train_features, (400, 3 * 3 * 512))
validation_features = np.reshape(validation_features, (100, 3 * 3 * 512))
test_features = np.reshape(test_features, (100, 3 * 3 * 512))

model = models.Sequential()
model.add(layers.Dense(256, activation='relu', input_dim=3 * 3 * 512))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(1, activation='sigmoid'))

model.compile(optimizer=optimizers.RMSprop(lr=2e-5),
loss='binary_crossentropy',
metrics=['acc'])

print("Starting training of NN on training, validation features...")
history = model.fit(train_features, train_labels,
epochs=30,
batch_size=20,
validation_data=(validation_features,
validation_labels))

print("Ending training of NN on training, validation features...")

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(1, len(acc) + 1)

57

plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.legend()
plt.figure()
plt.plot(epochs, loss, 'bo', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()
plt.show()

print("Starting evaluation of NN on test dataset...")
test_datagen = ImageDataGenerator(rescale=1./255)
test_generator = test_datagen.flow_from_directory(test_dir,
target_size=(112, 112),
batch_size=20,
class_mode='binary')

test_loss_and_metrics = model.evaluate(test_features, test_labels)
print("Ending evaluation of NN on test dataset...")
print('Available test metrics: ', str(model.metrics_names))
print('CNN accuracy on test dataset: ', str(test_loss_and_metrics))
print("Ending image classification process...")
print("Done")

58
Appendix D: DGF Console Output
I jw@Orion:~/Dropbox/MASTERS_THESIS_W/pits_v_sand5$ python3 listing_530.py
Using TensorFlow backend.
/usr/lib/python3.6/importlib/_bootstrap.py:219: RuntimeWarning: compiletime version 3.5
of module 'tensorflow.python.framework.fast_tensor_util' does not match runtime version
3.6
return f(*args, **kwds)
Starting image classification process...
2018-04-27 18:15:33.020201: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU
supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2
AVX AVX2 FMA
2018-04-27 18:15:33.021330: E tensorflow/stream_executor/cuda/cuda_driver.cc:406] failed
call to cuInit: CUDA_ERROR_UNKNOWN
2018-04-27 18:15:33.021350: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:158]
retrieving CUDA diagnostic information for host: Orion
2018-04-27 18:15:33.021355: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:165]
hostname: Orion
2018-04-27 18:15:33.021377: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:189]
libcuda reported version is: 387.34.0
2018-04-27 18:15:33.021389: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:369]
driver version file contents: """NVRM version: NVIDIA UNIX x86_64 Kernel Module 387.34
Tue Nov 21 03:09:00 PST 2017
GCC version: gcc version 7.2.0 (Ubuntu 7.2.0-8ubuntu3.2)
"""
2018-04-27 18:15:33.021397: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:193]
kernel reported version is: 387.34.0
2018-04-27 18:15:33.021401: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:300]
kernel version seems to match DSO: 387.34.0
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 112, 112, 3) 0
_________________________________________________________________
block1_conv1 (Conv2D) (None, 112, 112, 64) 1792
_________________________________________________________________

59
block1_conv2 (Conv2D) (None, 112, 112, 64) 36928
_________________________________________________________________
block1_pool (MaxPooling2D) (None, 56, 56, 64) 0
_________________________________________________________________
block2_conv1 (Conv2D) (None, 56, 56, 128) 73856
_________________________________________________________________
block2_conv2 (Conv2D) (None, 56, 56, 128) 147584
_________________________________________________________________
block2_pool (MaxPooling2D) (None, 28, 28, 128) 0
_________________________________________________________________
block3_conv1 (Conv2D) (None, 28, 28, 256) 295168
_________________________________________________________________
block3_conv2 (Conv2D) (None, 28, 28, 256) 590080
_________________________________________________________________
block3_conv3 (Conv2D) (None, 28, 28, 256) 590080
_________________________________________________________________
block3_pool (MaxPooling2D) (None, 14, 14, 256) 0
_________________________________________________________________
block4_conv1 (Conv2D) (None, 14, 14, 512) 1180160
_________________________________________________________________
block4_conv2 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block4_conv3 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block4_pool (MaxPooling2D) (None, 7, 7, 512) 0
_________________________________________________________________
block5_conv1 (Conv2D) (None, 7, 7, 512) 2359808
_________________________________________________________________

60
block5_conv2 (Conv2D) (None, 7, 7, 512) 2359808
_________________________________________________________________
block5_conv3 (Conv2D) (None, 7, 7, 512) 2359808
_________________________________________________________________
block5_pool (MaxPooling2D) (None, 3, 3, 512) 0
=================================================================
Total params: 14,714,688
Trainable params: 14,714,688
Non-trainable params: 0
_________________________________________________________________
VGG16 base summary: None
Starting feature extration process...
Using base directory for training, validation and test images:
/media/dropbox/Dropbox/MASTERS_THESIS_W/pits_v_sand5
Starting feature extration process on training images...
Found 400 images belonging to 2 classes.
Ending feature extration process on training images...
Starting feature extration process on validation images...
Found 100 images belonging to 2 classes.
Ending feature extration process on validation images...
Starting feature extration process on test images...
Found 100 images belonging to 2 classes.
Ending feature extration process on test images...
Ending feature extration process...
Starting training of NN on training, validation features...
Train on 400 samples, validate on 100 samples
Epoch 1/30
400/400 [==============================] - 0s 440us/step - loss: 0.7721 - acc: 0.5150 -
val_loss: 0.6698 - val_acc: 0.5000

61
Epoch 2/30
400/400 [==============================] - 0s 279us/step - loss: 0.7230 - acc: 0.5075 -
val_loss: 0.6338 - val_acc: 0.8600
Epoch 3/30
400/400 [==============================] - 0s 281us/step - loss: 0.6939 - acc: 0.5750 -
val_loss: 0.6118 - val_acc: 0.8400
Epoch 4/30
400/400 [==============================] - 0s 280us/step - loss: 0.6676 - acc: 0.6075 -
val_loss: 0.5898 - val_acc: 0.8700
Epoch 5/30
400/400 [==============================] - 0s 275us/step - loss: 0.6373 - acc: 0.6125 -
val_loss: 0.5660 - val_acc: 0.9000
Epoch 6/30
400/400 [==============================] - 0s 283us/step - loss: 0.6185 - acc: 0.6450 -
val_loss: 0.5474 - val_acc: 0.8900
Epoch 7/30
400/400 [==============================] - 0s 283us/step - loss: 0.5861 - acc: 0.7075 -
val_loss: 0.5280 - val_acc: 0.8900
Epoch 8/30
400/400 [==============================] - 0s 281us/step - loss: 0.5918 - acc: 0.6875 -
val_loss: 0.5108 - val_acc: 0.8800
Epoch 9/30
400/400 [==============================] - 0s 275us/step - loss: 0.5400 - acc: 0.7675 -
val_loss: 0.4956 - val_acc: 0.9000
Epoch 10/30
400/400 [==============================] - 0s 277us/step - loss: 0.5484 - acc: 0.7425 -
val_loss: 0.4813 - val_acc: 0.8900
Epoch 11/30
400/400 [==============================] - 0s 284us/step - loss: 0.5302 - acc: 0.7675 -
val_loss: 0.4640 - val_acc: 0.9000
Epoch 12/30
400/400 [==============================] - 0s 282us/step - loss: 0.5108 - acc: 0.7900 -
val_loss: 0.4501 - val_acc: 0.8900
Epoch 13/30

62
400/400 [==============================] - 0s 284us/step - loss: 0.4860 - acc: 0.8100 -
val_loss: 0.4364 - val_acc: 0.9000
Epoch 14/30
400/400 [==============================] - 0s 279us/step - loss: 0.4807 - acc: 0.8325 -
val_loss: 0.4254 - val_acc: 0.9000
Epoch 15/30
400/400 [==============================] - 0s 285us/step - loss: 0.4678 - acc: 0.8350 -
val_loss: 0.4181 - val_acc: 0.8900
Epoch 16/30
400/400 [==============================] - 0s 286us/step - loss: 0.4595 - acc: 0.8325 -
val_loss: 0.3995 - val_acc: 0.9000
Epoch 17/30
400/400 [==============================] - 0s 279us/step - loss: 0.4457 - acc: 0.8700 -
val_loss: 0.3889 - val_acc: 0.8900
Epoch 18/30
400/400 [==============================] - 0s 281us/step - loss: 0.4339 - acc: 0.8600 -
val_loss: 0.3797 - val_acc: 0.9000
Epoch 19/30
400/400 [==============================] - 0s 285us/step - loss: 0.4356 - acc: 0.8650 -
val_loss: 0.3695 - val_acc: 0.8900
Epoch 20/30
400/400 [==============================] - 0s 276us/step - loss: 0.4215 - acc: 0.8750 -
val_loss: 0.3655 - val_acc: 0.9000
Epoch 21/30
400/400 [==============================] - 0s 284us/step - loss: 0.4115 - acc: 0.8900 -
val_loss: 0.3556 - val_acc: 0.9100
Epoch 22/30
400/400 [==============================] - 0s 283us/step - loss: 0.3947 - acc: 0.8925 -
val_loss: 0.3431 - val_acc: 0.8900
Epoch 23/30
400/400 [==============================] - 0s 282us/step - loss: 0.3769 - acc: 0.8775 -
val_loss: 0.3381 - val_acc: 0.9300
Epoch 24/30
400/400 [==============================] - 0s 288us/step - loss: 0.3771 - acc: 0.8925 -
val_loss: 0.3293 - val_acc: 0.9200

63
Epoch 25/30
400/400 [==============================] - 0s 278us/step - loss: 0.3602 - acc: 0.9150 -
val_loss: 0.3211 - val_acc: 0.9100
Epoch 26/30
400/400 [==============================] - 0s 297us/step - loss: 0.3465 - acc: 0.9050 -
val_loss: 0.3147 - val_acc: 0.9100
Epoch 27/30
400/400 [==============================] - 0s 283us/step - loss: 0.3451 - acc: 0.9200 -
val_loss: 0.3074 - val_acc: 0.9100
Epoch 28/30
400/400 [==============================] - 0s 277us/step - loss: 0.3492 - acc: 0.8925 -
val_loss: 0.3065 - val_acc: 0.9200
Epoch 29/30
400/400 [==============================] - 0s 277us/step - loss: 0.3265 - acc: 0.9050 -
val_loss: 0.2961 - val_acc: 0.9200
Epoch 30/30
400/400 [==============================] - 0s 282us/step - loss: 0.3228 - acc: 0.9275 -
val_loss: 0.2895 - val_acc: 0.9200
Ending training of NN on training, validation features...
Starting evaluation of NN on test dataset...
Found 100 images belonging to 2 classes.
100/100 [==============================] - 0s 51us/step
Ending evaluation of NN on test dataset...
Available test metrics: ['loss', 'acc']
CNN accuracy on test dataset: [0.3282634776830673, 0.90000000000000002]
Ending image classification process...
Done

Abstract (if available)

Abstract Looting of archaeological sites is a global problem. To quantify looting on a nationwide scale and to assess the validity and scope of the looting reports and modern encroachment, satellite archaeologist have turned to mapping looting from space. High-resolution satellite imagery has become a powerful tool and resource for monitoring looting and site destruction remotely and proves to be an independent way to cross check and analyze against varied and unreliable reports from media and government agencies. It is estimated that over a quarter of Egypt’s 1100 known archaeological areas have sustained major damage and site destruction directly linked to looting. The organized looting and illicit trafficking of art and antiques, known as cultural racketeering, is a multi-billion dollar worldwide criminal industry that thrives in Egypt during times of political and economic turmoil and potentially funds drug cartels, armed insurgents, and even terrorist networks. This study analyzes methods used to monitor site looting at the archaeological site of al-Lisht which is located in the Egyptian governorate of Giza south of Cairo. Monitoring damage and looting over time has been largely dependent upon direct human interpretation of images. The manual image comparison method is laborious, time consuming, and prone to human-induced error. Recently, partially-supervised methods using deep convolutional neural networks (CNNs) have shown astounding performance in object recognition and detection. This study seeks to demonstrate the viability of using deep convolutional neural networks (CNNs) within the field of archaeology and cultural heritage preservation for the purpose of augmenting or replacing the manual detection of looting. It brings recent advancements from the field of Artificial Intelligence to an applied GIS challenge at the intersection of remote sensing and archaeology. The objective is to show that CNNs are a more accurate and expedient method for the detecting of looting with wide-ranging application beyond this specific research.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Object detection and digitization from aerial imagery using neural networks

PDF

A scenario based fire susceptibility approach for remote sensing platform comparison: Los Angeles county area, southern California, USA

PDF

Classifying ash cloud attributes of Eyjafjallajökull volcano, Iceland, using satellite remote sensing

PDF

An examination of close-range photogrammetry and traditional cave survey methods for terrestrial and underwater caves for 3-dimensional mapping

PDF

Assessing modern conflict to monitor human rights with remote sensing: Russia's war in Ukraine

PDF

Deep learning techniques for supervised pedestrian detection and critically-supervised object detection

PDF

Classifying the alpines: developing a methodology to track environmental changes in the alpines utilizing remote sensing and ecozone vegetation patterns

PDF

Exploring remote sensing and geographic information systems technologies to understand vegetation changes in response to land management practices at Finke Gorge National Park, Australia Between ...

PDF

From ruins to pixels: using remote sensing and GIS to analyze, document, and visualize archaeological sites and ancient roadways in Chaco Canyon, New Mexico

PDF

Evaluating machine learning tools for humanitarian road network mapping

PDF

Geographic object based image analysis for utility scale photovoltaic site suitability studies

PDF

An application of aerial drones in high definition mapping for autonomous vehicles

PDF

Efficient pipelines for vision-based context sensing

PDF

Surface rupture detection with support vector machine classification: case study from Ridgecrest, CA

PDF

Questioning the cause of calamity: using remotely sensed data to assess successive fire events

PDF

Mixed forest image classification of paper birch: using AVIRIS bandwidths ranging from 530 to 745 nm

PDF

Site location suitability analysis for a smart grid network

PDF

Using GIS to perform a risk assessment for air-transmitted bioterrorism within San Diego County

PDF

Creating a Web GIS to support field operations and enhance data collection for the Animal and Plant Health Inspection Service (APHIS)

PDF

Providing a new low-cost primary care facility for under-served communities: a site suitability analysis for Service Planning Area 6 in Los Angeles County, California

Asset Metadata

Creator Woolf, Timberlynn (author)

Core Title Deep convolutional neural networks for remote sensing investigation of looting of the archeological site of Al-Lisht, Egypt

School College of Letters, Arts and Sciences

Degree Master of Science

Degree Program Geographic Information Science and Technology

Publication Date 08/15/2018

Defense Date 08/15/2018

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

Tag Al-Lisht,archeology,deep convolutional neural networks,Egypt,image recognition,looting,OAI-PMH Harvest,remote sensing

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Fleming, Steven (committee chair)

Creator Email timberbarcelona@gmail.com,twoolf@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-71855

Unique identifier UC11671337

Identifier etd-WoolfTimbe-6767.pdf (filename),usctheses-c89-71855 (legacy record id)

Legacy Identifier etd-WoolfTimbe-6767.pdf

Dmrecord 71855

Document Type Thesis

Format application/pdf (imt)

Rights Woolf, Timberlynn

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