In keltoskytoi/iSEGMound: iSEGMound - A Reproducible Workflow for Mound Detection in LiDAR-derived DTMs

Exploitation of Automated Analysis Methods to detect burial mounds and structurally similar archaeological objects: a survey

The analysis carried out and communicated in this chapter reviews all published studies which are aimed at the detection of burial mound and mound-like structures in order to define a baseline to place present research. Numerous reviews of methods and applications used in Archaeological Remote Sensing have been carried out (e.g. @leiszOverviewApplicationRemote2013a, @agapiouRemoteSensingArchaeology2015; @davisObjectbasedImageAnalysis2019, @luoAirborneSpaceborneRemote2019, @fiorucciMachineLearningCultural2020): some broad and general, some more specific. The word survey was chosen specifically, because it was not a systematic search of databases as various ‘classical’ reviews did (such as @agapiouRemoteSensingArchaeology2015, @raunSystematicLiteratureReview2018, @magniniTheoryPracticeObjectbased2019 and @davisGeographicDisparityMachine2020a and @davisDefiningWhatWe2020). This survey is building on the experience of aforementioned systematic reviews (which show very diversified nomenclature and heterogeneous terms) and the scientific literature was collected based on these reviews, Researchgate, Academia and mainly considering open-access publications (if possible). As a starting point all published studies (at least available to the author via University access; number of studies in August 2021: 96) have been collected which deal with the automation of Archaeological Remote Sensing data set (practical studies which use any automation methods and analysis) but only those are discussed, which are concerned with the detection of burial mounds and structurally similar archaeological objects: altogether 31 publications. Burial mounds have been chosen as Objects of Interest for the review because they are relatively simple structures and a very common funerary custom throughout human history. Similarly structured archaeological objects like tell mounds or other monumental earthworks of circular form can be found in very different areas and time periods and are widely investigated and thus deliver good examples of automated analysis methods of these archaeological objects. The first archaeological objects to automatically detect were tell mounds and only then and with the dawn of the use of LiDAR data came burial mounds into the focus of research. While the early research carried on tell settlements can be connected to Ur & Menze, the monumental earthworks and mound shell rings can also be allocated to two research groups: Freeland et al. and Davis, Lipo & Sanger. The detection of burial mounds themselves shows more variability in the investigators (Figure 3). A complete list of references (which is by far not conclusive and is to be supplemented) is included as a supplement at the end of this thesis (Supplement 1).

library(dplyr)
library(ggplot2)
mounds_lit_2 <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds_short.csv'))
#read as tibble
mounds_lit_2 <- dplyr::as_tibble(mounds_lit_2)
mounds_lit_2_filt <- mounds_lit_2[, c(3:7)]
all_combined_2 <- table(mounds_lit_2_filt$Year, mounds_lit_2_filt$OoI,
                      mounds_lit_2_filt$Data, mounds_lit_2_filt$Software_type,
                      mounds_lit_2_filt$access)
methods_table_2 <- stats::ftable(all_combined_2)
methods_table_df2 <- as.data.frame(methods_table_2)
names(methods_table_df2) <- c("Year", "OoI", "Data", "Software_type", "access", "Freq")
ggplot(methods_table_df2, aes(fill=OoI, y=Freq, x=Year)) +
  geom_bar(position="stack", stat="identity") +
  ylim(0, 7) +
  scale_fill_manual(values=morris:::holland_palette) +
  ggtitle(expression(atop("Objects of Interest in Automated Archaeological Remote Sensing", atop("between 2006 and 2021", "")))) +
  theme_light() +
  xlab("Year") +
  ylab("Number of studies")

The main points considered in this survey were chosen based on the aforementioned reviews, which were conducted looking at citation indexes of different Remote Sensing methods in scientific periodicals (@agapiouRemoteSensingArchaeology2015, @raunSystematicLiteratureReview2018, @luoAirborneSpaceborneRemote2019, Davis 2020b), Institutional Affiliations of researchers (@agapiouRemoteSensingArchaeology2015, @raunSystematicLiteratureReview2018), the citation network (@raunSystematicLiteratureReview2018), OBIA by geographic region (@davisObjectbasedImageAnalysis2019), the developments and limitations of OBIA methods (@davisObjectbasedImageAnalysis2019), research goal using OBIA (@davisObjectbasedImageAnalysis2019), the development of different passive and active Remote Sensing methods (@luoAirborneSpaceborneRemote2019), the scale of ArchaeOBIA applications (@magniniTheoryPracticeObjectbased2019), datasets and methods applied in Archaeological Remote Sensing (@fiorucciMachineLearningCultural2020), different parameters, thresholds, methods and accuracy used in the automated detection of mound features (@davisDefiningWhatWe2020) and geometric disparity of Remote Sensing methods @davisGeographicDisparityMachine2020a). These systematic reviews stimulated questions such as: How can we learn from previous studies? How transparent is the workflow and if described, which variables and parameters (rule sets) were used? Which software was used? How reproducible or even replicable is the workflow in the studies?

Based on these questions the points investigated in this Master’s thesis are (Supplement 1):

(i) Location (only in Supplement 1) (ii) RS (Remote Sensing) data (iii) Methods (iv) Detail of methods (only in Supplement 1) (v) Variables/morphometric parameters (vi) OoI (Objects of Interest) (vii) Scale (viii) Software (ix) Access

The point “Location” (i) was documented, because different geographical and landscape conditions ask for different methods and objects can appear geographically and culturally diverse, which can explain the chosen method. The point “RS data” (ii) documents which type of remote sensing data was used and also it’s resolution if known. The point “Methods” (iii) organizes the studies in five main categories (Template Matching, Geometric Knowledge-based, GeOBIA-based and Machine Learning-based, Deep Learning; see Figure 4), based on @chengSurveyObjectDetection2016a (also see @lambersIntegratingRemoteSensing2019 and @rofflerStepStepDeveloping2020), instead of juxtaposing pixel-based and object-based methods. This differentiation points out the different object detection methods based on the approach to the dataset. Still, for the differentiation of Machine Learning-based methods, the expression pixel-based is going to be used for the classifier training-based methods, also because several studies use it (@sevaraPixelObjectComparison2016a, @niculitaBurialMoundDetection2020a).

knitr::include_graphics('C:/Users/kelto/Documents/iSEGMound/analysis/thesis/figures/Figure_4.png')

The point “Details of method” (iv) explains the workflow of the study in a nutshell, if comprehensible. Point v, “Variables/Morphometric parameters” highlights the variables and or morphometric parameters used in the study. “OOI” (vi) are the objects investigated. The point “Scale” (vii) implies the geographical range of the study, if it was applied to a broader region than the area on which the method was developed. The last two points (“Software” - viii, and “Access” - ix) investigated give us hints about the accessibility of the dataset and the code. Especially in earlier studies, information about the software used is not mentioned and sometimes can only be guessed. In the case of “Access”, apart from the information about accessible data and code it was also documented if there is a workflow (a comprehensible sequence of the steps of the workflow) or a flowchart (a very generalized workflow without clear steps) or if any of the equations used was published and or explained. It was also documented if any supplementary information or document was attached to the study which can guide to a better understanding for a possible reproduction of the study. Although all these points have been investigated, for practical reasons mainly the points (ii) RS (Remote Sensing) data, (iii) Method/s, (iv) Detail of method/s, (v) Variables/morphometric parameters, (vi) OoI/Object(s) of Interest, (viii) Software and (ix) Access is going to be elaborated in depth in this and the next chapter.

It is essential to discuss the basic traits of the used methods (iii, Supplement 2) in short, concentrating on their application in Archaeological Remote Sensing. It has to be stressed, that still no semantic lingua franca (between Archaeological Remote Sensing and Computer Vision and Remote Sensing Methods) exists when talking about automated methods in Archaeological Remote Sensing, thus methods are addressed in different ways (or most commonly in a very simplified way) and it is not always possible to determine which exact method has been utilised (recently it is improving, but it depends where the authors put the focus of their work). First, when shortly discussing the basic traits of the methods and their use in Archaeological Remote Sensing, it is only focused on the use of the different methods (what is used, where and when) since 2006, when tell mounds were first investigated using automated analysis methods. In the second part of this chapter the way of their use is investigated and analysed how the methods are used.

Template Matching-based methods utilize a template of the Object of Interest to be detected, which is then statistically matched to the input image. Two general directions can be distinguished: Rigid Template Matching, which requires a precise template which gets problematic when it comes to intra-class shape and size variations. Deformable Template Matching on the other hand can handle either free-form deformable templates or parametric deformable templates. (@chengSurveyObjectDetection2016a, 13-14). In the literature investigated Template Matching was used seven times for the detection of burial mounds, mound shell rings and tell mounds (@Boer2010UsingPR, @kramerArchaeologicalReactionRemote2015, @trierAutomaticDetectionMound2015a, @davisAutomatedMoundDetection2018a, @raunLIDARBasedSemiautomatic2019a, @davisObjectbasedImageAnalysis2019, @gholamrezaInvestigatingOLITIRS2021). Apart from @kramerArchaeologicalReactionRemote2015, Rigid Template Matching was preferred (Supplement 2). Template Matching itself is a method which is being revisited from time to time (Figure 5).

library(dplyr)
library(ggplot2)
mounds_lit <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds.csv'))
#read as tibble
mounds_lit <- dplyr::as_tibble(mounds_lit)
mounds_lit_filt <- mounds_lit[, c(3:7)]
all_combined <- table(mounds_lit_filt$Year, mounds_lit_filt$OoI, mounds_lit_filt$Method,
                      mounds_lit_filt$Data, mounds_lit_filt$Software_type)
all_combined_table <- stats::ftable(all_combined)
all_combined_df <- as.data.frame(all_combined_table)
names(all_combined_df) <- c("Year", "OoI", "Method", "Data", "Software_type", "Freq")
methods <- table(mounds_lit_filt$Year, mounds_lit_filt$Method)
methods_table <- stats::ftable(methods)
methods_table_df <- as.data.frame(methods_table)
names(methods_table_df) <- c("Year", "Method", "Freq")
names(methods_table_df)
TM <- methods_table_df %>% filter(Method=="Template Matching")
ggplot(TM, aes(x=Year, y=Freq)) +
  geom_bar(stat="identity", fill ="#A68A59") +
  ylim(0, 5) +
  ggtitle("Use of Template Matching-based methods between 2006 & 2021") +
  theme_light() +
  theme(legend.position="center") +
  xlab("Year") +
  ylab("Number of studies")

Geometric knowledge-based analysis works with specialized solutions for specific problems, applying rules based and established on knowledge of the regions of interest and it’s context. It uses either encoded prior geometric knowledge of the generic specificity of the Object of Interest and then e.g. applies hand-crafted mathematical morphometry rules for object extraction (morphometric derivatives such as Slope, Aspect, Curvature, Local Relief Model, vegetation indices etc.). Context knowledge based analysis uses knowledge about the relationship of the Object of Interest and the area it should be separated from (e.g. filters, textural analysis) (@chengSurveyObjectDetection2016a, 15). In the case of Automated Archeological Remote Sensing the two Geometric knowledge-based analysis approaches are often used in combination and form the data preparation step (Figure 6). Geometric knowledge based object analysis is only occasionally used per se as an automated analysis method for the detection of burial mounds, monumental earthworks and tell mounds: @rileyAutomatedDetectionPrehistoric2009a, @freelandAutomatedFeatureExtraction2016b, @romLandSeaAirborne2020b. (Lately) it is only occasional, that Geometric knowledge-based object analysis is not used as data preparation method (@Boer2010UsingPR, @caspariApplicationHoughForests2014a, @trierAutomaticDetectionMound2015a, @raunLIDARBasedSemiautomatic2019a, @caspariConvolutionalNeuralNetworks2019a, @kazimiObjectInstanceSegmentation2019b, @Kazimi2019SemiSL). In some cases, Geometric knowledge based analysis is included in the future work (compare works by @Kazimi2019SemiSL vs. @kazimiDETECTIONTERRAINSTRUCTURES2020a)(Supplement 2).

library(dplyr)
library(ggplot2)
mounds_lit <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds.csv'))
#read as tibble
mounds_lit <- dplyr::as_tibble(mounds_lit)
mounds_lit_filt <- mounds_lit[, c(3:7)]
all_combined <- table(mounds_lit_filt$Year, mounds_lit_filt$OoI, mounds_lit_filt$Method,
                      mounds_lit_filt$Data, mounds_lit_filt$Software_type)
all_combined_table <- stats::ftable(all_combined)
all_combined_df <- as.data.frame(all_combined_table)
names(all_combined_df) <- c("Year", "OoI", "Method", "Data", "Software_type", "Freq")
methods <- table(mounds_lit_filt$Year, mounds_lit_filt$Method)
methods_table <- stats::ftable(methods)
methods_table_df <- as.data.frame(methods_table)
names(methods_table_df) <- c("Year", "Method", "Freq")
names(methods_table_df)
GKNB <- methods_table_df %>% filter(Method=="Geometric knowledge")
ggplot(GKNB, aes(y=Freq, x=Year)) +
  geom_bar(stat="identity", fill ="#70795E") +
  ylim(0, 7) +
  ggtitle("Use of Geometric knowledge-based methods between 2006 & 2021") +
  theme_light() +
  theme(legend.position="none") +
  xlab("Year") +
  ylab("Number of studies")

Geographical Object-based Image Analysis (GeOBIA) dates back to the late 1970’s (@blaschkeGeographicObjectBasedImage2014a), but it was only around 2000 that Object-based Images Analysis became widely used in Remote Sensing studies due to the availability of high resolution Satellite data. This induced a paradigm shift not only in Remote Sensing but generally in GI Science, hence its new name: GeOBIA (@hayGeographicObjectBasedImage2008a). GeOBIA operates in two steps: images are divided into small segments, which are defined by the homogeneity of specific morphometric (shape, size, orientation), spectral, textural, context and neighborhood parameters (@hayGeographicObjectBasedImage2008a) and are then grouped together to meaningful homogeneous object candidates (the segmentation step), which are then classified by specific extracted object criteria in question (the feature extraction and classification step, @blaschkeGeographicObjectBasedImage2014a, 186, @chengSurveyObjectDetection2016a, 15, @hossainSegmentationObjectBasedImage2019, 115) or filtered by a rule-set. Various types of segmentation methods exist (as also their categorization: @blaschkeObjectBasedImage2010, @blaschkeGeographicObjectBasedImage2014a, @kumarREVIEWIMAGESEGMENTATION2014). Lately @hossainSegmentationObjectBasedImage2019 investigated the different segmentation methods from a Remote Sensing point of view, but here only the two main Segmentation methods used in Automated Archaeological Remote Sensing studies are described: Watershed Segmentation (or Transformation) (@niculitaGeomorphometricMethodsBurial2020b and @niculitaBurialMoundDetection2020a) and Multiresolution Segmentation (@kramerArchaeologicalReactionRemote2015, @sevaraPixelObjectComparison2016a, @freelandAutomatedFeatureExtraction2016b, @davisAutomatedMoundDetection2018a and @davisComparisonAutomatedObject2019b, @meyerAutomatedDetectionField2019a, @sarasanMappingBurialMounds2020a), the first being an Edge-Based Segmentation method, the latter a Region-Based Segmentation method. For an in-depth analysis of the methods see @hossainSegmentationObjectBasedImage2019.

Edge-Based Segmentation techniques are ‘top-down’ methods: first they locate edges in the image and then use contouring algorithms to close them. Edges are regarded as boundaries between objects where pixel properties are abruptly changing (@hossainSegmentationObjectBasedImage2019, 117). ‘Watershed Segmentation’ (implemented in open source software e.g. in SAGA or in the ForestTools package in R) simulates real-life flooding. It first transforms the image into a gradient (grey-scale), then identifies objects with clear segment boundaries, only to then create (fill) objects, for which it is also called a Region-Growing algorithm (@hossainSegmentationObjectBasedImage2019, 117, Table 1; Figure 7).

knitr::include_graphics('C:/Users/kelto/Documents/iSEGMound/analysis/thesis/figures/Figure_7.png')

Region-Based Segmentation techniques start from the complete opposite: they begin with an initial segmentation of the whole image (thus also called ‘bottom-up methods’) based on a specific rule-set. Regions are generated in two completely different ways: either by splitting the image into homogeneous regions based on inhomogeneity (region-splitting and then merging) or by region-growing from a so-called seed-region based on homogeneity (region-growing and then merging, @hossainSegmentationObjectBasedImage2019). ‘Multiresolution Segmentation’ (@baatzMultiresolutionSegmentationOptimization2000a, for which mainly eCognition is used in Automated Archaeological Remote Sensing) is a region-merging hierarchical segmentation method which starts with one pixel (seed) and applies pairwise merging of segments to build up hierarchical levels. The merging – or clustering (using local rule sets) is repeated (on multiple levels), until an object is recognized (@hossainSegmentationObjectBasedImage2019, 119, @rofflerStepStepDeveloping2020, 33-34; Figure 8).

knitr::include_graphics('C:/Users/kelto/Documents/iSEGMound/analysis/thesis/figures/Figure_8.png')

With regard to (Automated) Archaeological Remote Sensing, GeOBIA is sometimes also called archaeOBIA (@lamotteArcheOBIAConceptAnalyse2016a), which points to the fact, that (Automated) Archaeological Remote Sensing is in need of a completely different semantic ontology and rule sets than which is needed for customary GeOBIA methods used in Remote Sensing. Still recently the challenge in Remote Sensing has been to define segmentation parameters/rule sets which can be transferred to other images (@chengSurveyObjectDetection2016a, 16) – something archeOBIA is also struggling with (see Table 1; note the different variables used in the different studies). Burial mounds, monumental earthworks and mound shell rings have been investigated using GeOBIA methods nine times since 2015 (Figure 9). Apart from one case (@niculitaGeomorphometricMethodsBurial2020b: Watershed Segmentation, carried out with SAGA in R), Multiresolution Segmentation was applied (Supplement 2), almost exclusively using eCognition, (former Definiens), a software developed for and with the evolution of GeOBIA and thus is generally considered “the software” for GeOBIA (@blaschkeObjectBasedImage2010, Fig 4, @hossainSegmentationObjectBasedImage2019, Fig 1.), which is clearly reflected in this survey.

library(dplyr)
library(ggplot2)
mounds_lit <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds.csv'))
#read as tibble
mounds_lit <- dplyr::as_tibble(mounds_lit)
mounds_lit_filt <- mounds_lit[, c(3:7)]
all_combined <- table(mounds_lit_filt$Year, mounds_lit_filt$OoI, mounds_lit_filt$Method,
                      mounds_lit_filt$Data, mounds_lit_filt$Software_type)
all_combined_table <- stats::ftable(all_combined)
all_combined_df <- as.data.frame(all_combined_table)
names(all_combined_df) <- c("Year", "OoI", "Method", "Data", "Software_type", "Freq")
methods <- table(mounds_lit_filt$Year, mounds_lit_filt$Method)
methods_table <- stats::ftable(methods)
methods_table_df <- as.data.frame(methods_table)
names(methods_table_df) <- c("Year", "Method", "Freq")
names(methods_table_df)
GeOBIA <-methods_table_df %>% filter(Method=="GeOBIA")
ggplot(GeOBIA, aes(y=Freq, x=Year)) +
  geom_bar(stat="identity", fill ="#A9AF86") +
  ylim(0, 5) +
  ggtitle("Use of GeOBIA methods between 2006 & 2021") +
  theme_light() +
  theme(legend.position="none") +
  xlab("Year") +
  ylab("Number of studies")

Machine Learning-based methods are considered a subfield of Artificial Intelligence. Machine learning automates statistical methods to learn from input data either supervised, unsupervised or semi-supervised. Automated Archaeological Remote Sensing mainly utilizes supervised methods. Pixel-based Image Analysis, an image classification method, has been developed in the early 1970s for the digital analysis of Landsat Multispectral Scanner Systems (@phiriDevelopmentsLandsatLand2017 2017, 9) and is (still) widely distributed in Remote Sensing research, especially in land-use and land-cover classification. In contrast to GeOBIA, Pixel-based Image Analysis approaches an image on pixel basis, which are classified into different categories based on the information they carry (usually one variable). Given that the second step of GeOBIA can be a classification of the segmented image using variables of the image-objects best describing the Objects of Interest, GeOBIA is sometimes also seen as a form of Machine Learning (@davisObjectbasedImageAnalysis2019, 1). Random forest classifiers are supervised learning algorithms which consist of an ensemble of decision trees. Each (unrelated) decision tree is trained using a random subset of the training data. Each of these trees will give a prediction for a data point. Then, the prediction of all decision trees is averaged by a majority vote to a final prediction (Figure 9). The independence of the different decision trees increases the accuracy of the prediction and also eliminates problems that can be caused by outliers in the data set and works also well with small data sets, because of the facts just explained. These effects can be enhanced by resampling techniques and parameter tuning (@kuhnAppliedPredictiveModeling2013).

knitr::include_graphics('C:/Users/kelto/Documents/iSEGMound/analysis/thesis/figures/Figure_10.png')

Regarding the automated detection of burial mounds and similar structures, Pixel-based classification was the first method used in 2006 and has been more or less revisited since (Figure 11). When looking closely at the specific algorithms used, it is not surprising why the Random Forest algorithm was exploited in most Pixel-based Image Analysis studies (@menzeDetectionAncientSettlement2006a, @Menze2007CLASSIFICATIONOM, @Menze2007VirtualSO, @menzeMappingPatternsLongterm2012a, and @menzeMultiTemporalClassificationMultiSpectral2013b, @kramerArchaeologicalReactionRemote2015, @guyotDetectingNeolithicBurial2018a, @orengoAutomatedDetectionArchaeological2020a, @niculitaGeomorphometricMethodsBurial2020b, and @davisDeepLearningReveals2021a) and Mahalanobis Distance (@trierAutomaticDetectionMound2015a and @sevaraPixelObjectComparison2016a) and Support Vector Machine Classifiers (@caspariConvolutionalNeuralNetworks2019a) in less.

library(dplyr)
library(ggplot2)
mounds_lit <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds.csv'))
#read as tibble
mounds_lit <- dplyr::as_tibble(mounds_lit)
mounds_lit_filt <- mounds_lit[, c(3:7)]
all_combined <- table(mounds_lit_filt$Year, mounds_lit_filt$OoI, mounds_lit_filt$Method,
                      mounds_lit_filt$Data, mounds_lit_filt$Software_type)
all_combined_table <- stats::ftable(all_combined)
all_combined_df <- as.data.frame(all_combined_table)
names(all_combined_df) <- c("Year", "OoI", "Method", "Data", "Software_type", "Freq")
methods <- table(mounds_lit_filt$Year, mounds_lit_filt$Method)
methods_table <- stats::ftable(methods)
methods_table_df <- as.data.frame(methods_table)
names(methods_table_df) <- c("Year", "Method", "Freq")
names(methods_table_df)
PBIA <-methods_table_df %>% filter(Method=="PBIA")
ggplot(PBIA, aes(y=Freq, x=Year)) +
  geom_bar(stat="identity", fill ="#BCC0AF") +
  ylim(0, 5) +
  ggtitle("Use of PBIA methods between 2006 & 2021") +
  theme_light() +
  theme(legend.position="none") +
  xlab("Year") +
  ylab("Number of studies")

It was already suggested that with the development of remote sensing sensors and resulting new, higher resolution datasets the need for new analysis methods is constantly stimulated. Starting with pixel-wise analysis and followed by the object-level addressing of remote sensing imagery, lately an even bigger semantic step was taken: the analysis on the scene-level, which can be seen as the next paradigm shift (@chengSurveyObjectDetection2016a, @chengRemoteSensingImage2020a, 2, Fig. 2). The complex semantic structures of very high resolution images are addressed by Deep Learning, which is also a sub-field of Machine Learning. In contrast to Pixel- and Object-based Image Analysis, Deep Learning Algorithms consist of multiple stacked hierarchical layers (network architectures) which can handle complexity and abstraction. In the case of the identification of burial mounds and mound like structures Deep Learning is only in use since 2019 (Figure 12).

library(dplyr)
library(ggplot2)
mounds_lit <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds.csv'))
#read as tibble
mounds_lit <- dplyr::as_tibble(mounds_lit)
mounds_lit_filt <- mounds_lit[, c(3:7)]
all_combined <- table(mounds_lit_filt$Year, mounds_lit_filt$OoI, mounds_lit_filt$Method,
                      mounds_lit_filt$Data, mounds_lit_filt$Software_type)
all_combined_table <- stats::ftable(all_combined)
all_combined_df <- as.data.frame(all_combined_table)
names(all_combined_df) <- c("Year", "OoI", "Method", "Data", "Software_type", "Freq")
methods <- table(mounds_lit_filt$Year, mounds_lit_filt$Method)
methods_table <- stats::ftable(methods)
methods_table_df <- as.data.frame(methods_table)
names(methods_table_df) <- c("Year", "Method", "Freq")
names(methods_table_df)
DL <-methods_table_df %>% filter(Method=="Deep Learning")
ggplot(DL, aes(y=Freq, x=Year)) +
  geom_bar(stat="identity", fill ="#43595E") +
  ylim(0, 5) +
  ggtitle("Use of Deep Learning methods between 2006 & 2021") +
  theme_light() +
  theme(legend.position="none") +
  xlab("Year") +
  ylab("Number of studies")

To summarize the use of automated analysis methods to detect burial mounds and mound-like structures (Figure 13), it can be established that the first method used was Pixel-based Image analysis (@menzeDetectionAncientSettlement2006a), followed by Template Matching (@Boer2010UsingPR). Geometric knowledge-based analysis was used as an autonomous method only by @rileyAutomatedDetectionPrehistoric2009a, @freelandAutomatedFeatureExtraction2016b, @romLandSeaAirborne2020b, but as already concluded it is more often than not incorporated in workflows as data preparation method (e.g. @cerrillo-cuencaApproachAutomaticSurveying2017a). GeOBIA was first used in 2015 and remained a very effective analysis method, until recently when Deep Learning became widespread (including Semantic and Instance Segmentation).

library(dplyr)
library(ggplot2)
mounds_lit <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds.csv'))
#read as tibble
mounds_lit <- dplyr::as_tibble(mounds_lit)
mounds_lit_filt <- mounds_lit[, c(3:7)]
all_combined <- table(mounds_lit_filt$Year, mounds_lit_filt$OoI, mounds_lit_filt$Method,
                      mounds_lit_filt$Data, mounds_lit_filt$Software_type)
all_combined_table <- stats::ftable(all_combined)
all_combined_df <- as.data.frame(all_combined_table)
names(all_combined_df) <- c("Year", "OoI", "Method", "Data", "Software_type", "Freq")
ggplot(all_combined_df, aes(fill=Method, y=Freq, x=Year)) +
  geom_bar(position="stack", stat="identity") +
  ylim(0, 15) +
  scale_fill_manual(values=morris:::acanthus_palette) +
  ggtitle(expression(atop("Methods used in Automated Archaeological Remote Sensing", atop("between 2006 and 2021", "")))) +
  theme_light() +
  xlab("Year") +
  ylab("Number of studies")

Several studies compare different analysis methods or combine them. @freelandAutomatedFeatureExtraction2016b and @davisComparisonAutomatedObject2019b compare Geometric-knowledge-based analysis (iMound/Inverse Stochastic Depression Analysis) to GeOBIA, with success (this is also reflected in the fact that the original algorithm iMound by @freelandAutomatedFeatureExtraction2016b was reused by @davisComparisonAutomatedObject2019b and @romLandSeaAirborne2020b). Template Matching has been compared to GeOBIA (@kramerArchaeologicalReactionRemote2015, @davisComparisonAutomatedObject2019b: including Geometric knowledge-based method) and Pixel-based Image Analysis to GeOBIA (@sevaraPixelObjectComparison2016a) and to Deep Learning (@caspariConvolutionalNeuralNetworks2019a).

Although LiDAR data has been available for some time (see @Boer2010UsingPR, @rileyAutomatedDetectionPrehistoric2009a), it was only after 2010 that it made its way into everyday archaeological research, including various visualization methods (as Geometric knowledge-based analysis and data preparation method), making LiDAR visualisations a self-evident step for any archaeological project using ALS data (see also @kokaljAirborneLaserScanning2017a). It is mainly from 2015, when ALS data started to dominate and revolutionize Automated Analysis methods in Archaeological Remote Sensing and provoking new approaches, such as GeOBIA (Figure 14). Since the diffusion of LiDAR data, Satellite Imagery is mainly used in large-scale studies (@caspariConvolutionalNeuralNetworks2019a and @orengoAutomatedDetectionArchaeological2020a). Studies which require high resolution data and utilize LiDAR quickly shifted to Deep Learning solutions. Simultaneously also UAV data finds its way into the general data repertoire of Automated Archaeological Analysis (@sarasanMappingBurialMounds2020a).

library(dplyr)
library(ggplot2)
mounds_lit_2 <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds_short.csv'))
#read as tibble
mounds_lit_2 <- dplyr::as_tibble(mounds_lit_2)
mounds_lit_2_filt <- mounds_lit_2[, c(3:7)]
all_combined_2 <- table(mounds_lit_2_filt$Year, mounds_lit_2_filt$OoI,
                      mounds_lit_2_filt$Data, mounds_lit_2_filt$Software_type,
                      mounds_lit_2_filt$access)
methods_table_2 <- stats::ftable(all_combined_2)
methods_table_df2 <- as.data.frame(methods_table_2)
names(methods_table_df2) <- c("Year", "OoI", "Data", "Software_type", "access", "Freq")
ggplot(methods_table_df2, aes(fill=Data, y=Freq, x=Year)) +
  geom_bar(position="stack", stat="identity") +
  ylim(0, 7) +
  scale_fill_manual(values=morris:::peacock_palette) +
  ggtitle(expression(atop("Data types used in Automated Archaeological Remote Sensing", atop("between 2006 and 2021", "")))) +
  theme_light() +
  xlab("Year") +
  ylab("Number of studies")

When taking the points ‘Software’ and ‘Access’ into account, it has to be expressed that it is only a recent phenomenon that software and computation details have to be disclosed when publishing a study. Even though lately many journals require data and code availability statements, only a few studies provide openly accessible code and data: @orengoAutomatedDetectionArchaeological2020a, and @niculitaGeomorphometricMethodsBurial2020b. In the first case the data and the code are available using Google Earth Engine. In the latter case, the regulations of the local Cultural Heritage Management authorities require a signed form through the author of the study to be able to use the DEM on which the study was based on - nonetheless it can be accessed. Although with restrictions. Inspecting the information about the methodology details of the studies, nine cases have been identified (Figure 15): from not available (n/a) to workflow & code & data, many constellations can be observed. In many cases the workflow was published, in some cases only a flowchart. In this thesis as a workflow the clear explanation of the methodology in a chart form is defined (with a big probability of reproducibility if the observer knew the exact tools and software used or if those were made clear). As a flowchart on the other hand, a part of a workflow or a very schematized workflow was identified where only the main steps were arranged in chart form, thus making it impossible to trace back the specific steps and to reproduce the workflow or any part of it. Thus the only reproducible study, which published a really detailed workflow, the code and made the data set - although restricted by certain rules - available is Niculiță 2020.

library(dplyr)
library(ggplot2)
mounds_lit_2 <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds_short.csv'))
#read as tibble
mounds_lit_2 <- dplyr::as_tibble(mounds_lit_2)
mounds_lit_2_filt <- mounds_lit_2[, c(3:7)]
all_combined_2 <- table(mounds_lit_2_filt$Year, mounds_lit_2_filt$OoI,
                      mounds_lit_2_filt$Data, mounds_lit_2_filt$Software_type,
                      mounds_lit_2_filt$access)
methods_table_2 <- stats::ftable(all_combined_2)
methods_table_df2 <- as.data.frame(methods_table_2)
names(methods_table_df2) <- c("Year", "OoI", "Data", "Software_type", "access", "Freq")
ggplot(methods_table_df2, aes(fill=access, y=Freq, x=Year)) +
  geom_bar(position="stack", stat="identity") +
  ylim(0, 7) +
  scale_fill_manual(values=morris:::strawberry_palette ) +
  ggtitle(expression(atop("Access to any parts of studies investigated", 
                     atop("between 2006 and 2021", "")))) +
  theme_light() +
  xlab("Year") +
  ylab("Number of studies")

As already noted, eCognition is considered “the” software for GeOBIA (@hossainSegmentationObjectBasedImage2019, 122) thus it is evident that many will choose this simpler solution. At the same time, it must be emphasized that the first analyses were carried out in R (@menzeDetectionAncientSettlement2006a, @Menze2007VirtualSO, @Menze2007CLASSIFICATIONOM, @menzeMappingPatternsLongterm2012a, and @menzeMultiTemporalClassificationMultiSpectral2013b), i.e. with open source software. Using proprietary software not only marginalizes researchers and institutions who/which can’t afford rather expensive software, but it is also hard to comprehend the exact algorithm behind the software, and to understand to be able to apply it in another domain. It is also important to understand why certain tools or algorithms worked or did not work? This can only be done by creating reproducible workflows with open source software. In 2019 and 2020 more than 60% of the studies were conducted with FOSS (Free and Open Source Software) software. This number is only increasing (Figure 16).

library(dplyr)
library(ggplot2)
mounds_lit_2 <- read.csv(here::here('analysis/data/literature_analysis/burial_mounds_short.csv'))
#read as tibble
mounds_lit_2 <- dplyr::as_tibble(mounds_lit_2)
mounds_lit_2_filt <- mounds_lit_2[, c(3:7)]
all_combined_2 <- table(mounds_lit_2_filt$Year, mounds_lit_2_filt$OoI,
                      mounds_lit_2_filt$Data, mounds_lit_2_filt$Software_type,
                      mounds_lit_2_filt$access)
methods_table_2 <- stats::ftable(all_combined_2)
methods_table_df2 <- as.data.frame(methods_table_2)
names(methods_table_df2) <- c("Year", "OoI", "Data", "Software_type", "access", "Freq")
ggplot(methods_table_df2, aes(fill=Software_type, y=Freq, x=Year)) +
  geom_bar(position="stack", stat="identity") +
  ylim(0, 7) +
  scale_fill_manual(values=morris:::flowers_palette) +
  ggtitle(expression(atop("Software types of studies investigated", 
                     atop("between 2006 and 2021", "")))) +
  theme_light() +
  xlab("Year") +
  ylab("Number of studies")

This survey served the purpose to elaborate the development of Automated Analysis in Archaeological Remote Sensing, mainly with regard to the applied methods, the used software and access to information about the workflow. The aim was to show how little reproducible research has been published with regard to automated analyses and how much there is to be done in the future. Also, automated analysis in Archaeological Remote Sensing is being carried out on different scales with different algorithms, specialized to different research questions in mainly research contexts. This is also a reason why many critical voices see automated analysis methods as pastimes and not really something operational (see many AARGnews Editorials). To give way for the next chapter we can conclude that the results of automated analyses are compelling, but many studies start basically from the beginning, because when research is published without accompanying software, workflow, data and documented steps, it is a challenge and time consuming to understand, verify, expand and eventually surpass that research (@marwickComputationalReproducibilityArchaeological2017b, 425). This is one reason why many research projects start from the beginning, with a new idea instead of building on the previous knowledge. Another point is that automated analysis methods in Archaeological Remote Sensing are still in their infancy (@opitzRecentTrendsLongstanding2018 2018, 30) and best practices have not yet been established. Nonetheless it has to be stated that automated analysis itself is not a goal, but a method, a means to an end to further research and thus reproducible or even replicable best practices would help shift the focus on further development of methods than on continuous recommencement.

Chapter 2 discussed the use of automated analysis methods in Archaeological Remote Sensing for the detection of burial mounds and structurally similar archaeological objects. It was demonstrated that only 2 out of 31 studies have disclosed workflow, code and data and are thus reproducible (see chapter 3 for explanation). As a reproducible study in the R environment @niculitaGeomorphometricMethodsBurial2020b can be mentioned as the best example. It was also noted that even without the availability of code, there are workflows which are illustrated and also explained clearly enough that they can be reproduced, which on the other hand can take some time but is still possible to do when having enough experience with spatial tools. On this basis it was decided that the Geometric knowledge-based workflow ‘iMound’ (established by @freelandAutomatedFeatureExtraction2016b and reused by @davisComparisonAutomatedObject2019b and @romLandSeaAirborne2020b) is going to be utilized in R to detect burial mounds in LiDAR data in this Master’s thesis to create a reproducible workflow. During the implementation of the workflow it became clear that on the basis of the geomorphometry of the Train Area it was necessary to include another method to be able to effectively detect burial mounds (this is discussed in depth in Chapter 4). Thus, in addition to ‘iMound’ also parts of the reproducible workflow of @niculitaGeomorphometricMethodsBurial2020b, a GeOBIA method was used in this Master’s thesis.

keltoskytoi/iSEGMound documentation built on Dec. 21, 2021, 5:24 a.m.