You are on page 1of 14

Journal of Archaeological Science: Reports 5 (2016) 485–498

Contents lists available at ScienceDirect

Journal of Archaeological Science: Reports

journal homepage: www.elsevier.com/locate/jasrep

Pixel versus object — A comparison of strategies for the semi-automated


mapping of archaeological features using airborne laser scanning data
Christopher Sevara a,c,⁎, Michael Pregesbauer b, Michael Doneus b,c, Geert Verhoeven b,d, Immo Trinks b
a
Initiative College for Archaeological Prospection, University of Vienna, Franz-Klein-Gasse 1/III, 1190, Vienna, Austria
b
Ludwig Boltzmann Institute for Archaeological Prospection and Virtual Archaeology, Hohe Warte 38, 1190, Vienna, Austria
c
Department of Prehistoric and Historical Archaeology, University of Vienna, Franz-Klein-Gasse 1/III, 1190, Vienna, Austria
d
Vienna Institute for Archaeological Science, University of Vienna, Franz-Klein-Gasse 1/III, 1909, Vienna, Austria

a r t i c l e i n f o a b s t r a c t

Article history: Semi-automated approaches to archaeological feature detection can be invaluable aids to the investigation of
Received 20 October 2015 high resolution, large area archaeological prospection datasets. In order to obtain stable and reliable classification
Received in revised form 17 December 2015 results, however, the application of pre-processing steps to digital terrain data is needed. This study examines
Accepted 31 December 2015
semi-automated approaches to identification of archaeological features through a comparison of pixel-based
Available online 13 January 2016
and object-oriented data classification methods for archaeological feature detection in visualizations derived
Keywords:
from high-resolution airborne laser scanning data. In doing so, openness is presented as a suitable visualization
Classification for feature detection due to its illumination-invariant representation of convexity and concavity in terrain data.
Segmentation The methodology of both pixel-based and object-oriented data classification approaches is described and applied
OBIA to two datasets recorded over two archeological case study areas in Sweden and Austria. The diverse nature of the
ALS two datasets makes them ideal with regard to determining the robustness of the approaches discussed here. The
Birka obtained results are exported to a GIS environment and compared with manual visual interpretations and ana-
Kreuttal lyzed in terms of their accuracy. Therefore, this paper presents both a discussion regarding the merits of pixel-
Openness
and object-based semi-automated classification strategies with regard to archaeological prospection data as
well as practical examples of their implementation and results.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction Lasaponara et al., 2011). However, the density of archaeologically rele-


vant information produced from ALS data capture also renders its man-
Airborne laser scanning (ALS) is an established technology for dense ual analysis and interpretation a time-consuming task. Computer driven
and highly accurate measurements of the Earth's surface and thereby image analysis offers a way to rapidly extract archaeological informa-
for the mapping of archaeological sites and landscapes (e.g. tion from such very high resolution (VHR) datasets, and can even be
Ackermann, 1999; Challis et al., 2008; Doneus and Briese, 2006, 2011; used to detect objects of interest that are not visible to the naked eye.
Lasaponara et al., 2011; Sittler, 2004). For archaeological purposes, ALS Thus, such processes can act as a compliment to archaeological data in-
has so far been utilized to generate detailed digital terrain models terpretation, expediting and improving the image analysis workflow.
(DTMs) in order to depict subtle features of prehistoric and historic The wider archeological community is beginning to recognize the
structures in the topography, features that are sometimes hardly recog- potential benefit in employing computer-aided analysis techniques,
nizable for an observer located on the ground. ALS has considerable which can help in rapidly classifying large area, high-resolution archae-
methodological and economic advantages compared with traditional ological prospection data (Bennett et al., 2014; Cowley, 2012). While
terrestrial surveying for the time- and cost-efficient generation of the merits and extent to which such approaches are applicable continue
large-scale, high-resolution terrain models (see Opitz and Cowley, to be debated (e.g. Casana, 2014), the reality is that we now have the ca-
2013 for recent examples). This is especially true for forested areas, pability to produce so much data of such high spatial, spectral and tem-
whereby making use of latest full-waveform (FWF) laser scanner tech- poral quality that it is becoming difficult to interpret it all manually.
nology it has even become possible to conduct archeological reconnais- Therefore, instead of rejecting the notion of computer-aided detection
sance in areas that are covered with dense forest and low understory of archaeological information, the goal should be to find ways in
(Doneus and Briese, 2006; Hesse, 2014; Lasaponara and Masini, 2009; which we can use these techniques to our advantage, as informed by ex-
pert knowledge about the archaeological record. In this paper, we do so
⁎ Corresponding author at: Initiative College for Archaeological Prospection, University
through the examination of two approaches to semi-automated extrac-
of Vienna, Franz-Klein-Gasse 1/III, 1190, Vienna, Austria. tion of information that are commonly used today in the field of image
E-mail address: christopher.sevara@univie.ac.at (C. Sevara). analysis: (1) pixel-based classification and (2) object-oriented image

http://dx.doi.org/10.1016/j.jasrep.2015.12.023
2352-409X/© 2016 Elsevier Ltd. All rights reserved.
486 C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498

analysis (Baatz et al., 2008; Blaschke et al., 2000; Blaschke, 2010; Tso seem to be necessary. Here, a fundamental distinction can be made be-
and Mather, 2009). Pregesbauer (2013) demonstrated the applicability tween the use of feature detection routines for the targeted investiga-
of these approaches as applied to the semi-automated identification of tion and analysis of specific feature classes and their application in the
burial mounds on the island of Birka, Sweden. Sevara and Pregesbauer detection of more diverse archaeological content in a dataset for
(2014) developed an object-oriented model for the identification of ar- which potential feature types may be unknown.
chaeological features in ALS datasets in the Kreuttal region of Lower Although the first results of the abovementioned ALS feature detec-
Austria using a relational approach and linking feature detection to tion tests seem to be quite promising, an automated workflow for ar-
landform classification. chaeological interpretation still appears to be quite far away (see also
This study will further investigate the potential of these methods for Cowley, 2012). While a fully automatic approach is neither sought
the extraction of archaeological information contained in ALS-derived after nor advocated by the authors, the recent work presented above
DTMs and related visualizations. Therefore, the first part of this shows that advances in image analysis have advanced the field to a
study discusses openness-derived visualizations and their role in ar- point where semi-automated feature detection based on user-defined
chaeological feature detection and compares the principles of both the variables can reduce the time it takes to interpret and classify complex
pixel-based classification and the object-oriented image analysis ap- archaeological datasets.
proach. In the second part, both methods are applied first to an ALS
dataset derived from the case study area Birka-Hovgården in Sweden, 2.2. Data visualization and openness
a Viking Age settlement and Scandinavia's largest field of late Iron Age
burial mounds. Subsequently, each approach is applied to a second Spectral images with multiple bands are the common foundation for
ALS dataset covering the Kreuttal area in Lower Austria, an archaeolog- image analysis. In contrast to multi-band spectral images, ALS-derived
ical case study area with a range of feature types present in relief. In this data used in this study consist of only one layer, indicating height
way, the applicability of each approach can be evaluated in both a highly values, which can be used for classification. In this case, the feature
homogenous and a highly heterogeneous archaeological landscape. Ex- space for detecting and classifying archaeological objects can be either
tracted features are then exported to a GIS environment for further the filtered ALS point cloud, or the interpolated raster DTM created
analysis and comparison with manual interpretation, which serves as from that cloud, where any classification approach applied has to
a reference in order to compare the capabilities of the automatic classi- work with the relative spatial variance of the three dimensional points.
fication algorithm for archaeological purposes, and to estimate the In addition to the height values of the DTM, a visualization method is
achieved accuracy of both classification approaches. needed for the subsequent semi-automatic classification procedure.
This permits the clear, localized differentiation of relief features, regard-
2. ALS data, visualization and semi-automated feature detection less of the illumination source. There exists a broad range of possibilities
to derive models from the relative position of points in 3D space (Hengl
2.1. ALS data and archeological topography and Reuter, 2009), for example slope or local relief models (Hesse,
2010), the use of the sky-view factor (SVF) (Kokalj et al., 2011), positive
In order for archaeological structures to be visible in a DTM they and negative openness (Yokoyama et al., 2002; Doneus, 2013) or a com-
have to have survived in relief and the terrain model must be of suffi- bination of those methods. Using these models, micro-topography and
cient spatial resolution to represent the features. Preservation is often therefore archaeological structures contained therein can be enhanced,
best in constantly vegetated areas such as pastures and forests, where thereby increasing the potential for a successful classification.
the vegetation, to a certain degree, protects the surface from erosion. While there exists an abundant amount of literature on visualization
Here, archaeological structures are often still visible as variations of archaeological DTMs (e.g. Bewley et al., 2005; Bennett et al., 2012;
in micro-topography. These areas can therefore be regarded as archae- Devereux et al., 2008; Doneus and Briese, 2006; Humme et al., 2006;
ological archives providing abundant information about present re- Hesse, 2010; Kokalj et al., 2011; Opitz and Cowley (2013)), almost all
mains of past land uses that had an impact on the ground surface studies have concentrated on the advantage of DTMs for archaeological
(Kenzler and Lambers, 2015:73). Depending on the site and size of the feature selection. Mapping and delineation of features have been of less
surveyed area, thousands of features may be visible (Sittler, 2004; concern. In order for automated classification methods to be successful,
Doneus and Briese, 2011; Bofinger and Hesse, 2011), which have to be visualization techniques that help clearly indicate outline and geometry
mapped during archaeological data interpretation (Doneus and of individual features are necessary. Recently, openness (Yokoyama
Kühtreiber, 2013). et al., 2002) has been advocated as a suitable technique for this task
Interpretative mapping of archaeological structures contained in ALS (Doneus, 2013; Zakšek et al., 2011). The openness algorithm calculates
data can be a time-consuming process, especially when thousands of the mean value of the zenith angle for the determined horizons of a
features have to be drawn manually. Consequently, approaches for minimum of eight directions (N, NW, W, SW, S, SE, E, NE) (Yokoyama
semi-automatic detection or interpretation are sought-after alternatives et al., 2002:258) for a specific location in the elevation model. If the ze-
that aim to expedite the identification and interpretation process. While nith angle is used, a positive openness value is calculated (Fig. 1, red
several corresponding tests have been conducted for satellite and aerial lines). Accordingly, using the nadir, negative openness is the result
imagery (e.g. Casana, 2014; De Laet et al., 2007a; De Laet et al., 2007b; (Fig. 1, white lines). The search radius is controlled by modifying its
Figorito and Tarantino, 2014; Harrower et al., 2013; Jahjah and size, a manually determined parameter that is linked to the spatial res-
Ulivieri, 2010; Lambers and Zingman, 2013; Lasaponara and Masini, olution of the base DTM.
2006), examples for automated detection of archaeological features in As positive and negative openness can provide information on con-
ALS-derived terrain models are still rare (Boer, 2005; Pregesbauer, vexity and concavity of archaeological features (Doneus, 2013:6436),
2013; Schneider et al., 2014; Sevara, 2013; Sevara and Pregesbauer, these geometric characteristics can be used as an aid to computer
2014; Trier and Pilø, 2012; Trier et al., 2015). A number of the ap- aided classification of various feature types visible in relief. Though
proaches quoted above employ a technique in which geometrical tem- openness is similar to SVF (Zakšek et al., 2011:404), it has two main dif-
plates are used for the identification of geometrically simple objects ferences: (1) the process calculates an additional “negative” version
(e.g. mounds and pits) in ALS-derived DTMs. Approaches such as tem- using nadir angles, and (2) while the maximum angle when using the
plate mapping can be effective but may run into limitations when fea- sky-view factor can be a nadir of 90°, openness also includes angle-
tures of interest in VHR datasets do not correspond to pre-defined values that are larger (Doneus, 2013:6435; Zakšek et al., 2011:403). As
templates. Furthermore, when mining datasets for multiple feature a result, openness does not consider slopes, i.e. a location in an openness
types whose parameters may be unknown, more flexible approaches map will display the same value regardless of whether it was
C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498 487

Fig. 1. Positive openness (red lines) and negative openness (white lines). Numbers at angles indicate the azimuth of the angle.
After Yokoyama et al. (2002:259).

determined on an even slope or on a horizontal plane, while the sky- height, positive openness, sky-view factor, slope and the like). Thus,
view factors of these two locations would differ (because it is per defini- the aim of the pixel-based classification is to establish a per-pixel rela-
tion limited to the celestial hemisphere). Openness is therefore a trend tionship with a certain class category based on the attribute information
removal tool, and as a consequence, openness values within the bound- of itself and its spatial neighbourhood (Alt, 1990). The classes have ei-
ary of individual features are more homogeneous, which is an advan- ther been predefined by the investigator in the form of a supervised
tage for their classification. This differentiates openness from other classification approach, or identified by the software in an unsupervised
visualization calculations such as SVF that may, however, be more useful approach based on grouping the spectral properties of the respective
for visual interpretation. For these reasons, openness has been sug- bands of the entire dataset (Cheung, 2005).
gested as being particularly suitable for use in semi-automated feature When working with archaeological data, and especially when work-
detection applications (Doneus, 2013:6439; Zakšek et al., 2011:404). ing with DTMs containing micro-topographic structures in varying
In this case study, both terrain data and the derived models of posi- states of erosion, archaeological features generally express only a slight
tive and negative openness are stacked together into a multi-layered per-pixel variation compared with their surroundings. Therefore, only
dataset, taking a certain data redundancy into account. In that way, a the supervised pixel-based approach seems to be successful for archae-
multi-band dataset is approximated, functioning as a basis for the appli- ological feature detection, whereby the investigator uses known sample
cation of feature detection and description algorithms. The resulting areas based on prior knowledge to “train” the classifier to recognize de-
stacked dataset has been used to evaluate the potential of a pixel- cision boundaries in the feature space. Earlier attempts have shown that
based as well as of an object-oriented approach for the semi- unsupervised classification, in which the software-based classifier
automatic classification of the archaeological resources in two case “learns” from the input data, is only successful on very small datasets
study areas. containing a limited number of distinct features (Townshend et al.,
2000).
2.3. Semi-automatic classification approaches In the case of supervised classification, the quality of the training
data is the key factor for its success. Two factors seem to be most rele-
Classification techniques are commonly used in remote sensing and vant for many commonly applied classifiers: (1) the only objects that
can consist of a set of measurements of different features in the so- can be automatically classified are ones represented in the training sam-
called ‘feature space’. The feature space can be regarded as a k- ples; (2) the data included within the training samples should be as ho-
dimensional vector space, where individual features or a combination mogeneous as possible (Townshend 1981). All of the supervised
thereof are describing objects (e.g. a pixel as the most generic form, or statistical classifiers which have been developed over the years employ
a polygon shape as a representation of a certain entity). Descriptors three main steps in their process chain (Tso and Mather, 2009):
can be spectral properties such as reflectance, radiance, or transmit-
• Defining and selecting the desired target classes and collecting suffi-
tance, geometrical, and geographical (Alt, 1990). During the data classi-
cient training samples.
fication, a relationship has to be established between these objects and
• Using the collected training samples to estimate the statistical param-
pre-defined class categories. A comparative study on the difference be-
eters regarding the selected classification method.
tween pixel-based and object-oriented classification can be found in the
• Applying an appropriate decision rule to assign the pixel to the corre-
work of Kamagata et al. (2005).
sponding classes.
2.3.1. Pixel-based classification approach Datasets with a large number of attributes (i.e. data with high di-
Pixel-based classification approaches use the smallest entity within mensionality) will need more training samples to generate satisfactory
an image, the picture element (or pixel), in order to extract the feature results. According to the Hughes phenomenon (Hughes, 1968), any clas-
information in relation to one or more predefined classes. Each pixel can sifier becomes less effective above a certain data dimension. While the
be addressed by the x and y coordinates of the two dimensional image Hughes phenomenon is an issue for complex multidimensional datasets
space, and can be attributed with one or more values derived from the such as airborne imaging spectroscopy data, the issue of dimensionality
z axis, which is formed by the multiple layers of the dataset. In all does not seem to be a problem for ALS data.
cases the attributes of each pixel (in most cases, one or more spectral
values) function as the basis for the classification, in which an assembly 2.3.2. Object-oriented classification approach
of pixels with similar attribute values form an arrangement describing, In contrast to the pixel-based approach, which starts from the basic
for example, an archaeological feature to be classified. The attribute in- entities of an image, an object-based analysis uses the entire image or
formation can be any of the aforementioned model parameters (e.g. data set and breaks it down into meaningful segments. One essential
488 C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498

factor that distinguishes object-based approaches from the more tradi- Generally speaking, OBIA first uses a segmentation algorithm to di-
tional pixel-oriented approaches is the integration of the image content vide the image into “meaningful” objects that are non-overlapping.
into the classification procedure (Kettig and Landgrebe, 1976). While Image segmentation methods can be categorized into three groups:
data segmentation techniques have been used for the identification of (1) an image-driven approach applying statistical analysis, (2) a
objects since the 1970s, the last decade has seen rapid development of model driven approach, and (3) segmentation based on a homogeneity
more sophisticated techniques such as Object-based Image Analysis measure using spectral, spatial, textural, geometrical, contextual, tem-
(OBIA) (Blaschke, 2010; Blaschke et al., 2000), mainly due to an increase poral and prior knowledge as criteria (Dey et al., 2010). As stated
in computing power as well as the complexity and resolution of datasets above, this step often results in an over-segmentation, in which finer
(Platt and Rapoza, 2008). The subdiscipline of GEOBIA (GEographic segmented entities are often created because of the inherent inhomoge-
Object-based Image Analysis) has been established in order to differen- neity of the objects. The segmentation results form the basis of the
tiate OBIA applications that have a geographic component, such as ar- subsequent classification. Here, classes are defined and each individual
chaeological feature detection or land-use/land cover studies, from segment is assigned to a single class based on the employed target
applications such as automated detection of anomalies in medical imag- object's properties. Most commonly, spectral values, geometry,
ing (Hay and Castilla, 2008; Blaschke et al., 2014). neighbourhood relationships or semantic groupings are used. The selec-
Whatever the application, OBIA starts from the complete area under tion of the segmentation approach depends highly on the quality re-
investigation and consequently segments the contents in a way that fi- quired for the subsequent classification. The present case study uses
nally results in an image being divided into non-overlapping objects the homogeneity measure, where the homogeneity criteria are textural
(segments). In this way, the approach is opposite to that of a pixel- (openness, slope) and geometrical (roundness).
based one, which works from the smallest component of the image.
The aim of this approach is to subdivide the image into homogeneous 3. Case study areas and data acquisition
segments that describe the target features (pit, burial mound, etc.) as
correctly as possible. Thus, (GE)OBIA approaches have the advantage Two case study areas under investigation by the Ludwig Boltzmann
of producing image objects that more closely represent real-world ob- Institute for Archaeological Prospection and Virtual Archaeology (LBI
jects than individual pixels do (cf. Fisher, 1997:682), provided that the ArchPro) were chosen in order to test the applicability of similar
initial segmentation is appropriate (Platt and Rapoza, 2008:87). This is pixel- and object-based approaches to archaeological feature classifica-
achieved through the use of diverse segmentation methods, including tion. Both study areas have been covered by high-resolution ALS flights
point-based, edge-based, region-based, or combinations of these in periods suitable for maximum likelihood of ground returns, providing
(Schiewe, 2002). In an ideal application the resulting segments would highly detailed datasets. The site of Birka-Hovgården in Sweden (LBI,
correspond to real-world objects, i.e. the segmentation procedure 2015a) and the Kreuttal case study area in Lower Austria (LBI, 2015b)
would be able to map a target feature as a single image object. However, were chosen in order to demonstrate the applicability of each approach
none of the currently available segmentation methods are able to using the same data types in very different archaeological environ-
provide a perfect description of the features contained in an image. ments. As can be seen below, Birka has a short occupation sequence
Depending on the method applied, there will either be smaller or and relatively homogeneous feature types compared with the Kreuttal
larger resulting objects when compared with their counterparts in area, which contains traces of human activity visible in relief from as
reality. early as the Bronze Age up to the Modern Historic period. Thus, these
In most cases, segments will be smaller than the objects they classify, two highly contrasting case studies provide an opportunity to test
resulting in more than one segment describing a target feature both the applicability as well as the robustness of the approaches de-
(over-segmentation), which is due to the relative complexity of scribed in this study.
real-world features. Therefore, as a final step of OBIA, such segments
have to be assigned to a common class. When segments are larger 3.1. The Birka-Hovgården case study
than the real-world objects, the classification process ends up with
mixed classes, where each segment describes more than one object The archaeological site of Birka-Hovgården, a UNESCO-designated
(under-segmentation) (Liu and Xia, 2010:189) (Fig. 2). This case World Cultural Heritage Site located on the neighbouring islands of
should be avoided, since it would result in objects consisting of Björkö and Adelsö in Lake Mälaren in Central Sweden, was an important
more than one class, to which the dominant class would be assigned, Viking Age settlement and trading place between the years of ca. 750 CE
thus hampering a useful classification. and ca. 950 CE (Fig. 3). With a fortified hill-fort, an earthen town

Fig. 2. Segmentation example: unsatisfactory representation of a feature by under segmentation (left) in comparison with a feature appropriate segmentation (right).
C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498 489

rampart, several harbors and Hemlanden, Sweden's largest Iron Age as- archaeological research in the region (Fig. 4). The central topographic
sembly of burial mounds, Birka is considered to have been one of feature of the case study is a large, heavily wooded hill area whose
Scandinavia's most important Viking Age centers. Hemlanden measures ridge runs north–south along the center of the project area. Consisting
approximately 650 × 400 m NNW–SSE, and according to literature, it largely of a combination of evergreen and mixed deciduous vegetation,
consists of some 1600 ancient burials, including approximately 1500 the forested component of the case study area preserves a wealth of
burial mounds of 5 to 9 m diameter and a height of mostly 0.4 m to earthwork traces from multiple time periods below its canopy. The
1.0 m. Of those, circa 30 are surrounded by stone circles, circa 50 are lo- Kreuttal itself is a valley area that bisects this region and lies in a geolog-
cated within tri-cornered stone settings, 10 features are rectangular ically distinct zone at the interface of the Bay of Korneuburg
stone settings and five can be interpreted as smaller ship settings (Korneuburger Becken) area to the Viennese Basin (Wiener Becken)
(RAÄ Adelsö 118:1; Hermodsson, 2004). Hemlanden is covered to a (Wessely, 1998). The Kreuttal case study thus straddles two geologically
large extent by a light forest consisting mainly of birch trees. distinct regions, with the Kreuttal valley acting as a corridor between
In 1888 and 1889, surveyor J.J. Nordstrand mapped the burial two geographic zones. Although these landscapes can be considered
mounds on Björkö in great detail (Arbman, 1940), in the first recorded distinct, the Kreuttal valley corridor forms a link that can be seen as a
systematic modern investigation of the island. The first archaeological unifying factor in the area and a central point of study. These two land-
prospections involving terrestrial laser scanning in 2005, high- scapes have been culturally connected since prehistoric times, with nu-
resolution magnetic and ground penetrating radar (GPR) prospection merous settlements, funerary monuments, fortified hill-forts, and other
in 2006 (Trinks et al., 2014) and first large-scale motorized GPR evidence of occupation to be found in both the open and wooded areas
prospection in 2008 have resulted in detailed images of buried Viking of the region.
Age buildings, fortifications, track ways and graves (Trinks et al., Evidence suggests occupation in the region from at least the Early
2010). While no traces of Viking settlement architecture are visible Neolithic, with one of the largest known Early Neolithic enclosures
above the ground, largely undisturbed Viking Age stratification can be found at Grossrusbach-Weinsteig (Doneus et al., 2001; Löcker et al.,
found immediately below the plough layer at approximately 30 cm 2009:101), and a series of Middle Neolithic circular ring ditch systems
depth. Due to the distinct dark color of these rich archaeological layers, (Kreisgrabenanlage — KGA) identified in both the eastern and western
the central settlement area of Birka is called the ‘Black Earth’ (Svarta sections of the study area (Löcker et al., 2009:103 Neubauer,
jorden). 2012:154; Neugebauer, 1995:55; Trnka, 1991). Large hill-forts domi-
nate the mountainous region in the center of the study area, such as
3.2. The Kreuttal case study the presumed Bronze Age triple ditch hill-fort of Ochsenberg at the
north, and a possible Iron Age Oppidum in the center of the study area.
The Kreuttal case study (LBI, 2015b), is a 56 km2 area located north These locations provide both a defensive advantage as well as perfect
of Vienna, Austria and the subject of an integrated, landscape oriented positions for control over the resources of the region. Numerous other
archaeological prospection project that combines geophysical earthworks, particularly conflict remnants from World War II, dot the
prospection, remote sensing, survey and excavation methodologies in region and are woven into the fabric of the landscape, superimposed
the search for evidence of past land use, building on decades of on relict land use from the Bronze Age to the medieval period. For the

Fig. 3. Filtered digital terrain model of the northern half of the island of Björkö in Lake Mälaren. Viking Age Birka was located in the flat fields between the hill-fort, which is visible as a rocky
outcrop in the W, surrounded by a rampart in the NW. The grave field Hemlanden, the subject of this study, is located in the north central section of the island. Visualization: semi-
transparent sky-view factor calculation over a multidirectional hillshade, calculated using RVT (RVT, 2015).
Inset location map source: ESRI.
490 C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498

Fig. 4. Filtered DTM showing a central section of the Kreuttal case study area. A large prehistoric earthwork, provisionally dated to the Iron Age, can be seen in the center of the image. The
southern portion of the hillfort comprises the extent of the dataset used in this study. Visualization: semi-transparent sky-view factor calculation over a multidirectional hillshade,
calculated using RVT (RVT, 2015).
Inset location map source: ESRI.

purposes of this study, a 90 hectare segment of the case study area was Processing of the .las data was carried out by the LBI ArchPro (see
selected for analysis. Diverse archaeological features are located within Section 4.1).
this subsection, which is located in an afforested section of the case
study where preservation of objects in relief tends to be best (Fig. 4). 4. Data processing workflow and classification strategies

Classification strategies based on the principles outlined in


3.3. Data acquisition Section 2.3 were devised for each approach. The ALS point clouds
were first transformed into regularized raster DTMs from which open-
Full-waveform (FWF) ALS data were used in order to generate a ness visualizations were calculated, and then pixel- or object-based
high-resolution DTM of the entire island of Björkö, on which Birka is lo- image analysis steps were implemented in Trimble eCognition v8.7.1.
cated, as well as of the southern part of the neighbouring island of A workflow for each approach is outlined in Fig. 5. In both approaches,
Adelsö, where the rulers of Birka are believed to have resided at positive openness, negative openness and the DTM were stacked as de-
Hovgården (Nerman, 1918). Airborne Technologies (ABT), an Austrian scribed above. In the pixel-based approach, initial training classes of
partner of the LBI ArchPro and a commercial provider of ALS surveys,
conducted the survey on November 28th and 29th 2011, covering the Table 1
area in a matter of hours with a Riegl LMS-Q680i laser scanner Parameters of the ALS data acquisition at Birka-Hovgården and Kreuttal.
(Table 1). The time of survey was chosen in the off-leaf season, so that
Project Birka Kreuttal
the birch forest covering the NE part of Björkö would be largely
Purpose of scan Archaeology Land cover/environmental
defoliated, while the area was still free of snow. The operating altitude
Time of data 28th & 29th of November 11th of May 2012
was approximately 300 m above ground level, resulting in a mean acquisition 2011
point density on the ground of six points per square meter. Altogether, Point-density (pt/m2) 6 4–6
the entire project area was covered with 22 longitudinal strips with an Strip overlap 20% 20%
overlap of at least 20%, and two cross strips. The processing of the raw Scanner type Riegl LMS-Q680i Riegl LMS-Q680i
Full-Waveform Full-Waveform
data was carried out by ABT using Riegl's RiPROCESS software. Scan angle 45° 45°
For the Kreuttal case study area, FWF ALS data were acquired in LAS (whole FOV)
format from the government of Lower Austria, which had contracted Flying height above 300 m 350 m
data collection to a private supplier. A central part of the study area, in ground
Speed of aircraft (TAS) 105 knots 105 knots
a forested section containing previously identified prehistoric and his-
Laser pulse rate 400 000 Hz 400 000 Hz
toric earthworks, was identified as a suitable location. Flights took Scan rate 266 000 Hz 266 000 Hz
place on the 11th of May 2012, with the area being covered in 20 longi- Strip adjustment Yes Yes
tudinal strips, with an overlap of 20% (Table 1). The timing of this flight, Filtering Robust interpolation Robust interpolation
in early spring, also assured that vegetation would not be fully devel- (SCOP++) (SCOP++)
DTM resolution 0.5 m 0.5 m
oped, providing a high degree of penetration and ground point density.
C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498 491

Fig. 5. Data processing workflow for pixel- and object-based semi-automated feature detection strategies employed in this study.

roughly 250 pixels in size were defined in areas of high distinction, re- as similar processes do for the production of DTMs that are intended for
fined and added to until there was no increase in the classification accu- morphometric or physiographic modelling.
racy. In the object-based approach, the data stack was initially Once the DTMs were processed, positive and negative openness vi-
segmented and classified using radiometric and geometric attributes sualizations were created using the Openness module in the OPALS soft-
in order to describe distinct archaeological features and to separate ware package (Mandlburger et al., 2009). The ability to calculate
them from the background (non-archaeological features) using hard openness visualizations has also recently been implemented in the
value thresholds. The background can be seen as a mixed class contain- open-source Relief Visualisation Toolbox (RVT), developed by Zakšek
ing non-distinct features. In a subsequent classification step, objects in et al. (RVT, 2015). When using openness calculations, the size of the ra-
the remaining background are classified to split up the mixed classes dius should be adjusted for optimal feature delineation. This value is de-
using fuzzy thresholds. The resulting classifications could then be pendent upon the resolution of a particular DEM and relative size of
exported as either raster or vector objects for further use and accuracy features under investigation in a given study. For this study, openness
analysis in GIS applications. Although eCognition was used in this in- was calculated with a radius of 15 pixels (7.5 m), which was determined
stance, the relatively straightforward nature of both the pixel- and to be the best value for simultaneous generalization of feature values
object-based approaches presented here allows them to be implement- while preserving feature boundaries.
ed in other image analysis software. The use of the openness criterion turned out to be particularly infor-
mative in situations where archeological topographic features are very
4.1. DTM & visualization generation close to each other and extending over a large area. This is the case in
Birka's Hemlanden grave field, where hundreds of burial mounds are al-
In order to be able to identify micro-topographic structures of poten- most touching each other, and some are actually overlapping. Here, pos-
tial archeological relevance, the ALS point clouds needed to be convert- itive openness highlights topographic convexities and emphasizes the
ed to raster DTMs, which could then be used as a basis for visualization tops of the burial mounds (Fig. 6a). Negative openness emphasizes the
processing. In order to keep unwanted vegetation information from topographic concavities: the crests are highlighted by white lines and
appearing in the terrain model, point clouds were first classified into show the boundaries of the barrows quite clearly (Fig. 6b). Although
terrain and off-terrain points (Doneus and Briese, 2006). This was car- the feature density in the Kreuttal is somewhat lower and more hetero-
ried out using the hierarchical framework implemented in the software geneous, openness has proven to be applicable in this environment as
package SCOP ++ for point cloud filtering (Briese et al., 2002; Kraus well. Here, feature boundaries were also clearly delimited (Fig. 6 c, d).
and Otepka, 2005). The filtering strategy was adapted for archaeological
purposes in order to ensure that the archaeological objects were not re- 4.2. Classification procedure
moved during the filter and interpolation process (Doneus and Briese,
2006, 2011; Doneus et al., 2008). These parameters can also be found The resulting data stack of DTM, positive and negative openness pro-
online in the filtering parameters library at the website of the LBI vided the basis for the classification procedure, in which the raster im-
ArchPro (LBI, 2015c). The resulting DTMs, visualized in Figs. 3 & 4, ages are converted into discrete series of attributed features based on
have a ground sampling distance of 0.5 m. It should be noted that in the values mentioned above. The classification procedure differs signif-
order to preserve relief features of archeological interest, the filtering icantly for object and pixel-based approaches, with the pixel-based ap-
strategy employed here also preserves modern building footprints. proach relying solely on information contained in the pixel and its
Thus, it does not produce a ‘true’ bare-earth model of the earth's surface, surrounding neighbourhood, and the object approach relying on a
492 C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498

Fig. 6. Positive (a) and negative (b) openness calculated for a subset of the ALS area acquired over the grave field Hemlanden at Birka. A large number of burial mounds as well as a part of
the linear, NNW–SSE running town rampart can be seen in the image. Positive (c) and negative (d) openness calculated for a subset of the Kreuttal dataset. Multiple paths of varying depths
and alignment can be seen, as well as a section of the earthwork bank, ditch and gate which run through the center of the image from NW to SE. In figures (a) and (c), topographically
prominent areas and abrupt changes in slope are depicted as lighter in shade, while in figures (b) and (d) the deepest parts of features are displayed in a lighter shade.

combination of spatial and spectral values that are used to define the minimum distance between the pixel and the class center is calculated
boundaries and assign a class to objects of interest. using either the Euclidean distance or the Mahalanobis distance
A wide variety of pixel-based approaches have been developed over (Mahalanobis, 1936), which is a relative distance between data points
the past decades. For this study the fundamental technique, namely a in a multidimensional vector space. A quantitative comparison between
minimum distance classifier (Lillesand and Kiefer, 1999), has been both distance measures showed that the Mahalanobis distance per-
used to classify the archaeological features. In the pixel-based approach forms faster and the classification accuracy is slightly better (Fig. 7).
applied here, openness is the main variable used for feature definition. In The main advantage is that the classifier is mathematically and compu-
our cases, the addition of DTM-derived calculations such as slope do not tationally very efficient and therefore can be used on very large, high-
significantly increase classification stability. Therefore, in this applica- resolution datasets, as demonstrated by Benediksson et al. (1990).
tion slope is not used, although in other pixel-based workflows this var- As mentioned above, the object-oriented workflow typically starts
iable may be applicable. In order to determine the pixel class label, the with data segmentation and is followed by a classification step. The

Fig. 7. Comparison between Mahalanobis distance (left) and minimum distance (right) classifiers. The classification results using the minimum distance classifier shows a greater number
of smaller topographic objects derived from terrain structures classified as archaeological features than the Mahalanobis distance classifier.
C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498 493

initial objects are altered constantly in an iterative manner, progressing example from Birka the difference between object-oriented and pixel-
from the stage of object primitives toward “objects of interest” after based classification becomes evident when looking at the table of the
completion of the workflow. In the object domain, the archaeological classification accuracy (Table 2). Here, four datasets were compared:
features have been characterized by the attributes openness, slope and the map of burial mounds drawn by J.J. Nordstrand (Arbman, 1940),
roundness. In the current procedure, openness is derived from the posi- the manual interpretation of the ALS-derived DTM, object-oriented clas-
tive and negative values of the respective visualizations, slope is the rel- sification, and pixel-based classification (Fig. 10).
ative change in angle of terrain as derived from the height values of the Manual classification was regarded as the reference dataset, and the
terrain model, describing the orientation and steepness of an object, and other datasets were compared with this reference. To assess the accura-
roundness is computed from the similarity of the image object to that of cy of the automatically classified archaeological features, a distinction
an ellipsoid (Trimble, 2013:281). Each object is assigned to a class using had to be made between a proper recognition of a feature and its spatial
these parameters. Adjacent objects that meet the same class conditions matching in comparison with the manual interpretation map. Spatial
are grouped by an image-object fusion-procedure, where the fusion cri- matching was set to “true” when both classification results had a mini-
terion is the probability for belonging to the same object class (Trimble mum of 90% correspondence.
2013:89)(Fig. 8). The use of fuzzy transition between classes (Benz Table 2 shows that 91% of the features had been recognized with the
et al., 2004) enables us to deal with objects that would fail a classifica- object-oriented method, whereas the pixel-based method was able to
tion attempt using hard thresholds. By employing fuzzy logic the degree correctly classify 78% of the manually interpreted features. The areal
of membership to a class can have any value between one (true) and concordance with the object-oriented method is 85% compared with
zero (false). Thus, the membership function of a class allows the defini- the pixel-oriented method with 72%. A further analysis of these values
tion of relationships between the class feature (e.g. openness) and the showed that both methods had problems with small objects whose con-
degree of membership to a class on a sliding scale. trast to the surrounding environment is low. Overlapping objects have
been classified by both methods as one single object. This is a situation
5. Results where a template matching approach could have an advantage (e.g.
Trier et al., 2015).
5.1. Birka Regarding the spatial matching criterion of at least 90%, the classifi-
cation results can be transformed into type 1 (not classified) and type 2
At first glance, the results of both classification methods seem to re- (misclassified) error allocations as shown in Table 3. While the number
semble the manual visual interpretations (Fig. 8) for the Birka dataset. of type 2 errors is quite similar for both approaches, the main difference
The bank structure on the left part of the image is classified correctly, can be seen in the number of type 1 errors. Considerably fewer type 1
as well as most of the burial mounds. However, both automatic classifi- errors occur with the object-oriented classification approach than
cation methods show both possible types of false classifications: with the pixel-based classification approach. This difference is mainly
due to the application of fuzzy logic for class membership in the case
• Type 1: an existing burial mound is not classified. of the object-oriented classification approach. The use of fuzzy logic
• Type 2: an area classified as “mound” is actually not an archaeological for the class assignment seems to be a major advantage of object-
burial mound. oriented classification over the use of a pixel-oriented classifier.

5.2. Kreuttal
Both types of classification errors are indicative and reveal the bor-
derline between human interpretation and automatic classification. As discussed above, the archaeological context for the Kreuttal dif-
The results of the case study displayed less errors of type 1. The black fers greatly from that of Birka in both feature density and scale. In the
arrow in Fig. 9 points at one of the type 1 errors, where a burial section of the case study chosen to test the classification approaches,
mound was missed by the automated classification. The marked burial feature types include long linear earthworks such as banks, ditches
mound is hard to recognize in the data because neither its height values and path networks, as well as small pits, platforms and irregularly
nor its openness differ significantly from the surrounding. Nevertheless, shaped mounds (Fig. 11). In the Kreuttal classification, both approaches
in “manual” interpretation, an experienced archaeologist would not identify features, however even visual inspection shows that the suc-
hesitate to draw the outline of this feature and classify it as “mound”. Er- cessful detection rates are relatively low and the misclassification
rors of type 2 are mostly due to micro-topographic features such as rates are high. The accuracy assessment method used in the Birka exam-
rocks, outcrops, or filter residues that incorrectly have been classified ple is not entirely applicable, due in no small part to the increased num-
as mounds. These features exhibit a contrast to the surrounding terrain, ber of feature classes used in the Kreuttal interpretation. In the
but can easily be disregarded as a burial mound during subsequent visu- definition of this scene, the same sort of logic does not exist as it does
al interpretation. in the Birka example. Even discounting modern structures or features
The differences between the two tested classification methods be- of no archaeological interest, the feature classes and types are of varying
come obvious in a more detailed analysis (Fig. 9). In this specific scales and shapes, presenting an altogether more complex scene for

Fig. 8. Comparison of the manual (left), object-based (middle) and pixel-based (right) classification of the central Hemlanden area at Birka, which contains numerous burial mounds.
494 C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498

Fig. 9. Detail of the classification results; manual (left), object-based (middle), pixel-based (right).

interpretation. Furthermore, the manual classification has been con- analysis routines have difficulty in separating objects such as ‘paths’
ducted solely from the desktop and has yet to be verified in the field. Ad- from ‘ditches’. In reality, the features themselves may serve dual
ditionally, many of the features in the Kreuttal dataset are indistinct and functions, however this is often a matter of interpretation of function
show only a slight elevational and geometric variance to their surround- and semantics. The workflow does, however, identify sections of
ings. Again, both approaches had difficulty in these low contrast situa- prominent features, making it potentially useful as a tool for merely
tions, as in the Birka example. Similarly, overlapping objects were also identifying the presence of possible features of interest. Though the
classified as a single object. Nevertheless, using the manual classifica- workflow designed for Birka could be further modified to improve re-
tion as a reference, an estimate of the success rate of the pixel- and sults for specific feature classes, a more suitable answer would be to de-
object-based methods in overall detection of archaeological objects sign a different approach, which takes into account the multiplicity of
can be made for a subsection of the dataset (Table 4). The number of feature sizes and classes that are present in the Kreuttal dataset. In
corresponding features differs between the manual and semi- order to improve detection rates for such a variety of features, enhance-
automated approaches, as in a number of cases the semi-automated ments need to be made to the segmentation process and a hierarchical
classifications have identified numerous segments of a feature deemed logic needs to be implemented within the classification structure. This
a single object during manual classification. Due to the relatively low ac- was not necessary in the Birka example due to the relative homogeneity
curacy of feature detection already evidenced through visual inspection, of feature types. In order to maintain the same internal logic in the
spatial matching has been disregarded in favor of simply estimating the classification scheme for both case studies, a hierarchical approach
overall detection rate. was not implemented here. However, the main issues here are that ar-
From the manual interpretation, it can be seen that (Fig. 11a) that chaeological features present in the datasets vary greatly in terms of
there are many more classes of object present in the dataset than at scale, shape and prominence.
Birka, and likely from a much broader temporal sequence. Therefore
using the same approach as on the Birka dataset produces insufficient 5.3. Data export
results with regard to corresponding identification of feature types.
However, in both the pixel- and object-based applications, large and In order to perform the assessments described above, classified ob-
distinct linear features such as banks, ditches and main paths are jects were exported from eCognition as vector-based shapefiles, which
delimited to some extent (Fig. 11b, c). The pixel-based approach is could subsequently be opened in a GIS application for inspection and
even less successful here, due to the difficulty in discriminating homo- analysis. This process was included as an automatic, final step carried
geneous areas of interest in the dataset, leading to a difficulty in defining out after classification. Though a proprietary file format, shapefiles are
adequate training classes. Although some deep linear objects (paths) by now a de facto standard for spatial data transfer. In doing so, all object
are marginally better defined than in the object-based approach, only attributes, including classification and georeferencing information,
the top of banks are defined. Limited as it is, the GEOBIA approach al- were preserved and exported along with the object shapes. Options
ready appears more successful in this context (Fig. 11c), indicating also exist to export classification results directly to a spatial database
that openness, slope, and roundness exhibit potential as useful variables format, or as continuous raster surfaces. Thus, the classification results
in detecting objects of interest here as well. from both methods can be easily integrated into a geospatially based
For the Kreuttal dataset, the result clearly shows that a workflow analysis and review workflow. Additionally, classification results can
designed to identify and spatially define a specific range of targets (i.e. be attributed with values from the source datasets, allowing spatial
burial mounds and banks) does not produce a similar quantitative result data to be attributed with values from the layers in the source data
in a highly heterogeneous environment in terms of percentage of stack.
features detected when compared with human interpretation. The au-
tomation logic employed in the workflow developed in this paper is es- 6. Discussion: present limits and future possibilities
sentially binary (mound/not mound, or more generally feature/not
feature), which works well for a limited scalar and spatial range of fea- The results presented above represent a positive step forward for
tures. It is also unable to establish temporal boundaries of objects. Fur- semi-automated identification of archaeological features present in
thermore, there is a classificatory ambiguity present, in that the image ALS datasets. However, in archaeological terms, the examples presented

Table 2
Accuracy assessment of automatic classified object versus manual classification within the area under consideration at Birka.

Manual classification (reference) Nordstrand's map Object-based classification Pixel-based classification

Classified objects 1856 1731 1801 1572


Correctly classified objects – – 1694 1458
Spatial matching N90% – – 1581 1353
C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498 495

Fig. 10. Verification of manual classification results (left) and the object-based classification results (right) with the underlying map of J.J. Nordstrand. Mismatches are partly due to the
difficulties of rectification of the historic map.

here demonstrated that the use of an automated classification algorithm of the image analysis process (Belgiu and Lampoltshammer, 2013). One
as a substitute for manual human interpretation would result in a num- likely avenue for such a successful application could be through an on-
ber of missed or incorrectly classified features. This outcome is even tological framework, such as the one developed by Belgiu et al. (2014)
more critical considering that in the Birka example the archaeological for identification of diverse building structures in ALS data. Such an ap-
objects to be classified here belonged just to one class with a rather sim- proach could help improve feature detection rates for both case studies
ple shape. Here, object-oriented classification has the advantage of a sig- presented in this paper. Here, establishing a temporal sequence for ob-
nificantly smaller drop-out rate than a pixel-based approach; in the jects could also be helpful, and it may be possible to take advantage of
object-oriented approach fewer archaeological features remain unclas- the high resolution of ALS datasets similar to those used in this study,
sified. This lack of correspondence demonstrates that a fully automatic in order to identify feature superpositioning through automated
approach to archaeological feature detection is not advisable, and in- identification of feature cuts and abutments. Together with basic im-
deed such an approach should not be seen as a substitute for, or an at- provements in segmentation and classification rates, subsequent devel-
tempt to replace, expert human knowledge when it comes to the opment of such approaches and their reapplication to both datasets will
interpretation of archaeological prospection data. However, the results be the subject of future work.
also indicate that a large number of archaeological objects were detect- So, how far do computer-aided archaeological feature detection rou-
ed correctly. In the comparatively heterogeneous archaeological envi- tines need to go in order to be considered a useful application in the ar-
ronment of the Kreuttal case study the situation becomes even more chaeological interpretation toolkit? Overall, these approaches should
complex, as diversity of scale (target object size), feature type, and provide a functional benefit in the interpretation workflow. This can
environmental conditions provide more variables, which greatly reduce be in the form of a time-saving perspective, where accurate detection
successful classification rates for the workflows used in this study. and delimitation of archaeological features can both yield new discover-
Based on this evaluation, the application of semi-automated ap- ies and reduce the necessity to manually digitize features. In the case of
proaches to archaeological feature detection can be divided into two targeted detection of specific feature classes, this is a goal that already
broad categories: targeted detection and classification of individual, seems to be proving useful in a number of contexts. In these instances,
known feature classes with set parameters, and general detection of di- development generally focuses on improvement of spatial accuracy
verse archaeological features as a preliminary step in interpretation. and quantity of ‘correctly’ detected features. On the other hand, such ap-
As the presented approaches were designed with targeted detection proaches can also play an important role as tools in the general detec-
in mind, they clearly perform better in such environments. This is not tion of potential archaeological features of interest. In these cases,
to say that semi-automated approaches to archaeological feature detec- accurate delimitation of all spatial characteristics of a given object may
tion are not applicable in environments such as the Kreuttal, rather that not be a necessary outcome. Rather, rapid detection of potential objects
different strategies are needed for an altogether different application, of interest can act as an aid to more detailed, subsequent interpretation.
and one which may ultimately be more robust in nature. Furthermore, Developing feature detection routines within this framework would
difficulties in classifying archaeological features in such a context can also mean that a general structure for archeological feature detection
be viewed in part as a manifestation of the semantic gap (Smeulders could be subsequently modified in order to be applicable in a particular
et al., 2000:1353) in which there exists a disconnect between informa- context, without having to be completely rewritten. This may prove
tion extracted from target data and expert interpretation. Therefore, a especially useful as applied to datasets with a relatively low dimension-
prime task of future developments in GEOBIA-based feature detection ality. It is also worth noting that image analysis approaches have further
for archaeological applications such as this should focus on the bridging potential in the detection of features of interest that exist in spectra
of that gap through the connection of semantic information as informed invisible to the human eye. It is in these roles that semi-automated
by expert knowledge to the classification scheme. As OBIA approaches approaches to archaeological feature detection may have yet to be
rely on combinations of spatial, spectral and geometric characteristics explored to their fullest potential. Thus, while semi-automated ap-
and as such are context based, semantic data can be a viable component proaches have identified features in both case studies, it is clear that
a more complex, expert knowledge-driven classification schema
needs to be developed in order to deal with the diverse nature of
Table 3
the archaeological information present in the Kreuttal example
Accuracy assessment categorized in type 1 and type 2 errors within the area of consider-
ation at Birka. (Sevara and Pregesbauer, 2014). This sort of logic could also be
more effective in the attempt to identify ‘unknown’ types of archae-
Type 1 error objects Type 2 error objects
ological features, i.e. features and feature types that we do not al-
(not classified) (misclassified)
ready know of or are searching for. Therefore, the authors do see
Object-oriented classification 162 [8.7%] 107 [5.7%] the potential for GEOBIA-based approaches as aids to archaeological
Pixel-based classification 398[21.4%] 114 [6.1%]
interpretation for increasing the efficiency of identification, mapping,
496 C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498

Fig. 11. Comparison of the manual (a), object-based (b), and pixel-based (c) classification of the Kreuttal dataset. Images (b) and (c) are at a larger scale in order to better depict
classification results.

and classification of archaeological geocontent in both targeted and when a large number of similar, discrete features have to be mapped
general capacities. However, in the latter case the framework for feature or a statistical analysis has to be performed.
detection requires further development. Both pixel- and object-based methods offer the potential to apply
more sophisticated algorithms in order to increase the accuracy and
the reliability of the results. However, the aim of this study was to
7. Conclusion and outlook apply similar algorithms in diverse archeological environments in
order to examine their applicability in different environments and at
This paper has outlined two strategies for the semi-automated ex- different scales. In both studies, the object-based approach fared better
traction of archaeological features from ALS data, comparing the results than the pixel-based, largely due to the combination of parameters
of pixel- and object-based approaches to feature extraction in diverse available during image segmentation and classification. For the Viking
archaeological environments. While the logic of the feature detection Age burial mounds of Hemlanden at Birka, the classification algorithms
routines created for the detection of burial mounds was determined to compared here demonstrated that the developed methods are capable
be unsuitable for broad application in highly diverse archaeological en- of detecting typologically similar archeological features from a high-
vironments, it nevertheless proved effective for targeted feature detec- resolution ALS dataset and derived visualizations. In the case of the
tion and mapping. Furthermore, this study has illustrated the Kreuttal dataset, although both approaches failed to detect a number
applicability of the openness visualization for relief-based archaeologi- of feature types, the object-based approach still has the potential to
cal feature detection. Both approaches outlined here also have the abil- achieve more successful results for similar reasons. Additionally, the
ity to export results in vector-based, GIS compatible formats, allowing difficulty involved in creating adequate training samples in an environ-
for further refinement, analysis and integration of results with other ment like that of the Kreuttal all but eliminates a pixel-based approach
data. Though automated image classification techniques cannot and such as the one used here from future consideration. Overall, this illus-
should not take the place of an expert interpreter, it has been shown trates that feature detection routines designed for a limited, specific task
that they can nevertheless increase the efficiency of an interpretation should not be expected to function as a ‘magic bullet’ for archaeological

Table 4
General assessment of automated versus manual feature detection in the Kreuttal case study example.

Manual classification (reference) Object-based classification Pixel-based classification

Classified objects of interest 69 245 884


Correctly classified objects – 182 132
C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498 497

feature detection. While these examples do show that both demonstrat- Arbman, H., 1940. Birka — Die Gräber I, Tafeln. KVHAA, Stockholm.
Baatz, M., Hoffmann, C., Willhauck, G., 2008. Progressing From Object-based to Object-
ed approaches are capable of achieving classification results, a more oriented Image Analysis. Springer, Heidenberg.
complex classification approach is needed in environments with highly Belgiu, M., Lampoltshammer, T., 2013. Ontology based interpretation of very high resolution
heterogeneous feature types that manifest at different scales. This also imageries — grounding ontologies on visual interpretation keys. Proc. Agile, p. 2013.
Belgiu, M., Tomljenovic, I., Lampoltshammer, T.J., Blaschke, T., Höfle, B., 2014. Ontology-
demonstrates the fundamental difference between the use of semi- based classification of building types detected from airborne laser scanning data. Re-
automated approaches to detect objects which have potential as archae- mote Sens. 6 (2), 1347–1366. http://dx.doi.org/10.3390/rs6021347.
ological features with known parameters (such as burial mounds) and Benediksson, A.I., Swain, P.H., Esroy, O.K., 1990. Neural network approaches versus statis-
tical methods in classification of multisource remote sensing data. IEEE Trans. Geosci.
using these approaches to search for heterogeneous feature types with Remote Sens. 28, 540–552.
unknown parameters. Bennett, R., Cowley, D., De Laet, V., 2014. The data explosion: tackling the taboo of auto-
Nevertheless, as a tool for general identification of diverse archaeo- matic feature recognition in airborne survey data. Antiquity 88, 896–905.
Bennett, R., Welham, K., Hill, R.A., Ford, A.L.J., 2012. A comparison of visualization tech-
logical features, the object-based approach shows particular promise
niques for models created from airborne laser scanned data. Archaeol. Prospect. 19
and future research will focus on developing more complex hierarchical (1), 41–48.
schemes for feature classification in order to evaluate the applicability of Benz, U.C., Hofmann, P., Willhauck, G., Lingenfselder, M., 2004. Multi-resolution, object-
object-oriented approaches for these purposes. These include the han- oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS
J. Photogramm. Remote Sens. 58, 239–258.
dling of overlapping features on both spatial and temporal levels, the in- Bewley, R.H., Crutchley, S., Shell, C., 2005. New light on an ancient landscape: LiDAR sur-
clusion of non-archaeological feature classes to cope with large vey in the Stonehenge World Heritage Site. Antiquity 79, 636–647.
landscapes and non-contiguous features sets as well as the incorpora- Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS J. Photogramm.
Remote Sens. 65 (1), 2–16. http://dx.doi.org/10.1016/j.isprsjprs.2009.06.004.
tion of basic object semantics for an initial pre-interpretation of the clas- Blaschke, T., Hay, G.J., Kelly, M., Lang, S., Hofmann, P., Addink, E., Queiroz Feitosa, R., van
sified data (Sevara and Pregesbauer, 2014:140). As the proper der Meer, F., van der Werff, H., van Coillie, F., Tiede, D., 2014. Geographic object-
segmentation of a dataset is a necessary precursor to a successful classi- based image analysis — towards a new paradigm. ISPRS J. Photogramm. Remote
Sens. 87 (100), 180–191. http://dx.doi.org/10.1016/j.isprsjprs.2013.09.014.
fication, a multi-scale approach to image segmentation (Drăguţ et al., Blaschke, T., Lang, S., Lorup, E., Strobl, J., Zeil, P., 2000. Object-oriented image processing in
2014) may be a helpful initial step when dealing with data in which ar- an integrated GIS/remote sensing environment and perspectives for environmental
chaeological features at widely varying scales may be present. Further- applications. In: Cremers, A., Greve, K. (Eds.), Environmental Information for Plan-
ning, Politics and the Public. Metropolis Verlag, Marburg, pp. 555–570.
more, a comparative study of the techniques presented here with a Boer, A.G.d., 2005. Using pattern recognition to search LiDAR data for archeological sites.
template matching approach, such as outlined in Schneider et al. In: Figueiredo, A., Leite Velho, G. (Eds.), The World is in Your Eyes, Proc. of the XXXIII
(2014) and Trier et al. (2015), would also be extremely useful in order CAA2005 Conference. CAA2007: Tomar: Portugal, Tomar, pp. 245–254.
Bofinger, J., Hesse, R., 2011. As far as the laser can reach … laminar analysis of LiDAR de-
to evaluate these approaches in situations where targeted feature detec-
tected structures as a powerful instrument for archaeological heritage management
tion may be the desired application for image analysis. Data stacking in Baden-Württemberg, Germany. In: Cowley, D. (Ed.), Remote Sensing for Archaeo-
using multiple data types, such as inclusion of aerial imagery in various logical Heritage Management. Proceedings of the 11th EAC Heritage Management
spectra, may help increase detection rates. It will also be necessary to in- Symposium. Reykjavik, Iceland, 25–27 March 2010. 161–171. Archaeolingua; EAC
(Occasional Publication of the Aerial Archaeology Research Group, 3), Budapest.
vestigate morphometric variables, such as connecting classification re- Briese, C., Pfeifer, N., Dorninger, P., 2002. Applications of the robust interpolation for DTM
sults of landform elements (Drăguţ and Blaschke, 2006; Sevara, 2013; determination. In: Kalliany, R., Leberl, F., Fraundorfer, F. (Eds.), Photogrammetric
Verhagen and Drăguţ, 2012:699) with small scale archaeological fea- Computer Vision, International Archives of Photogrammetry. XXXIV, pp. 55–61 Graz.
Casana, J., 2014. Regional-scale archaeological remote sensing in the age of big data: au-
tures as described in this article, in order to gain a coherent picture of tomated site discovery vs. brute force methods. Adv. Archaeol. Pract. 222–233 (Au-
a landscape for subsequent archaeological interpretation. Though gust 2014).
there is much yet to be considered, the approaches presented here can Challis, K., Kokalj, Z., Kincey, M., Moscrop, D., Howard, A.J., 2008. Airborne LiDAR and his-
toric environment records. Antiquity 82, 1055–1064.
thus be considered as a useful contribution in the effort toward the for- Cheung, Y.-M., 2005. On rival penalization controlled competitive learning for clustering
malization of semi-automated approaches to archaeological feature de- with automatic cluster number selection. Knowl. Data Eng. IEEE Trans. http://dx.
tection in ALS data. doi.org/10.1109/TKDE.2005.184.
Cowley, D.C., 26–27 September 2012. In with the new, out with the old? Auto-extraction
for remote sensing archaeology. In: Bostater, C., Mertikas, S., Neyt, X., Nichol, C.,
Acknowledgments Cowley, D., Bruyant, J.P. (Eds.), Remote Sensing of the Ocean, Sea Ice, Coastal Waters,
and Large Water Regions 2012. Bellingham, Wash: SPIE (8532), Edinburgh, United
Kingdom 853206/1–853206/9 S.
The authors would like to thank the Government of Lower Austria De Laet, V., Paulissen, E., Waelkens, M., 2007b. Methods for the extraction of archaeolog-
GIS Services Division for providing airborne laser scanning data of the ical features from very high-resolution Ikonos-2 remote sensing imagery, Hisar
(southwest Turkey). JAS 34 (5), 830–841.
Kreuttal area, and Matthias Kucera, case study leader for the LBI ArchPro De Laet, V., Paulissen, E., Meuleman, K., Waelkens, M., 2007a. The effect of pixel resolution
Kreuttal case study. Many thanks to the two anonymous reviewers, and spectral characteristics on the extraction of archaeological features from very
whose comments and suggestions greatly improved the content of high-resolution remote sensing imagery: Sagalassos, Southwest Turkey. Remote
Sensing for Environmental Monitoring, GIS Applications, and Geology VII, Florence,
this paper. Additional data and support were provided by the Ludwig
Italy. 17 September 2007, p. 67490A.
Boltzmann Institute for Archaeological Prospection and Virtual Archae- Devereux, B.J., Amable, G.S., Crow, P., 2008. Visualisation of LiDAR terrain models for ar-
ology (http://archpro.lbg.ac.at), which is based on an international co- chaeological feature detection. Antiquity 82, 470–479.
Dey, V., Zhang, Y., Zhong, M., 2010. A review on image segmentation techniques with re-
operation of the Ludwig Boltzmann Gesellschaft (A), the University of
mote sensing perspective. In: Wagner, W. (Ed.), ISPRS TC VII Symposium – 100 Years
Vienna (A), the Vienna University of Technology (A), the Austrian ISPRS. XXXVIII, pp. 31–42 Vienna.
Central Institute for Meteorology and Geodynamic (A), the office of Doneus, M., 2013. Openness as visualization technique for interpretative mapping of air-
the provincial government of Lower Austria (A), 7Reasons (A), Airborne borne LiDAR derived digital terrain models. Remote Sens. 5 (12), 6427–6442. http://
dx.doi.org/10.3390/rs5126427.
Technologies GmbH (A), the Austrian Academy of Sciences (A), the Doneus, M., Briese, C., 2006. Full-waveform airborne laser scanning as a tool for archaeo-
Austrian Archaeological Institute (A), RGZM-Romano-Germanic Central logical reconnaissance. In: Forte, M. (Ed.), From Space to Place, 2. International Con-
Museum in Mainz (D), RAÄ-Swedish National Heritage Board (S), IBM ference on Remote Sensing in Archaeology. Archaeopress, Oxford, Rome, pp. 99–106.
Doneus, M., Briese, C., 2011. Airborne laser scanning in forested areas — potential and lim-
VISTA-University of Birmingham (GB), NIKU-Norwegian Institute for itations of an archaeological prospection technique. In: Cowley, D. (Ed.), Remote
Cultural Heritage Research (N), and Vestfold County Council (N). Sensing for Archaeological Heritage Management. Proceedings of the 11th EAC Heri-
tage Management Symposium, Reykjavik, Iceland, 25–27 March 2010.
Archaeolingua; EAC (Occasional Publication of the Aerial Archaeology Research
References Group, 3), Budapest, pp. 53–76.
Doneus, M., Kühtreiber, T., 2013. Airborne laser scanning and archaeological interpreta-
Ackermann, F., 1999. Airborne laser scannng: present status and future expectations. tion — bringing back the people. In: Opitz, R., Cowley, D. (Eds.), Interpreting Archae-
ISPRS J. Photogramm. Remote Sens. 54, 64–67. ological Topography. Airborne Laser Scanning, 3D Data and Ground Observation.
Alt, M., 1990. Exploring Hyperspace — A Non-mathematical Explanation of Multivariate Oxbow Books (Occasional Publication of the Aerial Archaeology Research Group, 5),
Analysis. McGraw-Hill, London. Oxford, pp. 32–50.
498 C. Sevara et al. / Journal of Archaeological Science: Reports 5 (2016) 485–498

Doneus, M., Briese, C., Fera, M., Janner, M., 2008. Archaeological prospection of forested Ludwig Boltzmann Institute for Archaeological Prospection and Virtual Archaeology (LBI),
areas using full-waveform airborne laser scanning. J. Archaeol. Sci. 35 (4), 882–893. 2015a. Archaeological prospection case study Kreuttal. Retrieved May 2015 from:
Doneus, M., Neubauer, W., Trnka, G., 2001. Die jüngerlinearbandkeramische http://lbi-archpro.org/cs/kreuttal/.
Grabenanlage von Großrußbach-Weinsteig in Niederösterreich — das größte Ludwig Boltzmann Institute for Archaeological Prospection and Virtual Archaeology (LBI),
Erdwerk der Linearbandkeramik. Preistoria Alpina 37, 145–159. 2015a. ALS filtering — filtering parameters library. Retrieved May 2015 from: http://
Drăguţ, L., Blaschke, T., 2006. Automated classification of landform elements using object- lbi-archpro.org/als-filtering/filtering-parameters-library.
based image analysis. Geomorphology 81 (3–4), 330–344. http://dx.doi.org/10.1016/ Mahalanobis, P.C., 1936. On the generalised distance in statistics. Proc. Natl. Inst. Sci. India
j.geomorph.2006.04.013. 2 (1), 49–55.
Drăguţ, L., Csillik, O., Eisank, C., Tiede, D., 2014. Automated parameterisation for multi- Mandlburger, G., Otepka, J., Karel, W., Wagner, W., Pfeifer, N., 2009. Orientation and pro-
scale image segmentation on multiple layers. ISPRS J. Photogramm. Remote Sens. cessing of airborne laser scanning data (OPALS)—concept and first results of a com-
88, 119–127. http://dx.doi.org/10.1016/j.isprsjprs.2013.11.018. prehensive ALS software. ISPRS Workshop Laser Scanning.
Figorito, B., Tarantino, E., 2014. Semi-automatic detection of linear archaeological traces Nerman, B., 1918. Kungshögarna på Adelsö och Sverigesäldsta konungalängder.
from orthorectified aerial images. Int. J. Appl. Earth Observ. Geoinf. 26, 458–463. Fornvännen 13, 65–77.
http://dx.doi.org/10.1016/j.jag.2013.04.005. Neubauer, W., 2012. Kreisgrabenanlagen — Middle Neolithic ritual enclosures in Austria.
Fisher, P., 1997. The pixel: a snare and a delusion. Int. J. Remote Sens. 18 (3), 679–685. In: Gibson, A. (Ed.), Enclosing the Neolithic: Recent Studies in Britain and Europe. BAR
http://dx.doi.org/10.1080/014311697219015. International, Oxford, pp. 147–163.
Harrower, M.J., Schuetter, J., McCorriston, J., Goel, P.K., Senn, M.J., 2013. Survey, automated Neugebauer, J.W., 1995. Archäologie in Niederösterreich. Poysdorf und das Weinviertel.
detection, and spatial distribution analysis of Cairn Tombs in Ancient Southern Ara- Niederösterreichische Pressehaus, St. Pölten.
bia. In: Comer, D.C., Harrower, M.J. (Eds.), Mapping Archaeological Landscapes Opitz, R., Cowley, D.C. (Eds.), 2013. Interpreting Archaeological Topography. Oxbow
From Space. Springer, New York, pp. 259–268. http://dx.doi.org/10.1007/978–1- Books, Oxford.
4614-6074-9_22. Platt, R.V., Rapoza, L., 2008. An evaluation of an object-oriented paradigm for land use/
Hay, G.J., Castilla, G., 2008. Geographic object-based image analysis (GEOBIA): a new land cover classification. Prof. Geogr. 60 (1), 87–100. http://dx.doi.org/10.1080/
name for a new discipline. In: Blaschke, T., Lang, S., Hay, G.G. (Eds.), Object-based 00330120701724152.
Image Analysis: Spatial Concepts for Knowledge-driven Remote Sensing Applications. Pregesbauer, M., 2013. Object versus pixel — classification techniques for high-resolution air-
Springer-Verlag, Berlin Heidelberg, pp. 75–89. borne remote sensing data. In: Neubauer, W., Trinks, I., Sailsbury, R., Einwögerer, C.
Hengl, T., Reuter, H.I., 2009. Geomorphometry. Concepts, software, applications. Elsevier (Eds.), Proceedings of the 10th International Conference on Archaeological Prospection,
(Developments in soil science, 33), Amsterdam. Vienna, 29 May–2 June 2013. Austrian Academy of Sciences Press, Vienna, pp. 200–202.
Hermodsson, Ö., 2004. Anmälan om specialinventering på Björkö och Hovgården 1997, Relief Visualisation Toolbox (RVT) 2015. http://iaps.zrc-sazu.si/node/67731#v accessed
RAÄ (RAÄ dnr 326–2743-2004. http://www.fmis.raa.se/(accessed 2013–04-10). May 2015.
Hesse, R., 2010. LiDAR-derived local relief models — a new tool for archaeological Schiewe, J., 2002. Segmentation of high-resolution remotely sensed data — concepts, ap-
prospection. Archaeol. Prospect. 17 (2), 67–72. plications and problems. ISPRS Commission IV Symposium: Geospatial Theory, Pro-
Hesse, R., 2014. Geomorphological traces of conflict in high-resolution elevation models. cessing and Applications.
Appl. Geogr. 46, 11–20. http://dx.doi.org/10.1016/j.apgeog.2013.10.004. Schneider, A., Takla, M., Nicolay, A., Raab, A., Raab, T., 2014. A template-matching ap-
Hughes, G., 1968. On the mean accuracy of statistical pattern recognizers. IEEE Trans. Inf. proach combining morphometric variables for automated mapping of charcoal kiln
Theory 14 (1), 55–63. sites. Archaeol. Prospect. http://dx.doi.org/10.1002/arp.1497 (Online Access).
Humme, A., Lindenbergh, R., Sueur, C., 2006. Revealing Celtic fields from LiDAR data using Sevara, C., 2013. Historic landscapes in relief: detection and interpretation of historic
Kriging based filtering. In: Maas, H.G., Schneider, D. (Eds.), Proceedings of the ISPRS landscape elements using gromorphometric and image analysis techniques. In:
Commission V Symposium ‘Image Engineering and Vision Metrology’, Dresden. Neubauer, W., Trinks, I., Salisbury, R.B., Einwögerer, C. (Eds.), Proceedings of the
Jahjah, M., Ulivieri, C., 2010. Automatic archaeological feature extraction from satellite 10th International Conference on Archaeological Prospection, Vienna, 29 May–2
VHR images. Acta Astronaut. 66 (9–10), 1302–1310. http://dx.doi.org/10.1016/j. June 2013. Austrian Academy of Sciences Press, Vienna, pp. 153–155.
actaastro.2009.10.028. Sevara, C., Pregesbauer, M., 2014. Archaeological feature classification: an object oriented
Kamagata, N., Akamatsu, Y., Mori, M., Li, Y.Q., 2005. Comparison of pixel-based and object- approach. South East. Eur. J. Earth Observ. Geomatics 3 (2 s), 139–144.
based classifications of high resolution satellite data in urban fringe areas. Asian Con- Sittler, B., 2004. Revealing historical landscapes by using airborne laser scanning — a 3D
ference on Remote Sensing (ACRS). AARS, Hanoi. model of ridge and furrow in forests near Rastatt (Germany). Int. Arch. Photogramm.
Kenzler, H., Lambers, K., 2015. Challenges and perspectives of woodland archaeology Remote. Sens. Spat. Inf. Sci. 36 (8), 258–261.
across Europe. In: Giligny, F., Djindjian, F., Costa, L., Moscati, P., Robert, S. (Eds.), Smeulders, A.W.M., Worring, M., Santini, S., Gupta, A., Jain, R., 2000. Content-based image
CAA 2014: 21st Century Archaeology, Concepts, Methods and Tools. Proceedings of retrieval at the end of the early years. {IEEE}. Trans. Pattern Anal. Mach. Intell. 22
the 42nd Annual Conference on Computer Applications and Quantitative Methods (12), 1349–1380. http://dx.doi.org/10.1109/34.895972.
in Archaeology. Archaeopress, Oxford, pp. 73–80. Townshend, J.R., 1981. The spatial resolving power of earth resources satellites. Progress
Kettig, R., Landgrebe, D., 1976. Classification of multispectral image data by extraction and in Physical Geography (5), 32–55. http://dx.doi.org/10.1177/030913338100500102.
classification of homogeneous objects. IEEE Trans. Geosci. Electron. 14 (1), 19–26. Townshend, J.R., Huang, S.N.V., Kalluri, R.S., Defries, R.S., Liang, S., Yang, K., 2000. Beware
Kokalj, Ž., Zakšek, K., Oštir, K., 2011. Application of sky-view factor for the visualisation of of per-pixel characteristics of land cover. Int. J. Remote Sens. 21, 839–843.
historic landscape features in LiDAR-derived relief models. Antiquity 85, 263–273. Trier, Ø.D., Pilø, L.H., 2012. Automatic detection of pit structures in airborne laser scanning
Kraus, K., Otepka, J., 2005. DTM modelling and visualization — the SCOP. Photogrammet- data. Archaeol. Prospect. 19 (2), 103–121.
ric Week. Wichmann, Heidelberg, pp. 241–252. Trier, Ø.D., Zortea, M., Tonning, C., 2015. Automatic detection of mound structures in air-
Lambers, K., Zingman, I., 2013. Towards detection of archaeological objects in high- borne laser scanning data. J. Archaeol. Sci. 2, 69–79. http://dx.doi.org/10.1016/j.
resolution remotely sensed images: the Silvaretta case study. In: Earl, G., Sly, T., jasrep.2015.01.005.
Chrysanthi, A., Murrietta-Flores, P., Papadopolous, C., Romanowska, I., Wheatley, D. Trimble Documentation (Trimble), 2013. eCognition v 8.7.1 Developer Reference Book.
(Eds.), Archaeology in the Digital Era: Proceedings of the 40th Conference in Comput- Trimble Germany GmbH, Munich.
er Applications and Quantitative Methods in Archaeology, Southampton, United Trinks, I., Johansson, B., Gustafsson, J., Emilsson, J., Friborg, J., Gustafsson, C., Nissen, J.,
Kingdom, 26–30 March 2012. Amsterdam University Press, pp. 781–791. Hinterleitner, A., 2010. Efficient, large-scale archaeological prospection using a true
Lasaponara, R., Masini, N., 2006. On the potential of QuickBird data for archaeological three-dimensional ground penetrating radar array system. Archaeol. Prospect. 17
prospection. Int. J. Remote Sens. 27, 3607–3614. (3), 175–186.
Lasaponara, R., Masini, N., 2009. Full-waveform airborne laser scanning for the detection Trinks, I., Neubauer, W., Hinterleitner, A., 2014. First high-resolution GPR and magnetic
of medieval archaeological microtopographic relief. J. Cult. Herit. 10, 78–82. http://dx. archaeological prospection at the Viking Age settlement of Birka in Sweden. Archaeol.
doi.org/10.1016/j.culher.2009.10.004. Prospect. 21 (3), 185–199.
Lasaponara, R., Coluzzi, R., Masini, N., 2011. Flights into the past: full-waveform airborne Trnka, G., 1991. Neolithische Befestigungen in Ost-österreich. Mitt. Anthrop. Ges. Wein
laser scanning data for archaeological investigation. J. Archaeol. Sci. 38 (9), 121, 137–155.
2061–2070. http://dx.doi.org/10.1016/j.jas.2010.10.003. Tso, B., Mather, P., 2009. Classification Methods for Remotely Sensed Data. CRC Press.
Lillesand, T., Kiefer, R.W., 1999. Remote Sensing and Image Interpretation. fourth ed. J. Verhagen, P., Drăguţ, L., 2012. Object-based landform delineation and classification from
Wiley & Sons, New York. DEMs for archaeological predictive mapping. J. Archaeol. Sci. 39 (3), 698–703.
Liu, D., Xia, F., 2010. Assessing object-based classification: advantages and limitations. Re- http://dx.doi.org/10.1016/j.jas.2011.11.001.
mote Sens. Lett. 1 (4), 187–194. http://dx.doi.org/10.1080/01431161003743173. Wessely, G., 1998. Geologie des Korneuburger Beckens. Beitr. Paläontol. 23, 9–23.
Löcker, K., Nau, E., Hinterleitner, A., Neubauer, W., 2009. Magnetic surveys of Early and Yokoyama, R., Shirasawa, M., Pike, R.J., 2002. Visualizing topography by openness : a new
Middle Neolithic settlements in Austria. ArchéoSciences 33 (suppl.), 101–104. application of image processing to digital elevation models. Photogramm. Eng. Re-
Ludwig Boltzmann Institute for Archaeological Prospection and Virtual Archaeology (LBI), mote Sens. 68 (3), 257–265.
2015a. Archaeological prospection case study Birka-Hovgården. Retrieved May 2015 Zakšek, K., Oštir, K., Kokalj, Ž., 2011. Sky-view factor as a relief visualization technique. Re-
from: http://lbi-archpro.org/cs/birka/index.html. mote Sens. Environ. 3 (2), 398–415.

You might also like