Object‐based image analysis: a review of developments and future directions of automated feature detection in landscape archaeology

Object‐based image analysis (OBIA) is a method of assessing remote sensing data that uses morphometric and spectral parameters simultaneously to identify features in remote sensing imagery. Over the past 10–15 years, OBIA methods have been introduced to detect archaeological features. Improvements in accuracy have been attained by using a greater number of morphometric variables and multiple scales of analysis. This article highlights the developments that have occurred in the application of OBIA within archaeology and argues that OBIA is both a useful and necessary tool for archaeological research. Additionally, I discuss future research paths using this method. Some of the suggestions put forth here include: pushing for multifaceted research designs utilizing OBIA and manual interpretation, using OBIA methods for directly studying landscape settlement patterns, and increasing data sharing of methods between researchers.


| INTRODUCTION
Researchers in many fieldsincluding computer science and geographyhave adopted machine learning algorithms to process remote sensing imagery (see Mountrakis, Im, & Ogole, 2011). In the late 1990s and early 2000s, a form of machine learning known as object-based image analysis (OBIA) was developed (Blaschke, 2010), but only recently have archaeologists utilized these methods (e.g. De Laet, Paulissen, & Waelkens, 2007;Menze, Ur, & Sherratt, 2006). In the last~15 years, archaeologists have used a variety of OBIA techniques that are highly successful in extracting features of interest from large-scale datasets at faster rates and lower costs than manual processing (Bennett, Cowley, & De Laet, 2014, 897).
Yet, there is much more that these methods can do to advance our understanding of the human past.
Today, there have been a number of significant studies using OBIA methods within archaeological contexts, but these studies are not evenly distributed geographically (see Table 1). Many focus on European localities, but fewer focus on the Americas, Asia, Africa, or island regions. Furthermore, most archaeological publications using OBIA methods are identifying potential sites, but they are not addressing potential settlement patterns that emerge from their results. This is important because future research should use these methods to answer archaeological questions concerning populations, socio-political organization, and past peoples at large. This article serves as a review of object-based machine learning methods that archaeologists have applied in landscape-scale remote sensing analysisincluding aerial and spaceborne data. I will detail the progress that has been made with these techniques as well as the avenues archaeologists are yet to travel. I begin by reviewing the basic concepts of OBIA and how it operates. I follow with a comprehensive summary of archaeological work that has been conducted using OBIA, paying particular attention to the successes and shortcomings of these studies. Then, I discuss possible future directions of OBIA and computational archaeology. I illustrate that while archaeologists have just begun to apply OBIA to research questions, the method offers unparalleled advantages that should be fully taken advantage of by future archaeological research.

| OBJECT-BASED IMAGE ANALYSIS (OBIA)
OBIA began to rise in popularity in the early twenty-first century, and since that time, uses of these methods have sharply increased (Blaschke, 2010). Blaschke (2010) attributes this rise to the development of a software called eCognition (Trimble, 2016). In addition to eCognition, several open-source platforms have been developed for OBIA analysis [e.g. GEODMA (Körting, Garcia Fonseca, & Câmara, 2013), InterIMAGE (InterIMAGE, 2009), Grass GIS (GRASS Development Team, 2018, also see Knoth & Nüst, 2017]. In its most basic definition, object-based analysis encompasses 'image-processing techniques that when applied either result in the segmentation (i.e. partitioning) of an image into discrete nonoverlapping units based on specific criteria, or are applied to define specific multiscale characteristics-from which segmentation may then be based' (Hay, Castilla, Wulder, & Ruiz, 2005, 340). Recently, remote sensing literature has used the term GEOBIA to refer to those applications of OBIA to Earth remote sensing imagery (Blaschke et al., 2014;Hay & Castilla, 2008). GEOBIA therefore constitutes a majority of OBIA applications within archaeology.
OBIA methods, in contrast to pixel-based methods, identify features using multiple variables. These include pixel value, object shape, textural information, neighbourhood analysis, and geographic context (Blaschke, 2010, 3;Blaschke et al., 2014). By utilizing multiple parameters simultaneously, OBIA is well suited for identifying features that are small, structurally homogeneous, and display differences with local topography (Davis, Sanger, & Lipo, 2018). OBIA builds on longstanding practices of remote sensing analysis including segmentation, edge detection, and classification (Blaschke, 2010, 3; see Kumar, Raj Kumar, & Reddy, 2014;Weng, 2010, for reviews of different types of segmentation and classification). As such, some have considered OBIA to be one of the greatest achievements in image processing of the twenty-first century (Arvor, Durieux, Andrés, & Laporte, 2013). One limitation of OBIA is that it requires very high-resolution datasets to work effectively (Blaschke et al., 2014, 181). However, as the spatial resolutions of remotely sensed data have improved, the accuracy and use of OBIA techniques have also increased (Hay et al., 2005).

| OBIA AND MACHINE LEARNING IN ARCHAEOLOGY
Object-based analysis of remote sensing data has only been extensively utilized by archaeologists for about 15 years. The number of peer reviewed publications using these methods within archaeological contexts is small (< 40) but growing (see Table 2). Additionally, a great deal of work has been presented on the use of OBIA at archaeological conferences.
Beginning in the first decade of the twenty-first century, researchers began to implement object-based computer algorithms to detect archaeological features in a systematic fashion. The first archaeological research implementing OBIA was primarily concerned with identifying large-scale linear features. For example, Bescoby (2006) used a mathematical function known as a Radon transform (which can determine the most common alignment and orientation of features within an image) and segmentation procedures to detect linear Roman structures in satellite imagery. Within a few years, more publications began to emerge using OBIA methods (e.g. De Laet et al., 2007;Jahjah, Ulivieri, Invernizzi, & Parapetti, 2007;Van Ess et al., 2006). All of these studies focus primarily on the detection of archaeological deposits, but Jahjah et al. (2007) also look at how OBIA techniques can monitor sites, document their preservation levels (also see Van Ess et al., 2006), and enhance the digitization of archaeological data acquired from remote sensing sources. As the resolution of

Subsequent research has implemented a slew of new variables
including topographic measurements such as hillshade, slope, and topographic openness (see Table 2). The results of these studies indicate a positive correlation between the number of factors accounted for during OBIA procedures and accuracy. However, if the parameters chosen do not match the features that are being sought after then the algorithm will not work. As such, an expert knowledge of the study area is an essential prerequisite for using automated detection methods.
The most recent archaeological uses of OBIA and machine learning yield highly accurate results (Freeland et al., 2016;Guyot et al., 2018;Lasaponara & Masini, 2018;Wang et al., 2017). Freeland et al. (2016) demonstrate the first use of hydrological depression algorithms for archaeological mound detection. In this instance, an inversed DEM was created and processed through an algorithm that looks for topographic depressions, effectively identifying and mapping mound features (also see Davis, Lipo, & Sanger, in press

| Limitations and criticisms
Despite the many successes of OBIA methods within archaeology, there are many who are skeptical of the feasibility of automated detection algorithmsspecifically for large-scale landscape analysis (e.g. Casana, 2014;Hanson, 2010;Parcak, 2009). Parcak (2009, 110) claims that automated archaeological site detection is impossible because every archaeological project is dependent on local variables. But local variables are precisely what OBIA can take into consideration when analysing remote sensing data, and regionally specific algorithms are essential for the success of automated prospection (see Davis et al., 2018). Parcak (2009,110) goes on to state that computers cannot pick up on the same subtleties in remotely sensed data as humans can by eye. However, the very fact that recent studies using automated means have detected sites that manual analysis has overlooked directly challenges this claim (e.g. Davis et al., 2018;Witharana et al., 2018).
In discussing the latest state of remote sensing research within archaeology, Opitz and Herrmann (2018) devote some of their attention to the methods involving automated detection of archaeological features. Part of their discussion revolves around a distrust of these methods, and they state: The reluctance to adopt automated feature extraction … is motivated by a combination of technological and social factors. On the technological side, machine learning approaches to automation remain in their infancy.
Automatic feature extraction for archaeological materials is still developing and has yet to match the efficiency of automatic feature extraction for targets with consistent appearance or for features in uniform environments. (Opitz & Herrmann, 2018, 30) The claim that these methods are still new and evolving is very much true, as this article indicates. Regardless, the infancy of the method is not a reason to stop developing and improving its ability to discern information of archaeological significance. OBIA and similar methods are imperfect and cannot replace manual evaluation completely, but at the same time, biases in knowledge by data analysts limit the accuracy of manual procedures and can lead to omission error (Bennett et al., 2014;Gheyle et al., 2018). It can never be our goal to completely automate the archaeological process, and to attempt such a feat would be a fool's errand. Nevertheless, improving automated methods to assist in the detection of archaeological deposits is not only an exciting avenue for future research, but also a necessary task.
Coastal and island regions that are under threat of destruction by climate change and rising sea levels cannot ever be fully surveyed using traditional means before their records are severely damaged. It is therefore imperative to document as much of these areas as we can before they are lost. By using OBIA and similar methods, we can conduct systematic surveys of entire areas and document landscapes efficiently. Thus, it is essential to utilize these techniques to study the archaeological record in a relatively complete form rather than limiting ourselves to small sample sizes of information.
Despite the benefits offered by OBIA, it is still far more common for archaeologists to use manual interpretation methods rather than semi-automatic means (Quintus, Day, & Smith, 2017, 352; also see Casana, 2014). Many researchers echo the earlier sentiments of Parcak (2009) by claiming that automated methods cannot account for the wide range of variability in the archaeological record. However, who is to say that one should only use one single automated method to scan an entire study area? Why not use a multitude of different algorithms to search for different parts of the record and then go through all the results by hand to fill in things that OBIA missed (sensu Bennett et al., 2014)? By using automated detection first, we can be sure that the entire study area is surveyed systematically without any lapses. Then by conducting a manual analysis, expert knowledge can assess the results and potentially identify nearby features that the automated method overlooked. Casana (2014) uses 'brute force', or manual extraction methods to survey an area covering 300 000 km 2 . This process took approximately 3-4 years. Using automated methods [which Casana (2014) attempted and stated to be successful], this process could have been sped up considerably. Quitnus et al. (2017) also illustrate the importance of manual evaluation, but highlight the fact that manual processing is imperfect, as there are still many false-positive and false-1 Although this article has focused exclusively on the use of OBIA for large-scale remote sensing data such as satellite imagery and LiDAR (GEOBIA), this method has also been used for other types of image analysis in archaeology. OBIA has been successful in classifying artefacts and features into statistically significant types (e.g. Lamotte & Masson, 2016;Ozawa, 1978), studying site formation processes (e.g. Sanger, 2015), testing the mineralogical classification of artefacts (e. negative identifications present using this approach. In the span of two weeks, Quintus et al. (2017) evaluated LiDAR covering approximately 5-10 km 2 , whereas semi-automated OBIA methods have allowed for the systematic evaluation of LiDAR datasets covering thousands of square kilometres in the same amount of time (e.g. Davis et al., 2018).
One thing that is certain is that ground-testing and manual analysis following automated detection algorithms is an essential step (Ainsworth, Oswald, & Went, 2013;Freeland et al., 2016, 72;Quintus, Clark, Day, & Schwert, 2015). Without groundsurveys, we cannot confirm the results of remotely sensed analysis and our knowledge cannot pass beyond a theoretical level.
Although in some instances automated survey may be inappropriate, we must remember that ground-survey, manual evaluation, and automated detection algorithms are all useful tools for archaeologists, and all possess their own benefits and drawbacks. As such, archaeologists must not exclude any of these methods out- There is a lot to gain from OBIA, especially in terms of understanding landscape-level archaeological patterns. However, there are certain avenues of research where these methods are yet to be fully invested: • First and foremost, the use of OBIA methods must be expanded into new geographic areas where they have been under-utilized or where they are yet to be introduced (e.g. North America, South America, Africa, coastal islands, etc., see Table 1). This is especially important for areas at risk of destruction from sea-level rises or currently experiencing violent conflict where cultural heritage is at risk.
• Second, we must continue developing new approaches that combine automated analysis with manual evaluation and subsequent field-testing to create a comprehensive landscape survey procedure. Each of these levels are essential for understanding the archaeological record. By combining them together, we can study landscape patterns at multiple scales, which is a vital component of landscape level archaeological research (e.g. Crumley, 1979;Millican, 2012;Robinson, 2010).
• Third, future work with OBIA should seek to compare different methods of automated feature detection (e.g. Davis et al., in review). By comparing different methods, researchers can best determine which methods are most appropriate for specific purposes and thereby adopt the successes and avoid the failures and setbacks of prior studies.
• Fourth, to improve the ability of OBIA to detect archaeological features, researchers must share their datasetsthis includes new algorithms, computer code, processing steps, and training data. By sharing this information, archaeologists around the world can contribute to and access different methods and necessary training data, thereby increasing and improving the use of OBIA for archaeological problems. By making code and data available to all, even the non-specialist can utilize some of these methods and contribute to the use of automated object detection.
• Finally, archaeologists should use OBIA for studies beyond the mere detection of features. Researchers can use detected objects to discuss broader spatial patterns of the archaeological record (e.g. Freeland et al., 2016). Although the discovery of new features is important, it is equally important to begin analysing this newly generated information to further our understanding of the human past.

| CONCLUSIONS
This article has sought to demonstrate the important advances that have occurred in applications of OBIA methods within landscape archaeology. It has also traced some possible paths for the future of these methods within the discipline. A lot of progress has been made, and yet there is still a great deal of untapped potential for OBIA to expand our understanding of the archaeological record. In the future, we should seek to incorporate (semi-)automated algorithms with manual analysis to ensure the broadest range of data is acquired. The importance of systematic documentation is vital in a world that suffers from cultural site destruction on a daily basis. OBIA is one method that can help to record, preserve, protect, and study the record of our collective human history.

FUNDING
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

DECLARATION OF CONFLICTS OF INTEREST
The author does not have any conflicts to declare, financial or otherwise.