On Recognizing Potential Archaeological Sites in Aerial Images
The objective of our research is to propose an automatic method capable of recognizing possible archeological sites from aerial images based on detection of areas with non-homogenous vegetation cover. In this post, we briefly describe our data, how new sites can be possibly discovered in aerial imagery, what are the steps we need to take to develop a solution, and our current progress.
Many archaeological mapping methods nowadays are dependent on extracting a precise Digital Surface Model (DTM) for possible site detection.
Data for DTM computation can be obtained either from crowdsourced and often inaccurate terrain documentation or precise but expensive remote sensing methods such as airborne laser scanning (ALS), hyperspectral imaging (AIS), UAV, sub-surface radars and others.
The combination of those methods provide considerably richer data and can uncover historical sites hidden below the surface or in areas covered by a build-up and dense vegetation. Still, there are archeological sites which can be recognized in aerial images with the naked eye. Those detectable areas are visible in locations with low vegetation cover as the underground remnants provide either deficient or improved nutrients to the soil resulting in distinct vegetation features.
However, manual search for such sites is both time-consuming, costly, or even impossible for larger areas, which makes the task an ideal candidate for automation. In our research, we hope to develop such a system capable of suggesting possible hidden underground archeological structures.
Background
Aerial photography has been systematically used for archeological research for the past century and even today remains, alone or in conjunction with other remote or ground-based sensors, the most applied method for site discovery. That is without a doubt largely because of its cost-effectiveness.
New potential sites can be detected as vegetation marks or terrain break-lines such as sharp edges, corners or circles, that wouldn’t naturally occur or do not fit its surroundings.
The easily available imagery has the advantage of abundant special detail, which enables to capture even small but important details in the vegetation cover such as burrows and pits, which would be often missed by other remote sensing methods. We have been using color aerial photography from Mapy.cz with the highest (when available) spatial resolution of 10 cm per pixel. That is exceptional in contrast to other remote sensing methods that have, in many cases, spacial resolution in meters per pixel. The portal also provides archived aerial photography from years ’15, ’12, ’06 and ’03, as each third of the Czech Republic is re-screened every year.
The main disadvantage of our data is its limited spectrum as the detection of crop marks in low contrast areas can become invisible in normal photography. Other problems connected to this method include detection possible only in uncovered areas and sensitivity to the seasonal changes in which the data is acquired.
Related work
As of today, most research is focused on using multiple remote sensing technologies and the combination of the datasets trying to get a precise Digital Surface Model representing the area. Such studies have been conducted in smaller areas (<50 ha) for cost-dependent reasons and often incorporate historical photographs from different time periods and other region-dependent documentation. For this reason, purposely developed algorithms may contain geographical bias and it’s recommended to use more widespread and generic algorithms for larger and varying areas [1].
In current studies [1, 2, 8] the general approach is to match found microrelief marks (elevation pattern standing out of its surroundings) or other sub-surface anomalies to similarly shaped crop marks. Found correlation between these two is then used as a tool to distinguish crop marks associated with archeological remains from other man-made crop marks created only in recent years.
Despite using other technologies than airborne photography, many limitations remain. In the current state of the art, the detection of crop marks is considered a difficult task since it depends on the type of sensor used, season and the type and status of the vegetation cover.
Similar in goals to our work is Aerial Laser Scanning in Archaeology [4] where the authors describe the use of low density (and lower cost) ALS data acquired from public services for terrain modifications detection. They conclude that their affordable data can serve not only for closer mapping of known sites but also in searching for new historic sites.
Comparable work in terms of image processing is a study on the discovery of ancient relics in historical aerial panchromatic photographs [3] from 2015. The study is directed at finding general rules for interpreting such images for hidden remnants. Another study [2] is centered around airborne multi-spectral imaging but also provides valuable information regarding the interpretation of vegetation marks in aerial photography.
Closer to our goals and means is a study from 2015 Towards an operative use of remote sensing for exploring the past using satellite data: The case study of Hierapolis (Turkey) [9] in which the authors focus on automatic detection of buried archaeological remains. They point out the merits of multi-date observations and exploitation of additional characteristics of remains like geometric patterns or regular shapes in unsupervised classification. Another relevant study [11] has been published earlier this year and explores the use of high-resolution images from a multispectral commercial satellite in conjunction with advanced classification algorithms to map surface archaeological features with distinct spectral signatures. The study has demonstrated that very high spatial resolution satellite images can be used to detect surface archaeological features directly and expresses the need to assess the possibility of using high-resolution images from non-commercial (and therefore free of charge) satellites.
Preliminary results
The flowchart of our proposed solution breaks down to two main parts.
The first part would be the classification of surfaces that can be used for further analysis. Those are areas with mostly homogenous low, green vegetation such as meadowlands, pasture lands, and fields, whereas build-up and forested areas would be filtered out.
The subsequent part is the automated archeological site detection on the potential areas suggested from the first classification step. This detection is based on unnatural shapes of the vegetation cover inhomogeneities, which signal possible underground building remnants.
We started our work in the second step — with detecting such shapes first since we’ve been provided with a restricted number of locations of already known sites. Later we will complete the method also with the preceding potential searchable areas detection step. So far, we have experimented with various approaches:
SIFT
Our first experiments were focused on the extraction of SIFT (Scale Invariant Feature Transform) [5] keypoints and descriptors from greyscale images. The question is whether we’re able to detect points of interest in low contrast, noisy data and subsequently match these points to our patterns. If so, we could detect shapes based on the amount of generated matches.
Coupled with basic image preprocessing like contrast correction and grayscale conversion (PCA or other methods) we have found that we can detect correct key points in most of our current data.
More experiments regarding further image preprocessing and matching descriptors to our patterns are underway.
Edge detection and generalized Hough transform
Our second approach is to use an edge detector to extract key image contours and then proceed to detect shapes such as rectangles, circles or other arbitrary shapes with the generalized Hough transform (GHT) [7]. A similar approach has been successfully used for the automatic extraction of circular archeological tops from Google Earth Images [10].
Unfortunately, using the GHT to detect shapes of different sizes and positioning leads to a higher computation cost than that of our previous idea. Furthermore, a number of stencils would have to be provided to detect more complicated shapes such as bastion forts and ramparts.
From experiments with various edge detectors, Canny detector applied to a grayscale image with upped contrast and Gaussian blur (denoising technique) yields best preprocessing results across our dataset for future application of GHT.
Adaptive thresholding and contour detection
Seeing quite clear “ditches” forming during the edge detection process led to another experiment. As most noise in the image results in quite small and discontinued shapes and we do have a general idea about the size of the shapes we are looking for, we can try to extract only shapes of a certain size from the images. Those shapes then need to be further classified.
To highlight the major contours present in the image we used further preprocessing to a grayscale image with upped contrast. Since having uneven levels of light in the image is a common occurrence, we used Otsu’s method to help even out the difference [6].
We further detected a number of largest contours (in yellow) and filter out those in our specified scale range (in green). This way we can get a binary mask showing major shapes present in the image that can be further analyzed.
As can be seen from our conducted experiments, most results lead to a rather clear distinction of foreign shapes present in the noisy aerial pictures. On the other hand, it has been quite impossible to differentiate small objects such as barrows and marks from wooden posts from the data noise.
Future work
We’ll continue to experiment with various approaches to improve our data quality. Once we decide on the most prospective technique, we will need to gather more data to configure all parameters and test the method on a broader dataset. We can then implement classifier and once we’ll have it tested on a bigger data sample, we will move on to the task of segmentation of the aerial photos.
After both parts of our implementation are complete we can conduct a test run on a chosen region in the Czech Republic and asses the results.
Acknowledgments: Our research is a collaboration with the Institute of Archaeology and the Institute of Information Theory and Automation of the CAS.
References
[1]: Calleja, Javier F. et al.,(2018). Detection of buried archaeological remains with the combined use of satellite multispectral data and UAV data,
International Journal of Applied Earth Observation and Geoinformation,
Volume 73, pages 555–573,https://doi.org/10.1016/j.jag.2018.07.023.
[2]: Doneus, M., Verhoeven, G., Atzberger, C., Wess, M., Ruš, M. (2014). New ways to extract archaeological information from hyperspectral pixels. Journal of Archaeological Science. 84–96. https://doi.org/10.1016/j.jas.2014.08.023.
[3]: Peng Lu et al., (2017) On the use of historical archive of aerial photographs for the discovery and interpretation of ancient hidden linear cultural relics in the alluvial plain of eastern Henan, China, Journal of Cultural Heritage, Vol. 23, Supplement, Pages 20–27, https://doi.org/10.1016/j.culher.2015.09.010.
[4]: Faltynova, M., Pavelka, K. (2011). Aerial Laser Scanning in Archeology. Geoinformatics FCE CTU. https://doi.org/10.14311/gi.6.14
[5]: Lowe, D.G. (2004). Distinctive Image Features from Scale-Invariant Keypoints.International Journal of Computer Vision. 60: 91 https://doi.org/10.1023/B:VISI.0000029664.99615.94
[6]: Sonka, M., Hlavac, V., Boyle, R. (2008). Image Processing, Analysis and Machine Vision. London: Chapman & Hall. https://doi.org/10.1007/978-1-4899-3216-7
[7]: D.H. Ballard, (1981). Generalizing the Hough transform to detect arbitrary shapes, Pattern Recognition, Volume 13, Issue 2, Pages 111–122,
https://doi.org/10.1016/0031-3203(81)90009-1.
[8]: Kucukkaya, Gulcin. (2004). Photogrammetry and remote sensing in archeology. Journal of Quantitative Spectroscopy and Radiative Transfer. 88. 83–88. https://doi.org/10.1016/j.jqsrt.2003.12.030.
[9]: Lasaponara R. et al., (2016). Towards an operative use of remote sensing for exploring the past using satellite data: The case study of Hierapolis (Turkey),
Remote Sensing of Environment, Vol. 174, Pages 148–164,
https://doi.org/10.1016/j.rse.2015.12.016.
[10]: Luo, Lei et al. (2014). Automated Extraction of the Archaeological Tops of Qanat Shafts from VHR Imagery in Google Earth. Remote Sensing. 6. 11956–11976. https://doi.org/10.3390/rs61211956.
[11]: Merlo, S., Thabeng, O. L., Elhadi A., (2019). High-resolution remote sensing and advanced classification techniques for the prospection of archaeological sites’ markers: The case of dung deposits in the Shashi-Limpopo Confluence area (southern Africa). Journal of Archaeological Science. 102. 48–60. https://doi.org/10.1016/j.jas.2018.12.003.
See more FIT CTU research blogposts.
