Picture yourself standing outside beneath the unrelenting Middle Eastern sun. The thermometer shows 120 degrees Fahrenheit, but it feels a lot hotter. There are no trees for refuge, just scorched earth. Not even the local villagers dare to go outside for long in these conditions. You are an archaeologist, though, and you have been tasked with identifying and making a map of as many ancient sites as you can in an 8-hour work day.
Site maps like the one you are producing are essential both in and outside the field of archaeology. Archaeological sites all over the world are destroyed every day through warfare, looting, as well as urban expansion. Local authorities use maps of site locations to prevent the irrevocable destruction of cultural heritage and data about the human past. In addition to their scientific value, the excavation and promotion of archaeological sites can feed tourism revenue via museum construction and site visits.
For much of the history of archaeology, the only way to accurately make such a map was to physically hike over the terrain, spotting archaeological sites with your eyes: concentrations of ceramic sherds, slight soil discolorations, elevated surfaces, as well as other tell-tale signs. As you might imagine, it can take years to systematically survey entire regions. Even then, archaeological site discovery is often slow and haphazard. Hope you brought a good hat and plenty of water!
Using Satellite Imagery for Archaeological Site Discovery
In many regions, however, it is simply impossible to perform this type of systematic survey from the ground alone within a reasonable time frame. Around the rich Mesopotamian settlement centers, for instance, there are hundreds of kilometers to survey. In Mesoamerica, jungle terrain is treacherous and makes it nearly impossible for archaeologists to find sites that have not already been found. As a result, large swaths of land are still relatively unexplored, with countless archaeological sites waiting to be discovered, protected, and excavated.
One promising recent approach for overcoming these hurdles has involved moving archaeological research from the earth to outer space. Satellite imagery has been productively used to solve a wide variety of problems in different domains--from predicting crop yields for commodity futures trading, to assessing environmental conditions for disaster mitigation.
In archaeology, the primary use of satellite images is for identifying archaeological sites. Satellite images are not only "pictures" in the visible color spectrum, but also capture varying wavelengths that can be used to identify different soil types (based on their moisture retention and vegetation growth), as well as elevations. ASTER images, for instance, contain 14 different wavelength bands, that can help make human-made, archaeological features pop out from the surrounding landscape.
Take the site of Teotihuacan in Mexico, for example (see the image below). Archaeological features such as temple platforms and ancient roads stand out from the surrounding landscape in this false color ASTER image:
Finding sites with satellite images can speed up archaeological discovery over large, unexplored regions, allowing archaeologists to pinpoint locations of interest before even setting foot in the field. "Space Archaeologist" Sarah Parcak, for instance, discovered a new possible Viking settlement in Newfoundland this past year that might not have been found without her manually searching over satellite imagery. For more on the promise of satellite imagery in archaeology, check out Parcak’s recent TED talk here.
While manually evaluating satellite imagery allows archaeologists to narrow in on possible archaeological sites, it still takes a great deal of skilled labor and time to pore over satellite images and identify the nuanced features of archaeological sites. For this reason, some space archaeologists have applied machine learning techniques in an effort to streamline the satellite survey process and make finding archaeological sites on the ground even faster and easier.
Automating Archaeological Site Prediction with Machine Learning Techniques
One of the most prominent recent examples of this machine learning approach is in war-torn Northeastern Syria. Bjoern Menze and Jason Ur recognized that archaeological site prediction using satellite imagery is not unlike other image-based machine learning domains. For instance, researchers who detect cancer tumors in medical imaging ask similar questions to those of space archaeologists. Are there site or tumor features in a given image? If so, where are they? Menze’s CV is a testament to this similarity; he works on both archaeological site prediction and tumor detection! For this reason, in 2012, Menze and Ur translated existing machine learning techniques into the domain of archaeological site prediction.
Archaeological site prediction is fundamentally about distinguishing satellite image pixels that contain sites from those that do not. To form a training dataset, Menze and Ur labeled archaeological site locations based on 1680 Syrian sites that had been identified through a combination of on-the-ground archaeological survey and manual inspection of high resolution CORONA satellite imagery. Then, using ASTER satellite images to predict whether or not there was a site at a given pixel, they trained a Random Forest classifier to recognize the spectral signature of these archaeological sites, assuming that identified sites from the region share common features with unidentified sites.
Specifically, Menze and Ur used the following method:
1. Cut each ASTER image up into 6x6 km spatial blocks (for a total of 36 blocks per ASTER image).
2. For each ASTER image, train a random forest classifier on the data from all the blocks but one. The predictive features used to train the classifier were the 14 ASTER wavelength bands, each providing a continuous numerical value for each pixel in the satellite image. These features were used to predict a binary response: each pixel either contains an archaeological site, or it does not.
3. Use the trained classifier to predict the location of archaeological sites in the one hold-out block and map the probabilities of an archaeological site at each pixel—resulting in a continuous value from 0 to 1 for every pixel.
4. Repeat the process for each of the 36 blocks within each ASTER image, so that each block will be held out from the remaining blocks once.
5. Combine the probability maps formed by each hold-out block, to form a single probability map of archaeological site locations for each ASTER image.
6. To account for seasonal variability, Menze and Ur produced probability maps for ASTER images from the same image region at different points in the year. They averaged the probability maps for each image region to form a single time-averaged probability map for each region.
Menze and Ur then outlined high-probability regions on the probability map for each image—their predicted site locations. When all was said and done, they correctly predicted 94% of sites that had been identified on the ground in previous archaeological fieldwork. Applying their method to novel regions, they predicted over 14,000 additional site locations—more than 18 times more sites than have ever been identified by archaeologists in the region! While ongoing conflict with ISIS currently makes intensive ground survey impossible, this map could help future Syrian archaeologists strategically protect and document remaining sites when the region stabilizes. In the meantime, the authors produced a Google Maps version of their results, so you can zoom in and check out outlines of the predicted sites for yourself:
Work is ongoing to decrease the classifier’s predictive error and generalize the use of these methods around the world. For instance, one successful approach involves ranking and weighting probability maps from a single ASTER image region instead of time-averaging all probability maps from an image region (Menze and Ur 2014). By assigning less weight to probability maps that poorly predict known archaeological sites and more weight to successful probability maps, Menze and Ur found they could decrease predictive error. No doubt, as machine learning in space archaeology expands to other world regions, this weighting approach and others will be necessary to account for location-specific variation and train successful classifiers.
Machine Learning is still in its infancy in archaeology, but the approach holds enormous promise. Just as in finance and disaster mitigation, the use of satellite imagery has become a vital resource for archaeology. As more archaeologists around the world automate the process of identifying archaeological sites with satellite images, sites will be found at a faster rate than ever before. The application of machine learning to archaeology has the capacity to usher in an exciting new era of great archaeological discoveries. Stay tuned!