Extracting archaeologically-relevant information from imagery in manual and (semi-) automated ways

Abstract

Today, there is a broad consensus among archaeologists that remote sensing encompasses a range of useful observational techniques for discovering and registering archaeological traces and landscape patterning. Since remote sensing is a generic name given to all methods that use propagated signals to observe the Earth’s surface from above, many different remote sensing systems exist, and they can all be characterised as imaging versus non-imaging, passive versus active, optical versus non-optical, and airborne versus spaceborne. Given the long history of aerial photography in archaeology, this talk will target only products from passive airborne optical imaging approaches. More in particular, the presentation will kick off by considering the multi-dimensional nature of images and illustrate how both the geometrical and spectral image dimensions can yield new data products that might enhance (or even enable) the manual archaeological interpretation of the scene depicted in the image. At this point, the presentation will bridge from this rather traditional interpretative mapping approach to the desire for more automation using pattern recognition techniques. Building upon the principles of the first part, the underlying rationale of pattern recognition and its reliance on image features will be explained. The talk will state the difference between (and strike a blow for) hand-crafted image features used in traditional pattern recognition approaches versus those that are concocted during deep learning within the ubiquitous artificial neural architectures. A handful of epistemological and methodological considerations concerning deep archaeological remote sensing learning will top of this talk.