||There is some debate whether or not automated procedures of object detection in images can be successfully applied to archaeology. In spite of recent advances of automated object detection in other fields – e.g., detecting people, faces, traffic signs etc. in different kinds of images – little progress has been made in detecting archaeological features in remotely sensed images. While there are some successful case studies, they are few and far between, and are usually limited to particular sites without attempts towards wider application. As it seems, the widely held belief that the great variation of the archaeological record prevents automated detection has led to a lack of investigations in this field.
Meanwhile, the amount and variety of remotely sensed images of potential interest for archaeology is increasing rapidly. As prices are decreasing, more and more of these images become affordable for archaeological projects. If they are to be efficiently used, visual image interpretation must be complemented by automated procedures for routine tasks, such as scanning large areas for typical, recurrent archaeological objects.
With such a purpose in mind, we decided to use recent fieldwork in the Silvretta Alps, on the Swiss-Austrian border, as a case study to explore the potential of high-resolution satellite images and automated object detection to assist archaeological fieldwork. An important goal of this project is to trace back alpine pastoralism, a specialized economic system involving summer grazing of livestock above the tree line, to its prehistoric origins. To this end, the archaeological survey in the Silvretta mountains (540 sqkm, 1500 to 3400 masl) focuses on ruins of huts, cattle compounds, and other remains of alpine pastoralism. The recorded structures serve as ground truth to develop automated methods to detect them in remotely sensed images. While each structure is unique, they share a limited range of geometries in terms of size and shape, making them a suitable test case for automated detection. Our goal is to identify sites highly likely to contain architectural remains of our interest to guide archaeological fieldwork. For the time being, other categories of archaeological objects, such as rock shelters and fire places, remain outside the scope of this study.
In September 2011 we acquired five Geoeye 1 images of our study area, featuring four spectral (VNIR) and a panchromatic band. After pansharpening, all bands have a spatial resolution of 0.5 m. The first step towards object detection is texture segmentation, i.e. filtering out textural regions. Our target objects occur in open areas, so that filtering out urban, forested and rocky regions greatly reduces the area to be searched. Segmentation is achieved by using mathematical morphology to measure texture contrast. In the second step local image features are extracted from the remaining areas of interest and grouped into larger curvilinear features. To each image point, the following search then assigns likelihood values of forming rectangular or convex structures corresponding to archaeological objects of our interest. This likelihood map is then used to visually validate archaeological objects in the images and in the field.