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)
Publication date: February 2019
Source: Journal of Archaeological Science, Volume 102
Author(s): Olaotse Lokwalo Thabeng, Stefania Merlo, Elhadi Adam
Archaeological prospection through remote sensing is based on the contrast between areas of archaeological interest and their surroundings. It has been used as the cheapest and the fastest way of locating and documenting areas of archaeological interest since the 1920s, initially with the aid of film-based aerial photographs. In recent years, there has been a shift towards the use of multispectral satellite data in prospecting for archaeological sites because of their ability to give information on spectral characteristics of archaeological material beyond the visible spectrum. However, spectral signatures for identifying archaeological sites are not universal, and an assessment of the applicability of remote sensing techniques in different archaeological landscapes is needed. This study tests the feasibility of prospecting for archaeological sites previously occupied by farming communities in the Shashi-Limpopo Confluence Area of southern Africa, using very high-resolution satellite WorldView-2 images. It also assesses the performance of advanced classification algorithms (support vector machine and random forest) and the contribution of new WorldView-2 bands in detecting archaeological sites. Two independent accuracy assessments were carried out, using a data set collected by Huffman (2011, 2009a) and a randomly generated holdout test dataset, respectively. The results demonstrate the potential of remote sensing methods in prospecting for archaeological sites previously occupied by farming communities using very high-resolution satellite images and advanced classification algorithms. Very high overall accuracies were achieved by: random forest, 95.29% using holdout sample and 97.71% using independent dataset; support vector machine, 88.82% using holdout sample and 95.88% using independent dataset, respectively. The new WorldView-2 bands were of least importance (compared to traditional bands) in detecting sites in Shashi-Limpopo Confluence Area. Despite high classification accuracies achieved by both classifiers, there were some misclassifications between vitrified dung sites and river sand. Therefore, to address this problem, this study recommends the use of robust classifiers such as object-based algorithms because of their ability to segment an image into homogenous objects and classify them using a combination of spectral, textual, sub-pixels, spatial, relational and contextual methods.