Prospecting archaeological mining sites using high resolution imagery and advanced classification algorithms in Ziwa, Zimbabwe

Abstract

Prospection of archaeological sites using remote sensing and machine learning technology is premised on the fact that the spectral signature of a similar archaeological feature is equivalent over a landscape defined by an archaeologist. Therefore, known sites can be used for the identification and location (prospection) of unknown sites. Archaeological prospection has been applied over different types of landscapes, initially with the aid of traditional field surveying, kites, aerial photography, satellites, and recently through drones mounted with high resolution sensors. Use of high-resolution satellite imagery from multispectral sensors on low earth orbit vehicles has taken the lead as it has scored success in fulfilling certain objectives in specific areas through its extensive surveying abilities and reduced costs in comparison with others. This study focused on the ability of advanced classification algorithms Random Forest (RF) and Support Vector Machine (SVM), methods that have proved effective for Iron Age (IA) archaeological sites in this region, and for prospecting at the archaeological site of Ziwa, an iron age site dated between 1500 and 1700A.D. found in the north-eastern highlands of Zimbabwe using free Sentinel-2 imagery. The performance of these machine-learning classification algorithms was assessed in order to ascertain which model performed better. Two independent accuracy assessments were done, one for SVM and another for RF using randomly selected test data that were collected at the Ziwa site. The results indicated a very high potential for using remote sensing methods in prospecting for archaeological mining sites as has been the case with the identification of kraals and vitrified dung (Thabeng et al. 2019). For the first set of classifications, satisfactory overall accuracies of 80.67% with a Kappa coefficient of 0.78, and 80.17% with a Kappa coefficient of 0.7724 were achieved for RF and SVM respectively. On comparing the performance of the two individual classifiers using the MacNemar’s test, a z value of 1.692 was obtained with a p value ≤ 1.96. This indicated that there was no significant difference between the two classifiers in discriminating the land use and land cover (LULC) classes built-up-area, bare-land, cultivated fields, enclosures, grassland, metalliferous-surface, non-metalliferous-surface, pit-structures, rock-outcrop, terraces, tarred road, water body, wooded vegetation and stone walls (terraces, pit structures and enclosures). In the second set of classifications, satisfactory overall accuracies 80.25% with a Kappa coefficient of 0,77, and 80.62% with a Kappa coefficient of 0.78 were achieved for RF and SVM respectively. On comparing the performance of the two individual classifiers using the MacNemar’s test, a z value of 2.289 was obtained with a p value ≥ 1.96. This indicated that after aggregating classes there was a significant difference between the performance of the two classifiers in discriminating the land use and land cover (LULC) classes. The ore grinding sites (approximately 60 cm in diameter), smelting areas/furnaces (<1-1.5m in diameter) and slag (approximately 3-5cm in diameter) however could not be detected due to their size falling way beyond Sentinel-2 10m resolution. To address this problem, this study recommends use of LiDAR or finer resolution imagery in combination with object-based classification as this has the ability to segment very small features with distinct textures, shape and signatures. While RF and SVM clearly indicated the landscape of the archeological interest (Area of interest ‘AOI’ around 36 km2) within the broader investigated area (surroundings of the AOI) and could discriminate between metalliferous surfaces and stonewall structures, it however did not precisely discriminate the individual archaeological features such as pit structures, enclosures and terraces within the 36 km2 under investigation. Remote sensing based archaeological prospection may not guarantee discrimination of the finer details of archaeological remains but may be used to understand location of a possible site. Furthermore, archeological prospection using the pixel-based methods has shown the ability to discriminate features of different physical composition (metalliferous surface and stonewalls) while it falls short in discriminating different features made from the same material such as pit-structures, enclosures and terraces