Comparison of SVM and boosted regression trees for the delineation of lacustrine sediments using multispectral ASTER data and topographic indices in the lake Manyara Basin

The lower member of the so called Manyara Beds is a distinct lacustrine sedimentary layer which indicates, with an elevation of more than 140 m above today's lake level, a high stand of the paleolake Manyara in the Monduli District in northern Tanzania. The Manyara Beds are rich in Pleistocene vertebrate fossils. In this study we focus on the delineation of this specific stratigraphic layer in order to yield new insights into paleontological settings, landscape evolution and to plan paleontological fieldwork. We compare the performance of a support vector classifier with a linear as well as a Gaussian kernel, with boosted regressiontree approaches to identify the lithostratigraphic layers of the Manyara Beds. For the identification of the lacustrine sediments, multispectral informationof ASTER satellite imagery and topographic indices derived from a digital elevation model were utilized as input feature sets. Acceptable classification accuracies were obtained with all methods. Thus, the Manyara Beds can be delineated and new sites with paleolake sediments were detected. The highest overall accuracy with 92% was provided by the support vector machine approach with a linear kernel for a binary classification problem. For a multi-class classification problem with three target classes the support vector classifier achieved 80% accuracy with a linear, as well as a Gaussian kernel. © 2015 E. Schweizerbart'sche Verlagsbuchhandlung, Stuttgart, Germany.