Semantic segmentation of UAV images in archaeological sites using deep learning

Abstract

Archaeological data plays an important role in defining the identity of the world, creating its memory, steering the economy of cities, recognizing the cultural and social origins of humanity, and building the future well. The globalization world process we are in and the unconscious human actions of this era have begun to cause major difficulties in the preservation of archaeological sites. Archaeological site detection, identification, and documentation are essential processes in preservation strategies. Archaeological research, preliminary preservation, excavation, and city planning require these processes to be effectively performed. Technological methods reduce the manual survey effort and even provide sufficient data for the pre-detection and documentation processes. However, computer-aided methods should be adopted to analyze, process, and interpret this ever-growing dynamic digital data. Deep Learning algorithms provide the ability to process masses of data, build models from this data and interpret subsequent data. In addition, the models obtained from this data allow the interpretation of subsequent data and reintegrating of the results into effective interpretations. Accordingly, in this study, in this study semantic segmentation is performed to detect archaeological structures utilizing the transfer learning-based deep Convolutional Neural Networks (CNNs) approach. The dataset from the Ancient City of Syedra, whose archaeological research continues today, is used. The dataset consists of UAV images and contains ancient Roman structures that were manually labeled and field-verified to a large extent. The architecture of CNNs has provided powerful detector algorithms for segmenting archaeological structures on the acquired dataset. In the study, U-net and LinkNet models, two CNN architectures, which are ResNet34 and ResNeXt50 were implemented as encoders to the U-Net and LinkNet CNN models. In order to support faster convergence and stronger generalization capability of the models, 6 different model architectures were prepared with the encoders used.