Deep learning to detect built cultural heritage from satellite imagery.-Spatial distribution and size of vernacular houses in Sumba, Indonesia

In Sumba Island - Indonesia, the implantation of vernacular houses, inside and outside traditional villages, is considered to be an efficient proxy for the on-going complex cultural transformations resulting from globalization. This study presents an easily reproducible workflow allowing buildings to be automatically detected from satellite imagery, demonstrating how modern computer vision methods based on deep learning can help in this task, which would be far too time-consuming when undertaken by hand. Eight deep learning architectures based on convolutional neural networks were compared in terms of ability to identify and locate precisely traditional houses from satellite images. By combining a Faster R-CNN ResNet 101 architecture with artificial data augmentation, the model was taught properly using 1033 instances of vernacular houses (AP@.50:.95 = 71.9). Once 14 952 traditional houses had been detected, the Histogram of Orientated Gradients (HOG) was computed and processed by several machine learning algorithms to assess their surface area, as this parameter conveys pertinent information about the economic and political position of the householder. The best classifier was found to be a support vector regressor (SVR, R 2 = 0.88), although the other classifiers tested also provided good results (R 2 > 0.76). Spatial analysis was used to draw conclusions from an anthropological / cultural identity point of view. More generally, these techniques not only offer a simple increase in recording capabilities for tangible cultural heritage, they open up new research perspectives, at greater scales. (c) 2021 Elsevier Masson SAS. All rights reserved.