Dwelling Extraction from Satellite Imagery in Refugee Camps: How Different Sample Data-sets Impact the Results of Convolutional Neural Networks (CNN)?

The presented study analyzes the impacts of different samples of training data-set on the results of the CNN-based technique for IDP camps mapping using World view image and provides an accuracy comparison with ground truth database. Different training data-set containing of two main sample images of our target object and the other objects as non-target samples were prepared. The target samples consists of three considered IDP camps namely large buildings, Tent (tunnel shape) and Tent (rectangular shape). Sample images of bare soil with sparse vegetation, other buildings and structures were considered as the non-target. Both target and non-target samples were used in order to train the structured CNN.