Analyzing the Impact of Optimization Techniques on U-Net-Based Building Detection Performance

Abstract. As a result of the growth of cities and the increase in migration from rural to urban areas, the urban population has grown remarkably. Building extraction is important in many practical and strategic areas. Automatic detection of buildings from images is important in terms of both speed and preventing interpretation errors arising from the differences in experience of experts. One of the main difficulties encountered in studies on automatic building detection is the generalization problem originating from the difference in the characteristics of roofs in complex environments. Recently, software and hardware systems have gained great importance due to the development of new technologies. As a result of these innovations, studies on deep learning architecture have increased. The training of deep learning architectures aims to minimize the loss function during learning. There are many optimization algorithms based on various mathematical principles; however, an optimization algorithm that can be generalized to all problems and is optimal for all conditions is still not fully defined. Therefore, the studies in literature continue to be experimental. In this study, the effects of optimization techniques on the automatic detection of buildings with different roof types from aerial images using the U-Net architecture are analyzed. In this study, Adam, Nadam, and RMSprop optimization techniques were used. The effects of optimization techniques on classification performance were investigated by examining computational costs and performance metrics.