Semantic Image Segmentation for Building Detection in Urban Area with Aerial Photograph Image using U-Net Models

Abstract

Detecting building location distribution in an urban area has been a concern of city government in many developing countries as a basis for city planning and development. In recent years, deep learning has gained research attention as the most attractive approach to address classification in the remote sensing field. One application of deep learning is a semantic image segmentation method whose aim is to classify each pixel in the image into a predetermined set of labels. In this experiment, the objective of semantic image segmentation is building detection in urban areas using a deep learning model in which each image pixel is categorized into either building or non-building label. Based on experimentation using aerial photograph imagery of Pasar Minggu Sub-District, South Jakarta City District, DKI. Jakarta Province and UNet model achieved 0.83 average training accuracy and 0,87 testing accuracy