Object-Based Integration Using Deep Learning and Multi-Resolution Segmentation in Building Extraction from Very High Resolution Satellite Imagery

Abstract

In recent years, there has been an increase in studies on the analysis of urban areas and the detection of changes in a fast and reliable way. In this respect, the classification of buildings is one of the prominent current issues of computer vision. As in many areas, the use of deep learning architectures is among the trending applications. Semantic segmentation applications have become widespread by using convolutional neural networks (CNN) to determine the building footprint. However, at the beginning of the problems encountered in the prediction images obtained after segmentation processes with deep learning, the noise formed by the effect of salt and pepper comes. In this study, the integration of the use of U-Net and SegNet algorithms, which are among the state-of-the-art CNN architectures, with the Object-Based Image Analysis (OBIA) and Multi-Resolution Segmentation (MRS) algorithm is used. Experiments were performed on the open shared Wuhan University Building Inference Dataset (WHUBED) consisting of very high-resolution satellite images (Gaofen-2, Worldview-2 and Ikonos). The model in the study, provides improvements in overall accuracy, F1 score, Dice score and Intersection over Union (IoU) metrics over the prediction results obtained using CNN alone. Building footprint maps obtained by building extraction are presented in the last section as comparative images. © 2025 Elsevier B.V., All rights reserved.