Comparison of Semantic Segmentation Deep Learning Methods for Building Extraction

Abstract

Urban planners use building extraction on satellite imagery to support government policies. However, the complex depiction of buildings in satellite imagery makes building extraction difficult. One way to extract buildings in satellite imagery is by semantic segmentation deep learning. This study aims to find a suitable deep learning semantic segmentation method by comparing the performance of UNet, UNet++, DeepLabV3, and DeepLabV3+ that combined with ResNet-101 and ResNet-50 as feature extraction algorithms and trained on two public datasets with different characteristics. UNet++ produces the highest performance for predicting both datasets, but with different feature extraction algorithms. MBD feature extraction is more suitable using ResNet-101 while AICrowd uses ResNet-50. However, if we consider time-consuming, DeepLabV3+ and UNet are more efficient for training building datasets because of consuming less time with quietly performance