Convolutional Neural Network for Building Extraction from High-Resolution Remote Sensing Images

Abstract

Buildings are one of the most important components of the city, and their extraction from high-resolution remote sensing images is used in a wide range of applications such as urban mapping. Due to the complex structure of highresolution remote sensing images, automatic extraction of buildings has been a challenge in recent years. In this regard, fully convolutional neural networks (FCNs) have shown successful performance in this task. In this research, a method is proposed to improve the famous UNet network. In classical UNet model high-level rich semantic features are fused with low-level high-resolution features with skip connection for pixel-based segmentation of images. However, the fusion of encoder features with features in corresponding decoder part causes ambiguity in segmentation results because low-level features produce high noise in high-level semantic features. We introduced the embedding feature fusion (EFF) block for enhancing the fusion of low-level with high-level features. For performance evaluation, a publicly available data provided with United States Geological Survey (USGS) high-resolution orthoimagery with the spatial Resolution ranges from 0.15m to 0.3m was used in comparison with several state-of-the-art semantic segmentation model. Experimental results have showed that the proposed architecture improves in extracting complex buildings from high resolution remote sensing images.