Building segmentation from satellite imagery using U-Net with ResNet encoder
Zhongwei Liu, Chen Baisong, Ao Zhang
Abstract
As one of the most important types of artificial ground features, building images and outlines are widely used in map updating, GIS analysis, urban planning, and environmental modelling etc. Fast and accurate segmentation of buildings is one of the hot and difficult topics in remote sensing image processing research for many years. This paper tries to apply the U-Net deep neural network model with ResNet encoder to remote sensing image segmentation for building extraction. First, we obtain some remote sensing images through some open datasets, and then we outline the building images to get the building mask, then we divide the dataset into eight to two to form two datasets for training and validation, and then we use this dataset to train and validate the U-Net model. The results show that the model's MIoU (Mean Intersection over Union) reached 0.83, and the model achieved a good building segmentation effect. This model can be used for building segmentation with clear boundaries, however, there are also a few improvements to be resolved in further studies, more accurate results, more straight outlines, less misclassification, etc.