Building extraction from high resolution remote sensing images based on improved U2-Net

Abstract

Buildings are one of the most important infrastructures in cities. Automatic extraction of buildings from high-resolution remote sensing imagery is of great significance for urban management and population estimation. Aiming at the problem of insufficient edge features and loss of details in the building extraction results, a network model combined with Canny edge detection and convolutional block attention module based on U²-Net was proposed in this paper. Adding the building edge feature map to the U²-Net to compensate for the problems of insufficient extracted building edge features and loss of detail. Using the convolutional block attention module to achieve effective feature extraction of buildings. Qualitative and quantitative analyses were performed on the WHU building datasets. The experimental results showed that improved U²-Net can accurately extract building regions from remote sensing images. And for the problem that deep learning network relies on training samples in the process of building extraction, this paper discusses the influence of the number of samples on the results. The experimental results showed that a reasonable setting of the number of building samples can improve the extraction accuracy of buildings, not that the greater the number of samples, the higher the accuracy.