MSBR-GNet: A High-Resolution Imagery Generative Optimization Model for Building Rooftop Boundary Guided by Interpretable Statistical Model in Spatial and Spectral Domain

Abstract

Automated extraction of building rooftop information is of great significance in remote sensing of land resources and other related applications. In this article, a building roof boundary generating optimization model called multiscale boundary regulation generative net (MSBR-GNet), guided by interpretability statistical model in the spatial and spectral domains, is proposed to solve the problem of inaccurate boundary segmentation caused by mixed pixel transition region of remote sensing images. Incorporate the boundary loss function guided by statistical models in the spatial and spectral domains into the generator loss calculation of MSBR-GNet, precisely constrain the regularized generation of building rooftop contours by interpretable mechanism. The experiments show that MSBR-GNet can extract more regular building rooftop contours, and the precision values in the INRIA, WHU, and Massachusetts public datasets reached 0.9275, 0.9228, and 0.8779, respectively, which can ensure the accuracy of building extraction while achieving optimal results in the boundary morphology evaluation index.

Keywords

• Building rooftops
• contour optimization
• edge transition zones
• generative networks
• instance segmentation
• interpretability