Using Deep Neural Networks for Automatic Building Extraction with Boundary Regularization from Satellite Images
MetadataShow full item record
The building footprints from satellite images play a significant role in massive applications and many demand footprints with regularized boundaries, which are challenging to acquire. Recently, deep learning has made remarkable accomplishments in the remote sensing community. In this study, we formulate the major problems into spatial learning, semantic learning and geometric learning and propose a deep learning based framework to accomplish the building footprint extraction with boundary regularization. Our first two models, Post-Shape and Binary Space Partitioning Pooling Network (BSPPN) integrate polygon shape-prior into neural networks. The other one, Region-based Polygon GCN (R-PolyGCN) exploits graph convolutional networks to learn geometric polygon features. Extensive experiments show that our models can properly achieve object localization, recognition, semantic labeling and geometric shape extraction simultaneously. The model performances are competitive with the state-of-the-art baseline model, Mask R-CNN. Especially our R-PolyGCN, consistently outperforms others in all aspects.