The importance of building footprints and their inventory has been recognised as foundational spatial information for multiple societal problems. Extracting complex urban buildings involves the segmentation of very high-resolution (VHR) earth observation (EO) images. U-Net is a common deep learning network and foundation for its new incarnations like ResUnet, U-Net++ and U-Net3+ for such segmentation. The re-incarnations look for efficiency gain by re-designing the skip connection component and exploiting the multi-scale features in U-Net. However, skip connections do not always improve these networks and context information is lost in the multi-scale features. In this paper, we propose three novel dual skip connection mechanisms for U-Net, ResUnet, and U-Net3+. This deepens the feature maps forwarded by the skip connections to find a more accurate trade-off between context and localisation within these networks. The mechanisms are evaluated on feature maps of different scales in the three networks, producing nine new network configurations. The networks are evaluated against their original vanilla versions using four building footprint datasets (three existing and one new) of different spatial resolutions: VHR (0.3m), high-resolution (1m and 1.2m), and multi-resolution (0.3+0.6+1.2m). The proposed mechanisms report efficiency gain on five evaluation measures for U-Net and ResUnet, and up to 17.7% and 18.4% gain in F1 score and Intersection over Union (IoU) for U-Net3+. The codes will be available in a GitHub link after peer review.

翻译：暂无翻译