The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B2-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2020, 383–390, 2020
https://doi.org/10.5194/isprs-archives-XLIII-B2-2020-383-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2020, 383–390, 2020
https://doi.org/10.5194/isprs-archives-XLIII-B2-2020-383-2020

  12 Aug 2020

12 Aug 2020

LONG-SHORT SKIP CONNECTIONS IN DEEP NEURAL NETWORKS FOR DSM REFINEMENT

K. Bittner1, L. Liebel2, M. Körner2, and P. Reinartz1 K. Bittner et al.
  • 1Remote Sensing Technology Institute, German Aerospace Center (DLR), Wessling, Germany
  • 2Computer Vision Research Group, Chair of Remote Sensing Technology (LMF), Technical University of Munich (TUM), Munich, Germany

Keywords: Conditional generative adversarial networks (cGANs), balancing hyper-parameters, long-short skip connections, 3D scene refinement, building geometry

Abstract. Detailed digital surface models (DSMs) from space-borne sensors are the key to successful solutions for many remote sensing problems, like environmental disaster simulations, change detection in rural and urban areas, 3D urban modeling for city planning and management, etc. Traditional methodologies, e.g., stereo matching, used to generate photogrammetric DSMs from stereo imagery, usually deliver low-quality results due to the matching errors in homogeneous areas or the lack of information when observing the scene under different viewing angles. This makes the tasks related to building reconstruction very challenging since in most cases it is difficult to recognize the type of roofs, especially if overlaid with trees. This work represents a continuation of research regarding the automatic optimization of building geometries in photogrammetric DSMs with half-meter resolution and introduces an improved generative adversarial network (GAN) architecture which allows to reconstruct complete and detailed building structures without neglecting even low-rise urban constructions. The generative part of the network is constructed in a way that it simultaneously processes height and intensity information, and combines short and long skip connections within one architecture. To improve different aspects of the surface, several loss terms are used, the contributions of which are automatically balanced during training. The obtained results demonstrate that the proposed methodology can achieve two goals without any manual intervention: improve the roof surfaces by making them more planar and also recognize and optimize even small residential buildings which are hard to detect.