Abstract
This study proposes a novel deep learning method for remote sensing imagery domain adaptation for the purpose of building vectorization. The method, which uses a Generative Adversarial Network to extract chromatic data from the target image and apply it to the source imagery tiles, was separately trained on the SpaceNet AOI 2 - Las Vegas dataset as well as the SpaceNet AOI 10 - Dar Es Salaam image and applied to process the CrowdAI dataset, which was then used to train the state-of-the-art HiSup building vectorization network. Finally, the SpaceNet AOI 2 - Las Vegas and a dedicated Dar Es Salaam evaluation dataset were used to test HiSup’s vectorization performance. When tested on Dar Es Salaam, the proposed method improves upon state-of-the-art domain adaptation approaches such as ColorMapGAN, Reinhard color transfer and CycleGAN by 17% in Average Precision, 14% in Average Recall, 7.2% in Intersection over Union (IoU) and 6.4% in Complexity Aware IoU (C-IoU). A combination of existing methods was outperformed by 14% in AP, 9% in AR, 4% in IoU and 4% in C-IoU.
Paper Type
Full Paper
DOI
10.62036/ISD.2026.200
Remote sensing imagery domain adaptation by chromatic data transfer for building vectorization
This study proposes a novel deep learning method for remote sensing imagery domain adaptation for the purpose of building vectorization. The method, which uses a Generative Adversarial Network to extract chromatic data from the target image and apply it to the source imagery tiles, was separately trained on the SpaceNet AOI 2 - Las Vegas dataset as well as the SpaceNet AOI 10 - Dar Es Salaam image and applied to process the CrowdAI dataset, which was then used to train the state-of-the-art HiSup building vectorization network. Finally, the SpaceNet AOI 2 - Las Vegas and a dedicated Dar Es Salaam evaluation dataset were used to test HiSup’s vectorization performance. When tested on Dar Es Salaam, the proposed method improves upon state-of-the-art domain adaptation approaches such as ColorMapGAN, Reinhard color transfer and CycleGAN by 17% in Average Precision, 14% in Average Recall, 7.2% in Intersection over Union (IoU) and 6.4% in Complexity Aware IoU (C-IoU). A combination of existing methods was outperformed by 14% in AP, 9% in AR, 4% in IoU and 4% in C-IoU.
Recommended Citation
Tadesse, S. & Kulawiak, M.(2026). Remote sensing imagery domain adaptation by chromatic data transfer for building vectorization. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.200