Generating Synthetic Satellite Imagery Using Generative AI and Geospatial Vector Reference Data for Fictitious Country Simulations

Read full text Summarise with VIRAT
Authors: Miguel Angel Lira Mijares
Cited by: 0
References: 22

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This work consists of a guideline on how to utilize Artificial Intelligence, more specifically, Generative Adversarial Network (GAN), to produce synthetic highresolution satellite imagery from vector maps of Land Cover-Land Use and additional objects or features in the shape of polygons. This workflow is based on the Zhu & Kelly (2021) work, with the difference that this thesis project was specifically focused on creating a realistic synthetic image from a place with no real refence unlike the original work were reference of existing places exists for their outputs. This research was developed within the Academia Partnership Program between UJI and the United Nations Global Service Centre (UNGSC), who provided technical input and expressed interest in the results. The proposed workflow successfully generates synthetic satellite image tiles and a coherent mosaic representation of a fictitious region. The resulting outputs are produced in standard image formats, their integration into common geospatial and remote sensing software environments for visualization, analysis, and further processing is possible by performing georeferencing of the fictitious area’s landmarks. Quantitative evaluation using non-referenced validation methods such as BRISQUE, CLIPIQA+, PI, and TOPIQ indicates that color-guided synthetic imagery (Z2) achieves consistently higher objective quality scores than non-color-guided outputs (Z1), resulting in visually clearer tiles and improved spatial consistency across mosaics. Qualitative differences were also observed, with Z1 preserving texture characteristics more closely aligned with real satellite imagery at comparable resolution, while Z2 produced sharper, perceptually higher-resolution outputs. However, a complementary user study indicated that participants were generally capable of distinguishing synthetic imagery from real satellite images, as reflected by consistent response patterns and systematic perceptual separation between image types. Nevertheless, as many projects do, this thesis project faced some challenges and limitations. To recall some of them, the seamless effect of the mosaics created with the output tiles was not possible, the open data sources were not updated or had some mismatches when combining the features to create the vector map, causing some misalignment and noise in the output synthetic tiles images. Recommendations for future improvements and research were suggested at the end of this current workflow.

Fields of Study & Subjects

Generative Adversarial Networks and Image Synthesis Satellite Image Processing and Photogrammetry Remote-Sensing Image Classification

Publication details

Journal / source
Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT)
Publication date
2 March 2026
Language
EN
Document type
Thesis