A carregar...
Publicação
Developing Fictitious Country Maps through Generative AI Techniques
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 4.37 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
This thesis explores the application of diffusion models to generate high-resolution maps for the fictional country of Carana, a scenario used by international peacekeeping organizations for training and strategic planning. The study addresses limitations of the manual creation of raster representations and limitations of existing imagery, which often lack the detail and adaptability required for various simulations. To achieve this, a comprehensive framework was developed, beginning with the acquisition of Sentinel-2 satellite imagery and preprocessing the data into 64x64 pixel tiles. A diffusion model, based on the U-Net architecture, was adapted to process these tiles, with training conducted on high-performance computing resources. Validation of the generated maps was performed using histogram-based analysis and Fréchet Inception Distance (FID) scores, with additional assessments focusing on spatial coherence and color consistency. While the framework successfully produced synthetic map tiles, challenges such as color mismatches and tile discontinuities highlighted areas for improvement. These challenges led to recommendations for future improvements, including testing various hyperparameters, advanced validation techniques and later on the integration of conditional generation using geospatial features.
Descrição
Dissertation submitted in partial fulfilment of the requirements for the Degree of Master of Science in Geospatial Technologies
Palavras-chave
Carana Diffusion models Generative AI Peacekeeping operations Sentinel-2