ConLoRA-Sat: High-Resolution Synthetic Satellite Imagery for UN Peacekeeping Using ControlNet and LoRA

Read full text Summarise with VIRAT
Authors: Seyed Amir Muhammad Qureshi
Cited by: 0

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

Satellite imagery is essential for United Nations peacekeeping operations, yet data scarcity in regions such as the Horn of Africa limits the effectiveness of training programs for mission personnel. This thesis presents ConLoRA-Sat, a framework for generating synthetic satellite imagery by combining Low-Rank Adaptation (LoRA) and ControlNet architectures with Stable Diffusion XL (SDXL). The framework addresses the dual challenges of adapting to satellite imagery characteristics and enabling precise spatial control through land cover segmentation maps. Trained on image chips from Mogadishu and Addis Ababa at 0.5m resolution, ConLoRA-Sat achieves strong performance across multiple evaluation dimensions. Domain-specific evaluation demonstrates 88.8% segmentation consistency, indicating that generated imagery preserves the semantic structure specified by input conditioning maps. Downstream task evaluation shows that models trained exclusively on synthetic imagery retain 90.1% of real-data performance, while augmenting real data with synthetic imagery improves baseline performance by 2.0%. In low-resource scenarios with only 25% real training data, synthetic augmentation recovers 92.5% of full-data performance. Human evaluation with domain experts revealed a 20.8% fooling rate in distinguishing synthetic from real imagery, with rural scenes achieving the highest fooling rate of 27.1%. The framework enables generation of unlimited training scenarios with known ground truth for the fictitious country Carana used in UN peacekeeping exercises, addressing operational security concerns while maintaining regional authenticity. Requiring only 24 GPU hours for training and 12.5 GB storage for deployment, ConLoRA-Sat demonstrates that parameter-efficient adaptation of foundation models offers a practical solution for generating synthetic satellite imagery in data-scarce humanitarian contexts.

Fields of Study & Subjects

Generative Adversarial Networks and Image Synthesis Advanced Neural Network Applications Remote-Sensing Image Classification

Publication details

Journal / source
Repositori UJI (Universitat Jaume I)
Publication date
2 March 2026
Language
EN
Document type
Thesis