ConLoRA-Sat: Parameter-Efficient Synthetic Satellite Imagery Generation for Data-Scarce Humanitarian Contexts

Seyed Amir Muhammad Qureshi, Carlos Granell , Sven Casrteleyn

Data scarcity constrains Geospatial Artificial Intelligence (GeoAI) in regions such as the Horn of Africa. One example is United Nations peacekeeping training, which needs fictitious yet geographically plausible terrain rather than identifiable real locations. This article presents ConLoRA-Sat, a framework that generates satellite imagery with controllable land cover, together with a segmentation map for every generated image. The framework pairs Low Rank Adaptation (LoRA) with ControlNet on a Stable Diffusion XL backbone, adapting a general-purpose generative model to 0.5 m imagery from Mogadishu and Addis Ababa using only 322 training images. A segmentation model trained exclusively on synthetic imagery retains 90.1% of the real-data baseline, and supplementing 200 real images with 600 synthetic ones recovers 92.5% of full-data performance. In a perceptual study with 53 satellite imagery experts, synthetic images were misclassified as real in 20.8% of trials, and all three peacekeeping practitioners consulted stated that they would use the imagery to design training scenarios. The framework is released alongside an open pipeline and a curated dataset of more than 4,300 Horn of Africa image chips with paired segmentation maps and structured captions, offering a pathway
for GeoAI in under-represented regions.