Pipeline for synthetic remote sensing data generation using Unreal Engine
Linus Ziesel, Lilli Belle Wanner, Jonas Mispelhorn, Marit Zenker, Andreas Michel, Jannick Kuester · 2025
The effectiveness of AI models in remote sensing critically depends on the availability of large, annotated datasets. However, in many application areas such as defense, real-world data is expensive, restricted, or unavailable. To address this limitation, we present a highly customizable synthetic data generation pipeline based on the Unreal Engine. The workflow enables the creation of photorealistic satellite imagery and automatically provides annotations in formats such as COCO Panoptic. Key features include a modular spawn point system for dynamic scene composition and a dedicated graphical user interface (SimUI), specifically designed for configuring data generation. Our pipeline not only provides a solution for generating synthetic data of high diversity and visual fidelity, but also directly addresses the critical bottleneck of usability. The complexity of the native Unreal Engine interface represents a substantial barrier to effective use, particularly for researchers without prior experience in game engines. By lowering this barrier, the proposed system makes advanced synthetic data generation more accessible and practical. In a user study with ten participants, the pipeline reduced the time required to generate a dataset of 800 images with diverse aircraft formations to an average of under ten minutes, compared to several hours of training typically needed for the native Unreal Engine interface. These results demonstrate that the proposed pipeline combines scalability, realism, and usability, thereby offering an effective solution to overcome the data scarcity bottleneck in remote sensing object detection.