Synthetic AI-Generated Satellite Imagery to Improve Earth Observation-Based Neural Networks

Recent advances in satellite technology have significantly progressed, yet acquiring high-quality images with meaningful labels for Earth observation missions remains a costly and time-intensive process. Furthermore, captured scenes frequently exhibit defects such as misaligned color channels, extensive cloud cover, or repetitive patterns in similar environments. Fortunately, the evolution of generative artificial intelligence offers a solution by enabling the creation of realistic synthetic scenes, simulating the characteristics of any targeted imager, and thereby mitigating the scarcity of authentic data. This paper demonstrates the feasibility of transferring knowledge from specialized AI-generated datasets to Earth observation missions. Leveraging a novel dataset of Spanish map tiles, Pix2Pix, CUT, and ControlNet models were implemented to synthesize satellite imagery. To analyze structural and topological generalizability, identical U-Net instances were trained on the resulting collections for building, road, and water segmentation tasks, and subsequently tested on independent authentic imagery. The results reveal a clear decoupling between visual realism and functional utility. Incorporating synthetic samples into hybridized training datasets successfully surpassed the limitations of using real data alone, increasing maximum Dice scores by 0.9% (to 54.1% for buildings), 2.3% (to 38.6% for roads), and 4.1% (to 46.5% for waterbodies). This systematic validation establishes structural-guided synthetic data augmentation as a robust, adaptable strategy for Earth observation applications across diverse sensors and geometric objectives.