Synthetic data augmentation for earth observation object detection tasks

Neural networks have transformed remote sensing, making it easier to derive insights from satellite images for various Earth Observation (EO) applications. Yet, their potential is often limited by the lack of labeled data. Traditional data augmentation methods, while attempting to address this, require significant manual input and lack visual diversity, compromising model performance. We introduce an innovative data augmentation strategy that automates the generation and integration of objects into satellite imagery, enhancing datasets for object detection. Our method notably improves car detection model performance, surpassing traditional augmentation techniques.