Enhancement Strategies For Copy-Paste Generation & Localization in RGB Satellite Imagery

Satellite imagery is widely used for various applications, such as land-cover classification, field delineation, and environmental monitoring. However, satellite images can also be subject to malicious manipulation, such as copy-paste attacks, where a region from one image is pasted onto another to create a fake scene. Due to the different processing chains that characterize their lifecycle, the multimedia forensics community developed specific tools for analyzing RGB satellite images. Among the characteristics that differentiate satellite images from standard digital pictures, their dynamic characteristics have still not been investigated. In this paper, we study the effect of different data normalization techniques for the generation and analysis of copy-paste attacks on RGB satellite imagery. We show that these techniques boost deep learning tools developed for copy-paste localization, as they promote the extraction of satellite attribution traces. However, they could also help hide these clues, making the detection task more challenging. Our results suggest that careful attention and precise consideration of the data lifecycle should be given when analyzing data modalities different from standard digital pictures.