Generative Adversarial Networks for Artificial Satellite Image Creation and Manipulation

In the research conducted and reported herein, we demonstrate the ability to generate and manipulate synthetic satellite images by employing a set of deep learning architectures (DL),Synthetic satellite images offer a wide range of potential applications, including the generation of massive, labeled datasets for artificial intelligence (AI) applications, manipulated image detection, and natural disaster monitoring and detection. Although deep learning (DL) architectures have proven to be successful in creating and manipulating natural images and multimedia content, their application in generating synthetic satellite images has not been thoroughly investigated. This can be attributed to variations in semantic content, number of bands, bit, and spatial resolution. To tackle this problem, we will concentrate on two distinct methods of manipulation: global image modifications and local splicing. With an appropriately trained version of cycle GAN architecture, the objective of the land cover transfer is to effectuate a change in the land cover from modified images of vegetation to barren land and vice versa. By employing the pix2pix GAN architecture, the seasonal transfer technique achieves the transformation of a winter-acquired (summer-acquired) satellite image to its summer (winter) counterpart. We use the Alps dataset for seasonal transfer and the land cover dataset for land cover transfer to test the effectiveness of these methods. We describe two types of transformer-based image generation algorithms for local tampering: the image generative pre-trained transformer (iGPT) and the vision transformer. The techniques mentioned above utilize synthetic splices that are incorporated into non-uniform areas of the target image to create modified images, while also ensuring that the splices’ boundaries are not visible. Through the utilization of the SEN12MS and World databases, we have verified the effectiveness of said techniques. We use Sentinel-2 images to demonstrate the validity of the suggested methodologies and emphasize their possible applications and limits.