EXPLORING THE CAPABILITIES OF GENERATIVE ADVERSARIAL NETWORKS FOR SENTINEL-2 IMAGES GENERATION

Deep learning is a branch of artificial intelligence (AI) focused on developing algorithms and models that can learn from vast amounts of data to perform tasks that were previously exclusive to humans or difficult to manage with classical algorithms. Deep learning algorithms have been used for various applications, including image and speech recognition, language translation, and more. One of the latest models of deep learning is the Generative Adversarial Network (GAN). GANs are used for unsupervised learning to generate new data samples that resemble an existing training data distribution. GANs have primarily been used for generating 8-bit grayscale or RGB images of everyday subjects such as faces, animals, and more. However, the generation of Earth-observed data is still a less explored field. This thesis investigates the application of GANs to multispectral satellite images, focusing on Sentinel-2 images. Generated satellite images can be used for various purposes, including data augmentation, image-based analysis, simulation, and testing. For instance, synthetic images can be used to augment available training datasets, leading to improved performance of other machine-learning models. They can also be used to analyze the distribution of features or patterns in real data. Moreover, GANs can be used to simulate different scenarios for testing and evaluating machine-learning models, as it can be difficult to find real data with specific characteristics. In conclusion, synthetic satellite images generated by GANs can be useful in various contexts, including military applications, where they can be used to create highly realistic fake imagery to conceal important targets or regions of interest