GENERATION OF SATELLITE IMAGES FROM TEXT USING GAN NEURAL NETWORK COMBINED WITH AN SBERT TRANSFORMER

Abstract

Satellite image generation is a complex task that involves not only processing large volumes of data but also analyzing additional factors that affect image quality and detail, such as geographic diversity, human activity, seasonal changes, and weather conditions. The use of GAN networks enables consideration of these factors and allows for modeling multi-layered, complex satellite imagery. The relevance of this study lies in examining how the application of GAN networks in combination with modern language models impacts the diversity, level of detail, and fidelity of generated satellite images relative to real-world data. This is especially important for tasks in monitoring, geospatial analysis, and urban planning, as it facilitates the rapid production of high-quality images that closely approximate real data for areas that are missing or difficult to collect, and also helps expand training datasets for other models.

The main objective of this paper is to develop and experimentally verify the accuracy of GAN models with an SBERT-based text encoder, to assess their controllability, and to evaluate how effectively this approach processes semantic information from textual descriptions and transforms it into realistic images. The results demonstrate that modifications of convolutional GAN networks (DCGAN and WGAN-GP), when combined with SBERT, can convert textual data into satellite imagery with high accuracy. It was also determined that increasing the number of training epochs is critically important, given the substantial complexity and diversity of satellite images and the variability of text descriptions in terms of style and context. The proposed method of integrating the SBERT-based encoder is a universal solution that can be applied to train various types of GAN models.