Multi-Task Generative Adversarial Networks for Semantic-Guided Remote Sensing Image Generation

Semantic layout synthesis based on Generative Adversarial Networks (GANs) has made significant advances, however, the generation of high-quality real remote sensing images through semantic layouts remains a significant challenge. In this paper, we propose a Multi-Task Generative Adversarial Network (MTGAN) that utilizes semantic layouts to generate remote sensing images featuring five common geographic objects. Our approach takes into account the unique characteristics of remote sensing images, such as multiple scales, multiple objects, low inter-class separability, and high intra-class heterogeneity. The MTGAN utilizes a global-local GAN structure, with the global generator producing global context information and the local generator generating specific class information. By combining global macro information and local detail information, the MTGAN is able to generate remote sensing images with close contextual connections and clear details of geographical objects.