Evaluation of synthetic aerial imagery using unconditional generative adversarial networks

Image generation techniques, such as generative adversarial networks (GANs), have become sufficiently sophisticated to cause growing concerns around the authenticity of images in the public domain. Although these generation techniques have been applied to a wide range of images, including images with objects and faces, there are comparatively few studies focused on their application to the generation and subsequent evaluation of Earth Observation (EO) data, such as aerial and satellite imagery. We examine the current state of aerial image generation by training state-of-the-art unconditional GAN models to generate realistic aerial imagery. We train PGGAN, StyleGAN2 and CoCoGAN models using the Inria Aerial Image benchmark dataset, and quantitatively assess the images they produce according to the Fréchet Inception Distance (FID) and the Kernel Inception Distance (KID). In a paired image human detection study we find that current synthesised EO images are capable of fooling humans and current performance metrics are limited in their ability to quantify the perceived visual quality of these images.