A machine learning approach to improving quality of atmospheric turbulence simulation

Abstract

The availability of large quantities of labeled data is key to the successful application of artificial intelligence and machine-learning (AIML) algorithms. However, live field data is expensive to collect and the atmospheric conditions that affect image quality are impossible to control. The capability to generate realistic, synthetic training data under operational conditions represents a significant opportunity to operationalize AIML algorithms at a significant cost-savings compared to live field collections. This work seeks to improve upon prior NVESD efforts to simulate the degrading effects of atmospheric turbulence on long-range, ground-to-ground imagery. Specifically, we implement a novel Generative Adversarial Network (GAN) architecture – which we named Noise-GAN – that is capable of detecting and then emulating spatially-varying signals directly from image data. We present results from a proof-of-concept study in which we demonstrate the capability of the Noise-GAN learn the spatial statistics of blur and distortion from turbulence-degraded imagery.