Enhancing the Resolution of Satellite Imagery Using a Generative Model

Abstract

Recent breakthroughs in deep learning algorithms introduced the image super-resolution technique that maps the low-resolution image to generate a high-resolution image. These techniques increase various surveillance applications by providing finer spatial details than data from original sensors. Satellite images obtained from Moderate Resolution Imaging Spectroradiometer (MODIS) observation offer essential information about the earth’s landscape, ocean, and ecosystem, contributing to monitoring various applications in the scientific field. The spatial resolution of satellite images has a significant impact on image accuracy. This paper focuses on improving image resolution by training a convolutional neural network to produce super-resolution images from low-resolution images. We present an implementation of Super Resolution Generative Adversarial Network (SRGAN), a GAN-based approach that uses a perceptual loss function that includes an adversarial loss and a content loss. Using a discriminator network that is designed for discerning between super-resolved images and original photo-realistic images, the adversarial loss drives the solution of this architecture to natural images. Moreover, the content loss is driven by perceptual similarity rather than pixel space similarity. We used this architecture to satellite images collected from NASA MODIS devices and found satisfactory results. Our key finding is that our system’s result can now be used to improve a variety of low-resolution images.