All Projects

Cloud Removal in Hyperspectral Satellite Images using Generative Adversarial Networks

Apr. 2018 - Apr. 2019

Won Best Solo prize and Best Satellogic data use at Stanford Big Earth Hackathon.

Satellite imagery can be used to monitor the environment or predict disasters and enable quick responses. Clouds bring uneven illumination, blurring and occlusion of the target. Satellite images are traditionally multispectral, i.e. include only a few wavelengths. Hyperspectral images include numerous wavelengths, including near-infrared, and are becoming more widely available. Finally, Generative Adversarial Networks (GANs) are among the most successful unsupervised techniques for generating realistic images by training 2 networks in competition (generator vs discriminator).

Image Credit: NASA Goddard Space Flight Center Image Credit: NASA Goddard Space Flight Center

I proposed to apply GANs to hyperspectral satellite images in order to generate the missing patches from below the clouds. I devised and implemented the whole data pre-processing pipeline which includes classifying satellite images patches as cloudy/clear and synthesizing cloud masks.

Details on DevPost / Invited poster at ICME Xpo 2018