GANs for Satellite Multi-Spectral Imaging
Thesis submitted for my Masters of Science in Computer Vision, Robotics, and Machine Learning in 2019.
Imagine a world where you can do more with less.
Harnessing the mass of previously captured images from existing earth observation LANDSAT-8 satellites.
Taking cutting edge Deep Learning & GAN techniques from research to solve a real world problem.
With the guidance of Dr Raffaella Guida’s extensive experience in earth observation, I providing a proof of concept that presented an avenue to reduce the costs and recover from data loss.
The problem I saw…
Earth observation satellites use
Specialized Multi-Spectral Imaging
to capture data at different wavelengths, to ascertain different information about the world it sees.
Temporal gaps in data
and outages in service due to maintenance or interference is a debilitating side effect.
Long term maintenance of such systems,
ensuring data continuity
is prohibitively expensive, yet unbelievably useful for the scientific community.
The opportunity I sensed…
While the field of Earth Observation looked at strictly mathematical ways of combining data, they rely on the panchromatic band images to produce
Pan-Sharpened Images
which sometimes don’t exist due to temporal gaps, and become more and more expensive to build as satellites need to be replaced every few years (in the order of millions of dollars for a launch and commissioning).
Looking at everything I’d learnt during my postgraduate studies about the effort going into high-realism, and accurate 1-to-1 approximations of real images based on
Generative Adversarial Networks
I decided to apply these techniques to be able to approximate a panchromatic image based on the standard RGB image channels alone.
My solution…
By using GANs to approximate the Panchromatic channel for any instance of an RGB image, we decouple the problems of temporal reliance, potentially even eliminating the need for a dedicated panchromatic sensor. Ultimately this solution presented the potential for a significant saving in design time, effort, and costs, as well as operational costs to launch (by reducing launch weight) and by minimizing ongoing maintenance costs.
