Training a Machine Learning Algorithm to Create More Insight From Low-Resolution Images
The advent of machine learning and access to readily available Earth satellite imagery is fostering big data solutions that unlock valuable insights.
When we learned of the competition, we identified an overlap between the issues that the Hackathon Challenge was addressing and the issues that we were investigating as part of our Honours Project at The University of Adelaide. The project is a requirement for the final year of Bachelor of Electrical and Electronic Engineering (Honours) and was sponsored by our employer, Consilium Technology. Although the Hackathon asked for solutions in the mining, agricultural and/or environmental industries, we decided our project focused on an issue that affects all of those categories – the availability and frequency of high-resolution Earth observation imagery.
Australia faces a unique set of challenges, unlike any other continent or country, due to its geographical diversity, scale, and variable climate. Because Australia is one of the largest nations on Earth, it is expensive to capture high-resolution images of its entirety at a frequency desired by many leading geospatial insight services, such as Consilium Technology’s GAIA (Geospatial Artificial Intelligence for Agriculture). However, low-resolution satellites collect new imagery of the entire continent much more often – and it’s typically free of cost, like ESA’s Sentinel imagery.
Our project aims to create a machine learning algorithm that can predict what a low-resolution satellite image would look like if it was a higher resolution image. We artificially generate the higher resolution image by training the algorithm to compare DigitalGlobe’s WorldView imagery and Sentinel-2 imagery – and creating an output that is a higher resolution version of the Sentinel-2 image.
To put it simply, the rationale for our approach was to combine the information from the two satellite constellations, and utilise machine learning to recreate what a higher resolution image might look like at a time when a high-resolution image is not available (the generated image is based on an existing low-resolution image).
The results that have been derived from our Honours Project to date have been impressive. Our project differs from traditional super-resolution problems in that a machine learning algorithm is improving the resolution of Sentinel-2 satellite imagery from 10 m to 45 cm. By applying the algorithm, which was trained on high-resolution imagery, to a low-resolution image, we can predict a lot of the finer-grain details in the image. The outcomes of our systems are only expected to improve over time, as additional data will enable our machine learning algorithms to continually improve its predictions.
All of the hackathon submissions were assessed against various criteria, including the use of machine learning and contribution to open algorithms. We were very excited to find out that our project had won the Innovation Award.
High-resolution imagery is used to monitor crop health in the agriculture industry, generate road maps automatically, and monitor mine site safety in the mining industry. The innovation of our project is that our results are applicable across a variety of domains, as the need for frequent access to high-resolution imagery is industry agnostic.
Editor’s note: The DigitalGlobe 2018 Australia Sustainability Hackathon aimed to address Australia’s most conflicting issues surrounding mining, agriculture and environmental sustainability using machine learning and satellite imagery. This blog post was originally posted by DigitalGlobe after Consilium Technology Graduate Software Engineers, Lucas Sargent and Liam Mellor, won the Innovation Award for developing a solution that resonated across the three categories.