Google Earth Engine: a smarter way to make false colour composites

False Colour Composites (FCC)

What do you do when you need to map large features in difficult to reach places? You use satellite imagery of course. Satellite imagery is great, it reveals insights that would otherwise go unnoticed during geological field mapping by allowing you to see Earth’s surface in new ways. At Umvoto, our geologist’s regularly use various satellite imagery to better understand and map the earth’s surface. However, at Umvoto we don’t only use Google Earth, we also prefer something more informative (with the bonus of creating art).

False colour composites (more details here) are not merely fancy looking images but have a wide range of applications, from mapping and investigating geological features, such as the Great Dyke and Vredefort Dome to assessing agricultural water use, all with the bonus of a growing set of custom scripts one can use. At Umvoto we regularly use the wonderful, and publicly free, European Space Agency Sentinel-2 imagery (see infographic below) to investigate the earth, from landuse mapping to highlighting lithological variations.

Key highlights and facts Copernicus Sentinel-2 mission (courtesy of the ESA)
Key highlights and facts Copernicus Sentinel-2 mission (courtesy of the ESA)

A big (data) challenge

One of the challenges of processing and analysing satellite imagery is downloading, processing, and storing of enough imagery to have adequate spatial and temporal coverage to create a product that is useful to Geologists, such as this false colour composite (FCC) of Eritrea.

False colour composite (FCC) of Eritrea constructed using Sentinel-2 bands 12, 11, 8, 4, 3. By comparing the FCC to the true colour base image, the usefulness of FCC imagery in assisting in identifying different landscape features become evident – although it does take while to adjust to a FCC world.
False colour composite (FCC) of Eritrea constructed using Sentinel-2 bands 12, 11, 8, 4, 3. By comparing the FCC to the true colour base image, the usefulness of FCC imagery in assisting in identifying different landscape features become evident – although it does take while to adjust to a FCC world.

Searching for and downloading imagery is a very time-consuming process, and without guarantee of being totally cloud free. Perhaps, more importantly, image processing can be just as, if not more, time consuming and (to many people) a boring task. Depending on your internet speed, it can take 2 hours to download one Sentinel-2 scene, followed by 2 hours to process and package that scene into an easily useable format – that’s 4 hours per scene. Once assembled, a single scene’s file size is approximately 650 MB (only for the raw files) with this size easily doubling during analysis and soon your hard-drive is filled.

For a relatively small country like Eritrea (smaller than the Western Cape province of South Africa) it takes 37 scenes of Sentinel-2 imagery to map it. That means, if all runs smoothly, from image search to a final FCC snapshot of Eritrea, it will take 148 hours, and 50 GB of storage space. No small task. But what if we are interested in changes and dynamic patterns and to include all Sentinel-2 images of Eritrea for the year 2020? This would equate to 3 879 images, and approximately 5 TB of storage space and we would require 15 516 hours to process the country. This would mean working for nearly 2 years, without any recess, to pull off such an analysis. More importantly this leaves us with no time for any fun, exploration, and interpterion of the imagery.

Working smarter not harder

As we attempt to leverage new and better imagery, so too do we need to leverage the technologies that allow us to spend more time interpreting and exploring the data, rather than the mind-numbing work of handling big data. For me, one of these promising technologies is Google Earth Engine (GEE), not to be mistaken for the already epic Google Earth (Pro). GEE is a cloud computing platform, effectively dealing with the challenge of big data analysis. Already incorporating a massive, and an ever growing, collection of datasets GEE is not only limited to image processing but has been used applications ranging from land cover/land use classifications, hydrology, urban planning, natural disasters, and climate analyses (all activities that Umvoto undertakes). The ability of GEE to process petabytes of remote sensing data, covering large geographic areas and long temporal series, makes GEE an ideal tool to explore the Earth’s surface. Enhancing its usability, GEE has a large library of functions and algorithms to assist in data exploration and analysis (see Amani et al (2020) for a comprehensive review of GEE). Now, with a few lines in the GEE’s code editor (javascript or python), we can undertake something that once seemed near-impossible in a couple of hours.

Using GEE we can shift from dealing with the mundane task of data assembly, to following our curiosity and playing with different datasets. By focusing on asking questions and exploring data we can find solutions to the problems posed to us. Looking again at our Eritrean FCC, a product of 3 879 Sentinel-2 images, it took me a day, not years, to complete, giving me more time to analyse the image. The real question now in remote sensing analysis is no longer “how to?”, but rather “where to?”.

FCC zoomed into a northern region of Debubawi Keyih Bahri zoba.
FCC zoomed into a northern region of Debubawi Keyih Bahri zoba.
FCC zoomed into a central region of Debubawi Keyih Bahri zoba.
FCC zoomed into a central region of Debubawi Keyih Bahri zoba.
FCC zoomed into a southern region of Debubawi Keyih Bahri zoba.
FCC zoomed into a southern region of Debubawi Keyih Bahri zoba.