Algorithms uncover game-changing details in new images of Earth's surface
Researchers are using satellite images of our planet to create maps that can help predict the spread of bushfires and floods.
Call to mind an iconic image of Earth from space: blue and green, shrouded in white. Today, Earth-observing satellites such as the European Space Agency's (ESA) Sentinel-2 capture startlingly high-resolution images of the entire planet's surface every five days and make these available to researchers (and the general public) for free.
But just how detailed is the information researchers can glean from images taken 800 kilometres above Earth's surface? And why do they want to know what it looks like?
Dr François Petitjean and his team from the Monash Data Science and AI group are creating algorithms that allow them to harvest this satellite image data, combine it with expert knowledge and data from fieldwork, and create a better map than we have ever seen.
This map shows in unprecedented detail the variation in vegetation across the state of Victoria, from crops to vast eucalyptus forests.
A project whose time has come
Dr Petitjean points to three developments in satellite imagery that have made his work possible: greatly improved resolution, the collection of data at more points over time, and free access to the satellite images.
ESA's Sentinel-2 satellites produce images with a spatial resolution that is 625 times that of NASA satellite Terra.
The difference can be seen in the images above. The one on the left is how it appeared in the data used to create the current map of Victoria, while the image on the right shows the higher resolution of Sentinel-2 data.
Sentinel-2 provides a full picture of Earth, every 5 days. Having a series of satellite images of the same area over, say, a year, instead of a single point in time is a big deal when mapping vegetation.
It's the difference between being able to distinguish between different types of crops, versus lumping them together in a more general category, for example, because despite the impressive resolution of the images captured by a satellite in orbit 800 km away, it's impossible to distinguish canola from wheat in a single image.
"They look exactly the same, there's no difference in colours," Dr Petitjean explained. "The only difference is when they’re sown, when they grow, and how fast and when they're harvested."
Over time, changes in the growth of different types of vegetation allow Dr Petitjean to make fine-grained distinctions, and at a high resolution. Not only that, but prior to the Landsat and Sentinel satellites, he wouldn’t have been able to afford the cost of one image, let alone a series of them.
"Twenty years ago, if you wanted one picture of Victoria at the resolution we have now, it would cost you about a million dollars. Only massive businesses or big government organisations could use this data, because it was way too expensive,” he said.
"Now we have all these images available, and they distribute it for free.”
Instead, the bill goes mostly to European citizens, whose taxes fund the Copernicus program to which Sentinel-2 belongs.
Come fire or flood waters
According to Dr Petitjean, there are thousands of applications in Australia for the kind of maps he is pioneering. One is forecasting the yield of various crops at a national level and using that to project annual export-import figures.
“If you can say, 'I know how many hectares of wheat and canola and oats we have', and you know how much you're going to consume, then you can predict how much you're going to export," he said.
Other uses of the map are particularly relevant to Australia's changing climate and the more frequent and severe bushfires and floods that come with that change.
When modelling where and how a fire is likely to spread, you need data about the weather conditions, but you also need to know how many tonnes of fuel there are per unit of land. Why? Because a fire moving through a forest of old gum trees is going to gain far more momentum than a fire moving across a lawn. That's exactly the kind of information the land map provides.
It's similarly informative for modelling the spread of flood waters, because the map provides data on how much water will be absorbed by the soil, and how much will spread further across surfaces such as roads.
Considering the contribution of Dr Petitjean's work to these models puts the importance of accuracy into perspective. The higher the resolution of the map, the more accurate the model, and the better equipped we will be to prepare for and manage bushfires and floods.
"The resolution in this new land map is 25 to 625 times greater than what you'll see in the maps that are currently used in Victoria to model the spread of fire, and plan and carry out controlled burns."
Seeing the fruits of machine learning labour put to use in these ways is a reminder that AI has a lot to offer beyond commercial applications, which is something Dr Petitjean feels very strongly about.
"Machine learning is sitting at the tip of the spear of capitalism. To me, it's very important that we consider AI for social good."
This is borne out in his current research interests, which apply machine learning to solve problems in areas as diverse as health and public libraries.
But it’s the land map that’s taking up most of Dr Petitjean's time for now. His project, funded by an ARC Discovery Early Career Researcher Award, will wrap up in mid-2020, and by then he hopes to have demonstrated to Australian government agencies just how much they could gain from increasing the resolution of their existing maps.
"I'm developing a prototype for Victoria in 2019, so that they can take it on board, use it to update the map every year, and make it scale to a country level," Dr Petitjean said.
Read more about Dr Petitjean’s work at MonashVegMap.