How Deep Learning could revolutionise Earth Observation
Earth Observation has been a little overlooked this year but integrating Deep Learning technologies could power rapid developments across the entire field. Here, we look at two interesting ways this sub-domain of artificial intelligence could aid industry and international agencies alike.
It’s hard to argue with the enthusiasm for looking upwards and outwards that has swept the globe in recent months; we have witnessed some truly historic scientific landmarks, including the discovery of water on Mars and the emphatic flyby of Pluto (NASA is releasing new images of the dwarf planet’s blue sky as I type). Of course, it’s worth remembering that those achievements came about from innovative technologies converging over the past decade, and similar forward leaps look set to revolutionise Earth Observation in the near future.
Earth Observation and image recognition
When looking down from above, one longstanding challenge with managing satellite data concerns hyperspectral data classification. For a moment, however, let’s look at a broader level and consider image recognition technology more generally.
Historically, programming computers to classify images effectively has been difficult. Google Image Search demonstrates this point well; this search engine does not currently recognise the image itself, but finds matches by analysing metatags and page content, the origins of which have at some point been manually user-generated. Looking to the future, however, these search functions will be able to match images by actually understanding what an object looks like and matching that with others it understands to be similar. In fact, Baidu, the Chinese search giant, and others, have already arrived at this point. This area of work, known as Deep Learning, is most commonly based on creating convolutional neural networks (if you’re an outsider to the field, they might sound incredibly complicated, but don’t worry, to an insider they seem even more complex). In essence, though, they are systems by which computer agents learn to recognise image segments at a minute scale as a stepping stone to understanding features and then finally objects and scenes, without additional inputs beyond raw data.
Deep Learning and hyperspectral data classification
Returning to an Earth Observation context, hyperspectral data are essentially images that reveal added context about the materials pictured. Classification systems for these images are ineffective at present and huge mountains of data are being constantly harvested. Fortunately, we seem to be right at the upswing of exponential improvements here. And, such is the nature of Deep Learning, once engineers reach a certain level of success with algorithm development for the agent itself, raw data is all that is required for further improvement. Usefully, raw data is not something satellites struggle to collect, they do so in abundance, which leads neatly on to Deep Learning applications and data storage.
Can Deep Learning help with data storage?
So what about data storage? Well here’s where things get a little more theoretical, but also more fascinating with regards to artificial intelligence. Some researchers, have described a new type of deep neural network – a Perpetual Learning Machine – that uses statistical recall biases to mimic human memory to a limited extent, creating a ‘use it or lose it’ stimulus driven memory process. The suggestion is that researchers could build algorithms that make decisions about what to use their memory for, essentially ‘forgetting’ information they have learnt to deem unimportant through experimentation on huge datasets. The implication is that the huge challenges in data storage that currently exist could be overcome not by us, but by computerised agents themselves.
Going deeper
This only scrapes the surface of some very complex fields of course, but it does go some way to highlighting just how exciting near-future technology convergence could be for Earth Observation applications. If hyperspectral data classification or storage are important to your project or business, but you don’t work in these research fields directly, I’d suggest seeking out the best minds in Deep Learning right now, as there are still so few who really grasp the subject area at both broad and detailed scales.
Keep up to date
Twitter is a great place to keep up to date with Earth Observation developments as they happen. Follow us here:
Tweets by @spacecongress
To save you time, we’ve also curated a list of other top Earth Observation Twitter accounts to follow:
| • Aberystwyth University: @AU_EarthObs | |||||||||
| • African Physicists: @AfricanPhysics | |||||||||
| • Canadian Space Agency: @CSA_ASC | |||||||||
| • Disasters Charter: @DisastersChart | |||||||||
| • Earth Observing Lab: @ncareol | |||||||||
| • ESA EarthObservation: @ESA_EO | |||||||||
| • ESA Italy: @ESA_Italia | |||||||||
| • ESA Science: @esascience | |||||||||
| • European Space Agency: @ESA | |||||||||
| • Group on Earth Observations: @GEOSEC2025 | |||||||||
| • India Space: @India_inSpace | |||||||||
| • Japan Aerospace Exploration Agency: @JAXA_en | |||||||||
| • NASA Astronauts: @NASA_Astronauts | |||||||||
| • NASA EO: @NASA_EO | |||||||||
| • NASA ESTO: @NASAESTO | |||||||||
| • NASA Goddard: @NASAGoddard | |||||||||
|
• NASA Landsat Program: @NASA_Landsat
NASA Soil Moisture Active Passive mission @NASASMAP
|
|||||||||
| • NASA: @NASA | |||||||||
| • National Weather Service: @NWS | |||||||||
| • NOAA: @NOAA | |||||||||
| • Sensing Our Planet: @sensingrplanet | |||||||||
| • Space Generation Advisory Council: @SGAC | |||||||||
| • The World Meteorological Organization: @WMOnews | |||||||||
| • UK Space Agency: @spacegovuk | |||||||||
| • World Resources Institute: @worldresources | |||||||||