Everyone can use AI on satellite imagery
One of the most powerful applications of AI, the exploration of satellite imagery, is accessible to almost every business.
Launched 80 at a time, the low-cost “cube” satellites stream data to AI models that mix image recognition and analytics to document the world. Together they give access to a detail of information at a speed and cost unthinkable a few years ago.
The proliferation of satellite imagery
Since 1992, firms like Digital Globe can legally commercialize images at 0.31 pixel meters. A car on a road will be 10 pixels and an open door 3 pixels — still detectable. Satellites like Worldview3 provide multispectral or infrared imagery with a rich documentation. As Digital Globe puts it: someone standing under the Hollywood sign in Los Angeles can count people on the Golden state bridge in San Francisco.
These satellites scan 60% of the planet every month, but a specific application remains a project. That has changed with nanosatellites.
Founded in 2009, Planet labs has launched 233 of its doves. The size of a shoe box, each cost less than $100K, vs $300 million for a satellite like Worldview 3. Each cube sends 10,000 pictures a day, covering the size of a US state. The resolution is “only” 3 meters pixel, but with so many satellites working together the revisit rate is weekly or even daily.
A Planet Labs Dove; each is only 34 centimeters long (14 inches). Credit: Planet Labs.
With the cubes always looking, a business can detect and track activities on a real-time basis.
To forecast agricultural production, Descartes Lab looks at every field from every farm- ingesting 5Tb of data every day. Machine learning algorithms derive chlorophyll level from spectral images and issue weekly forecast more accurate than the inspectors on the ground. Its founder, Mark Johnson, notes:
“It took Landsat over 40 years to collect under a petabyte of data with 7 satellites. Planet will easily produce over a petabyte a year.” (Mark Johnson, The Verge, Aug. 2016).
Automation with deep learning
Gaining access to the images is one thing, but until recently the exploitation was at best semi-automated and often manual. Since 2012 deep learning has opened the way to a full automation.
In the 2012 ImageNet competition, the Toronto team used a deep learning model to knock down the image recognition error rate from 25% to 16%. It’s now below 6% — in theory at the human level (not a fair comparison, the competition uses a few thousand categories only).
In the past few years, deep learning techniques have exploded. The community is releasing open source libraries to refine the main technique used, convolutional neural network(CNN). A development can use medical research libraries to separate buildings in an urban setting.
The U-net model, developed at the Universty of Freiburg, is available in open source.
ImageNet competitions are now far less exciting, but SpaceNet competitions still are. In 2017 the goal was to recognize — and separate — houses in four cities: an easier task in Las Vegas (88% success) than in Karthoum (55%)!
The explosion of business opportunities
Before deep learning, this type of application was a major project — or was not considered at all. Businesses can’t hire hundred of experts to look at satellite imagery. It’s not a trivial task and crowdsourcing like Amazon Mechanical Turk is not an option either.
But with doves always looking and machine learning always detecting “We are able to detect things before we even know about them”, says a national security expert (G. Marks in Bloomberg News).
A business can analyze real-time market insights and see what’s going to happen — instead of building a plan on a high-level forecast and find out from the sales report what has already happened. The shift is at the core of AI business value: induction instead of a deduction.
AI models are famously used to count cars on parking lots. Orbital Insights has counted over 1.2 billion cars, gaining insights on business results before any public estimates. A team revised the estimation of China oil reserves by reading the shadows left by the floating lids on tanks.
Some part of the world have no reliable economic data — or even census: a Stanford team has used city lights to measure development. Facebook combines Digital Globe images with World Bank data to determine precise rates of internet access.
Previously a company could check the progress of a competitor new plant in China, but had no way to find out if some equipment was sent to Costa Rica. Now an AI application can detect the unloading in a port and follow the truck traffic on the road, especially if the engineers tweet about the warm weather they found in Latin America.
Building a prototype
A small team can deliver a prototype in a few months at a minimal cost. It will, however, need access to CNN expertise. Building a complex model in a neural network remains the hard part.
Some challenges are not trivial. Objects of interest can be rare in a picture (class imbalance) — forcing a complex set-up. Most of the time a project requires to mix and match different techniques for the object classes.
The dust has not settled on the toolset side either — the best option is to use a combination of tools like Caffe 2 or Theano without getting too attached to any. A team cannot expect to find standard solutions and be able to re-use templates — each project is still ad hoc.
A team will need special skills in imagery and deep learning. Kaggle is a good place to start.
An option to test a business hypothesis is a Kaggle competition — keeping in mind that robustness is not the goal there. Another option to gain access to imagery and deep learning expertise is to partner with a startup of the sector.
A model can start small then add other sources — mixing low and high-resolution images, recognition with IoT data from the ground, or insights mined from public data and social media.
The output will be a depth of information unthinkable before, updated every week instead of every quarter. The best part? A team of a few people can start today.