AI for Earth Create an Artificial-Intelligence Platform for the Planet, Urges Lucas N

COMMENT SCIENCE FICTION Six giants of the POLLUTION A call for Pakistan TECHNOLOGY AI developers need LAB LIFE Here’s to coffee- genre reflect on their role in a and India to craft air-quality a checklist for human-rights room chat, catalyst of world unhinged p.329 policy together p.334 violations p.334 collaborations p.334 COURTESY OF MICROSOFT COURTESY

Microsoft, in collaboration with others, is using algorithms to convert satellite images into information about categories of land cover, such as forests. AI for Earth Create an artificial-intelligence platform for the planet, urges Lucas N. Joppa.

arlier this year, I became Microsoft’s business. And the time is right: applications at least once a month. Nearly one-fifth do first chief environment scientist. I’ve of the type that I and others have long been so daily (see go.nature.com/2bwmejp). been tasked with deploying the com- developing can now work at scale. That is Those data are fed to increasingly power- Epany’s deep investments in artificial intelli- why, in 2017, Microsoft put US$50 mil- ful algorithms that link people to others, gence (AI) research and technology to help lion into a 5-year programme called AI products or information. Yet we are flying people around the world monitor, model and for Earth (see ‘Microsoft for Earth’ and blind when it comes to understanding the ultimately manage Earth’s natural systems. www.microsoft.com/aiforearth). natural world. Most people I meet are surprised that one Scientists still struggle to predict the of the world’s leading technology companies FLYING BLIND effects of climate change at the resolution has a role such as mine. Yet I believe that in Today, we know more than ever about of cities or regions, or over timeframes of the next few years, every major tech firm will human activity. More than one-quarter of months or weeks — largely because they be working on applying AI to sustainability. the 7.6 billion people on Earth post detailed don’t have the kinds of data needed to It is the ethical thing to do. It is good for information about their lives on Facebook make such predictions, or because they

COMMENT

lack the algorithms to convert data into understand, and at appropriate spatial and together with where and how often it’s useful information. In the United States, the temporal resolution. With enough data photographed, these algorithms are provid- best available data sets on land cover, at a on which to train, and with human feeding new ways of producing global popula- resolution of 30 metres, were last updated back, these systems can learn to tag photos, tion estimates for endangered species4. nearly 7 years ago. Globally, the picture is acoustic recordings and genetic informa- Environmental applications of AI are much less complete. Yet without accurate tion with species names; or to convert sat- already attracting the attention of invest- information, housing developers, foresters ellite imagery into information on water ment firms, particularly in the agriculture or other land planners can’t make evidence- availability at a landscape scale. sector. Using data from high-resolution based decisions about which parcels of land Various organizations are already making satellites orbiting Earth, the tech company to use for which purposes, and how much impressive advances in Earth applications. Descartes Labs in Santa Fe, New Mexico, to leave untouched. iNaturalist and eBird, for instance, are iden- is monitoring crop production globally. Meanwhile, almost 95% of oceans — tifying species using communities of citizen Through the application of neural-network which cover more than two-thirds of Earth’s scientists. So far, iNaturalist’s 575,000 mem- algorithms to more than three petabytes of surface — remain unexplored. And scien- bers have recorded nearly 7 million obser- satellite and weather data (1 petabyte is tists have described only around 1.5 mil- vations of more than 128,000 species (see 1015 bytes), users can obtain information lion of the estimated 10 million species on ‘Assisted identification’). about projected yields that outperform all Earth, and know little more than the names iNaturalist relies on experts to validate other available forecasts in resolution and and collection locations for most of those1. users’ recordings, but deep convolutional accuracy. Likewise, Blue River Technology, AI systems could help in all of these neural networks are reducing the amount a California start-up, is using AI algorithms domains. In fact, after seven years of work- of expert input required — one of the big- combined with high-resolution cameras ing at the intersection of environmental and gest bottlenecks in attached to tractors and other field equip- computer science, I’m convinced that the the organization’s “The global ment to produce automatic weed-detection technology is now mature enough and the growth. Currently, environmental and weed-removal systems. In my view, global environmental crisis acute enough for more than crisis is acute the recent acquisition of Blue River by the to justify the creation of an AI platform for 5,000 species of enough to justify US agricultural company John Deere for the planet. plant and animal, the creation of $300 million represents just the beginning What I’m envisioning is a portfolio of its algorithms are an AI platform of the agricultural industry’s AI transfor- AI-infused ‘Earth applications’ available to able to make an for the planet.” mation. people in diverse domains, from forestry to accurate identifi- Others are using AI to inform land-use fisheries. These would be analogous to the cation four out of management decisions — such as how application programming interfaces (APIs), every five times, on average. And when the to establish wildlife corridors for species such as those for searching or mapping, that algorithms’ top five predictions are consid- such as lynxes and wolverines across the have enabled people to build software ser- ered for each sighting, the correct species is US Rocky Mountains as effectively and vices using components already made by included 95% of the time3. efficiently as possible. This is a difficult engineers at technology companies. The same types of deep-learning algo- computational problem because there rithm now used by iNaturalist’s vast com- are so many possible solutions. But AI EARTH APPLICATIONS munity of users are also helping ecologists advances from operations research, devel- AI and environmental-science researchers to classify millions of underwater snapshots oped for instance to work out how best to are now applying algorithms to topics as of corals. Other non-profit organizations, route traffic along a network of roads with varied as pollution modelling, agricultural- such as WildBook, are scouring the pho- the least cost and delay, are offering guid- yield optimization, animal-migration tographs of a particular species posted ance5. Likewise, game-theory researchers tracking and Earth-system modelling. on social-media platforms such as Flickr. are using AI to help law-enforcement offic- (Many of them will come together next By identifying each individual animal, ers to efficiently monitor the vast protected month in Austin, Texas, at the 17th Annual areas they are typically assigned to cover6. Conference on Artificial and Computa- (AI algorithms identify what monitoring tional Intelligence and its Applications to ASSISTED IDENTIFICATION strategy will maximize the probability of the Environmental Sciences.) Experts validate species identi cations made patrols detecting illegal activity in a way by users of the iNaturalist app, which Numerous advances in these areas are launched when the organization was founded that minimizes the probability of criminals coming from AI breakthroughs in non- in 2008. Arti cial intelligence is reducing the working out what that strategy is.) environmental ones. For instance, the amount of expert input required. But these inspirational examples need to SOURCE: INATURALIST increasing demand for low-cost cam- 8 be the norm, not the exception. Too often, era systems for smartphones and other researchers publish exciting results about devices has necessitated cheaper algorith- the application of AI to an environmental mic (instead of hardware) approaches to problem and those results are never trans- 6 improving image resolution. These ‘super Algorithms are now lated into applications. Or an AI system helping members resolution’ AI techniques can be used to to identify more is handed to a non-profit organization or improve Earth-system models by statisti- than 5,000 species government agency that lacks the resources 4 with an average of cally ‘downscaling’ low-resolution projec- 81% accuracy. and expertise to take advantage of it. Worse, tions of around 100 square kilometres to traditional AI innovators in industry and high-resolution ones of around 12 square academia rarely consider working on km that are more relevant to local land-use 2 environmental applications. planners2. Numbers of observations (million) Numbers of observations In short, AI systems can now be trained TWO QUESTIONS to classify raw data from sensors on the 0 I believe that for every environmental prob- ground, in the sky or in space, using cat- 2010 2015 lem, governments, non-profits, academia egories that both humans and computers and the technology industry need to ask two

326 | NATURE | VOL 552 | 21/28 DECEMBER©2017 Mac m il2017lan Publishers Li mited, part of Spri nger Nature. All ri ghts reserved. ©2017 Mac millan Publishers Li mited, part of Spri nger Nature. All ri ghts reserved.

COMMENT

Computer scientists generally publish their work in conference proceedings. (Conference organizers use a peer-review process similar to that used by editors at traditional academic journals.) Awarding prizes at leading conferences for the best solution, rather than for the best paper, could motivate students and faculty members to invest more efforts in engineering. Currently, AI researchers tend to test their algorithms on a few standard data sets. For instance, image-recognition software is generally tested on ImageNet, a database of around 14 million photographs. (Sub- jects include people, scenes and objects, as well as plants and animals.) Earlier this year, iNaturalist made its data set of 5,000 photographs of birds, mammals, amphib- ians and other taxonomic groups available for attendees of the Computer Vision and Pattern Recognition Conference in Hono- lulu, Hawaii. More environmental scientists should be prioritizing efforts to collate key data sets and make them available. Com- puter scientists can also help by communi- Artificial-intelligence systems can be trained to classify data, such as that from this motion-activated cating what kinds of data they need. camera, into categories that humans and computers can understand. There is growing interest in funding AI research for environmental applica- questions: ‘how can AI help solve this?’ and environmental monitoring. For instance, tions. Since 2008, the US National Science ‘how can we facilitate the application of AI?’ the creation of a national, sustained eco- Foundation has funded the Computational Governments. The public sector must system assessment in the United States Sustainability Network, a collection of AI ensure that environmental-data collection — as originally called for in 2011 by the researchers working on environmental platforms continue to be produced, and President’s Council of Advisors on Sci- sustainability. Submission tracks on sus- that the data are made broadly and freely ence and Technology7 — would motivate tainability are now included at leading AI available in formats easily ingestible by AI government agencies, academic research- conferences. And environmental journals NICOLAS ASFOURI/AFP/GETTY algorithms. Various projects are already ers, non-profits and the tech industry to are dedicating more space to the topic. making key data broadly available. One explore scalable solutions for monitoring Fully engaging the academic AI commu- is the European Space Agency’s (ESA’s) the nation’s natural resources8. nity in environmental issues, however, will Copernicus programme, the world’s larg- Non-governmental organizations. The require the creation of dedicated academic est single Earth-observation programme, implementation of environmental sustain- centres focused on translating research to and its associated Sentinel missions for ability projects often falls to resource-con- applications that could be used at scale. land, ocean and atmospheric monitoring. strained non-governmental organizations These could be attached to leading AI Another is a dedicated satellite system for (NGOs). Thus, the recognition within research institutes, such as the University tracking wildlife, supported by ESA along NGOs of the power of AI to aid their work of Southern California Center for Artificial with the German and Russian governments — and of the importance of investing in Intelligence in Society, and involve institu- (see go.nature.com/2jwanje). technological solutions — is critical. tions around the world. Governments, which tightly regulate A set of core AI technologies, such as Technology sector. The full participation the use of wireless communication chan- species-recognition algorithms, could of the technology sector is needed in efforts nels, can also provide the funding for pro- enable multiple organizations to work more to provide key data in appropriate formats, jects that are focused on the efficient use cheaply and efficiently. Currently, what’s as well as the algorithms, the infrastructure of microwave (cellular), radio and other emerging largely depends on the priorities to train those algorithms on the data, and spectra. Efforts such as the US Defense and interests of the tech sector. Domain- the means of making the end services avail- Advanced Research Project Agency’s Spec- specific NGO consortiums (including those able to as many people as possible. But why trum Collaboration Challenge, the world’s focused on water, say, or on biodiversity) should companies whose primary business first “machine intelligence competition could help to identify which problems is not environmental sustainability engage? to overcome spectrum scarcity”, could funders, researchers and the private sector At Microsoft, striving to democratize expand our ability to collect environmen- should prioritize. Such consortiums could access to our technology is part of our tal information from remote locations (see also provide guidance on how to develop culture. The initial mission to ‘put a com- go.nature.com/2bbhuo8). Ultimately, ‘intel- general infrastructure that multiple organ- puter on every desk and in every home’ ligent’ sensors that can identify which spec- izations could build to create their own has become a mission to ‘empower every trum to tap into at any moment, depending specific applications. person and organization on the planet to on availability and cost, could transform Academia. Several changes would achieve more’. the efficiency of data-collection efforts. encourage more AI researchers in academia More broadly, market trends indicate Finally, government-led projects can to focus on environmental sustainability, that demonstrating social responsibility incentivize investigators from diverse and to translate their findings into applica- is good for business. Investors and own- domains to work on AI approaches to tions that others can use. ers of a collective $100 trillion in assets are

©2017 Mac millan Publishers Li mited, part of Spri nger Nature. All ri ghts reserved. ©2017 Mac millan Publishers Li mited, part of Spri nger Natu21/28re. All ri g DECEMBERhts reserved. 2017 | VOL 552 | NATURE | 327

COMMENT now requesting that companies provide Microsoft is one of several tech and others founded in 2016 that came information on climate change, water and companies working to provide technology together this year to decide on best prac- forest issues through the reporting platform to help solve environmental challenges. tices — should agree on a set of standards run by the charity CDP and through the Google, Amazon, the software company for producing algorithms for Earth obser- United Nations Principles of Responsible Esri in Redlands, California, and the data vations. The partnership could draw on Investment. Also, for all major corpora- provider Planet Labs in San Francisco, Cal- work from the 2015 Transformations con- tions, any type of uncertainty — financial, ifornia, are among those striving to make ference in Stockholm. This produced a pre- political or environmental — is generally their software, services and data available liminary Biosphere Code Manifesto, with bad for business. Conversely, in an increas- for environmental applications. the aim of writing ten guiding principles ingly resource-constrained world, sustain- As companies ramp up their efforts, how- for the application of algorithms for envi- ability is key to resilience; reducing the ever, they must address the fact that today’s ronmental sustainability. likelihood of environmental disasters safe- AI technologies often come at a price and Some will argue — rightly — that it was guards supply chains, for instance. require computational expertise that puts largely the technology of the first and sec- The opportunity to work on societally them out of reach of many. Schemes to ond industrial important topics can be a huge pull for address access and education are crucial. “Artificial revolutions that potential employees in a competitive Good examples of these include Microsoft’s intelligence caused the envi- industry. Since Microsoft launched AI multi-day instructor-led AI for Earth educa- is not a ronmental issues for Earth, employees have approached me tion courses (see ‘Microsoft for Earth’) and panacea for of today, and that from all corners of the company, especially Google’s annual Geo for Good workshop, environmental we already know data scientists and engineers, wanting to designed to help non-profits and other problems.” what we need to apply their skills to issues that they care organizations make better use of Google’s do: reduce the about. map data. emission of greenhouse gases, the destruc- Finally, there are major market opportu- tion of forests and the overfishing of oceans. nities in this area. High-resolution maps of BEST PRACTICE But decisions about what actions to take natural resources are crucial to those work- Researchers and developers must ensure will be easier to make — and less vulner- ing on environmental issues. They are also that Earth applications are trustworthy, able to politicization — if we know what is useful to the military, urban planners and transparent and fair. The Partnership on AI happening on Earth, when and where. AI providers of insurance, among others. — a consortium of technology companies can help to provide that information. Time is too short, and Earth’s resources too important, for companies such as MICROSOFT FOR EARTH Microsoft to ignore what is likely to be humanity’s biggest challenge yet: mitigating A US$50-million programme for the environment and adapting to changing climates, ensur- ing the resilience of water supplies and sus- Microsoft’s AI for Earth programme, When complete, this system should tainably feeding a rapidly growing human announced this month, is dedicated allow organizations such as the population, all while stemming an ongoing to finding solutions to challenges Chesapeake Conservancy to gain detailed and catastrophic loss of biodiversity. relating to climate, agriculture, water insights into how the lands they care about AI is not a panacea for environmental and biodiversity — in which users of are changing, and so achieve a type of problems. But history will judge the suc- the resulting services and applications precision conservation that is desperately cess of the Information Age by our ability to could grow to tens or even hundreds of needed. Knowing exactly where agricultural deploy its resulting technology in steward- millions of people. (Such services may run-off is entering the Chesapeake Bay, ship of the planet. It’s a big challenge, but an be produced by Microsoft or by other for instance, could indicate which square even bigger opportunity. ■ organizations with Microsoft’s support.) metres of land should be replanted. It involves pairing the company’s In another collaboration, with Lucas N. Joppa is chief environment technical expertise on artificial researchers at Johns Hopkins University scientist at Microsoft, Redmond, intelligence (AI) projects for in Baltimore, Maryland, and the University Washington, USA. environmental monitoring and modelling of Pittsburgh in Pennsylvania, among e-mail: [email protected] with that of other organizations. For others, we are working on ways to use 1. Mora, C., Tittensor, D. P., Adl, S., Simpson, A. G. B. example, six months ago, we partnered organisms such as mosquitoes as self- & Worm, B. PLOS Biol. 9, e1001127 (2011). with the California firm Esri, an powered data-collection devices. The aim 2. Vandal, T. Preprint at https://arxiv.org/ international supplier of geographic is to create an AI-powered metagenomics abs/1703.03126 (2017). information system software, and the pipeline, and to gain insights about an 3. Van Horn, G. et al. Preprint at https://arxiv.org/ abs/1707.06642 (2017). Chesapeake Conservancy in Annapolis, ecosystem from information about the 4. Berger-Wolf, T. Y. et al. Preprint at https://arxiv. Maryland. In this collaboration, we aim animals that mosquitoes feed on. org/abs/1710.08880 (2017). to apply deep-learning techniques to We have made our AI technologies 5. Le Bras, R. et al. In Proceedings of the 27th AAAI Conference on Artificial Intelligence 1305– convert freely available high-resolution available to more than 35 organizations 1312 (2013); available at http://go.nature. imagery from the US Department of in more than 10 countries in the past few com/2nxgycl Agriculture’s National Agricultural months, for applications ranging from 6. Nguyen, T. H. et al. In Proceedings of the 2016 Imaging Program into land-cover species-abundance modelling to live International Conference on Autonomous Agents & Multiagent Systems 767–775 (2016); available at categories (forests, fields, water and poacher detection in drone imagery. We http://go.nature.com/2atkzyi impervious surfaces) at 1 metre have also hosted multi-day courses to 7. PCAST. Report to the President: Sustaining resolution — all in a system that can be ensure that environmental scientists and Environmental Capital: Protecting Society and the Economy (White House, 2011); available at easily updated when new images are managers are able to take full advantage http://go.nature.com/2btene2 acquired. of the tools. L.N.J. 8. Jackson, S. T. et al. Science 354, 838–839 (2016).