REPORT

Track 3:
The Eye in the Sky:
Space, AI & Satellite

In partnership with

ENTER
 > >
B/B

Contents

Page 02
Session A
Page 10
Session B
State of Play Poverty

Page 19
Session C
Page 22
Session D
Page 26
Conlusion
Deforestation Agriculture
 > >
1/31

The satellite track introduced the challenges and the opportunities of satellite
imagery, described what satellite data is and its use and applications for
dealing with three particular tasks: poverty, deforestation and its impact on the
environment and climate, and agriculture.
 > >
2/31

A
Session

State of Play
 > >
3/31

Stuart Russell, Professor at University of California, Berkeley,

moderated this Session.

In his introductory remarks, he observed there are hundreds of satellites in space, producing huge amounts of data.

The world is imaged every day. We are now working at the intersection between satellite imaging, geospatial analysis

and AI to maximize the full potential of these technologies. With AI and satellites, we can see the whole world at once

and do many things that benefit humanity – predicting famines, developing models of Earth systems – the potential

is amazing. According to Professor Russell, we are still some way off using satellite data for achieving an accurate

count of a population/census [although Facebook, UNDP and others have analyzed night-time lights to determine

population distributions].
 > >
4/31

observed that “space businesses” (comms, positioning systems) are

Einar Bjørgo,
becoming more fully integrated, but suggested we are not yet seeing
Manager at
the big impact expected from Earth Observation (EO). Today, we are at a
UNOSAT,
game-changing stage, with potential for a tremendous impact, especially
for developing countries. But we need more capacity development. He
described UNOSAT (powered by CERN and based in Geneva) and the work
of UNOSAT’s team of expert analysts who examine satellite imagery and
train people from different backgrounds – NGOs, governments, even private
business – to use satellite imagery.

It is important we include everyone to help them benefit from this technology.

 > >
5/31
suggested we don’t need to use AI to do the things we do
well already, we need to use it to do things we currently
cannot do. He presented statistics on the number of
satellites – Governments were launching roughly ten
satellites a year until 2015. Now there is an explosion in the
number of satellites – although he noted we must be careful
about counting numbers of satellites, as their capabilities
also matter and/or using different types of satellites in
combination (e.g. layered data and hypercubes). Space
Mark Doherty, agencies are working together using Earth Observation
Head of EO satellites to try to understand climate action and disaster
Development risk reduction. However, various obstacles exist, including:
Division at ESA restrictive data access; lack of standardization; data that are
not fit enough for purpose; Lack of analysis of ready data;
insufficient frequency of observations; and long-term trend
analysis. He distinguished between validation data (which
are expensive to collect and difficult to obtain from third-
parties) and analysis-ready data, which involves correcting
data from different sensors. He described the use of AI to
correct for cloud cover and new hyper spectral systems.
Observational audits of CO2 emissions are needed for
global stocktaking to support compliance with the Paris
agreement. The big challenge is how we can put together
AI and satellite data with models of the planet to identify
signals, model them and get even greater predictive power
than today.
 > >
6/31
observed the tremendous achievement of imaging the
whole Earth every day was accomplished in 2017 using
180 dove satellites at 3m per pixel. There are 13, soon to
20, Planet satellites capable of imaging 70-80 cm per pixel.
It will be necessary to put more AI out in space, unless
we are planning on downloading all the data generated
continuously. Mr. Zolli showed how solar installations in
China are visible from space. He presented his 4 Is model
– we are moving from (1) information to (2) insight to (3)
Andrew Zolli, Big Indicators to (4) Big Instruments. Extracting signals is
VP becoming harder from the deluge of information, but the
at Planet Labs “insight revolution” – machine learning / AI, computer vision,
and basic statistical analysis – can help with this. Information
is becoming more real-time – for example, imagery of the
burning of Rohingya villages in Myanmar was analyzed
in real-time to try to lobby and change behavior. This will
also help us move from retroactive analysis to real-time,
to predictive analytics. Mr. Zolli showed models of urban
growth for Dar es Salaam in Tanzania growing at 8-15% a
year, and deforestation in Bolivia. We need it to happen
in a more organized way and all work on it as a collective
common good. Mr. Zolli called for a data strategy and an AI
strategy for the UN system as a whole.
 > >
7/31
observed that the number of satellites is doubling every
one to two years For people to look at the earth every
day using satellite images, 8 million people would be
needed doing nothing else every day. Orbital Insight’s
clients include Wall Street making investment decisions,
insurance companies trying to decide insurance rates for
disaster recovery and humanitarian agencies on a range
of problems. He described examples of: deforestation of
virgin forests and managed/planted forests versus urban
James Crawford, areas in Malaysia; poverty mapping in Sri Lanka and Mexico;
CEO of Orbital
Insight
and collaborations with the U.S. government agencies and
insurance agencies after Hurricane Harvey struck Texas.
The models for damage and zones vulnerable to flooding
proved useful for modelling losses and insurance pay-outs,
as well as planning for future extreme weather events.
James Crawford described other examples of trends in
deforestation of virgin forests and managed/planted forests
versus urban areas in Malaysia (see figure) and poverty
mapping in Sri Lanka and Mexico (see figure below).
 > >
8/31

Deforestation of virgin forests and

managed/planted forests versus
urban areas in Malaysia and poverty
mapping in Sri Lanka and Mexico

Image: Forestation & Land Use in Malaysia Image: Poverty Rates in Mexico
Source: Orbital Insight, World Bank.
 > >
9/31

pointed out that tech companies have the capacity to use
AI with planetary data which vastly outstrips governments
and citizens. He asked whether AI can be used to reduce
information asymmetries? It is not clear that it can – satellite
imagery is being used to create information asymmetries
with regards to oil and other important commodities.
We need to define what planetary data is, and what it
consists of, and figure out the global standards that allow
that planetary system to function. He suggested we need
David Jensen to put in place the safeguards to mitigate the risks of
of UNEP
information asymmetries – not just in fragile States, but in
many different countries. What can the UN do to advance
the use of planetary data and AI? The UN is exploring
the frontier technologies already being tested. The UN
Global Pulse initiative aims to harness big data, and the UN
Environmental Assembly in March 2019 will focus on tech
and innovation.
 > >
10/31

B
Session

Poverty
 > >
11/31

Lars Bromley of UNOSAT moderated this Session

He described UNOSAT’s work – UNOSAT specializes in the use of satellite technologies in many day-to-day

operations of the UN system and recently, it has been developing the AI aspects of its work. He acknowledged that

the SDGs are an immense challenge, as they represent the “eternal problems of the human condition boiled down

into 17 Goals”.
 > >
12/31

stated that it is not just income poverty that matters, we some jobs in future. He described how satellite data can be
Bernhard Kowatsch,
need to think about [future] economic opportunities, which combined with price information to inform people about
Head of Innovation
are disrupted by displacement and war. AI is an opportunity, the fair price for their crops, access to information points,
Accelerator at
but not the single silver bullet that changes everything; weather data, farming advice, remote sensing and voice
World Food
there are other technologies we can leverage. He described recognition to realize the full benefits of access to the
Programme
poverty in the context of satellites and AI – where roads are knowledge economy. There is a big difference between
constructed, agriculture may develop and an increase in whether you make data available publicly, or publicly
incomes may follow infrastructure development. AI can be available data that has been sanitized, raising important
used to identify structures, and the wealth of neighborhoods data privacy considerations about what are you able to
based on roof type. He was unsure whether it is possible to share, with whom and how? We need further data sharing
identify trucks from the shapes and patterns of the truck’s agreements as to who can share data with whom, and
roofs. He acknowledged that AI can have both positive and privacy.
negative effects, especially for poverty – AI may replace
 > >
13/31
described the work of the Sustainability in AI Lab at
Stanford as using sources of non-traditional data to
Marshall Burke,
measure development outcomes on the ground. We can
Assistant Professor
use satellite data for poverty prediction because access to
at Stanford,
ground data is very infrequent. Humans have historically
been very good at distinguishing well-being in imagery,
but computers have got much better at image recognition
tasks (the best computer system surpassed the accuracy of
human recognition abilities in 2014). There are many new
sources of satellite data, and greater quantities of data.
He described the use of convolutional neural network and
transfer learning to investigate plot-level maize productivity,
electricity consumption and poverty in sub-Saharan
Africa (Nigeria, Tanzania, Uganda, and Malawi), as well as
poverty in India (at district and village levels). However,
data availability can be a problem – the measurements you
pull off the sensors have improved significantly recently.
We have to start with countries where we have data – the
fundamental constraint is not imagery, it is labelled data.
Nigeria has nearly 200 million people, but we only have
a labelled dataset of 2000-3000 observations for Nigeria.
Stanford wants to work with policy-makers and practitioners
to operationalize satellite-based estimates.
 > >
14/31
observed satellites can help map the spread and land use
of cities. There are currently 55 million children displaced
due to war or violent conflicts. Climate change is affecting
millions of people, including almost 500 million kids, with
food and security issues. She presented UNICEF’s work
in combining mobile phone data and satellite imagery to
map poverty – poverty is a complex thing and does not
depend on any one thing. We can also map nutrition in Iraq, UN Innovation Network to enable conversations between
how can satellite images help us map the nutrition piece?
Naroa Zurutuza UNICEF is working with FAO in Malawi, combining satellite
of UNICEF
and drone imagery can help map crop yields. We can also
map refugees on the move, or displaced to new locations,
and map the growth of refugee camps. We are testing new
models like crowd-sourced science that different researchers
can contribute to. She described UNICEF’s initiative to
map schools combining satellite imagery with real-world
data in Liberia and Nigeria and ultimately, worldwide. The
different UN agencies.
 > >
15/31

Image: Al Zataari Refugee Camp in Jordan – before and after

Source: UNOSAT (Lars Bromley).

 > >
16/31

stated that Global Pulse is interested in augmented methods any one example does not apply elsewhere, depending on
& cognitive learning to use minimum amount of information different climactic conditions, different times of day – what
from experts to use human expert time most efficiently. looks like a building could be a stone somewhere else!
Poverty is a complex phenomenon, the type of roof of a Discussions focused on issues with the imagery and the
building is an important proxy indicator for the wellbeing of resolution needed for images – high-resolution images are
a household and people who dwell within it. A 5% and 10% costly and difficult to obtain, and may only be obtained
error rate might sound low, but in vulnerable regions, a small every few days. Really high-res satellite images run
number of pixels might represent an unacceptably high North-South in 200 km strips, but if your road runs East-
rate of error – especially in a crisis situation. How can we West, you will need a lot of images! Data download may
John Quinn reduce this error rate still further? He described the use of become a barrier, and data sharing. At the 50 cm level,
of UN Global Pulse satellite data for slum analysis and water supply in Kampala, there are important privacy concerns, including individual
Uganda, as well as big data analysis of radio talk and social identification, and the large untapped datasets still out
media in Uganda. We have to consider context, there is there.
huge variation inherent in satellite data – what works well in
 > >
17/31

Interesting examples cited of the use

of Satellite Imagery for Development included:
• Agricultural crop monitoring in Germany and how it relates to the Common • Mapping the growth and development of Dar es salaam (growing at 8-15% a
Agricultural Policy (Mark Doherty) year, no trees left in reserves, flood patterns have changed (Andrew Zolli)

• Crop yield rates in Mali in November 2016, single crops (Mark Doherty) • Deforestation in Bolivia (Andrew Zolli)

• Nitrogen dioxide mapping for urban centres is monitored daily, with hourly
frequency coming soon (Mark Doherty)
• Detect ships and put these together and get a global picture of global
emissions – ozone, CO and GHG methane gas (Mark Doherty).
• Solar installation in China (Andrew Zolli)
• Mapping poverty in Mexico and Sri Lanka (James Crawford)
• Mapping flood depth map of Houston to help insurance companies and the
State of Texas for future disaster preparedness (James Crawford)
• Mapping crop yields, poor harvests and famine in Iraq (James Crawford)
• Mapping steel furnaces and ship building in China (James Crawford)

• Mapping oil reserves and unequal negotiating power (David Hensen, UNEP)
 > >
• Location and growth of refugee camps (Einar Bjørgo, UNOSAT)
• Burning of Rohingya villages in Myanmar (Andrew Zolli)
• Analysis of the drought that preceded the crisis in Syria (Mark Doherty)
• Neighborhood bulldozed by speculators in Brazil ahead of the Olympics
(Andrew Zolli)
• Crop productivity in Nigeria and India (Marshall Burke).
• Measuring poverty or constituents of poverty in Sub-Saharan Africa
(Marshall Burke)
• Land use in Iraq (Naroa Zurutuza)
18/31
• School mapping initiative to map schools round the world – Liberia and
Nigeria (Naroa Zurutuza)
• Combining satellite and drone imagery to map crop yields in Malawi
(Naroa Zurutuza)
• Roof type of buildings is an important proxy indicator for the wellbeing
of a household and people who dwell within it in Uganda (John Quinn)
• Slum analysis – Kampala water supply in Kampala city (John Quinn)
• Satellite imagery can be used to explore why Lake Dracana in Western
Kenya is shrinking (audience, Human Rights Watch).
 > >
19/31

C
Session

Deforestation
 > >
20/31

Dominique Herman described recent developments in the use of satellite technology

of The Forest Trust to map deforestation and coverage. Monitoring forestation
(TFT) depends significantly on the definition of forest used – see Figure.
Global Forest Watch does a worldwide measure of forests, but it
does not separate plantations and planted forests from natural
forests, therefore a version for virgin forest is needed, such as
the Starling service (Airbus and TFT) - doing this systematically
across the tropics would be great. There is a move from real-time
monitoring to predictive modelling of what will happen (e.g.
deforestation around roads), especially if this is then combined
Figure: Different Definitions of Forests & Tree Cover
with real-world information (e.g. information about forestry
concessions).
 > >
21/31

Figure: Global Trends in Forest Cover Figure: The Causes of Tropical Deforestation Figure: Land-use

Radar (SAR) imagery is very useful because it can cut through cloud cover, but difficulty with slopes and accuracy. However, issues and barriers for
automation include large amounts of data to process and often satellite images are taken from various viewing angles of imagery (e.g. there is a big
difference between 0-10%). Many other factors can change the aspect of satellite imagery and make automation challenging such as: time of day, shadows,
cloud shadows, seasonality, trees do lose their leaves, atmospheric factors (haze, fog). AI can be used to separate out planted and virgin forests.
 > >
22/31

D
Session

Agriculture
 > >
23/31

described recent developments in the use of satellite About 1/3 of all food is wasted in both developed and
technology to map agricultural yields and address hunger. developing world. Poor infrastructure contributes to food
He described WFP’s work and how AI can help food security, waste in developing world. Lack of access to markets is
hunger, and smallholder agriculture. 1/9 people are living difficult and exacerbates poverty. AI can improve predictive
in hunger; conflict, natural disasters and extreme climate models and capacity-building. They can allow for different
events all contribute. The largest number of people living in kinds of crop mapping, including infrastructure assessment,
hunger are those living in chronic hunger. Malnutrition and and automated monitoring for 13 countries. Mr. Opp
food deprivation in the first one thousand days can stunt described innovation in agriculture for smallholder farms
children’s long-term growth. The world was making progress in East Africa. Key product areas include: input insurance;
Robert Opp against hunger – figures for extreme hunger were falling, credit extension; training and agronomic information;
of World Food despite the growth in world population, but this trend supply chains; de-risking behaviours; risk diversification. AI
Programme (WFP) reversed and we are now adding 20 million hungry people will unlock new forms of mobile-delivered services.
every year. Cash assistance is helpful to avoid disrupting
local markets.
> >
24/31

Figure: Impact of Conflict on Agriculture

Source: WFP/VAM.
 > >
25/31

described how it may be possible to lower cost of RCTs (randomized control trials).
Using remote technologies can help reduce costs, but building out AI quickly may lock
Mohanty
people in. CrowdAI comprises the following:
Sharada
• Crowdsourcing AI solutions for various issues (e.g. making models of human
skeletons)

• Creating forums for collective mapping of housing

• Standardized evaluation metrics and open satellite imagery

• Meaningful annotations to models (information on the ground) and accessible

baselines (low barrier to entry)

• Resources (host datasets and computer resources to evaluate submissions) and

expertise on the ground.
 > >
26/31

Satellites, satellite data and satellite imagery offer many exciting possibilities for mapping, charting and
exploring the Earth and all its natural assets, as well as the inhabitants and changing geologies of the Earth. We
may be able to model and learn more about terrestrial processes, as well as human development. However,
there are also risks and trade-offs involved in sustainable development. The same roads that enter into a
formerly natural territory can facilitate the growth of agriculture and help raise local incomes, but at the expense

Conclusions of forests and the natural habitat. The same data, mapping and imagery that can be used to explore our natural
resources and indicate promising sites for mining and drilling can also create information asymmetries in
markets, where information is unequally shared or accessed (for example, when negotiating mining contracts).
> >
27/31

My View of the Earth from Space, by

Women Who Have Flown in Space
 > >
28/31

“Here is the best view of the world. The view from space
Anousheh Ansari, station is amazing, from up here, everything is peaceful.
XPRIZE Foundation You can’t see the borders, you can’t see different races or
Board of Directors, religions, you just see one Earth, very peaceful and very
Space Ambassador, beautiful rotating and I think that’s part of the reason they
and the first female wish more and more people would be able to experience
private space this firsthand. Because from talking to astronauts, people
explorer have a different perspective on life and how important it is
for us to do everything in our power to preserve the only
home we have in the Universe”.
 > >
29/31

“Although I had been well prepared and had imagined

Liu Yang, it a thousand times, the image of space, I was still deeply
Astronaut and Pilot astonished. I could hardly describe its beauty because words
and the first fail to express how beautiful and miraculous it is. Looking
Chinese woman in back on Earth from 340 kilometers away, the planet where
space we humans and other creatures have lived for millions of
years, it looks so glamorous.

Samantha “It’s an exciting time where opportunities will multiply to go

Cristoforetti, to space.”
Astronaut and Pilot
 > >
30/31

Key Takeaways from the Satellite Track

Since 2017, it is now possible to image the entire Earth’s surface every day. However, there is a trade-off between the availability of
high-res data (which is available every few days) and low-res data (potentially every day), with high-res data obviously more costly and
expensive to obtain.

Global satellite monitoring might even prove the equivalent of global ‘closed-circuit TV’, and can be used to hold Governments and
other parties accountable for their actions – e.g. mapping burning villages in conflict areas, or potentially mapping carbon dioxide
emissions for future conformance/compliance with the Paris climate accords or other agreements.

We are moving from historical analysis to real-time information (e.g. retreat of the ice caps) and we shall increasingly move to predictive
modelling. Using AI means we can also economize on human analysis time and save the human analyst time for the tasks really needed.
 > >
31/31

We are moving from data asymmetries/poverty to information asymmetries with a potential impact on whole markets – e.g.
foreknowledge of steel production depending on furnace heat, or size of potential oil reserves in a new discovery.

The purposes we think imagery (and the associated data) may be used for are not always how they get used – we think they may be used
to achieve the SDGs, when in fact it may make exploitation even more efficient. There are a number of major users of satellite data not
even at the development table or talking with the development community (military on truck movements, Wall Street for crop yields and
taking positions on the futures market).

For deforestation, a road going into an area is the best predictor of the development of agriculture (electrification and a potential
increase in local incomes) within five years, as well as predictor of deforestation. This means that while satellite imagery could help
with the monitoring of certain SDGs (e.g. for tracking climate change), it could also help humankind become even better at finding and
exploiting the Earth’s natural resources (including fish reserves). There may literally be nowhere to hide from the eye(s) in the sky!

Regional learning in predictive models – algorithms trained in Africa may not always work in India – there are dangers in generalizing,
while models can only be as good as the data they are trained with – the quality of the data can determine the quality of the model.