FDL Europe 2018

In 2018 we expanded the Frontier Development Lab from the US to Europe. Working in partnership with the European Space Agency (ESA) we hosted a research sprint at the University of Oxford. With the support of NVIDIA and the Satellite Applications Catapult, two teams of researchers tackled challenges under the mission area “Mission Control for Planet Earth”. We wanted to test the idea that it might be possible to create a virtual Mission Control Centre that would enable us to monitor, predict and be better placed to respond to threats to the planet and its inhabitants.

The research sprint was a great success, and the outputs from the teams showed that the idea of Mission Control for Planet Earth has great potential to be further developed in future.

team one: mapping informal settlements

Challenge: To use AI and Earth observation data to identify, and automate mapping of, informal settlements to enable governments and aid agencies to better monitor and support these communities.

One-third of the world’s urban population lives in informal settlements (source: World Bank, 2009 data). Yet it is not uncommon that the locations and sizes of these settlements are simply unknown. People living in these areas often have no security of residence, often lack basic services, and housing is unlikely to comply with planning, building and safety regulations.

Detecting and mapping informal settlements is of importance to a number of the United Nations Sustainable Development Goals (SDGs). This is due to that fact that it is the most socially and economically disadvantaged people who often call these informal settlements “home”.

However, data regarding informal and formal settlements is hard to come by, and if it exists at all, is often incomplete. Identifying where these settlements are is of paramount importance to both government, and non- government organisations (NGOs), such as the United Nations Children’s Fund (UNICEF), who can use this information to assist in delivering effective social and economic aid. Traditional methods of gathering this data are costly, preventing efforts to gather such information at the required scale.

Our team of researchers had just eight weeks to see if they could use the combination of Earth observation data and AI techniques to map informal settlements - no easy challenge!

The FDL Informal Settlements team developed an effective end-to-end framework that can detect and map the locations of informal settlements using only low-resolution, freely available, Sentinel-2 satellite imagery. This is a leap forward from previous studies that used much more costly costly “very-high resolution” (VHR) satellite (and aerial) imagery.

Their research also demonstrated a deep learning approach to detect informal settlements with VHR imagery for comparative purposes, allowing agencies to monitor changes in settlements over periods of time.

In addition to this, the team showed how AI approaches can detect informal settlements by combining both domain knowledge and machine learning techniques, to build a classifier that looks for known roofing materials used in informal settlements.

Ongoing Impact
The team has presented their work at prestigious AI conferences (NeurIPS and AAAI), and it was featured by FDL Europe partner NVIDIA and GeoConnexion. They are seeking partners to continue working on the project and support its deployment.

Mapping Informal Settlements in Developing Countries using Machine Learning and Low Resolution Multi-spectral Data (arXiv.org, 2019)

Generating Material Maps to Map Informal Settlements (arXiv.org, 2018)

Mapping Informal Settlements in Developing Countries with Multi-resolution, Multi-spectral Data (arXiv.org, 2018)

Presentation Details
NeurIPS 2018 Workshop on Machine Learning for the Developing World | 2018
AAAI ACM Conference in AI, ethics and society | 26th - 27th January 2019, Honolulu, Hawaii, USA


TEAM TWO: Machine Learning to Aid Disaster Response

Challenge: To use state-of-the-art AI to dramatically reduce the time and effort required to create effective disaster impact maps.

Disaster events such as earthquakes, hurricanes, and floods cause loss of human lives and create substantial economic damage. Lack of information about the level of damage immediately following these events restricts first-responder efforts, hinders efficient response coordination by authorities, and delays can lead to additional loss of life.

The minutes, hours and days after a natural disaster are critical and insights into the scale and type of damage can help to inform the rescue effort. The faster that this can be achieved, the faster an informed response can be put together.

The FDL Disaster Response team developed a novel approach to performing rapid segmentation of flooded buildings by fusing multi-resolution, multi-sensor, and multi-temporal satellite imagery in a convolutional neural network (CNN).

The method consists of multiple streams of encoder-decoder architectures that extract temporal information from medium-resolution images and spatial information from high-resolution images before fusing the resulting representations into a single medium-resolution segmentation map of flooded buildings.

This enables significantly expedited generation of satellite imagery-based flood maps, which are crucial for first responders and local authorities in the early stages of flood events.

By incorporating multi-temporal satellite imagery, rapid and accurate post-disaster damage assessment can be gained, helping governments to better coordinate medium and long-term financial assistance programs for affected areas.

This method exceeds state-of-the-art models for building-footprint segmentation as well as alternative fusion approaches for segmentation of flooded buildings, and can be performed well using freely available medium-resolution data instead of significantly more detailed (and expensive) very high-resolution data used in previous methods.

Ongoing Impact
The team have presented a poster and a paper at leading AI conferences NeurIPS and AAAI. They are currently working on spin-off projects stemming from their initial work. UNICEF has expressed an interest in deploying this work in real-world situations and the team is currently working on a feasibility study to see how this could be implemented.

Presentation Details
Thirty-Third AAAI Conference on Artificial Intelligence | 27th January - 1st February 2019, Honolulu, Hawaii, USA

“One of the great things about the FDL Europe experience has been getting to interact with AI specialists from computer vision. Computer vision offers us a tremendous new tool to be able to understand changes over time and begin running predictions for the future.”    Dr Ramona Pelich, FDL 2018 Researcher

“One of the great things about the FDL Europe experience has been getting to interact with AI specialists from computer vision. Computer vision offers us a tremendous new tool to be able to understand changes over time and begin running predictions for the future.”
Dr Ramona Pelich, FDL 2018 Researcher


The results of FDL 2018 were presented to experts from the Earth Observation community at ESA ESRIN's 'Phi week' event in Frascati, Rome, 2018 as well as NeurIPS and AAAI - leading AI conferences.

More background on the results and process of FDL 2018 can be read here: