Quantifying the effects of climate and conflict on forced displacement
A new tool to automate the download of satellite imagery and quantify drought patterns — particularly in areas where violent conflict and displacement are prevalent — will give UNHCR and others insights to inform proactive and responsive humanitarian measures.
By Amy Lynn Smith — Independent Writer + Strategist
And the Innovation Service
Forced displacement is intimately tied to climate change and violent conflict. Better understanding those complicated relationships is essential to enabling UNHCR to proactively fulfil its mandate. In Somalia, for instance, drought and conflict can be key factors determining when and where people move. Accurate data-driven models representing that dynamic could enable a more effective humanitarian response.
Tracking and anticipating the movement of people in Somalia was the aim of Project Jetson, a predictive analytics experiment started in 2017. Although no longer in operational use, Jetson gathered extensive evidence on how climate and violent conflict exacerbate forced displacement, and how artificial intelligence (AI) could be used to inform programming. The project was ultimately focused on predicting arrivals, however, rather than identifying the root causes of displacement.
Climate is a key piece of this puzzle — but humanitarian agencies looking to develop their own modelling around specific research questions often encounter difficulties finding a historic dataset that describes climate-related conditions, according to Sofia Kyriazi, a data and AI engineer with UNHCR’s Innovation Service.
The Innovation Service has been collaborating with the University of Essex — specifically the Human Rights, Big Data and Technology initiative — to develop a solution. (If you’re a data expert keen to skip ahead and get straight to our open-source code for this new tool, you can find it here, and we welcome your input!)
Computer science PhD student Bruno Arcanjo began working with the Service in May 2021, having picked up a project initially started by another researcher, Grigorios Kalliatakis, in 2019. Arcanjo’s research focuses on computer vision — in other words, how artificial intelligence can automatically classify images according to certain patterns. Arcanjo says:
The ultimate purpose is to help people in Somalia. The way we get there is by detecting patterns in drought in the different regions of Somalia, and trying to correlate it with violent conflict.
Having established a mechanism to accurately detect these patterns, systems can be established to respond — supporting UNHCR’s preparedness and enabling more preventive actions.
Looking more closely at climate
Arcanjo credits his research predecessor for putting all the pieces in place to gather data on climate. Kalliatakis established a solid foundation for the research by creating a tool to automate the download of high-volume files of satellite imagery. Arcanjo has enhanced his predecessor’s code to make the automation more accurate, delimiting the download by region and automating colour-based classification of the imagery.
The software Arcanjo is building with the Innovation Service uses two well-established metrics to assess the plant cover captured in these satellite images: the Vegetation Condition Index (VCI) and Normalised Difference Vegetation Index (NDVI). These numeric indexes assess an area’s vegetation status, from good to poor, as well as how it might be changing. In most parts of the world, a direct relationship can be inferred between this data and rainfall patterns. Arcanjo notes:
If a place has good vegetation, represented by a certain colour, and suddenly the colour changes — meaning the vegetation condition is worsening — we can start to assume there might be a drought. But it’s not just the data but also the history of the region [that matters], because if an area always has poor vegetation, it’s not necessarily a drought — you have to look at how the VCI changes over time.
Using AI to solve a satellite imagery challenge
Satellite imagery collection has a huge challenge, stemming from the fact that both the photographer (the satellite) and its muse (the Earth) are moving.
The satellite passes over a geographic location at a certain time, and takes a picture of that area from a certain distance and angle. The next time the satellite passes through the same point in its trajectory, its position — or the Earth’s rotation — might mean it takes the picture from a different angle. The picture could be straight on, at a vertical angle to the Earth’s surface (resulting in what is known as a “nadir image”), or its perspective might be slightly off-kilter (an “off-nadir image”).
Modern satellites are able to collect off-nadir images to improve revisiting time (the time interval between two identical satellite flights), but such images typically suffer from decreased spatial resolution and distortions.
There are several research studies into how AI can solve the off-nadir problem (i.e., the fact that the images are of subtly different areas) without compromising image quality. However, few of these studies are conducted a) with applications to vegetation imagery in Sub-Saharan Africa or b) in forced displacement settings (given the issues of collecting satellite imagery in areas of active violent conflict).
This is where Arcanjo comes in.
Layering big data to map vegetation change
In data-science speak, Arcanjo has built an earth-observation/remote-sensing longitudinal analysis map to analyse trends and seasonality, using AI computer vision techniques, based on previous Project Jetson research work.
In other words, this software enables users to:
- Automatically download satellite images from earth-observation open data repositories (for instance, LANDSAT) based on desired coordinates.
- Crop each squared image to extract the part that corresponds with at least a portion of a region of Somalia. Since most regions are big enough to require several images, these cropped squares must be pulled together into a composite to form a complete region.
- Automatically categorise different images from the same region according to the state of their vegetation, assigning colours based on VCI and NDVI values.
- Build a colour-coded visual map of vegetation changes over time.
Using data to forecast the future
This climate-related satellite imagery is just one dataset — and Somalia is just one setting this process could be applied to.
The tool we’re working on will employ satellite imagery gathered over the last decade in addition to other datasets (which all fed into Project Jetson). These include: complementary climate-related data (for instance on weather anomalies or climate sensitive commodities) from development organizations in the region, such as FAO SWALIM or FSNAU; UNHCR’s wealth of data on forcibly displaced populations; and violent conflict data from organizations like ACLED and the Uppsala Conflict Data project.
The next step for the project is to combine all of this data to train human-supervised machine-learning models. Once computers have been trained to recognize what specific data means (for instance, how certain colours correspond to plant cover), they will be able to create forecasts. Kyriazi explains:
We want to know, for example: If drought is increasing, does it make conflict increase or decrease in a specific region? How many months does it take for conflict to rise in areas where you have dramatic changes in the levels of vegetation? For this project, we’re specifically interested in how climate indicators affect conflict, which has a direct effect on displacement.
An accessible, flexible tool
The collaboration between the University of Essex — via Arcanjo — and the Innovation Service has generated a flexible open-source tool that can be used, edited and improved on by humanitarian actors, software developers, data scientists, data engineers and computer scientists. The tool doesn’t support a user interface, so its main audience for now is skilled users who are able to interact with the command line.
If you’re a data scientist interested in helping to build out the frontend, we welcome your input! You can find the tool’s open-source code here.
With accurate data inputs, this tool’s potential is expansive — with positive implications for the world’s most vulnerable. Arcanjo says:
We will ideally be able to use the end dataset (a matrix of VCI values per region) to make connections between this data and other factors — such as violent conflict — to predict if there are going to be incidents down the line, which can lead to displacement. Then that information can be used to do something about it to prevent people being hurt. That is, in my view, the ultimate goal: to help humans at the end of the line.
Explore the open-source code here and find out more about Project Jetson here. For more on UNHCR’s Data Innovation workstream, visit our website.
