Introduction
The United Nations High Commissioner for Refugees’ (UNHCR) 2024 mid-term report estimates that forced displacement has exceeded 120 million worldwide. The humanitarian sector is grappling with immense challenges in effectively addressing it. Recognising the imperative for innovative approaches, the Danish Refugee Council (DRC) is using AI technology to forecast displacement to be better prepared for the needs of displaced communities.
Background
The increasing human displacement due to conflicts, natural disasters, and climate change is challenging to address. Alexander Kjærum, Global Advisor at DRC, says that the use of AI was inspired by the insights of DRC’s country director in Myanmar. After returning from the Rohingya crisis, the country director highlighted the extensive data collection efforts undertaken by DRC and the broader humanitarian sector. However, in spite of the significant time spent, unexpected crises continued to surprise the humanitarian sector, as demonstrated by the 2015 refugee crisis. This realisation prompted a reflection on how to get more value from the data collected.
At the same time, there was a surge in Artificial Intelligence (AI) interest and discussion within the academic and business sectors. This raised a key question at DRC: Could AI be employed to enhance the ability to forecast future humanitarian crises?
AI Development
Why AI?
DRC utilises AI to process vast amounts of data and run algorithms to create forecasts based on many different indicators including conflict, health, environment, food insecurity, and socio-economic indicators. By analysing the data along these indicators, the models predict future displacement scenarios, helping humanitarian actors plan and respond more effectively to support affected populations.
Different models for different aspects
DRC has developed four predictive models: The Foresight model predicts national-level forced displacement; the AHEAD model focuses on predicting displacement at sub-district level; the Slow-Onset Drought-Related Displacement (SODRD) model forecasts displacement, with a focus on pastoralists, due to drought; and the SPIN model forecasts risk levels of pastoralist communities.
Approach
DRC began its journey in developing predictive models by first creating the Foresight model to predict displacement at the national level. This initiative was carried out in collaboration with International Business Machines Corporation (IBM). During this process, displacement drivers and credible data sources were identified, such as World Bank (WB), UN agencies, NGOs and academic institutions, resulting in a comprehensive dataset comprising over 120 indicators related to violence (e.g., conflict incidents and violence against civilians), governance (e.g., political stability and corruption levels), economy (e.g., unemployment rates and income inequality), environment (e.g., natural disasters and food security), socio-demographics (e.g., population growth and vulnerable groups).
The Foresight model is valuable for long-term planning and advocacy; however, there was a need for more operational tools that provide precise details on when and where displacement will actually occur, thereby helping staff on the ground to prepare. Therefore, the AHEAD model was developed to operate at the sub-district level, predicting three months into the future. Collaboration with country teams and regional offices was crucial to ensure the model accurately reflected local conditions and needs. Thus, a significant portion of the initial AHEAD model was developed by DRC’s West Africa regional office to account for local contexts.
Both models primarily address conflict-induced displacement. DRC then developed two additional models to also cater for climate-related displacement. The first, SODRD model, applied in Somalia and the Somali region of Ethiopia, analyses the interdependencies between rainfall, livestock, land structure, population movements and socioeconomic parameters. This model simulates potential displacement among pastoralists based on seasonal weather forecasts and has been utilised for proactive response planning in Somalia.
DRC attempted to replicate the SODRD model in the Sahel region; however, it proved too challenging due to the complexity of the model and lack of available data This led to the development of the SPIN model to predict risk levels in the Sahel countries' pastoralist corridors on the basis of historical security incidents. This model leverages the early warning system of a regional pastoralist network in West Africa, which collects alert messages on various issues including conflicts. The model predicts future alerts, enabling the mapping of safe corridors and passages for pastoralists to mitigate climate risks.
All four models were rigorously tested by applying them to historical data and real-time scenarios in various regions. The models have shown remarkable accuracy based on comparison between historical forecasts and the actual displacement that happened.
Operationalising AI
The Foresight model forecasts displacement at the national level across 26 countries. The AHEAD model is actively applied at the district level in specific regions, including the Liptako-Gourma region in the Sahel, which encompasses Burkina Faso, Mali, and Niger, as well as in South Sudan and Somalia.
The SODRD model is used in districts within Somalia.The primary users of the models are DRC program staff and humanitarian actors who are involved in displacement response efforts. They utilise user-friendly dashboards that display displacement numbers and trends. To assist staff who are not data scientists, snapshots and reports of the model results are provided. Regular updates and training sessions are also part of the user support strategy.The models are managed in-house by DRC’s data scientists who are handling modeling, updates, and information extraction.
Learnings
A critical learning point emerged regarding the importance of a structured data collection process. The initial ambition was to leverage internal data for the modelling, but it turned out that much of the internal data lacked historical depth and systematic collection, which inhibited the use for model development. Reflecting on this, Alexander Kjærum, Global Advisory at DRC, remarked:
“If we had the opportunity to start over, we would prioritise a more structured approach to data collection from the beginning.”
Another key lesson was the difficulty in helping the users of the models understand and trust the predictions made by AI models. The models are complex and can sometimes seem like a "black box," meaning it is not always clear how they arrive at their predictions. DRC have tried to address this by developing the ability to do scenario-based forecasts, where users can tweak the underlying parameters of the model to generate their own forecasts. This helps to create a better understanding for the end-users around the assumptions and trends that contribute to calculating the forecast.
Plans for the future
DRC is committed to expanding the geographical coverage of its predictive models to include more countries. A significant step includes a three-year grant to extend the AHEAD model initiative to at least 12 more countries. It is also exploring the integration of social media data (potentially from X and Facebook) to enhance analysis capabilities and is considering remote sensing to improve the accuracy and depth of its models.
Plans are underway to introduce new functionalities, such as enhance the users’ ability to create scenario-based forecasts. This feature will allow users to adjust parameters and generate various versions of predictions, giving the users a clearer understanding of the forecasting process of the models. The organisation is also considering the development of a specialised model focused on predicting significant surges in displacement. While the current models are good at forecasting typical displacement patterns, they often struggle with unprecedented events.
To further strengthen capabilities to carry out the tasks, DRC is expanding the team by adding more data scientists. For long-term financial sustainability, efforts are being made to maintain and expand partnerships with organisations like the International Organization for Migration (IOM), tech companies and secure funding from donors and philanthropists.
Where to learn more
- Alexander Kjærum, Global Advisory at DRC, alexander.kjaerum@drc.ngo
- DRC Predictive analysis