How can AI be used to improve predictions of humanitarian need?

For years, DI has provided a world-leading, data-driven assessment of the global financing landscape for humanitarian response. However we’ve seen a growing emphasis on anticipatory action in humanitarian programmes. Start Network data indicates that while 55% of humanitarian funding targets somewhat predictable crises, only 1% is pre-arranged. The game-changing potential of machine learning represents a huge opportunity to harness the power of data and evidence to enhance anticipatory action in the face of crisis and support those most in need.

Although the frequency of climate events has been increasing in recent times, they are relatively rare in the historical data, and anticipating humanitarian needs is a formidable task. Against this backdrop, AI is emerging as a valuable tool, capable of learning from sparse data and recognising intricate patterns to improve predictions of humanitarian need. It might even be used to account for interrelationships between crisis drivers and feedback loops (for example, when displacement due to conflict creates additional pressure on climate which in turn exacerbates climate crisis).

We used a machine-learning model to project the impact of climate change on humanitarian need. To do this, we used a set of potential future climate-change scenarios (proposed by the Intergovernmental Panel on Climate Change) that capture both the climatic effects of varying greenhouse gas emissions from now until 2100 and the socioeconomic conditions that they might create.

1. Best-case scenario: stringent reduction in carbon emissions with minimal socioeconomic barriers to climate change adaptation and mitigation.
2. Middle of the road: intermediate reduction in carbon emissions with moderate barriers to climate change adaptation and mitigation.
3. Baseline scenario: no reduction in carbon emissions and significant challenges to climate change adaptation and mitigation
4. Fossil-fuelled development scenario: an increase in carbon emissions, significant challenges to climate mitigation, few challenges to climate adaptation.

What impact might these scenarios have on humanitarian needs?

Change in US$ values from 2023

What impact might these scenarios have on humanitarian needs

In each combined scenario, the projection consistently indicates a rise in humanitarian need until at least 2080, though this is likely an underestimate due to modelling constraints. The impacts of each scenario range from a plateauing 27% increase (our best-case scenario) to a 97% and rising increase (in a world that continues to rely heavily on fossil fuels). However, the crucial insight lies in the distinctions between scenarios, particularly the notable divergence between the best-case scenario and others. Further exploration of the interplay between drivers like conflict, displacement and climate is essential as their complex interactions are likely contributors to increasing and worsening protracted humanitarian crises.

Conclusion: The value of machine learning when projecting

Machine learning's flexibility provides a dynamic understanding of how elements within a scenario influence outcomes, improving the precision and adaptability of projections. As conditions shift – encompassing both single variables and their interactions – projections adjust. It's important to understand that scenarios aren't direct variables for specific outcomes; they define the context within which a projection is plausible. This nuanced approach allows machine learning models to adapt and offer insights into the complex interplay of factors shaping the impact of climate change on humanitarian needs.

While limited, the projections underscore the significance of a triple-nexus approach to crisis response and long-term development, especially in fragile and conflict-affected states that typically receive less development and climate investment. Prioritising anticipatory action, disaster risk reduction, and climate adaptation finance can alleviate the burden on humanitarian actors and finance as crises escalate in both scale and cost. Coordinated action is pivotal to preserving development gains and ensuring a more sustainable future.

You can read more about our approach below, or read our full methodology note. We welcome questions and feedback on our approaches and are always looking for new partners to help ensure data-driven evidence and analysis are used effectively in policy and practice. Join our mailing list, follow us on Twitter or LinkedIn, or get directly in touch.

Learn more about this research

Why we chose machine learning to address this question

Machine learning was critical to our approach on humanitarian projections because of three characteristics of the data we were drawing on: endogeneity, sparseness and scope complexity.

Endogeneity means that while we are ultimately trying to predict humanitarian needs, it is likely that these needs could themselves be a predictor of our displacement and conflict inputs.

Sparseness comes about because most countries don’t experience humanitarian need, climate disasters, conflict and displacement on an annual basis; the events we are trying to predict are relatively rare in the global context.

Lastly, scope complexity refers to the fact that actions taken at the subnational level can have regional impacts, but we’re measuring it all with national data. Natural disasters seldom affect entire countries at once, but they can precipitate man-made disasters that can have international spill-over effects.

As compared to traditional estimation techniques like autoregression, machine learning addresses many of these data challenges. It also provides some key lessons for machine-learning while evaluating humanitarian needs.

How we went about the modelling

We created this machine learning model using the Intergovernmental Panel on Climate Change (IPCC)’s Shared Socioeconomic Pathways (SSPs) which address the potential challenges for climate mitigation and adaptation, and the Representative Concentration Pathways (RCPs), which are centred on a metric known as ‘radiative forcing’: a measure of how much energy is retained in the atmosphere due to the greenhouse effect.

The five Shared Socioeconomic Pathways (SSPs) and seven Representative Concentration Pathways (RCPs) together form scenarios that match socioeconomic conditions with greenhouse gas emissions. From these, we derived four baseline scenarios offering a practical and comprehensive range of possibilities and used these model-projected changes in the incidence of conflict and numbers of people displaced until 2100. We used this data to create the four comprehensive models of projected changes in total humanitarian need.

While we tested several types of machine learning including random forest, gradient boosting, and support vector machines, a transformer was found to achieve the highest accuracy rate in fitting historical data. Machine-learning models do require larger amounts of data for training and some of this data can be procedurally generated by systematically joining adjacent countries into ‘mini-regions’. Furthermore, model hyperparameters can be tuned to better account for causal complexity: while our transformer model was able to predict simple relationships with 80%+ accuracy through 8 dimensional embeddings and a transformer depth of 1, more complex relationships were better estimated with 32 dimensional embeddings and a transformer depth of 4.