AWS Government, Education, & Nonprofits Blog

Helping to End Future Famines with Machine Learning

In 2017, more than 20 million people across northeastern Nigeria, Somalia, South Sudan, and Yemen faced famine or famine-like conditions, the result of a complex intersection of conflict, poverty, climate change, and food prices. Today, 124 million people live in crisis levels of food insecurity, requiring urgent humanitarian assistance for their survival.

The United Nations, World Bank, and International Committee of the Red Cross (ICRC) with support from Amazon Web Services and other technology companies, recently launched the Famine Action Mechanism (FAM). The FAM is the first global mechanism dedicated to preventing future famines. In the past, responses to these devastating events have often come too late, once many lives have already been lost.

The FAM seeks to change this by moving towards famine prevention, preparedness, and early action—interventions that can save more lives and reduce humanitarian costs by as much as 30%. The initiative will use the predictive power of data to help trigger pre-arranged funding, working closely with existing systems.

“The fact that millions of people—many of them children—still suffer from severe malnutrition and famine in the 21st century is a global tragedy,” World Bank Group President Jim Yong Kim said. “We are forming an unprecedented global coalition to say, ‘no more.’ The Famine Action Mechanism is a preventative approach that knits together innovative technology, early financing, and strong partnerships on the ground in an effort to prevent famine. It will help us deploy our combined resources to protect the poorest and most vulnerable, and it will allow us to refocus our collective attention on the millions of chronically food-insecure people who suffer each year.”

Machine Learning on AWS

Predicting famine is difficult due to many factors, including differing international standards for famine data collection and diverse causes of famine (climate, conflict, economic policy). Although Machine Learning (ML) and AI models can learn complex patterns and make accurate predictions, often the potential of these technologies is limited by the quality of the data.

Food security assessments have been championed over the last few decades, and several prominent early warnings systems and data providers are in place that help inform on-the-ground programming efforts. One clear challenge, however, is the frequency at which data is produced and analyzed as major food security reports usually come out only twice a year. These reports are the result of intense surveys and require significant investments in time and energy.

Innovative technology and modeling approaches can help address this challenge by helping to inform the level of food insecurity occurring between major reporting cycles, establishing a continuous, real-time stream of information. Additionally, data points that ultimately are verified on the ground can be predicted ahead of time with the aid of machine learning. This has tremendous implications for the ability of decision makers to identify trends earlier and ultimately unlock resources for preventative action that can stop risks from becoming crises. These efforts also have the potential of bringing in significantly more data from a wider range of stakeholders to reveal connections to other critical development and humanitarian concerns.

Drawing from World Bank and UN work, AWS has built an ML pipeline from data ingestion and storage in Amazon Simple Storage Service (Amazon S3) to model deployment using Amazon SageMaker. To create a model that takes into account the multidimensional causes of famine, AWS is using a diverse dataset processed by the World Bank team together with UN partners, including features from satellite imagery, conflict data, weather forecasts, local food prices, and agricultural production and is able to train multiple machine learning models. The models have so far been able to identify significant and distinctive patterns across geographical regions and countries.

Future efforts include deeper analysis on conflict data to take into account the spatial and temporal nature of violence. This will enable even more accurate predictions in regions where conflict is the primary driver of famine, such as Afghanistan.

Using AWS tools like Amazon SageMaker and Amazon S3, AWS has built an ML model, which presents further improvements to initial models with an 11% increase in accuracy for Somalia and South Sudan. Employing the built-in features of SageMaker, such as automated model tuning and auto-scaling of compute capacity, allows for rapid model iteration and experimentation, producing a model with increased accuracy in a shorter period of time.

Increased Accuracy and Performance

In practice, this increase in accuracy means that more IPC Phase 3 events and IPC Phase 4 events can be correctly identified. This will enable the UN and World Bank to address crisis and emergency events more quickly. Rapid response is critical, as IPC Phase 3 is officially designated as a crisis event, where 1 in 5 households are not able to meet nutrition needs. In addition, Phase 4 is an emergency phase that includes large food consumption gaps.

“With advances in artificial intelligence in the cloud, we’ve been able to take these capabilities and put them into the hands of large and small organizations alike and regardless of expertise in machine learning. As we feed innovation at all levels, what’s become abundantly clear is that this powerful technology has the profound potential to do good in the world,” said Swami Sivasubramanian, Vice President of Amazon Machine Learning. “I’m especially energized by the work underway with the United Nations, World Bank, and others to leverage AWS machine learning services like Amazon SageMaker in the FAM to help more accurately and quickly identify areas that are at risk for famine-like conditions and deploy aid before it is too late. Thanks to artificial intelligence, for the first time in human history, we have the power to inhibit the plague of famine and save lives before situations reach a crisis level.”

The FAM is being introduced to a small group of vulnerable countries initially, with the aim of ultimately providing global coverage.

Listen to the podcast titled “Before famine strikes: heeding the warnings to save lives.” And watch the video featuring Franck Bousquet, Senior Director, Fragility Conflict and Violence, World Bank from re:Invent.