Forecasting transitions in the state of food security with machine learning using transferable features

Area of Science:

• Agricultural economics
• Environmental science
• Data science

Background:

• Food insecurity is a growing global concern exacerbated by conflict, climate change, and economic instability.
• Current methods for assessing food security, such as expert consensus and surveys, are resource-intensive, time-consuming, and costly.
• Accurate forecasting is crucial for enabling timely interventions by humanitarian organizations.

Purpose of the Study:

• To develop and evaluate an extreme gradient-boosting machine learning model for forecasting monthly food security transitions in Ethiopia.
• To achieve spatial granularity at the livelihood zone level and provide forecasts with lead times from one to 12 months.
• To utilize open-source data for a scalable and transferable forecasting solution.

Main Methods:

• An extreme gradient-boosting machine learning model was employed to predict transitions in food security status, represented by Integrated Food Security Phase Classification (IPC) data.
• 130 variables derived from 19 categories of open-source data were used as predictors, including climate, land, market, conflict, infrastructure, demographics, and livelihood zone characteristics.
• Model performance was evaluated against baseline methods, focusing on forecasting deteriorations and improvements in food security.

Main Results:

• The model identified food security history and surface soil moisture as the most relevant predictors.
• The machine learning model significantly outperformed baseline methods, achieving at least double the performance (F1 macro score) for forecasting deteriorations.
• The model demonstrated better performance for longer-term forecasts (7 months; F1 macro average = 0.61) compared to short-term forecasts (3 months; F1 macro average = 0.51).

Conclusions:

• Combining machine learning with IPC ratings and open data offers a valuable enhancement to existing food security forecasting approaches, providing longer lead times and more frequent updates.
• The developed forecasting model shows potential for transferability to other countries due to the widespread availability of predictor data from global repositories.
• This data-driven approach can support more effective and proactive food security interventions, particularly in regions vulnerable to climate change and conflict.