« Back to Results

Big Data and Near-Real-Time Monitoring of Food Emergencies

Paper Session

Monday, Jan. 4, 2021 10:00 AM - 12:00 PM (EST)

Hosted By: Agricultural and Applied Economics Association
  • Chair: Rob Vos, International Food Policy Research Institute

Predicting Food Insecurity with Machine Learning

Kathy Baylis
University of Illinois
Erin Lentz
University of Texas-Austin
Hope Michelson
University of Illinois


identification of food insecurity crises can enable faster and more effective humanitarian
responses to mitigate casualties from hunger and save lives. Using machine learning, we
develop a predictive model of food security based on readily available, spatially granular data
on prices, geography, and demographics. As with any rare event, one challenge with predicting
food crises is the low rate of severe food insecurity in existing data that could be used to train a
model data. We use several different approaches to address this imbalance to allow us to
capture a higher fraction of these rare events. We apply our procedure to forecast food
security in three sub-Saharan African countries: Malawi, Tanzania, and Uganda. Bearing in mind
the possible spatial-temporal correlations between observations in the training and testing sets,
we use previous years’ data to predict later years food insecurity. Combined with cost-sensitive
learning and sampling, the machine learning models constantly outperform the logistic
regression models in detecting the food-insecure villages. The machine learning models are
able to identify close to 100% of the food insecure villages and the majority of the most food
insecure villages compared to the logit model. To demonstrate how this model can be used for
real-time assessment, we apply a convolutional neural network (CNN) with transfer learning to
use satellite imagery to predict food security in Malawi. Using only the satellite imagery, we are
able explain up to 60% of the village-level variation in food security outcomes. Our findings
show that a data-driven model with the help of machine learning methods can significantly
improve a model’s ability to identify food insecure households even when the data are
imbalanced. Our paper demonstrates that this approach could be used in a scalable,
automatically-updated prediction model that could enhance current famine early warning

Forecasting Correlated Poverty and Malnutrition Indicators for Targeting, Monitoring and Evaluation Purposes

Linden McBride
St. Mary's College
Christopher B. Barrett
Cornell University
Christopher Browne
Cornell University
Leiqiu Hu
University of Alabama
Yanyan Liu
International Food Policy Research Institute


Recent extreme weather events, the COVID-19 pandemic and east African locust infestation of 2020, outbreaks of violence in various places, and food price shocks have all vividly demonstrated that food emergencies can arise quickly, demanding concerted policy response. High quality, subnational maps with tolerably accurate estimates of recent and current poverty and malnutrition conditions can be incredibly valuable to humanitarian and development agencies designing, targeting, monitoring and evaluating interventions in such settings. However, fielding surveys to generate the necessary data takes considerable time and money. So researchers and policymakers have been looking to earth observation and other publicly available, near-real-time data streams for information that might prove useful in providing such estimates. Recent research has established that machine learning methods applied to spatially precise remotely sensed and/or call data can predict some poverty measures with reasonably good out of sample forecasting accuracy (Blumenstock et al. Science 2015; Jean et al. Science 2016; Pokhriyal and Jacques PNAS 2017). To date, however, those methods and models have generally not performed well in predicting indicators of malnutrition. Simultaneously, poverty prediction has been moving away from theory-informed asset dynamics and related models towards atheoretic methods that simply offer the lowest prediction error. It is important that those who rely on such models to inform the distribution of benefits understand the tradeoffs and limitations of each approach. We report on a new machine learning based forecasting method that uses a suite of publicly available data series to estimate ensembles of poverty and malnutrition data to generate national poverty and malnutrition maps. We demonstrate, using nationally representative panel data from 12 low-income countries, that joint prediction o future poverty and malnutrition prevalence at a fairly local spatial scale (survey enumeration area) generates reasonably good predictive accuracy.

Digital Breadcrumbs & Dietary Diversity: Can Mobile Phone Metadata Predict Food Security

Oscar Barriga Cabanillas
University of California-Davis
Joshua Blumenstock
University of California-Berkeley
Daniel Putman
University of California-Davis
Travis Lybbert
University of California-Davis


Understanding whether a given product, program or intervention improves livelihoods is as
important as it is challenging. While established impact evaluation techniques can provide
credible evidence of impact, they can also be expensive and lack representativeness at scale
because they require active survey-based data collection – often for a spatially clustered subset
of beneficiaries. Building on encouraging research that uses mobile phone records (CDRs) in
developing countries to predict household wealth, we test whether CDRs might offer a big data
substitute for survey-based measures of food security and thus enable CDR-based impact
evaluation. We use a World Food Programme emergency unconditional cash transfer in
response to a severe drought in Haiti combined with CDRs from the dominant mobile network
operator as the basis for this test. A conventional survey-based regression discontinuity
evaluation of the transfer shows that it measurably improved food consumption, food security
and dietary diversity outcomes of beneficiaries. With these benchmark impacts in hand, we
then turn to machine learning to predict these outcomes using household CDRs. We are unable
to predict these outcomes with any meaningful degree of accuracy. With noisy predicted
outcomes, we are unable to replicate our conventional impact estimates. We conclude with a
post-mortem discussion of our failure in this context and application to extract any real
predictive content from CDRs. While these data clearly offer potent new insights about
development processes and outcomes, appreciating their limitations in such applications is a
critical prerequisite to tapping their full promise and possibilities as a novel data source.

Dynamic Poverty Prediction with Vegetation Index

Binh Tang
Cornell University
Ying Sun
Cornell University
Yanyan Liu
International Food Policy Research Institute
David Matteson


Accurate and timely estimates of economic status are critical to policymakers in the world’s
poorest countries. Previous work has applied convolutional neural networks (CNNs) to highresolution
satellite imagery to perform poverty prediction (Jean et al. 2016). Although
promising, such imagery has limited access and lacks a temporal signal. In this paper, we
leverage the continuous streaming of the normalized difference vegetation index (NDVI), one of
the widely known satellite measurements of Earth’s vegetation greenness, to estimate two key
poverty indicators, consumption expenditure and asset, in Malawi, Nigeria, Rwanda, Tanzania,
and Uganda. As ultra-poor regions heavily depend on agriculture, NDVI provides an
uninterrupted signal for crop heath and poverty tracking in general. We show that publicly
available, moderate-resolution vegetation index can be used with CNNs to produce equally
accurate poverty estimates as satellite images for developing countries that are heavily
dependent on agriculture. In contrast to previous work, the continuous streaming of wellknown
vegetation indices also allows us to update our estimates in light of weather shocks,
opening up the possibility of making dynamic poverty mapping at minimal cost. We perform
poverty prediction for an out-of-sample period and capture changes in poverty measures for
ultra-poor regions in Uganda.
JEL Classifications
  • A1 - General Economics