Files in this item



application/pdfZHOU-DISSERTATION-2020.pdf (7MB)Restricted to U of Illinois
(no description provided)PDF


Title:Three essays on machine learning and food security
Author(s):Zhou, Yujun
Director of Research:Baylis, Kathy
Doctoral Committee Chair(s):Baylis, Kathy
Doctoral Committee Member(s):Michelson, Hope; Winter-Nelson, Alex; Brunner, Robert
Department / Program:Agr & Consumer Economics
Discipline:Agricultural & Applied Econ
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):food security, machine learning, development economics
Abstract:This dissertation contains three chapters studying the interactions between weather shocks, food prices, food policy and food security. Food insecurity has resurfaced to be a vital issue in the past two decades in sub-Saharan Africa. Income and price shocks that endanger food security could from different aspects, including drought, flood, increased volatility in international and domestic market prices, and changes in agricultural policies. Knowing when and where food insecurities happen and the reasons why they happen matters for improving humanitarian and aid responses, as well as making the right agricultural policy. My dissertation focusses on the factors that are central to food security, including the market price of grains, weather shocks in grain production areas, and agricultural policies. The first chapter focuses on the effect of stockholding policy on food prices based on evidence from Zambia. This chapter evaluates the effect of the purchase and sales activities of the Zambian Food Reserve Agency (FRA) on maize market prices across more than thirty markets in Zambia using monthly price data from 2003 to 2008. To deal with the endogeneity of the purchases and sales activities on the prices, I use predicted FRA purchase and sales targets as instrumental variables. Results show that the purchase activities help to raise producer prices during harvest seasons and that selling of maize stocks at subsidized price help to decrease maize prices during lean seasons. Controlling for other policies, I find evidence that FRA activities stabilize retail prices in major district markets within the cropping year. The reduced volatility in price from the policy would help households to become food secure yet at the cost of substantial financial burdens to the government. However, there is limited evidence found that the FRA is able to reduce price volatility between years. The second chapter builds a framework for food security predictions using readily available public data and machine learning methods that focuses on improving the accuracy of identifying the food insecure households. Rapid and accurate identification of food insecurity crises can enable humanitarian responses to mitigate casualties from hunger and save lives. With a focus on capturing the food insecure households as much as possible, I combined machine learning algorithms, cost-sensitive learning, and data sampling techniques to improve the model’s ability in identifying the minority class. The framework is tested on three different countries in Sub-Saharan Africa with 10% to 60% higher area under the ROC curve (AUC) than the baseline model. With an increase in the amount of false-positive cases that can be solved, I set a threshold that consistently successfully identify more food insecure villages with a lack of training data. The third chapter expands the food security prediction in the second chapter to a broader level by making use of satellite images. Predicting food insecurity crises ahead of time can allow humanitarian agencies to position resources to help those places where they are needed the most, mitigating casualties from hunger and saving lives. I utilize satellite imagery to develop a predictive model of food security. Using spatially granular data, the model is able to generate food security prediction at the “grid” level in Malawi with reasonable performance on the validated testing dataset. Combined with readily available secondary information from market price data and environmental variables, the model explains up to 61% of the variations in food security measures. This remote sensing-based approach, when combined with other types of secondary data, does well in predicting food security outcomes. This chapter opens door for a scalable, generalizable, automatic framework that can generate grid level, real-time prediction of food security that could greatly enhance the current famine early warning systems.
Issue Date:2020-07-13
Rights Information:Copyright 2020 Yujun Zhou
Date Available in IDEALS:2020-10-07
Date Deposited:2020-08

This item appears in the following Collection(s)

Item Statistics