Using earth observation products to improve maize price predictions in sub-Saharan Africa

In rain-fed agricultural systems, extreme weather events and shifts in weather patterns can dramatically reduce agricultural productivity. Simple economic models of supply and demand project that reductions in agricultural productivity lead to a decline in food supply and an increase in food demand, which often results in increasing food prices. For millions of low-income households, high food prices limit both food availability and accessibility and decrease household food security. Given this chain of events, policymakers and aid programs often monitor local prices as an indicator of onset food insecurity crises. To improve the monitoring of food insecurity, we examine the ability of several earth observation (EO) products, which are often used to predict or explain agricultural productivity, to predict monthly maize prices in several markets throughout sub-Saharan Africa.

Our work is motivated by three factors: 1) Many regions across sub-Saharan Africa are experiencing changes in weather patterns which are affecting agricultural productivity and increasing the frequency of food insecurity crises. 2) EO products are easily accessible and freely available at fine scale spatial resolutions and high-dimensional temporal scales. Yet, they have not been fully utilized and implemented in routine international food price outlooks. 3) Price movements provide important information on the demand and supply of staple foods and are key insights to the onset of food insecurity. However, in developing countries, price data is often difficult to obtain, infrequently collected, and often has several missing observations.

In this paper we use EO products that capture temperature, precipitation, evaporative demand, and the density of vegetation as model inputs. We incorporate these inputs in two types of unsupervised machine learning models to predict market level monthly maize prices, namely tree-based methods and the Least Absolute Shrinkage and Selection Operator (LASSO). We compare the performance of these models to the univariate commonly used Autoregressive Integrated Moving Average (ARIMA) model. We find that the incorporation of EO products in some markets outperforms the univariate prediction models. We also find that the use of EO products has a varying degree of influence on predictive accuracy. To further understand these results and determine which EO products are most predictive of market prices we analyze the associations between predictive errors, model parameters, and spatial characteristics of the environment surrounding each market.