Satellite-based multi-annual yield models for major food crops at the household field level for nutrition and health research: A case study from the Nouna HDSS, Burkina Faso

Abstract

Increasing frequencies of climate change-induced extreme weather events like prolonged droughts pose significant challenges for small-scale subsistence farmers in sub-Saharan Africa, who rely on the yearly harvest by more than 80 % of their nutritional needs. However, we do not have a good understanding of yield estimates at the field and household level (with a mean field size of < 2 ha) to understand nutritional priorities in vulnerable communities due to their scarcity in the literature, particularly yield estimates that do not require re-collection of in-situ data. Statistical models for estimating regional food crop yields based on high-resolution satellite data at the field level may provide better insights into how to address health risks such as child malnutrition. Especially in low-resource contexts, where the burden is greatest and expected to worsen in future climate projections. Our study developed crop-specific, satellite-based yield models using a novel three-year data set of in-situ yield measurements as exemplified for a rural region in Burkina Faso. The aim of the model is to reduce the need for in-situ field data collection while still assuring accurate yield estimates at the field level. The model employed LASSO regression and was based on monthly vegetation index composites from Sentinel-2 and weekly accumulated Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) rainfall data. Our yield modeling results showed that there was less overfitting when there was more training data over three years that demonstrated a wider range of yields, which also led to better model fits. R² values ranged from 0.62 (Maize) to 0.3 (Sorghum) for the three-year yield models, with normalized root mean square error (nRMSE) values ranging from 12 % − 16 %. An additional plausibility check confirmed the validity of our models, as we compared the magnitude of our yield estimation with national yield statistics for Burkina Faso. We demonstrated that our models based on three years of in-situ data may capture some of the inter-annual variability in yields, which could be a step toward minimizing the necessity for in-situ measurements in the future. Our advances in predicting yield estimates at the field level enable a linkage between household-level yields, socioeconomic indicators, nutritional status of children, and the health status of the household members. A further application is linking high-resolution yield data to farmers’ productivity losses from increasing heat under climate change.