Maize yield estimation in West Africa from crop process-induced combinations of multi-domain remote sensing indices

Remote sensing data, crop modelling, and statistical methods are combined in an original method to overcome current limitations of crop yield estimation. It is then tested for timely estimation of maize grain yields and their year-on-year variability in Burkina Faso. Outputs from the SARRA-O crop model were used as a proxy for observed data for calibration. The final remote sensing-based yield model was constructed on the interaction between aboveground biomass at flowering (AGB-F) and crop water stress (Cstr) over the reproductive and maturation phases. Various vegetation and drought-related indices were derived from different spectral domains and tested. Model performance was evaluated by cross-validation against (a) simulated yields and (b) independent yields from ground surveys aggregated at village level. The results showed that the RF (Random Forest) model outperformed the MLR (Multiple Linear Regression) model for year-on-year yield estimation at the end of the season when compared to simulated yields. Surface soil moisture (SSM) information, as a proxy for soil water available for plant growth, together with information on the temperature of the canopy cover, helped to improve the RF maize yield model, impacting more particularly the estimation of crop water stress. Lastly, two months before harvest the RF model predicted 46% of the observed end-of-season maize grain yield variability. The combined remote sensing, crop model and machine learning method is thus an effective approach for estimating and forecasting inter-annual maize crop yields in environments where field data are scarce, such as in most parts of the African continent. However, more research is needed to better retrieve the spatial variability of yields in order to strengthen current agricultural monitoring systems, and to address societal challenges, such as declining food security.