Developing Maize Yield Predictive Models from Sentinel-2 MSI Derived Vegetation Indices: An Approach to an Early Warning System on Yield Fluctuation and Food Security

Annual crop yield fluctuation due to natural and anthropogenic factors is a major concern of the Ethiopian Government. For an immediate response to drastically changing crop yields and resulting harvest failures and to enhance the country's food security in general, extensive area crop growth monitoring and early prediction of production are needed. In this study, we developed an early maize (Zea mays) yield forecasting model using Sentinel-2 MSI (Multispectral Instrument); the study was carried out in the Abaya district of the Oromia Regional State in Ethiopia. The model consists of the following components: (1) Sentinel-2 image-based cropland identification for different crop development stages, (2) extraction of time series Sentinel-2 vegetation indices for crop growth monitoring, and (3) a simple linear stepwise forward regression approach for yield prediction. We tested different spectral indices regarding their performance in describing the crop development and eventually predicting the expected yield. The result showed that (1) the linear Red-Edge Enhanced Vegetation Index (Red-Edge EVI), (2) the combination of the Enhanced Vegetation Index (EVI) and the Green Vegetation Index (GVI), (3) the combination of the Red-Edge EVI and Soil Adjusted Vegetation Index (SAVI), and (4) the combination of the Normalized Difference Vegetation Index (NDVI), Red-Edge EVI and SAVI offer the best predictive model about 2 months before harvesting with the highest coefficients of determination (R-2) of 0.73, 0.80, 0.84, and 0.88, respectively. The correlation for the GVI was generally lowest compared to established models, and no evidence of a peak correlation for NDVI was observed. Our approach showed a high accuracy of detecting maize fields, detecting crop phenology, and early predicting of grain yield for the study year 2018. Our simple model may generate early warning information, which may support in-time decision-making regarding food supply when critical yield fluctuations are to be expected.