Mapping Field-level Maize Yields in Ethiopian Smallholder Systems Using Sentinel-2 Imagery

Remote sensing offers a low-cost method for estimating yields at large spatio-temporal scales. However, the use of high spatial and temporal resolution remote sensors to map field-level yields in heterogeneous smallholder systems in the developing world is not well understood. Here we examined the ability of Sentinel-2 satellite imagery to map field-level maize yields across smallholder farms in two regions in Oromia district, Ethiopia. We specifically evaluated how effectively different indices, MTCI, GCVI, and NDVI, and different models, linear regression and random forest regression, can be used to map field-level yields. We also examined how generalizable our models were if trained in one region and applied to another region, where no data were used for model calibration. We found that linear regression models that use MTCI led to the highest yield prediction accuracies (R2 ranging from 0.24 to 0.47), particularly when using only localized data for training the model. These models were not very generalizable, especially when models were applied to regions that still had significant haze remaining in the imagery. Our results highlight the ability of Sentinel-2 imagery to map field-level yields in smallholder systems, though accuracies are limited in regions with high cloud cover and haze.