Estimation of Maize grain yield using multispectral satellite data sets (SPOT 5) and the random forest algorithm

Crop yield estimation is a very important aspect in food production as it provides information to policy and decision makers that can guide food supply not only to a nation but also influence its import and export dynamics. Remote sensing has the ability to provide the given tool for crop yield predictions before harvesting. This study utilised canopy reflectance from a multispectral sensor to develop vegetation indices that serve as input variables into an empirical pre-harvest maize (Zea mays) yield prediction model in the north eastern section in Free State province of South Africa. Some fields in this region that were grown of maize under rain-fed conditions were monitored and the grain harvested after 7-8 months with actual yields measured. The acquisition of suitable medium resolution SPOT 5 images over this area was in March and June before the grains were harvested in July of 2014. A number of well known spectral indices were developed using the visible and near infrared bands. Through the random forest algorithm predictive models, maize grain yields were estimated successfully from the March images. The accuracies of these models were of an R-2 of 0.92 (RMSEP = 0.11, MBE = -0.08) for the Agnes field and for Cairo the R-2 was 0.9 (RMSEP = 0.03, MBE = 0.004). These results were produced by the SAVI and NDVI respectively for both fields. It was therefore evident that the predictive model applied in this study was site specific and would be interesting to be tested for an optimal period during the plant life cycle to predict grain yields of maize in South Africa.