Predicting within-field variability in grain yield and protein content of winter wheat using UAV-based multispectral imagery and machine learning approaches

Prediction of crop yield and quality is an essential component of successful implementation of precision agriculture. Given the recent commercialization of low-cost multispectral cameras mounted on unmanned aerial vehicles and advances in machine learning techniques, prediction systems for crop characteristics can be more precisely developed using machine learning techniques. Therefore, the model performances for predicting wheat grain yield and protein content between the machine learning algorithms based on spectral reflectance and plant height (e.g. random forest and artificial neural network) and the traditional linear regression based on vegetation indices were compared. Although the machine learning approaches based on reflectance could not improve the grain yield prediction accuracy, they have great potential for development in predicting protein content. The linear regression model based on a 2-band enhanced vegetation index was capable of predicting the yield with a root-mean-square error (RMSE) of 972 kg ha(-1). The random forest model based on reflectance was capable of predicting the protein content with an RMSE of 1.07%. The reflectance may have been linearly correlated with total biomass; thus, it was also linearly correlated with grain yield. There was a nonlinear relationship between the grain yield and protein content, which may have resulted in the higher model performance of the machine learning approaches in predicting protein content. However, this relationship would be variable according to the environment and agronomic practice. Further, field-scale research is required to assess how this relationship can be varied and affect the model generality, particularly when predicting protein content.

Automated summary

Comparison was made between machine learning algorithms based on spectral reflectance and traditional linear regression models for predicting wheat grain yield and protein content, showing machine learning approaches have potential for predicting protein content but not for improving yield prediction accuracy.