Spectral features fusion of effective criteria on wheat yield prediction

The yield of the wheat crop is affected by the climate and soil parameters such as moisture and nutrients, plant pests and diseases. The main objective of this research is the feature level fusion of multiple effective criteria on the wheat yields using linear and machine learning regression models. The effects of vegetation condition, moisture, nutrients and pests on wheat yield are represented by spectral indices those are extracted from remotely sensed data. Optimum spectral indices are selected as the input features to each of the multiple linear and machine learning regression models such as decision tree, support vector regression and generalized regression neural network. The evaluation of the experimental results in eight wheat fields indicates that the wheat yield prediction based on spectral features fusion show the mean improvement of 0.81 in RMSE comparing with considering only one vegetation index in all regression models. Moreover, all investigated machine learning regression models have about 0.03 more performance than the multiple linear regression model as indicated by R2 coefficient. The generalized regression neural network model with the least RMSE error 0.0063 has the best results compared with other machine learning regression models and MLR.