UAS-based multispectral imaging and feature selection for yield prediction

Abstract

Yield prediction of wheat using unmanned aerial systems requires the extraction of a few meaningful features from the images. To identify these features, a total of 52 different spectral indices and 14 statistics at 19 dates during the growth season were extracted from multispectral images. Texture features were found more useful for yield prediction during the stem elongation stage compared to vegetation indices. Building a random forest model using the features selected by the feature selection algorithm Boruta gave an RMSE of 34.7 g/m2 compared to 50.7 g/m2 by the model built with all available features. Feature selection therefore improved the stability of yield prediction.