*2.7. Feature Selection*

There are multiple approaches for feature selection, feature subset selection, and feature weighting. Filter approaches like Relief are very fast and provide a weight for each feature. In contrast, wrapper

approaches have the big advantage of dealing well with high levels of redundancy and selecting the best subset with minimal size [38]. A major drawback is the high computational load. Feature selection at all scales (on ground-canopy and UAV images) was performed using a wrapper approach comprising a SVM or SVR, respectively. A sequential forward feature selection (Statistics Toolbox, Matlab2013a) was used, and the called criterion function minimizing the prediction error was implemented based on LIBSVM 3.18. For the SVM, the accuracy was maximized and for the SVR, the RMSE was minimized. Due to the limited number of samples in the UAV data set, a leave-one-out cross-validation was performed to generate the test predictions to calculate the criterion.
