Selecting a representative training set for the classification of demolition waste using remote NIR sensing