NON-INVASIVE ESTIMATION OF TRASH LEVELS IN BILLET CANE USING IMAGE ANALYSIS

Image analysis of mixed billet cane and trash found that cane, leaf, and tops show characteristic patterns in the amount of light reflected in each band (R, G, B, & NIR) but that there was considerable overlap between cane and trash categories. Including region based radiometric and spatial descriptors in cane/trash classification schemes improved classification from about 60% accuracy to about 80% accuracy. The study found that surface coverage proportions of leaf and cane for a given cane/leaf weight fraction are highly variable. Images of about one metre square are required to usefully discriminate trash levels for cane weight fractions between 80% and 100%.