Hyperspectral Feature Selection Ensemble for Plant Classification

Interest in hyperspectral imaging systems has increased recently substantially for studying and monitoring plant properties and conditions. The numerous financial (i.e. improve breeding process) and environmental (i.e. reduce usage of herbicide) advantages of such systems have been a driving force behind the latest surge. This paper aim to differentiate different plant species using hyperspectral image analysis. Main contribution of the work lies in the use of combined output of multiple feature selection algorithms, as compared to the use of single feature selection algorithm. Two independent hyperspectral datasets, captured by different instrumentations, were used in the evaluation. In total, six different feature selection algorithms (relief-f, chi-square, gini index, information gain, FCBF, and CFS) were used in the experiment. Experimental results show significant improvements in classification accuracy with the ensemble version of multiple feature selection algorithms compared to with the individual feature selection algorithms. Keywords—ensemble learning; feature selection; hyperspectral imaging; support vector machine