Combining binary classifiers to improve tree species discrimination at leaf level

Abstract

This paper focuses on the discrimination of seven different savannah tree species at leaf level using hyperspectral data. The data is small in size, high-dimensional and shows large within-species variability combined with small between species variability which makes discrimination between the tree species (hereafter referred to as classes) challenging. We focus on two classification methods: K-nearest neighbour and feed-forward neural networks for the discrimination of the classes. For both methods, direct 7-class prediction results in high misclassification rates. We therefore construct binary classifiers for all possible binary classification problems and combine them using Error Correcting Output Codes (ECOC) to form a 7-class predictor. ECOC with 1-nearest neighbour binary classifiers result in no improvement compared to a 1-nearest neighbour 7-class predictor whereas ECOC with neural networks binary classifiers improve accuracy by 10% compared to neural networks 7-class predictor, and error rates become acceptable.