Signifying the information carrying bands of hyperspectral imaging for honey botanical origin classification

Abstract

Abstract The development of a robust honey classification model based on hyperspectral imaging requires finding significant wavelength bands describing honey botanical origins. Significant wavelength bands could be discovered by using feature selection methods that each method is commonly used as a standalone method. This paper proposes a strategy of combining some feature selection methods to maximise the reduction degree which utilise as a minimum number of bands as possible with marginal performance degradation. The proposed strategy successfully found relevant bands from wavelength bands spans within 400–1000 nm to classify and qualify 21 types of honey coming from different botanical origins. The proposed feature selection methodology provided two relevant feature sets, named the maximum and the balanced performance feature sets, giving a comprehensive option for system development. The experimental results showed a significant reduction of band numbers while maintaining the classification performance intact.