Feature-Ensemble-Based Novelty Detection for Analyzing Plant Hyperspectral Datasets

Recently, there has been a significant increase in the use of proximal or remote hyperspectral imaging systems to study plant properties, types, and conditions. Numerous financial and environmental benefits of using such systems have been the driving force behind this growth. This paper is concerned with the analysis of hyperspectral data for detecting plant diseases and stress conditions and classifying crop types by means of advanced machine learning techniques. Main contribution of the work lies in the use of an innovative classification framework for the analysis, in which adaptive feature selection, novelty detection, and ensemble learning are integrated. Three hyperspectral datasets and a nonimaging hyperspectral dataset were used in the evaluation of the proposed framework. Experimental results show significant improvements achieved by the proposed method compared to the use of empirical spectral indices and existing classification methods.