Hyperspectral waveband selection algorithm based on weighted maximum relevance minimum redundancy and its stability analysis. (26th May 2020)

• Record Type: Journal Article
• Authors: Liu, Yao; Li, Ming; Wang, Shuwen; Wang, Runtao; Jiang, Wei
• Is Part Of: Measurement science & technology. Volume 31:Number 8(2020)
• Publication Date: 2020-05-26
• Subjects: hyperspectral imaging -- waveband selection -- maximum relevance -- minimum redundancy -- rough set; Physical measurements -- Periodicals; Scientific apparatus and instruments -- Periodicals; Equipment and Supplies -- Periodicals; Science -- instrumentation -- Periodicals; Technology -- instrumentation -- Periodicals; Mesures physiques -- Périodiques; Physical measurements; Scientific apparatus and instruments; Periodicals; 502.87
• DOI: 10.1088/1361-6501/ab816d
• Languages: English
• ISSNs: 0957-0233

Abstract

Abstract: Owing to the highly dimensional nature of hyperspectral imaging datasets, waveband selection has become an important step in processing. In this work, we propose a novel weighted maximum relevance minimum redundancy waveband selection algorithm. The relative importance between redundancy and relevance is better balanced by introducing a weight coefficient. The mutual information between wavebands and target classes and wavebands based on the neighbourhood rough set theory were calculated using the proposed algorithm. Using the forward greedy search algorithm, the wavebands with maximum relevance to target classes and minimum redundancy to previously selected wavebands were selected. In the classification of soybean hyperspectral imaging datasets, weighted maximum relevance minimum redundancy algorithms with an equal weight and an unequal weight both performed well in terms of classification accuracy. The classification performances of the extreme learning machine classifiers are satisfactory. The average classification accuracy mostly exceeds 95%, when the neighbourhood size is greater than 0.12. In addition to classification accuracy, the stability of the algorithm under small perturbations was studied. The stability decreases as the perturbation level increases. The proposed algorithm is more stable with an unequal weight than with an equal weight. By applying the proposed algorithm, the weight coefficient can be selected flexibly to achieve optimum performance (the best stability, best classification performance, and smallest size of subsets), instead of limiting the algorithm to a fixed equal weight for relevance and redundancy.