Spectral characteristics and species identification of rhododendrons using a discriminative restricted Boltzmann machine

• Published in: Spectroscopy letters 2017-02, Vol.50 (2), p.65-72
• Main Authors: Xue-juan, Chen, Xiang, Wu, Zhong-qiang, Yuan, Xiang, Chen, Yu-wu, Zhang, Chun-xiang, Cao
• Format: Article
• Language: English
• Subjects: Data processing; Derivatives; Discriminative restricted Boltzmann machine; Feature extraction; Feature recognition; Flowering; Germplasm; Image classification; Inversions; Karst; Machine learning; Mathematical models; Mountains; Parameters; Remote monitoring; Remote sensing; rhododendron; Species classification; Spectra; spectral recognition; Spectrometry; Spectroscopy; vegetation spectrum
• ISSN: 0038-7010; 1532-2289
• DOI: 10.1080/00387010.2017.1278709

Rhododendrons are an important genus of alpine flowering plant used ornamentally worldwide. The purpose of this study is to improve the application of remote-sensing technology for investigating and monitoring mountain rhododendron germplasm. Research area is the Baili Rhododendron National Forest Park located in the karst region of Guizhou Province, China. Field spectrometry was used to acquire spectral data for 20 samples extracted from eight rhododendron species. A deep-learning algorithm from a discriminative restricted Boltzmann machine was used with the original spectral data from the different rhododendron species to obtain the optimal parameters for the model. Simultaneously, the data processing methodology from the discriminative restricted Boltzmann machine was used to recognize the original spectra, the noise smoothed spectra, and the first- and second-order spectral derivatives with accuracies of 88.54%, 88.54%, 93.75%, and 90.62%, respectively. The results show that the discriminative restricted Boltzmann machine is effective in recognizing spectral information for different rhododendron species. Changes in the first-order derivative gave the most accurate classification, but changes in the second-order derivative significantly reduced the sample training time. Changes in both derivatives therefore proved useful in recognizing and extracting particular features of the plant species. This research may therefore further support the use of hyperspectral remote-sensing imagery for investigating and monitoring germplasm, species classification, and physiological parameter inversions for rhododendrons from various mountain regions of China.