Classification of Chinese cabbage and radish based on the reflectance of hyperspectral imagery

Abstract

In this research, the ground based hyperspectral reflectance of chinese cabbage and radish depending on the vegetation growth stages was compared to each other. The classifiers namely decision tree, random forest and support vector machine were tested to check the feasibility of classification depending on the difference in hyperspectral reflectance. The ability of classifier was compared with the overall accuracy and kappa coefficient depending on the vegetation growth stages. The spectral merging was applied to find out the optimal spectral bands to make new multispectral sensor based on the commercial band pass filter with full width at half maximum (FWHM) such as 10nm, 25nm, 40nm, 50nm and 80nm. It was ascertained that the pattern of hyperspectral reflectance varied in chinese cabbage and radish and also found a certain disparity of pattern in different vegetation growing stage. Although the classifying ability of support vector machine with linear method was higher than the other six methods, it was not suitable for new multispectral sensor. Hence, the decision tree with Rpart method is advantageous as a best classifier to make new multispectral sensor in order to separate the hyperspectral reflectance of chinese cabbage and radish depending on the vegetation growth stages. The substantiates two alternative aggregate of bands 410nm, 430nm, 700nm and 720nm with 10nm of FWHM or 410nm, 440nm, 690nm and 720nm with 25nm of FWHM were suggested to be the best combinations to make new multispectral sensor without the overlap of FWHM.