A study on nitrogen concentration detection model of rubber leaf based on spatial-spectral information with NIR hyperspectral data

Abstract

Rapid and non-destructive detection of nitrogen concentration in rubber tree is of great significance for estimating rubber yield and precise fertilization of rubber trees. The characteristic spatial distribution of nitrogen is uneven. So the weighted average spectrum of NIR hyperspectral data obtained based on the proportion of different regions contains more spatial information than the average spectrum, which is related to the accuracy of the rubber leaf nitrogen content detection model. In this study, a hyperspectral data clustering method based on spatial-spectral information was proposed to establish the nitrogen model of rubber leaves. The KPCA-GMM method is used to cluster the hyperspectral data of each rubber leaf into C categories, and the area ratio of each category on the blade is obtained based on the clustering result, thereby calculating the weighted average spectrum of each blade. As the number of clustering centers increases, the weighted average spectrum will gradually approach the average spectrum, and the leaf spatial spectrum information will be lost. The experiment obtains different weighted average spectra by setting different numbers of cluster centers, and uses different weighted average spectrum data sets to establish models. The data with the clustering center parameter of 8 set up the best prediction model. The experimental results show that the average spectral data set obtained when the clustering center parameter is 8 is more suitable for estimating the nitrogen content in rubber leaves. The unsupervised and fast performance of KPCA-GMM provides guidance for the area division during the online sample spectrum acquisition process.