Optimal Spectral Wavelengths for Discriminating Orchard Species Using Multivariate Statistical Techniques

Abstract

Sustainable management of orchard fields requires detailed information about the tree types, which is a main component of precision agriculture programs. To this end, hyperspectral imagery can play a major role in orchard tree species mapping. Efficient use of hyperspectral data in combination with field measurements requires the development of optimized band selection strategies to separate tree species. In this study, field spectroscopy (350 to 2500 nm) was performed through scanning 165 spectral leaf samples of dominant orchard tree species (almond, walnut, and grape) in Chaharmahal va Bakhtiyari province, Iran. Two multivariable methods were employed to identify the optimum wavelengths: the first includes three-step approach ANOVA, random forest classifier (RFC) and principal component analysis (PCA), and the second employs partial least squares (PLS). For both methods we determined whether tree species can be spectrally separated using discriminant analysis (DA) and then the optimal wavelengths were identified for this purpose. Results indicate that all species express distinct spectral behaviors at the beginning of the visible range (from 350 to 439 nm), the red edge and the near infrared wavelengths (from 701 to 1405 nm). The ANOVA test was able to reduce primary wavelengths (2151) to 792, which had a significant difference (99% confidence level), then the RFC further reduced the wavelengths to 118. By removing the overlapping wavelengths, the PCA represented five components (99.87% of variance) which extracted optimal wavelengths were: 363, 423, 721, 1064, and 1388 nm. The optimal wavelengths for the species discrimination using the best PLS-DA model (100% accuracy) were at 397, 515, 647, 1386, and 1919 nm.