Italian Journal of Agronomy (Jul 2012)

Combined approach based on principal component analysis and canonical discriminant analysis for investigating hyperspectral plant response

  • Anna Maria Stellacci,
  • Annamaria Castrignanò,
  • Mariangela Diacono,
  • Antonio Troccoli,
  • Adelaide Ciccarese,
  • Elena Armenise,
  • Antonio Gallo,
  • Pasquale De Vita,
  • Antonio Lonigro,
  • Mario Alberto Mastro,
  • Pietro Rubino

DOI
https://doi.org/10.4081/ija.2012.e34
Journal volume & issue
Vol. 7, no. 3

Abstract

Read online

Hyperspectral (HS) data represents an extremely powerful means for rapidly detecting crop stress and then aiding in the rational management of natural resources in agriculture. However, large volume of data poses a challenge for data processing and extracting crucial information. Multivariate statistical techniques can play a key role in the analysis of HS data, as they may allow to both eliminate redundant information and identify synthetic indices which maximize differences among levels of stress. In this paper we propose an integrated approach, based on the combined use of Principal Component Analysis (PCA) and Canonical Discriminant Analysis (CDA), to investigate HS plant response and discriminate plant status. The approach was preliminary evaluated on a data set collected on durum wheat plants grown under different nitrogen (N) stress levels. Hyperspectral measurements were performed at anthesis through a high resolution field spectroradiometer, ASD FieldSpec HandHeld, covering the 325-1075 nm region. Reflectance data were first restricted to the interval 510-1000 nm and then divided into five bands of the electromagnetic spectrum [green: 510-580 nm; yellow: 581-630 nm; red: 631-690 nm; red-edge: 705-770 nm; near-infrared (NIR): 771-1000 nm]. PCA was applied to each spectral interval. CDA was performed on the extracted components to identify the factors maximizing the differences among plants fertilised with increasing N rates. Within the intervals of green, yellow and red only the first principal component (PC) had an eigenvalue greater than 1 and explained more than 95% of total variance; within the ranges of red-edge and NIR, the first two PCs had an eigenvalue higher than 1. Two canonical variables explained cumulatively more than 81% of total variance and the first was able to discriminate wheat plants differently fertilised, as confirmed also by the significant correlation with aboveground biomass and grain yield parameters. The combined approach proved to be effective, being able to synthesise the redundant radiometric information in a reduced number of indicators of plant nutritional status, which could be utilized to delineate homogeneous within-field areas to be submitted to site-specific fertilization.

Keywords