Proceedings of the XLVII Italian
Society of Agricultural Genetics - SIGA Annual Congress
Verona,
Italy - 24/27 September, 2003
ISBN 88-900622-4-X
Poster
Abstract - 2.12
NEW METHODS
FOR THE CLASSIFICATION OF POPLAR CLONES BY MEANS OF MORPHOLOGICAL DESCRIPTORS
F.M.STEFANINI**,
A. GIORCELLI***, F.PICCO***, A.CAMUSSI*
*) Dipartimento
di Biotecnologie agrarie and
**) Dipartimento
di Statistica Università di Firenze
***) Istituto di
Sperimentazione per la Pioppicoltura, Casale Monferrato (AL)
poplar clones, classification methods, descriptors,
Random Forest
Tests for
distinctness, homogeneity and stability of Poplar clones are still based on the
use of proper descriptors of the main morphological and phenetic
characteristics of the plant. Although the importance of a precise
identification of clones is widely acknowledged, no sound technique has yet
come into wide use. Many descriptors related to economic and productivity
traits show a reduced repeatability or within clone variability. The poor
discriminant power of single traits can be counterbalanced by a multivariate
approach.
In the frame of a
research project to develop models for the joint analysis of descriptors with
different statistical properties, we have analysed experimental data related to
30 Poplar clones, representative of the germplasm collection. The experimental
design consists in ten replications in two localities (Casale Monferrato and
Mantova). In each replication 18 descriptors (both qualitatitive and
quantitative) were recorded on two plants. The individual discriminant power of
descriptors were estimated by usual statistical techniques (linear model and
logistic model).
A multivariate
approach by means of parametric procedure is ineffective due to the joint
presence of variables with different sampling properties.
We have applied
some new numerical techniques based on computer simulation approaches to
overcome difficulties due to the probability distribution of different traits.
Among others, the procedure known as Random Forest was particularly suitable for clones
discrimination. Random Forest, proposed by Leo Breiman (University of
California, Berkeley,USA), is based on the building of a large set (Forest) of
classification trees, generated at random, which are allowed to evolve
generation by generation on the basis of computer simulations. Some internal
estimates are produced and they are useful to describe the classification
process and the relative importance of single traits. The procedure can include
data of different distributive properties (quantitative traits, categorical
variables and so on).
With our data, we
received a good classification performance with a mean misclassification rate
of 0,13 and with 22 clones with a rate under this value.
The procedure
allowed to individuate the best variables according to classification ability.
Final aim of the work is to individuate simple rules that can be easily applied
in the typical condition of nursery practice.