Abstract:
An RT model (RT1) was constructed using 35 agro-morphological characters for 45 mustard
(Brassica juncea) accessions. Based on the ‘variable importance’ of the model RT1, another
model (RT2) was developed. These models were developed using classification and
regression tree algorithms. The classificatory performance of the RT1 model was compared
with RT2 model. RT1 and RT2 models classified the mustard accessions with
misclassification rates of 2.3% (98% accuracy) and 4.3% (96% accuracy), respectively. The
variable importance of RT1 and RT2 explained that leaf length (LLCM), hypocotyl length
(HLCM), hypocotyl-anthocynin coloration (ACH) and leaf width (LWCM) at seedling stage
and main inflorescence length (LMICM), silique length (SLMM) and seed yield/plant
(SYDIVPG) at maturity stage play an important role in classifying mustard accessions.
Comparison of RT1 with RT2 revealed that accuracy of classification made by RT1 is higher
in predicting class memberships among mustard accessions. A large degree of variability
within and between Sri Lankan mustard accessions has been observed for agromorphological characters with respect to LLCM, HLCM, ACH, LWCM, LMICM, SLMM and
SYDIVPG. The genetic diversity of certain mustard accessions such as Accession Numbers
346, 8658 and 9726 is too high and RT models failed to classify them correctly with
acceptable accuracy.
