Plant Phenomics
5
Cross-Validation 2 (CV2)
Cross-Validation 1 (CV1)
Me
th
o
d
1
Me
th
o
d
2
Training Data
Testing Data
Data not used
Environment
Environment
Environment
Environment
Environment
Environment
1 2 3 4 5 6
1
2
3
-
-
292
Ge
n
o
ty
pe
1 2 3 4 5 6
1
2
3
-
-
292
Ge
n
o
ty
pe
1 2 3 4 5 6
1
2
3
-
-
292
Ge
n
o
ty
pe
1 2 3 4 5 6
1
2
3
-
-
292
Ge
n
o
ty
pe
1 2 3 4 5 6
1
2
3
-
-
292
Ge
n
o
ty
pe
1 2 3 4 5 6
1
2
3
-
-
292
Ge
n
o
ty
pe
BLUPs
BLUPs
Figure 1: Cross-validation scenarios (CV1 and CV2) and preprocessing methods (Methods 1 and 2) used to assess phenomic prediction model
performance. Method 1 and Method 2 differ in BLUP computation, while CV1 and 2 depict two plant breeding scenarios for prediction
in multienvironment tests. These CV scenarios represent a combination of different preprocessing (to handle missing data) methods and
prediction challenges native to plant breeding practices. In Method 1, for both CV scenarios, individual environment BLUPs are computed
and subsequently used in model training and testing the model. In Method 2, combined environment BLUPs are computed and subsequently
used in training the model, while individual environment BLUPs are used in testing the model.
used for model training, while 20% of accession for that
environment was used for testing; i.e., for Environment#2,
model training was done on 80% of random accessions from
Environments# 1, 3, and 4, and testing was done on 20% of
remaining accession from Environment#2. For CV1 and CV2,
the training and testing procedures were repeated 10 times
and the mean accuracy for each CV-Method combination is
reported. Training and testing sets were compiled for each CV
iteration and training data used to parameterize model and
prediction made onto the test set following model training.
Two preprocessing methods were used to parameterize
RF prediction models (see Statistical Model section), and we
then tested two CV scenarios to emulate prediction chal-
lenges faced by breeders in field trials with unbalanced data.
From a practical application viewpoint, the CV1 strategy is a
scenario where phenomic data is collected on all genotypes
while yield is collected on a subset of lines and breeders may
wish to estimate the rank performance of untested genotypes
not phenotyped for yield but with available physiological
trait data. The CV2 strategy is deployable where breeders
are interested in predicting rank performance of untested
accessions (no seed yield data) and untested environments
(unseen environment) with no seed yield but with phenomic
traits. The CV2 strategy is an improvement to leave-one-
environment-out [47] situation as we excluded test genotypes
from model training.
Model prediction accuracy is reported using Spearman
rank correlation coefficient between observed values and
predicted values of the test set computed by recording the
mean values across all 10 training-testing iterations and all
folds of CV. Cross-validation schemes were developed in R
using in-house script.
Do'stlaringiz bilan baham: