The two continent-wise ANN models, each trained on
data from three field sites, correctly predicted disease
severity class at field sites in the other continent on >77%
of days; the overall prediction error was 21·9% for the
Australian and 22·1% for the South American model
(Table 1). The prediction of the Australian ANN model
was most accurate for the Carimagua site and least accurate
for Planaltina. The South American model predicted
severity at Southedge and Springmount with an accuracy
>93%, but the prediction accuracy for the Samford site
was 85% of days. The
ANN model without the Carimagua data correctly predicted
severity on only 54% of days at this site. The other
four models were accurate on >73% of days (Table 2).
The prediction errors of these models were considerably
higher than for the continent-wise ANN models (Table 1)
for all sites except Samford and Planaltina. The percentage
of type I, II or total error was not generally influenced
by the number of observations (days) in the training or
test data sets in either continent-wise or cross-continent
ANN models. The overall prediction error was between
14 and 45% for the cross-continent models developed
with 110–139 observations (days) in the training data sets;
this was between 4 and 37% for continent-wise models
with either 73 or 77 observations in the training data set.
A sensitivity analysis of the continent-wise models
showed that RAIN, LWP and RAD for the day of disease
severity assessment and RAIN on the previous day are the
four most important weather attributes in the Australian
ANN model. RAIN and RAD on the day of severity
assessment and RAIN and LWP on the previous day were
also the four most significant input attributes in the South
American ANN model, although RAIN and RAD were
more important and LWP was less important in the South
American than in the Australian model (Table 3). Similarly,
RAIN, RAD and LWP on the day of severity assessment
and/or the previous day were the most important
weather attributes in each of the six cross-continent models,
and RAIN during the day and/or the previous day was
the single most important of all attributes in all models
except the one trained on data that excluded Planaltina,
where another moisture-related variable, LWP on the previous
day, was more important than RAIN (Table 3).
Prediction and sensitivity of REG models
Overall, the REG models developed using data from all
Australian or South American sites were not as effective as
ANN models in predicting disease severity for sites in the
other continent, and the prediction errors of the REG
models were higher for two out of three sites in each continent
(Table 1). The overall prediction error of 31·5% for
the Australian REG model was higher than the ANN
model, but the 20·8% error for the South American model
was lower than the ANN model. When data for five Australian
and South American sites were pooled to predict