The two continent-wise ANN models, each trained ondata from three field sites, correctly predicted diseaseseverity class at field sites in the other continent on >77%of days; the overall prediction error was 21·9% for theAustralian and 22·1% for the South American model(Table 1). The prediction of the Australian ANN modelwas most accurate for the Carimagua site and least accuratefor Planaltina. The South American model predictedseverity at Southedge and Springmount with an accuracy>93%, but the prediction accuracy for the Samford sitewas <63%.In the final series of ANN models, data from all Australianand South American sites were pooled and modelswere trained on data from five sites to predict the severityclass for the remaining sixth site. Of the six cross-continentANN models developed in this way, the model developedwithout the data from Planaltina was the most accurate,successfully predicting severity on >85% of days. TheANN model without the Carimagua data correctly predictedseverity on only 54% of days at this site. The otherfour models were accurate on >73% of days (Table 2).The prediction errors of these models were considerablyhigher than for the continent-wise ANN models (Table 1)for all sites except Samford and Planaltina. The percentageof type I, II or total error was not generally influencedby the number of observations (days) in the training ortest data sets in either continent-wise or cross-continentANN models. The overall prediction error was between14 and 45% for the cross-continent models developedwith 110–139 observations (days) in the training data sets;this was between 4 and 37% for continent-wise modelswith either 73 or 77 observations in the training data set.A sensitivity analysis of the continent-wise modelsshowed that RAIN, LWP and RAD for the day of diseaseseverity assessment and RAIN on the previous day are thefour most important weather attributes in the AustralianANN model. RAIN and RAD on the day of severityassessment and RAIN and LWP on the previous day werealso the four most significant input attributes in the SouthAmerican ANN model, although RAIN and RAD weremore important and LWP was less important in the SouthAmerican than in the Australian model (Table 3). Similarly,RAIN, RAD and LWP on the day of severity assessmentand/or the previous day were the most importantweather attributes in each of the six cross-continent models,and RAIN during the day and/or the previous day wasthe single most important of all attributes in all modelsexcept the one trained on data that excluded Planaltina,where another moisture-related variable, LWP on the previousday, was more important than RAIN (Table 3).Prediction and sensitivity of REG modelsOverall, the REG models developed using data from allAustralian or South American sites were not as effective asANN models in predicting disease severity for sites in theother continent, and the prediction errors of the REGmodels were higher for two out of three sites in each continent(Table 1). The overall prediction error of 31·5% forthe Australian REG model was higher than the ANNmodel, but the 20·8% error for the South American modelwas lower than the ANN model. When data for five Australianand South American sites were pooled to predict
การแปล กรุณารอสักครู่..
