A simple Bayesian network to interpret the accuracy of armyworm outbreak forecasts

Forecasting of outbreaks of armyworm (larvae of the moth Spodoptera exempta) employs information from rain gauges and moth traps. Rainfall is an independent variable, but moth catch is affected by rainfall, and outbreak risk is affected by both moth catch and rainfall. A simple Bayesian network was used to describe these relationships and so derive conditional probabilities. The data were from a new initiative, community-based forecasting of armyworm in Tanzania, in which outbreak risk for a village is determined locally from a single moth trap and rain gauge located within the village. It was found that, following a positive forecast, an armyworm outbreak was approximately twice as likely to occur as would be expected by chance. If the forecast was negative because of insufficient moths, outbreaks were half as likely as would be expected by chance. If the forecast was negative because of insufficient rain, however, the outbreak probability remained similar to chance: an aspect of the forecast that requires improvement. Overall, a high forecasting accuracy can be achieved by village communities using simple rules to predict armyworm outbreaks. © 2006 Association of Applied Biologists.