In time series of insect populations, it is generally common to observe peaks and dips in population size, which can often become recurrent. This is sometimes character-ized by a wide spectrum, resulting in abrupt population outbreaks for several reasons. For these cases, it is particularly important to carrying out analyses to project the risk of occurrence of these population levels, in order to apply measures to prevent undesirable events. The periods prior to outbreaks can be characterized as "Alert Zones" which may contain recurrent ecological patterns useful for predictions. The detection of zones can result in population diagnostics important for decision making in highly-oscillated systems that commonly occur in insects, especially pests. A few years ago, a new way for carrying out pattern analysis, the Alert Zone Procedure (AZP), was proposed. The essence of this idea is absolutely compatible with insect outbreak forecasting, but to do that new improvements have to be completed allowing for a bet-ter predictive power. Currently, the tasks of forecasting and classification have been improved by the new Machine Learning (ML) algorithms that are applied in many do-mains of science based on the purpose of understanding how algorithms can "learn" these tasks. For the outbreak context, in general, many methods have been developed to enhance predictions such as Recurrent Neural Networks, Random Forests, Sup-port Vector Machines and others. The results in terms of prediction of these methods seem to be very promising to aid in predicting pest outbreaks. Therefore, the context of the current project is based on pest management, Machine Learning and statis-tics. This interdisciplinary project has two main goals: (1) to improve the AZP and (2)to compare the pest outbreak predictive power of the improved AZP and different ML methods.
News published in Agência FAPESP Newsletter about the scholarship: