Aprendizado de máquina aplicado à previsão de infestações de pragas através de mudanças meteorológicas

Abstract

The use of pesticides for controlling the population growth of a given pest is the most commonly used technique. One of the reasons for this is the response time that such method has, acting quickly and eliminating the threat to the crops. However, such pesticides are known for their health risk to the consumers of the products, the rural workers, and the pesticide appliers. As infestations are caused by insects, and many of these insects have characteristics that are strongly influenced by weather factors, for example, being ectothermic, which makes them fragile to temperature changes in the region. Based on this, it is possible to state that the behavior of insects is predictable, capable of being determined by the climatic changes in the region. In this context, this work proposes the creation of a model for prediction of an infestation based on the climatic changes in the region. More specifically, this work aims at predicting the insect population in a crop one week in advance. The meteorological data used in predictions were the temperature and rain weekly average. Besides that, the insect population information current in the crop was also used as an input for the model. To create and validate the model, a toy dataset was created from a LSTM network with data from Campinas Biological Institute. For the prediction, a new online learning LSTM network was designed. However, it was split into a training dataset and a test dataset. At the end of experiments, it could be attested that the online model achieved a Mean Squared Error less than the tradicional model.