AGROBA Y E S - Pacote R para agricultura de precisão com redes Bayesianas

Abstract

Many historians consider agriculture as the technology that shaped human development, allowing population concentration and the emergence of modern societies. Food security depends heavily on agriculture meeting the challenges posed by population growth and climate change. Information technology is an important ally for overcoming these difficulties in a sustainable way, increasing productivity with better use of available resources. Proper resource management depends on the availability of necessary data for producers and public agents. In this context, techniques for forecasting and estimating agricultural productivity are relevant tools for decision-making. However, in addition to generating productivity preview values, a model must be able to describe the main factors that affect this productivity and verify the space-time distribution of these factors. Probabilistic inference models are techniques that can provide the best results, as they are explainable models and can incorporate new data as they arise. This work presents a study of the main techniques applied to estimating and predicting agricultural production results and develops an R package for production data inference with spatially distributed static Bayesian networks and dynamic Bayesian networks, capable of predicting agricultural production results. In tests performed with data generated in a controlled manner, the models created by inference algorithms did not always reflect the expected structure. The networks created from sets of pre-defined arcs presented a superior performance in relation to the networks inferred by data, even so with less than desirable performance for a model intended for use in actual cases. The developed work introduces a new method to deal with a problem that is both important and difficult, which shows that there is still a need for improvements in the models used.