Modelagem do CO2 evoluído de argissolo tratado com dejetos de suíno por modelos de regressão não lineares: prioris de máxima entropia

Abstract

Pig breeding is an important agribusiness activity in Brazil. In recent years pig production has increased and, consequently, the amount of liquid waste. Swine manure has a high polluting potential and an appropriate way to dispose of manure is on agricultural soils, since it is rich in organic matter and nutrients. Organic matter can accelerate the activity of microorganisms responsible for decomposition and, as a consequence, release of minerals contained in waste. The amount of CO2 released at the beginning of decomposition is greater, since substances that are easily degraded are mineralized, over time the amount of CO2 released decreases, due to the decomposition of only the most resistant substances. The decomposition dynamics can be described by nonlinear regression models, however, the theory for regression models is asymptotically valid and, in general, in research, small samples are used. An alternative is to use the bayesian methodology which has been shown to be efficient even with small samples. But criticism has been made of the bayesian approach for the effect that a subjective prior distribution can have on the posterior distribution. One way to determine priors objectives is through maximum entropy prior distributions. Thus, the work aims to use maximum entropy priors to the parameters of the nonlinear models Stanford & Smith and Cabrera in the description of the mineralization data of CO2 of swine manure applied on the soil surface. In addition, using simulated data, to understand the effect that the hyperparameters of the distribution a prior have on the curve a posterior of the Stanford & Smith and Cabrera models. Both nonlinear models using the bayesian methodology with maximum entropy prior were efficient in studying the data of carbon mineralization of swine manure on the soil surface, in addition, the bayesian method proved to be a viable alternative to circumvent the problem of sample size. It was shown how the values of the hyperparameters of the a prior distributions influence the posterior curve of the Stanford & Smith and Cabrera models. The nonlinear Cabrera model promoted greater information gain based on the Kullback-Leibler measure.