Implementação e análise comparativa entre modelos de regressão para estimação de nutrientes em camas de aviário a partir de espectroscopia NIR

Abstract

The growth of the Brazilian poultry industry in relation to the industrial market of the sector is indeed an important factor for the country’s economic growth, considering Brazil as one of the largest producers and exporters of chicken meat in the world. However, the increasing of residues from this practice can be a source of damage for the environment, due to the wrong discard of these substances on the ground, which may increase the concentration of certain chemical elements, wherein at a certain level can cause damage to the exposed environment. Thereby, an alternative for better use of the recurrent residue of poultry litter, is the use of this one as an organic fertilizer, after its correct chemical characterization. However, the classical methods of chemical concentration analysis isn’t attractive to the producers, once that this method has an elevated cost and its long time of analysis. Thus, statistical and mathematical methods start to gain space among the chemometrics studies, with the objective of training regression models that recognize patterns of chemical concentration levels in samples from the analysis of NIR spectroscopy, once that the designed model becomes useful for several new measures, avoiding future expenses. The standard technique used for this kind of training is known as Partial Least Squares, which seeks to find the solution of the linear relationship between the variables, using linear techniques for the data regression. Based on the assumption that data analysis in the real world is better related to nonlinear propagation characteristics, this work proposes the application of the technique of Support Vector Machines, which uses a nonlinear approach as a way to solve the linear equation that correlates the variables. As intended, predictions models of N, C, K and P concentration in 160 samples of poultry litter were created. From which the comparison between the two techniques demonstrate that the Support Vector Machines is more efficient than the Partial Least Squares method, however, presents in its composition biggers challenges in the training stage. These work also presents the construction of a methodology of data preprocessing, allowing to create a sequence of standard operations for both of the training applications, which allows both to build an optimized model to achieve the best performance of the system.