Redes neurais artificiais para predição dos teores de matéria orgânica e argila do solo na região dos Campos Gerais utilizando espectroscopia de reflectância difusa

Abstract

Determining the soil organic matter and clay are important to obtain indicators of soil quality. Such measurements are help the agronomic management providing support for the recommendation of lime and fertilizer. For this quantification, analysis of soil becomes a "tool" indispensable, being increasingly used, especially when associated the the precision farming technology, in which the producer performs a higher number of analyzes aiming to identify soil variability of the property. However, laboratory tests bring some disadvantages, such as the time required for the analyses, and also the generation of waste. An option to perform the analyzes of organic matter and clay soil, quickly and without chemical residues, is by the use of visible to infrared spectroscopy and near (vis-NIRS - Visible and Near Infrared Spectroscopy). The aim of this work was to propose a methodology for predicting the soil organic matter and clay, by combining the use of Regression Analysis and Artificial Neural Networks in order to develop models to estimate these attributes. A database with information about soil analysis obtained by the conventional method and the method vis-NIR was used. The first step was to select the spectral bands that presented a better correlation with the response variables (clay and organic matter) by means of multivariate regression model. In order to improve the estimation of soil organic matter and clay, the group that presented the highest coefficient of determination was used as input of the Artificial Neural Networks. The quantity of 111 soil samples were used for calibration the models of soil analysis, and their spectra were obtained on a spectrophotometer FOSS NIR model XDS. The results were evaluated by the coefficient of determination (R2), considering the significance level of 5%. Coeficients of 0,89 and 0,94 were obtained in the prediction of organic matter and clay respectively, with indices highly significant (P <0,001), indicating the proposed methodology could be useful to predict the attributes studied.