Study of the chemical composition of Urochloa brizantha using the SPAD index, neural networks, multiple linear models, principal components and cluster analysis

The objectives of this study were to explore the relationship between plant variables using correlation and principal component analysis; to explore the chemical composition patterns in a subgroup of plants using cluster analysis; and to compare the prediction ability between a linear model using only the SPAD index as a predictor with multiple linear regression and neural networks with the SPAD index, morphological and climatic measurements as predictors of the chemical composition of Urochloa brizantha leaves and stems. The experimental design was in blocks (three blocks) with five treatments, totaling 15 experimental units. Chlorophyll measurements and forage sampling were performed every 28 d. The variables used in the statistical analysis were: percentage of leaves, stems and dead plant material; plant height; relative chlorophyll (SPAD); percentage of acid detergent fibre in leaves and stems (ADF.L, ADF.S); percentage of neutral detergent fibre in leaves and stems (NDF.L, NDF.S); lignin percentage in leaves and stems (LIG.L, LIG.S); and nitrogen content in leaves and stems (N.L, N.S). The climatic variables were monthly average minimum and maximum temperatures and monthly rainfall. The correlation between SPAD with N.L and N.S was 0.56 and 0.49, respectively, and between N.L with N.S was 0.87. The correlation between the observed and predicted responses using simple linear regression, with SPAD as the predictor, ranged from 0.198 for ADF.L to 0.577 for N.L. However, the correlations ranged from 0.497 for LIG.L to 0.759 for N.S when multiple regression was used with other predictors, besides SPAD. The prediction accuracy using neural networks ranged from 0.501 for LIG.L to 0.863 for N.S and was higher than multiple regression for all characteristics except LIG.L and LIG.S. Principal component analysis efficiently condensed the most important information of the 13 original variables measured in the plants into three principal components due to the redundancy of information of the variables. According to the cluster analysis, plants with higher nitrogen content in their leaves and stems presented lower fibre contents and dead plant material, and were denser than those with lower nitrogen content. Multiple linear regression can be used to predict lignin content in leaves and stems and neural networks must be used to predict nitrogen and the other fibre contents. SPAD is an important predictor of nitrogen content in tropical pastures, but it is not the only predictor that must be used in regression models and neural networks; morphological characteristics and climatic conditions increase the prediction accuracy of the models.