Calibração do NIR a partir de um banco de dados de Urochloa brizantha

Abstract

This study aimed to analyze the particularities of the NIRS calibration processfromadatabase composed of Urochloa brizantha samples, elucidating the best processforobtaining calibration equations that best fit data with wide variation in composition. Atotal of 368 Urochloa brizantha samples were collected, grouped in 4databasesreferring to water, drought, transition and total periods. The samples weresubmittedto the analyzes of DM, MM, CP, NDF, ADF and iNDF and scannedinspectraAnalyzer (Zeutec) spectrometer. The calibrations were performedwiththeaidof software (Worx 2.1.5.0) in automatic mode for all periods and manual selectingfrom 1 to 10 filters for the parameters of CP and NDF of the total period. Descriptiveand statistical analysis of each database was also performed. The Mahalanobisdistance excludes values that do not present a similar pattern to theothersandaspects such as different climatic conditions, cultivars and fertilizationlevel causevariations in the NIR spectra contributing to the increase in the number of outliers.More data can increase dissimilarity between values and increase thenumberofoutliers. The total period database is the most efficient because it has alarger datarange. The increase in the number of filters little changed thedescriptivecharacteristics of the variables CP and NDF. The minimumand maximumvalueshave changed little, but the number of outliers tends to increase as thenumberoffilters increases. Inclusion and removal of samples from the calibrationdatasetdirectly influences the statistical calibration parameters. The CP variablepresentedexcellent determination coefficients in the calibrations of all periods, evidencingthehigh precision of the NIRS method for the prediction of CP. Lower valuesinthedetermination coefficients of MM can be attributed to the reduced predictivecapacityof the NIRS method for the evaluation of this fraction. Fiber is composedof several functional groups that are not chemically well defined reducing thepredictioncapacity by the NIRS methodology. The equations developed using1factorpresented the worst calibration parameters. The best calibration equationfor thevariables CP and NDF is the one developed with 2 filters, being the samechosenbythe calibration software in automatic mode. Grouping all periods increasedthe database's amplitude and hence the efficiency of the calibration model withoutsharply reducing the statistical calibration parameters. Although automatic calibrationpresents good statistical parameters, it is recommended to observeall theparticularities provided by the increase in the number of filters until reachinglowparameters and, then, to observe which equation is best to choose, as often several equations are satisfactory.