Developing Robust Faecal near Infrared Spectroscopy Calibrations to Predict Diet Dry Matter Digestibility in Cattle Consuming Tropical Forages

Abstract

Various studies, mainly from temperate areas, have reported calibrations developed from the near infrared (NIR) spectra of faeces (F.NIRS) for predicting diet digestibility in ruminants and there has been substantial variation in predictive accuracy as indicated by calibration and validation statistics. The present study was conducted to develop and examine the reliability and robustness of F.NIRS calibration equations to estimate dry matter digestibility (DMD) of forage diets ingested by cattle grazing in the rangelands of northern Australia. A large and diverse calibration data set of matched diet–faecal pairs was obtained over 10 years using three sampling methods: (1) grazed pasture with diet samples collected from oesophageal fistulated steers and faeces collected from resident cattle; (2) in vivo digestibility experiments with penned cattle fed forage hays; and (3) penned cattle fed pasture freshly harvested from the field. Estimated in vivo DMD reference values were determined using pepsin–cellulase in vitro analysis of diet samples. The final calibration set of 1052 samples represented 264 diets with DMD ranging from 38% to 75%. Calibration statistics for DMD% were: standard error of calibration = 1.87, standard error of cross validation = 1.91, and the coefficient of determination, r2 = 0.90. Factors of particular importance, with regard to the accuracy of DMD reference values, are identified and discussed and recommendations made for minimising reference errors. A comprehensive series of independent validation tests was conducted by selecting validation sample sets from the entire sample set according to a range of criteria. Each validation sub-set was tested using the calibration calculated from the remainder of the sample set. These tests showed that sampling method and experimental site often had important effects on calibration statistics and performance and also that the standard error of performance of the overall calibration would likely be <2.5 DMD percentage units when applied to samples sourced from regions and pasture types represented in the calibration. Despite the large size and diversity of the calibration data set it was concluded that robustness would likely be improved by expansion of the calibration data set.