Faecal predictions by near infrared spectroscopy for Ibex : overcoming the problem of sample quantities

The difficulty of direct observation and the protected status of some mountain wild and domestic ungulate species (Capra i. ibex (CI), Rupicapra rupicapra (RR) , Ovies aries (OA)) confer to faecal analyses a great advantage in the study of foraging strategies, as chemical composition of faeces can provide information on quality and quantity of animal feed intake. However the weight of a sample is sometimes too low to allow reference chemical analyses. In the framework of a study on ungulate nutrition in the French Alps, a total 233 faeces samples was collected; most of them from CI (191 samples) and some of RR (10) and OA (31). Only 24 samples reached the weight required for chemical analysis, which did not allow to build a NIRS calibration. So it was decided to apply the following procedure: 1/ take the spectra of all samples, 2/ mix some of the samples in order to get sufficient quantity for analysis, 3/ build a calibration by gathering data on these mixes + the 24 individual samples, 4/ apply this calibration to all samples. All 233 samples were dried, ground and spectra in diffuse reflectance were measured on a FOSS NIRSystem 6500 spectrometer in ring cups. A hierarchical classification (non supervised) of samples was performed on the basis of spectral data. Faeces within groups were mixed. This procedure ensured to maintain a high variability in the calibration database. Then chemical analyses were performed on 54 samples: the parameters measured were Mineral Matter (MM), Crude Protein (CP), Crude Fibre (CF), Van Soest fibre fractions (NDF, ADF, ADL). A first important feature was to test if the mixed samples represented well the total population. A Principal Component Analysis was performed on the calibration database, and the spectra of all (233) individual samples were projected on the principal component axes. Only 12% of samples from CI had Mahalanobis (H) distance greater than 3. This confirmed that the selection of analysed samples was representative. Then calibrations equations were built for chemical components with PLS (partial least square) regression. SECV obtained were 0.84%, 0.48%, 1.27%, 1.86%, 1.84%, 1.31% respectively for MM, CP, CF, NDF, ADF and ADL. These results allowed the prediction of the chemical composition of the whole database. The conclusion of this study is that the strategy of mixing samples with a procedure ensuring the minimal loss of variability can help to overcome problems of insufficient quantity of matter for chemical analysis. In the present study, the use of this strategy allowed to have a complete chemical profile on all samples collected. An additional benefit was the lower analytical cost compared to a reference analysis of all samples. [Résumé d'auteur]