To optimise grazing livestock nutrition, it is necessary to know both the available dry matter yield and the nutritive characteristics of pasture at the farm-scale in near real time. Previous studies have shown the potential of using field spectrophotometers that measure the reflectance of light across the visible to near infrared spectrums to gather information on pasture dry matter yield (DMY) and nutritive characteristics. This study sought to calibrate and validate new mathematical models for ten parameters including dry matter yield and nine nutritive characteristics of relevance to ruminant nutrition. As a part of the analysis process, two innovative approaches were tested: the use of a hybrid modelling approach where partial least squares regression (PLSR) outputs were used as support vector regression (SVR) inputs; and, the inclusion of covariate data. These approaches were compared with traditional stand-alone PLSR and SVR modelling approaches without covariates. The study was undertaken in six predominantly perennial ryegrass pastures on a single farm in the temperate zone of South-Eastern Australia. A total of 204 pasture samples were scanned with a field spectrophotometer over several spring growth stages in late 2019 and subsequently analysed by wet chemistry to obtain reference nutritive values. The raw reflectance spectra were initially pre-processed using a variety of techniques and then used to test the four kinds of chemometric models. In cross validation, hybrid models showed a superior fit for all variates in comparison to the other model types tested. However, the differential was reduced in independent validation where, out of 10 best-performing models for dry matter yield and key nutrient properties, six were produced by the hybrid modelling, three from SVR and one from PLSR. For every hybrid model that was built, adding covariate(s) consistently improved model performance but the increase was small (a reduction in normalised root mean square error (RMSE) of -0.36 % on average for all properties considered). The best performing models were comparable with other published literature with normalised RMSE of prediction ranging from 1.7 – 23.1 % (a mean of 9.7%). Well-predicted variates included metabolisable energy, digestible energy, DMY, and crude protein. Fibre fractions, ash and dry matter were less well-predicted but still had acceptable normalised RMSE values (< 10 %) while carbohydrate fractions were the poorest predicted variates. It was concluded that hybrid modelling in chemometric analyses can modestly improve accuracy and shows promise as an alternative to more traditional approaches. Using covariates also improved accuracy, but the additional time and effort to gather such information outweighed the minor benefits of inclusion.
Read full abstract