Machine learning approaches for early prediction of adult wool growth and quality in Australian Merino sheep

Abstract

Wool production and its quality play important roles in determining the total income received by Australian sheep producers. Enabling accurate and early prediction of wool production and quality traits for individual and groups of sheep can provide useful information assisting on-farm management decision-making. Robustness and high performance of modern prediction methods, namely Machine Learning (ML) algorithms, make them suitable for this purpose. In this research, flock specific environmental data and phenotypic information of yearling lambs were combined to identify the most effective algorithm to predict adult Greasy Fleece Weight (aGFW), adult Clean Fleece Weight (aCFW), adult Fibre Diameter (aFD), adult Staple Length (aSL), and adult Staple Strength (aSS). Algorithms were evaluated and compared in terms of prediction error, the correlation between predicted and actual phenotype in a test set, and for uncertainty in prediction.Artificial Neural Networks (NN), Model Tree (MT) and Bagging (BG) were used to carry out these predictions and their performance was compared with Linear Regression (LR) as the gold standard of prediction. The NN method had the poorest performance in all five traits. MT and BG had very similar performance and for a number of practical reasons, our method of choice was MT for early prediction of adult wool traits. The correlation coefficients of MT predictions were 0.93, 0.90, 0.94, 0.81 and 0.59 with Mean Absolute Error of 0.48 kg, 0.41 kg, 0.92 µm, 6.91 mm and 6.82 N/ktex, for predicting aGFW, aCFW, aFD, aSL, and aSS respectively.