Machine learning regression algorithms for predicting muscle, bone, carcass fat and commercial cuts in hairless lambs

The growth in demand and demand for quality in the sheep chain has generated the need for automation techniques in the meat industry and the need to obtain responses with greater speed and standardization. The research aimed to predict tissue characteristics of the carcass and commercial cuts based on measurements obtained by VIA – see oimage analysis, carried out on cold carcasses of hairless lambs, using machine learning employing regressive techniques for variable selection. Information from 72 carcasses of castrated male lambs, aged between 8 and 11 months, with an average cold carcass weight of 16.13 ± 3.98 kg, was used. Images of the right side of the carcasses were captured from the dorsal and lateral views using a digital camera. From the ImageJ2 software, VIA data, measurements and shape descriptors (areas, perimeters, widths, lengths, convexities, solidities) were obtained, combined with cold carcass weight and used to generate four sets of data, called descriptor sets (DSs). Obtaining DS1, DS1’, DS2, DS2’, DS3, DS3’, DS4 AND DS4’. To generate these sets, a database was formed and divided into a training bank (with 70% of the observations) and a test bank (30% of the observations). Multiple linear regression models were developed using Stepwise, LASSO, and Elastic Net regression methods, combined with k-fold cross-validation, to evaluate the performance of the models. The accuracy of the estimates was based on RMSE, R², Pearson correlation and bias metrics. For the variables tested in this study, the proposed shape descriptors were mostly efficient in predicting tissue and weight variables. DS1' with the LASSO technique presented the best adjustments for variables total muscle and fat followed by shoulder, loin and rib cuts. The descriptors tested by this study were able to predict with quality the vast majority of the characteristics tested, the variable cold carcass weight (CCW), introduced as additional predictor, promoted a consistent improvement in the fits of all models. DS1 presented greater constancy for the twenty-three predicted characteristics and Stepwise presented the worst predictive performance, in relation to LASSO and Elastic Net. Despite close adjustments between the generated models, in general, Elastic Net presented lower performance than LASSO.