Ultra-wideband microwave precisely and accurately predicts sheepmeat hot carcase GR tissue depth.

A portable ultra-wideband microwave system (MiS) coupled with an antipodal slot Vivaldi patch antenna (VPA) was used as an objective measurement technology to predict sheep meat carcase GR tissue depth, tested against AUS-MEAT national accreditation standards. Experiment one developed the MiS GR tissue depth prediction equation using lamb carcasses (n = 832) from two slaughter groups. To create the prediction equations, a two layered machine learning stacking ensemble technique was used. The performance of this equation was tested within the dataset using a k-fold cross validation (k = 5), which demonstrated excellent precision and accuracy with an average R2 of 0.91, RMSEP 2.11, bias 0.39 and slope 0.03. Experiment two tested the prediction equation against the AUS-MEAT GR tissue depth accreditation framework which stipulates predictions from a device must assign the correct fat score, with a tolerance of ±2 mm of the score boundary, and 90% accuracy. For a device to be accredited three measurements captured within the same device, as well as measurements across three different devices, must meet the AUS-MEAT error thresholds. Three MiS devices scanned lamb carcases (n = 312) across three slaughter days. All three MiS devices met the AUS-MEAT accreditation thresholds, accurately predicting GR tissue depth 96.1-98.4% of the time. Between the different devices, the measurement accuracy was 99.4-100%, and within the same device, the measurement accuracy was 99.7-100%. Based on these results MiS achieved AUS-MEAT device accreditation as an objective technology to predict GR tissue depth.

Predicting lamb carcase composition from tissue depth measured at a single point with an ultrawide-band microwave scanner.

This study evaluated the ability of portable ultra-wide band microwave system (MiS) to predict lamb carcase computed tomography (CT) determined composition % of fat, lean muscle and bone. Lamb carcases (n = 343) from 6 slaughter groups were MiS scanned at the C-site (45 mm from spine midline at the 12th /13th rib) prior to CT scanning to determine the proportion of fat, muscle and bone. A machine learning ensemble stacking technique was used to construct the MiS prediction equations. Predictions were pooled and divided in 5 groups stratified for each CT composition trait (fat, lean or bone%) and a k-fold cross validation (k = 5) technique was used to test the predictions. MiS predicted CT fat% with an average RMSEP of 2.385, R2 0.78, bias 0.156 and slope 0.095. The prediction of CT lean% had an average RMSEP of 2.146, R2 0.64, bias 0.172 and slope 0.117. CT bone% prediction had an average RMSEP of 0.990, R2 0.75, bias 0.051 and slope 0.090. Predictions for CT bone% met AUS-MEAT device accreditation error tolerances on the whole range of the dataset. Predictions for CT lean% and fat% met AUS-MEAT error tolerances on a constrained dataset.