Ovine Observer

Predicting lamb composition from carcase weight and tissue depth

Andrew Williams, Murdoch University, WA
Corresponding author:


Lean meat yield is an important profit driver for the sheep meat industry. The current industry standard for determining carcase composition is based on carcase weight and a measurement of fat depth by palpation of the GR site (located over the 12th rib 110 millimetres from the mid line), however this has been shown to be a highly imprecise estimate of lean meat yield.

This precision can be markedly improved by manually measuring grade rule (GR) tissue depth in millimetres (mm), however reliance upon a single point measurement is still likely to introduce significant bias in genetically diverse populations of lambs. Furthermore, there is concern within the Australian lamb industry that these measures are prone to bias due to human operator error as well as variation in abattoir processing.

GR is measured 110mm from the top of the fillet
GR tissue depth measurement site

There is little data available to quantify this error, with a key limitation being the method for determining carcase composition. Historically this has been reflected through carcase bone-out data, yet this is problematic due to varying bone-out specifications across data sets, as well as large human-imposed operator effects.

In Australia, with the introduction of computed tomography (CT) scanning methodologies, datasets are now available to assess the efficiency of predicting carcase composition using carcase weight and GR tissue depth.

This study assesses the capacity of carcase weight and GR tissue depth to predict carcase CT fat% (the percentage of fat in a carcase, as measured by CT) across multiple datasets. It was hypothesised that the precision and accuracy of this prediction would vary between different groups of animals. 

Materials and methods

This study made use of 28 different datasets totaling 2289 lambs where CT estimates of carcase fat%  (CT fat%), carcase weight and GR tissue depth measurements had been collected over a nine year period. Each dataset contained recordings for between 48 and 99 lambs.

One of these data sets consisted of lamb carcases that were sourced over a 45-minute period immediately following slaughter from a commercial abattoir near Bordertown, South Australia. These lambs were selected randomly across a broad range of fatness and carcase weight, hence their parentage is unknown.

CT image of lamb carcase

The remaining 27 of these datasets were individual slaughter groups of lambs from the Meat and Livestock Australia Resource Flock experiment, or from the Sheep Cooperative Research Centre Information Nucleus Flock experiment.

The lambs (Merino, Maternal x Merino, Terminal x Merino and Terminal x Border Leicester-Merino) were the progeny of 433 industry sires, representing the major sheep breeds used in the Australian industry. 

The siretypes included Terminal sires (Poll Dorset, Suffolk, Texel, White Suffolk), Maternal sires (Border Leicester, Coopworth, Dohne Merino), and Merino sires (Merino, Poll Merino).

Each dataset represented a slaughter group which was balanced for sire breed. In all cases tissue depth at the GR site and hot carcase weight were measured within one hour of slaughter. CT scanned data was captured between two and five days post-mortem.

An equation to predict CT fat%, based on carcase weight and GR tissue depth, was trained in one dataset and then validated in the other 27 datasets. This process was then repeated 28 times to test the transportability of the different prediction equations, providing 756 different tests.

Results and discussion

Within the training data across each of the 28 datasets the error associated with predicting CT fat% (difference between predicted CT fat% and actual CT fat%) from carcase weight and GT tissue depth varied markedly. The root mean squared error (RMSE) of the differences ranged between 1.67 to 3.10 CT fat% units. Therefore within the training data, two thirds of the predictions will fall within 1.6 to 3.10 CT fat% units of the actual CT fat%. The models described as little as 15% and as much as 77% of the variation within these populations.

When the 28 trained models were validated across each of the other datasets the precision and accuracy indicators showed marked variation. Across the 756 validation tests the RMSE values averaged 2.36, yet ranged between 1.67 to 3.33 CT fat% units (Table 1). The R2 values (a measure of the variation explained by the model) averaged 0.52, yet ranged between 0.12 and 0.77. An R2 of 1 would indicate a perfect correlation between the predicted and actual CT fat% values.

These results highlight the substantial variation in precision when using carcase weight and GR tissue depth to predict composition. The accuracy indicators also varied, with bias values (indicating over and under prediction) ranging between 6.83 to -6.95 (Table 1). 

Table 1 Precision and accuracy estimates for the relationship between actual CT fat % and predicted CT fat % from models containing hot carcase weight in kilograms (kg) and GR tissue depth (mm). Precision estimates include R-squared and RMSE, and accuracy is indicated by the bias. Values are shown for the mean, standard deviation, minimum and maximum from testing 28 models across 27 datasets, a total of 756 validation tests. (Note: *The average of the absolute values of bias is reported.)
Value Mean Standard deviation Minimum Maximum
R2 0.52 0.15 0.12 0.77
RMSE 2.36 0.30 1.67 3.33
Bias 1.60* 2.08 -6.95 6.83

Figure 6 provides an example of the process of training the model, and then validating it across the remaining datasets. In this example, validation tests are shown for only one of the remaining 27 datasets.

In support of the hypothesis, these results highlight the substantial variation in prediction precision and accuracy when using carcase weight and a single point measure of tissue depth to reflect carcase fatness. This work aligns well with previous studies where carcase weight and GR tissue depth demonstrated poor precision and limited ability to differentiate between genetically diverse lines of sheep.

In part some of the variability would be driven by differences in data range between the datasets, potentially causing some extrapolation beyond the range of the training data. However all of the datasets used had broad variation with data ranges that substantially overlapped, hence in most cases there was relatively little extrapolation. As such, the variability in prediction is likely due to a number of other factors. 

Actual CT% fat increases along with predicted CT% fat
Figure 6 Example of the relationship between actual CT fat % and predicted CT fat % from a model using hot carcase weight (kg) and GR tissue depth (mm). Dashed line represent a perfect prediction; solid line show the average prediction in the validation dataset

Use of a single measure of tissue depth to reflect carcase fatness relies upon accurate and consistent measurement. Although these are experimental datasets in which great care has been taken during the collection of GR tissue depth, there are still likely to be processing and operator effects which may vary between datasets. Furthermore, under commercial conditions within those abattoirs that measure GR tissue depth there is likely to be even greater variability due to operator error when working at speed.

It should be noted that most Australian abattoirs don’t measure GR tissue depth directly, instead using palpation of the carcase to estimate this measure. Hence under commercial conditions the prediction of carcase fatness would be even more variable than that demonstrated in this study.

The prediction of carcase fat composition relies upon a strong correlation between GR tissue depth at its site of measurement with fat composition elsewhere in the carcase. However, there is evidence that suggests that genetics can strongly influence this correlation, redistributing bone, muscle and fat within the carcase. Therefore the genetic differences present across datasets may offer a source of error contributing to the bias and variable precision.

In the present study this effect may be limited as the datasets used are derived from nucleus flock slaughters, with strong genetic linkage between these groups through common sires or common dams. The only exception to this was for dataset 28 which consisted entirely of randomly sourced animals of unknown parentage, thus it is not surprising that this group demonstrated substantial bias (example shown in Figure 6). 


These results demonstrate the variability in estimating carcase fat composition from carcase weight and GR tissue depth. Based on the genetic variation present in Australian sheep flocks, and the likely increase in measurement error under commercial conditions, we conclude that the variation in bias and precision demonstrated in this study could well be understating the bias that is present in a commercial setting.

This illustrates why the Australian sheep industry has little confidence in these measurements to reflect carcase composition and highlights the need for a whole carcase composition measurement that is independent of breed, processing and operator error.