Improving the efficiency of beef production through the use of precision farming technology and machine learning techniques

Abstract

In recent years, the livestock industry has been under increasing pressure to reduce greenhouse gas (GHG) emissions, in an attempt to mitigate global warming and climate change. Along with reducing emissions, productivity must be increased to maintain food security for the ever-growing global population. To achieve this, efficiency improvements are necessary. This thesis examines how incorporating precision livestock farming (PLF) technologies into current production systems can support producers in reaching their efficiency goals by automating practices and enabling data-driven decision making. As PLF solutions develop and improve, they become more affordable, and the resources needed for storing collected data are also improving. This, along with advancements in data processing methods, allows for real-time analysis of big data, providing further opportunities and incentives for the uptake of technologies both on farm and post-slaughter.

For beef production, a clear market need for the adoption of automated technologies to reach efficiency goals has been demonstrated. An identified key area of improvement would be the estimation and quantification of carcass quality traits, both on-farm and in-abattoir. Imaging technology has the potential to objectively estimate current carcass classification grades (EUROP fat and conformation classes). Along with this, there is also the potential for additional carcass traits of value, such as the meat yield from the whole carcass, specific carcass regions, or primal cuts, to be incorporated into assessment systems. This improves objectivity due to the estimation of quantifiable traits. Advancements in data processing are also allowing for suitable analysis of data, with techniques such as machine learning (ML) modelling providing an opportunity for timely analysis of high-dimensional data, which may be less suited to traditional statistical methods.

This thesis therefore investigated the potential for imaging technology, both on-farm and postslaughter, to objectively predict carcass classification grades and additional carcass traits of value, using both traditional statistical methods (multiple linear regression, MLR) and ML techniques (random forest and artificial neural networks). Data were obtained through the OPTI-BEEF project (UKRI project reference: 10096696).

In the first study (Chapter 2) an imaging system requiring limited infrastructure was used to extract 3-dimensional (3D) measurements (widths, lengths and volumes) from in-abattoir images of beef carcasses (n = 17,250). The 3D measurements were used to predict cold carcass weight (CCW) and EUROP conformation and fat classes. Traditional statistics (MLR using stepwise feature selection) were used to build prediction models. The best models predicted CCW with high accuracy (R² = 0.70), conformation class with moderate accuracy (R² = 0.50), and fat class with low accuracy (R² = 0.23), indicating the potential for 3D technology to objectively estimate these traits.

The second study (Chapter 3) explored ML techniques (random forest and artificial neural networks) for prediction of the previously assessed traits (EUROP conformation class, fat class, and CCW) using the same dataset as Chapter 2. This allowed for a direct comparison between traditional statistics and ML techniques. The best performing ML models predicted CCW with high accuracy (R² = 0.72), conformation class with high accuracy (71% of classes predicted correctly), and fat class with moderate accuracy (57% of classes predicted correctly), indicating that the ML techniques are not only suitable for the prediction of these traits, but also perform better than the MLR models developed in Chapter 2.

Chapter 4 explored the same 3D measurements of beef carcasses and their ability to predict meat yield traits, using either MLR or ML techniques. Estimating saleable meat yield traits mechanically allows for objective prediction, and as these traits can be quantified, it therefore allows for true assessment of the monetary value of a carcass in terms of saleable meat, rather than an assumption based on the subjective assessment of the carcass shape and fat coverage. Carcasses (n = 484) were butchered, with the weight of individual primal cuts, trim and fat being recorded. These weights were used to calculate four saleable meat yield (SMY) weights (SMY of the hindquarter, flank and forequarter, and the total SMY (sum of hindquarter, flank and forequarter SMY)), and weights of four primal cuts (sirloin, ribeye, topside and rump). Additional MLR models were also built using the EUROP conformation and fat classes as predictors, allowing for comparisons between the accuracy of prediction using either automated 3D measurements or current subjective data conventionally recorded at the abattoir. The ML and MLR models built using the 3D measurements were able to predict the SMY traits with high accuracy (R² = 0.72-0.95) and the primal cuts with moderate-high accuracy (R² = 0.52-0.74) when combined with additional animal and carcass details (e.g. breed type, sex, CCW, abattoir). Generally, the MLR models resulted in the highest accuracy for model validation. The accuracies were similar to MLR models built using the EUROP classification grades and carcass details (SMY traits R² = 0.76-0.96, primal cuts R² = 058-0.79). This demonstrates that automated measures predict the traits with similar accuracies than current data conventionally recorded at the abattoir, indicating the potential for these traits to be assessed objectively.

Finally, in the fourth study (Chapter 5), a series of 3D measurements, extracted from images of live beef animals, were combined with animal details to predict the EUROP carcass classification grades and CCW for 762 suckler beef cattle. The images were captured on commercial farms during the finishing period. Linear regression (MLR using stepwise feature selection) and ML techniques (random forest models) were explored. The MLR models predicted carcass traits with low-moderate accuracy (conformation class R² = 0.37, fat class R² = 0.24, CCW R² = 0.38). The best performing ML models resulted in higher accuracy, predicting 55% of conformation classes correctly and 46% of fat classes correctly. The random forest model for the prediction of CCW resulted in identical accuracy to the MLR model (R² = 0.38). The results indicate the 3D measurements, and the models created are capable of predicting the EUROP conformation class and CCW with moderate accuracy and EUROP fat class with low accuracy. The benefits of accurately estimating these traits on farm has been widely recognised, therefore further work is required to improve predictions allowing for accurate assessment.

The information presented in this thesis demonstrates the potential for imaging technology with limited infrastructure to predict carcass traits. The on-farm system resulted in low-moderate accuracies for the prediction of EUROP conformation and fat class, as well as CCW, suggesting further work is required for this system. It has previously been noted that breed specific models could be of value for the prediction of these traits in the live animal. Therefore, a dataset with large enough numbers covering suitable representation of each classification grade for each specific breed could be of value in aiding predictions. It is clear, however, that there could be significant benefits, both economically and environmentally, in predicting these traits on farm, and thus avoiding over/under-finished cattle being sent to slaughter. Objective, on-farm classification should therefore be further explored.

The in-abattoir system indicates higher potential for implementation commercially, by either providing an objective system for the prediction of current EUROP classification traits and CCW or augmenting current classification systems by estimating additional traits of value (weight of saleable meat or individual primal cuts). The accuracies resulting from the models predicting the primal cut weights, however, were slightly lower than the models predicting the SMY traits regardless of the technique used, suggesting further work may be required for suitable estimation of carcass primal cuts, such as a larger dataset for analysis to aid the prediction of the smaller portions of the carcass. The 3D measurements that were automatically measured were not originally selected during the development of the system based on their potential relationship to individual primal cuts, but rather their relationship with areas of the carcass that were susceptible to changes across the different carcass classification grades. Therefore, additional carcass measurements could be explored for estimating these traits and increasing accuracy.

Superiority of the ML techniques over the traditional statistical methods varied. The ML models tended to result in higher accuracies for the prediction of categorical traits (EUROP fat and conformation class), thus suggesting their suitability for objective classification. When predicting continuous traits (CCW, weight of SMY traits and primal cuts), however, the MLR models tended to outperform the ML models. This suggests that although ML techniques may indeed be more suited for classification over MLR, this is not the case for continuous traits, with MLR still being a suitable option.

The findings of this thesis should be verified using an alternative dataset with increased numbers. A larger dataset holds the potential for suitable numbers of each classification grade to be present, particularly the extreme classes, allowing for more robust models. Similarly, suitable numbers could allow for abattoir- or breed-specific models, which has the potential to further improve predictions. The current thesis, however, indicates that regardless of the technique used, there is the potential for imaging systems requiring limited infrastructure to objectively predict carcass quality traits on-farm and post-slaughter. The economic and environmental benefits of estimating these traits are important and thus there is great potential for this commercially within the industry.