Body composition estimation in breeding ewes using live weight and body parameters utilising image analysis : A thesis submitted in partial fulfilment of the requirements for the Degree of Doctor of Philosophy at Lincoln University

Abstract

Farmers are continually looking for new reliable, objective and non-invasive methods for estimating ewe body condition. Live weight (LW) in combination with body condition score (BCS) are used by farmers as a basis to determine the condition of the animal. Where LW is a crucial indicator of body composition, body condition can be evaluated by determining the amount of fat in the animal. This amount plays a key role in ewes’ health condition and animal productivity. The body condition score is used to monitor animals to ensure the best condition and is a measure between 1 (low condition) and 5 (high condition). If an ewe has a condition below 2 this is considered poor, whereas above 3 is regarded as good condition and ready for breeding. The current method is subjective (relies on professional judgment from farm handlers) as such it can introduce an element of error when estimating the fat, which makes it difficult to monitor the animal condition. A quick, objective, and accurate method of body composition estimation is required to improve farm management. If such a method could be devised, many farmers around the world would utilize it to assist in managing sheep farms. In addition, image processing and body measurements have not been used before to estimate body composition for ewes during the production cycle using a comprehensive, repeatable, non-invasive method. Image processing and body parameter measurements have been widely used to estimate ewe body size and weight. The objective of this thesis was to establish a relationship between body parameters of body length, width, depth and height as independent variables and body fat, lean, bone and carcass (total weight of body fat, lean and bone) as dependent variables. The aim was to use these easily obtained body parameters to predict body composition. Two full experiments at weaning and pre-mating were conducted to establish the relationship between body composition and body parameters using measurements automatically determined by an image processing application at Lincoln University sheep farm for 88 Coopworth ewes. Computerised Tomography (CT) technology was used as a benchmark to validate the predicted body composition. A trial run, wool test, uncertainty test, repeat test and carcass test were also conducted to minimise uncertainty and test the experiment setup. The image processing application used techniques from OpenCV library such as image extraction, convert to HSV colour, erode, dilate and smooth filter to remove the ewe’s head, legs (for side image) and extract the body to calculate the body parameters in an automated method. Multivariate linear regression (MLR), artificial neural network (ANNs) and regression tree (RT) statistical analysis methods were used to analyse the relationship between independent and dependent variables to predict body fat, lean, bone and carcass. The artificial neural network method was found to be the best method to show how much variance of the dependent variables is explained by a set of independent variables. The result showed a correlation between fat, lean, bone and carcass weight determined by CT and the fat, lean, bone, carcass weight and percentage of fat–carcass weight estimated by live weight and body parameters calculated in an automated method using the image processing application with R2 values of 0.88, 0.85, 0.72, 0.97 and 0.94, respectively for the training data of 138 ewes with a root mean square error (RMSE) less than 2.5. A new set test data of was used to test the accuracy of the results of multivariate linear regression, neural networks and regression tree. The neural networks model provided the highest R2 for total fat prediction with R2=0.90 and RMSE=1.01 with a maximum difference of 2.7 kg and a minimum difference of 0.018 kg between the predicted value and the actual value, lean prediction with an R2 of 0.72 and RMSE=1.03, bone prediction with an R2 of 0.50 and RMSE=1.21, carcass prediction with an R2 of 0.95 and RMSE=1.31 and percentage of fat – carcass weight with an R2 of 0.90 and RMSE=1.30, respectively. ANNs also showed the lowest RMSE for fat with a value of 1.01 and for carcass with a value of 1.31. The image processing application calculations showed an uncertainty of -9.43 to 9.22 mm for chest width. The result also showed that many ewes had the same body condition score but different fat and chest widths, which confirmed that the body condition score may not provide an accurate indication of fat. The results showed an optimal fat of 9.37% of LW for ewes during the production cycle. If the percentage of fat is less than or more than 9.37%, farmers must take action to improve the conditions of the animals to ensure the best performance during weaning and ewe and lamb survival during the next lambing. The new method can be used to determine body composition on sheep farms as an alternative to BCS since it showed more accurate results. This method can also lead to the use of new image analysis technologies and more research on using image processing on-farms. The accuracy of the new method is slightly less than CT but it takes less time and cost than the CT. It can be used on-farm at any stage during ewe production cycle and can be applied on animals with wool or after shearing the wool.