Advanced search
Start date
Betweenand


Use of Biometric Images to Predict Body Weight and Hot Carcass of Nellore Cattle

Full text
Author(s):
Alexandre Cominotte
Total Authors: 1
Document type: Master's Dissertation
Press: Jaboticabal. 2018-11-01.
Institution: Universidade Estadual Paulista (Unesp). Faculdade de Ciências Agrárias e Veterinárias. Jaboticabal
Defense date:
Advisor: Otavio Rodrigues Machado Neto
Abstract

The work was divided in two studies. The objective of this study was to predict the body weight (BW) and the average daily gain (ADG) of Nellore cattle by 3-D images and to compare four prediction models: Multiple Linear Regression (MLR), LASSO Regression, Partial Least Squares (PLS) and Artificial Neural Networks (ANN). A total of 234 images of bovine Nellore were collected. Data collection was performed in four stages throughout the life of the animal: Weaning at 244 days of age and 202.3 kg (± 27.1), Stocker at 457 days of age and 213.9 kg (± 25.1), Initial of Termination at 590 days of age and 334.5 kg (± 29.2) and Finish of Termination at 763 days of age and 449.5 kg (± 47.5). In the first three phases images of 62 Nellore cattle were collected, while in the last phase only 48 images were collected. The ADG was measured: 1: Weaning - Stocker, 2: Weaning – Initial of Termination, 3: Weaning – Finish of Termination, 4: Stocker - Initial of Termination, 5: Stocker - Finish of Termination and 6: Initial of Termination - Final of Termination. In study 2, four hundred and fifty images of Nellore cattle were collected in four experiments for prediction of BW and hot carcass weight (HCW). Four experimental sets were considered: Set 1 includes experiments 1, 2 and 3 for training and experiment 4 for validation; Set 2 includes experiments 1, 2 and 4 for training and experiment 3 for validation; Set 3 includes experiments 1, 3 and 4 for training and experiment 2 for validation; Set 4 includes experiments 2, 3 and 4 for training and experiment 1 for validation. Experiments 1, 3 and 4 were conducted at centro de pesquisa do Pólo Regional de Desenvolvimento Tecnológico dos Agronegócios da Alta Mogiana, em Colina - SP, Brazil. For the experiment 1, 48 male Nellore bulls, with a mean of 24 ± 2 months and a mean live weight of 449.4 (± 46.9 kg) were used. For experiment 2, 228 Nellore male bovines were used, with an average of 22 ± 2 months and initial mean live weight of 588.4 (± 35.5 kg). Experiment 3 was carried out at the Confinamento Experimental do Departamento de Melhoramento e Nutrição Animal da Faculdade de Medicina Veterinária e Zootecnia, UNESP – Câmpus de Botucatu, in which 83 male Nellore bulls were used, with an average of 22 ± 2 months and weight live weight of 516.6 (± 38.0 kg). For the experiment 4, 91 male Nellore bulls were used, with an average of 22 ± 2 months and mean live weight of 589.4 (± 30.1 kg). The Microsoft Model 1473 Kinect sensor was used to capture images from both studies. Concomitantly with the acquisition of images, each animal was weighed in an electronic scale. The analysis of the images was divided into three stages: first, the estimation of the distance from the ground to the camera, second, a segmentation stage and a final step to extract the characteristics. Characteristics with correlation below P <0.05 were removed from the model. Thus, for the statistical analyzes in study 1, we used: Body area and volume, six widths and heights along the animal's back and length. To compare the evaluated models, the method of cross validation Leave-One-Out was used. For the statistical analyzes in study 2 were used: Area and body volume, eleven widths and heights along the animal's back, length, eccentricity and two measures of the curvature of the spine. In study 1, the ANN method increased PC accuracy and accuracy compared to the other methods: Weaning: (concordance correlation coefficient (CCC) = 0.94; root mean square error of prediction (RMSEP) = 8.6 kg; R2 = 0.72). Stocker: (CCC = 0.87, RMSEP = 11.4 kg, R 2 = 0.79). Initial of Termination: (CCC = 0.96, RMSEP = 7.7 kg, R2 = 0.92). Thus, as for the BW, the use of the ANN method increased the precision and accuracy of the predictions for ADG compared to the other models used. For study 2, the predictive ANN method for body weight obtained better results in sets 1, 2, 3 and 4 compared to PLS, LASSO and RLM (CCC = 0.73, 0.66, 0.70, 0.74; RMSEP = 19.6, 27.2, 27.2, 33.7 kg and R2 = 0.58, 0.53, 0.53, 0.59, respectively). As for warm carcass weight, the RLM predictive method presented better results for Set 1 and 4, LASSO obtained higher results for Set 2, for Set 3 the ANN predictive method presented better results. Set 1 (CCC = 0.14, RMSEP = 8.7 kg and R2 = 0.05, respectively), Set 2 (CCC = 0.53, RMSEP = 20.1 kg, and R2 = 0.40, respectively) = 20.1 kg and R2 = 0.44, respectively), Set 3 (CCC = 0.56, RMSEP = 20.6 kg, and R2 = 0.44, respectively). Set 4 (CCC = 0.69, RMSEP = 22.7 kg, and R2 = 0.52, respectively). This study indicates that 3-D images obtained from the Kinect sensor have potential to be used to predict body weight, weight gain and hot carcass weight in Nellore cattle. (AU)

FAPESP's process: 17/02057-0 - Use of visual scores and biometric image in Nellore cattle
Grantee:Alexandre Cominotte
Support type: Scholarships in Brazil - Master