Uso de imagens biométricas para predição do peso corporal e de carcaça quente de bovinos Nelore
Carregando...
Data
2018-08-13
Autores
Cominotte, Alexandre [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
O trabalho foi dividido em dois estudos. No estudo 1, objetivou-se predizer o peso corporal (PC) e o ganho médio diário (GMD) de bovinos Nelore por meio de imagens tridimensionais e comparar quatro modelos de predição: Regressão Linear Múltipla (RLM), Regressão LASSO (LASSO), Partial Least Squares (PLS) e Artificial Neural Networks (ANN). Foram coletadas 234 imagens de bovinos Nelore. As coletas de dados foram realizadas em quatro fases da ao longo da vida do animal: Desmame aos 244 dias de idade e 202,3 kg (± 27,1), Recria aos 457 dias de idade e 213,9 kg (± 25,1), Inicio da Terminação aos 590 dias de idade e 334,5 kg (± 29,2) e Final da Terminação aos 763 dias de idade e 449,5 kg (± 47,5). Nas três primeiras fases foram coletadas imagens de 62 bovinos Nelore, enquanto que na última fase apenas 48 imagens foram coletadas. O GMD foi medido: 1: Desmama - Recria, 2: Desmama – Inicio da Terminação, 3: Desmama – Final da Terminação, 4: Recria – Inicio da Terminacao, 5: Recria – Final da Terminação e 6: Inicio da Terminação – Final da Terminação. No estudo 2, quatrocentas e cinquenta imagens de bovinos Nelore foram coletadas em quatro experimentos para predição de PC e peso de carcaça quente (PCQ). Quatro conjuntos experimentais foram considerados: Set 1 inclui os experimentos 1, 2 e 3 para treinamento e experimento 4 para validação; Set 2 inclui os experimentos 1, 2 e 4 para treinamento e experimento 3 para validação; Set 3 inclui os experimentos 1, 3 e 4 para treinamento e experimento 2 para validação; Set 4 inclui os experimentos 2, 3 e 4 para treinamento e experimento 1 para validação. Os experimentos 1, 3 e 4 foram conduzidos no centro de pesquisa do Pólo Regional de Desenvolvimento Tecnológico dos Agronegócios da Alta Mogiana, em Colina - SP, Brasil. Para o experimento 1 foram utilizados 48 bovinos machos químicamente castrados da raça Nelore, com média de 24 ± 2 meses e peso vivo médio de 449,4 (± 46,9 kg). Para o experimento 2 foram utilizados 228 bovinos machos não castrados da raça Nelore, com média de 22 ± 2 meses e peso vivo médio inicial de 588,4 (± 35,5 kg). O experimento 3 foi realizado no Confinamento Experimental do Departamento de Melhoramento e Nutrição Animal da Faculdade de Medicina Veterinária e Zootecnia, UNESP – Câmpus de Botucatu em que foram utilizados 83 bovinos machos não castrados da raça Nelore, com média de 22 ± 2 meses e peso vivo médio de 516,6 (± 38,0 kg). Para o experimento 4 foram utilizados 91 bovinos machos não castrados da raça Nelore, com média de 22 ± 2 meses e peso vivo médio de 589,4 (± 30,1 kg). O sensor Kinect modelo 1473 da Microsoft foi utilizado para a captura das imagens de ambos estudos. Concomitantemente à aquisição de imagens, cada animal foi pesado em balança eletrônica. A análise das imagens foi dividida em três etapas: primeira, estimativa da distância do chão à câmera, segunda, uma etapa de segmentação e uma etapa final para extração das características. As características com correlação abaixo de P <0,05 foram removidas do modelo. Assim para as análises estatísticas no estudo 1 foram utilizadas: Área e volume corporal, seis larguras e alturas ao longo do dorso do animal e comprimento. Para comparar os modelos avaliados, foi utilizado o método de validação cruzada Leave-One-Out. Para as análises estatísticas no estudo 2 foram utilizadas: Área e volume corporal, onze larguras e alturas ao longo do dorso do animal, comprimento, excentricidade e duas medidas da curvatura da espinha dorsal. No estudo 1, o método de ANN aumentou a precisão e acurácia de PC comparado aos outros métodos: Desmama: (concordance correlation coeficiente (CCC) = 0.94; root mean square error of prediction (RMSEP) = 8.6 kg; R2 = 0.72). Recria: (CCC = 0.87; RMSEP = 11.4 kg; R2 = 0.79). Início da Terminação: (CCC = 0.96; RMSEP = 7.7 kg; R2 = 0.92). Assim, como para o PC, o uso do método ANN aumentou a precisão e acurácia das predições para GMD comparado aos outros modelos utilizados. Os resultados obtidos para foram: 1 (CCC = 0.74; RMSEP = 0.02 kg/d e R2 = 0.67), 2 (CCC = 0.89; RMSEP = 0.02 kg/d e R2 = 0.85), 3 (CCC = 0.84; RMSEP = 0.03 kg/d e R2 = 0.80), 4 (CCC = 0.54; RMSEP = 0.10 kg/d e R2 = 0.51) , 5 (CCC = 0.82; RMSEP = 0.06 kg/d e R2 = 0.76), 4 (CCC = 0.90; RMSEP = 0.09 kg/d e R2 = 0.82). Para o estudo 2, o método preditivo de ANN para peso corporal obteve melhores resultados nos sets 1, 2, 3 e 4 comparados com PLS, LASSO e RLM (CCC = 0.73, 0.66, 0.70, 0.74; RMSEP = 19.6, 27.2, 27.2, 33.7 kg e R2 = 0.58, 0.53, 0.53, 0.59, respectivamente). Já para peso de carcaça quente o método preditivo RLM apresentou melhores resultados para o Set 1 e 4, LASSO obteve resultados superiores para o Set 2 e para o Set 3 o metodo preditivo ANN apresentou melhores resultados. Set 1: (CCC = 0.14; RMSEP = 8.7 kg e R2 = 0.05, respectivamente), Set 2 (CCC = 0.53; RMSEP = 20,1 kg, e R2 = 0.40, respectivamente), Set 3 (CCC = 0.56; RMSEP = 20.1 kg e R2 = 0.44, respectivamente), Set 4 (CCC = 0.69; RMSEP = 22.7 kg, e R2 = 0.52, respectivamente).Este estudo indica que imagens tridimensionais obtidas a partir do sensor Kinect apresenta potencial para serem utilizadas para predizer peso corporal, ganho de peso e peso de carcaça quente em bovinos Nelore.
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.
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.
Descrição
Palavras-chave
Análise de imagens, Bovinos de corte, Visão computacional, Modelos preditivos, Beef cattle, Predictive models