Uso de imagens digitais para predição do peso corporal em suínos cruzados

Abstract

The technologies development made it possible to estimate morphometric measurements by digital images captured by depth sensors, that can be used in the development of mathematical models for the prediction of body weight in animals. The objective of the present work was to develop mathematical models to predict body weight in crossbred pigs from morphometric measurements obtained with two-dimensional digital images. The data collection was carried out in the swine sector of IF Goiano, Urutai campus, in 52 crossbred pigs housed in the growing-finishing phase. The animals were randomly weighed and filmed by the Microsoft Kinect® v1 depth sensor camera in the dorsal view. Subsequently, they were conducted for the measurement of manual morphometric traits, without restraint, being the dorsal length, heart girth and flank circumference, heart girth and flank width, heart girth and flank height and depth. And, finally, the lateral images of the animals were captured by the mobile phone. Following the image quality criteria, only one frame per animal was selected and, together with the lateral images, a total of 52lateral images and 52 dorsal images were selected. The morphometric traits measured from the lateral images were the lateral length, heart and flank height, depth, lateral area, and perimeter. The morphometric measurements from the dorsal images were the dorsal length, heart, flank, abdominal width, the dorsal area, and perimeter. Pearson's correlations between morphometric variables and body weight and mathematical models were developed using Multiple Linear Regression (MLR) and Support Vector Regression (SVR) methodologies for bodyweight prediction. After identifying the most suitable models to predict body weight, the average weight predicted by each of the equations selected by the MLR and SVR method was estimated and validated with the real average body weight. All Pearson correlations between real bodyweight and real and predicted morphometric measurements by digital images were moderate to high magnitude positive and statistically significant. The difference between the real bodyweight and predicted by the SVR and MLR through the real measurements was -400 g and -1.95 kg, respectively. The best model from the MLR analysis and closest to the real one was the one obtained by the morphometric measurements of the dorsal images, which explain 91% of the variance of the predicted body weight and with 3.83 kg above the real one. As for the analysis by SVR, the best model for predicting body weight presented 98% of the determination coefficient. The mathematical models developed from the morphometric measurements predicted by the digital images proved to have the potential to estimate the bodyweight of crossbred pigs from measurements such as dorsal length, depth, and heart width. The SVR method proved to be more efficient in predicting body weight through morphometric measurements of length and dorsal area, height, and heart girth width.