Recipient Organization
UNIV OF WISCONSIN
21 N PARK ST STE 6401
MADISON,WI 53715-1218
Performing Department
DAIRY SCIENCE
Non Technical Summary
Nearly all cows will experience negative energy balance (NEB) to support the high energy demands of lactation during the transition period. This NEB can lead to a variety of metabolic disorders. Body condition score (BCS) is a commonly used tool to monitor and manage NEB in lactating cows. However, BCS is a periodic, subjective measurement that cannot detect small changes in body shape or composition. Consequently, the development of a computer vision system to assess BCS in real-time will play a crucial role to precisely detect body tissue mobilization. Computer vision systems have emerged as a powerful technology to identify animals and monitor complex behavioral traits. The integration of computer vision systems is vital to automate real-time collection of cow-level phenotypes, such as body tissue mobilization, body weight, and biometrics. Such automated systems would allow the development of predictive analytics for early detection of health issues during the critical transition period in dairy cows. The objective of this project is to develop a high-throughput phenotyping platform using computer vision systems based on deep learning to integrate phenotypes from body shape, biometrics measurements, and animal identification. Such phenotypes will be used to create predictive modeling for precise and real-time detection of NEB and associated health problems. This project will have a major economic impact on the dairy industry in Wisconsin and beyond by producing a real-time framework for early diagnosis of health issues during the transition period at the individual-cow level. Application of our resulting framework will generate economic value by saving treatment costs and reducing culling rates in lactating dairy cows, which will decrease the negative economic impact of transition cow health disorders.
Animal Health Component
50%
Research Effort Categories
Basic
20%
Applied
50%
Developmental
30%
Goals / Objectives
The proposed project will change the paradigm of disaggregated cow-level information and create a state-of-the-art artificial intelligence system to effectively diagnose metabolic disorders during the transition period by: (1) developing a ground-breaking computer vision technology for monitoring cows during transition period, and (2) taking advantage of integrating cow-level information into a single database to create predictive analytics tools. The following specific objectives are proposed in this project:Aim 1: Development of a computer vision system for real-time assessment of body tissue mobilization and body weight.Aim 2: Assessment of 3D dorsal images from lactating dairy cows as a potential tool for animal identification.Aim 3: Integration of computer vision-based features and development of analytical tools for early detection of health issues in transition cows.The possibility of high-throughput phenotyping will make breakthroughs andnew discoveries in dairy science (e.g. genetics/genomics, nutrition, physiology, and reproduction) and improve the management of livestock operations. Our project is low-risk, yet innovative, and it will provide tools and strategies that will have an impact well beyond the current proposal.
Project Methods
In Outcome 1 and 2, we will develop a computer vision system capable of identifying individual animals based on body shape and detecting body tissue mobilization. The output of the system will be a set of image-based features related to cow body shape (e.g., volume of thurl area). We will develop an algorithm to monitor changes in cows' body shape by evaluating each specific image-feature across the transition period. The expected result would be a reliable and non-invasive system for animal identification, and a significant decrease in the number of cows with metabolic disorders and a reduction in severity among those that are affected.Aim 1: A CVS will be developed to extract 3D image-features related to cows' body shape and BW. RFID-camera systems will be installed at Emmons Blaine Dairy Cattle Research Center (EBDCRC). Our objective is to real-time monitor cows from -60 to +60 days relative to calving (DRTC). The CVS will have two cameras (Intel® RealSense™ Depth Camera D435), in which one will acquire a top view from the cows' dorsal area, and the other will collect a rear-view image of the cow. Data from 200 animals will be collected, stored, and analyzed during Year 1 and 2. After data collection, three important steps will be performed: image segmentation, feature extraction, and model development. An automated image segmentation and best frame selection process will be implemented to isolate the cows' body from the image background. Next, feature extraction will be implemented using two approaches: 1) biological features, and 2) computational features. The biological features, here called biometric body measurements, are known to be associated with BW and shape, such as dorsal area, hip height, dorsal width, dorsal length, volume, eccentricity, and Fourier shape descriptors. Three anatomical landmarks related to BCS [(1) area between hook bone, spinous process and transverse processes; (2) thurl area; and (3) pin bone or pin area] were defined as regions of interest, and for each region all features noted will be extracted according to Fernandes et al. (2019) and using deep learning algorithms, specifically variational autoencoders (Kingma and Welling, 2014). Cows will be classified as high plasma NEFA if one or more samples between 5 to 14 DIM exceed 600 μEq/L; otherwise, the cow will be classified as low NEFA. The body features extracted by CVS will be associated with high and low NEFA classes, and we will also investigate the capability of the body features (shape and biometrics) to predict plasma NEFA as a continuous (μEq/L) and discrete (high/low) variable. The features extracted (biological and computational) will be used as potential predictors of BW and plasma NEFA concentration using different approaches. Baseline methods like Partial Least Squares (Dórea et al., 2018b) and Bayesian Regression Models (Toledo-Alvarado et al., 2018), will be evaluated against state-of-art approaches using deep learning like Artificial Neural Networks (Dórea et al., 2018b), and Recurrent Neural Networks (RNN, Donahue et al., 2014). Prediction quality of all models will be assessed by Leave-One-Cow-Out Cross-Validation. Model precision and accuracy will be evaluated as described in Dórea et al. (2017).Aim 2: The depth images and its respective metadata (date, time, and animal ID) collected from Aim 1 will be used to develop the computer vision system for animal recognition based on surface body shape. The depth image will be transformed into point cloud data, and subsequently encoded as voxels. A 3D convolutional neural network named will be implemented to generate the predictions. Images will be stored by day as a way to evaluate the predictive ability of the algorithm to recognize animals within day as well as across days, as cows' body shape changes.Outcome 3 will allow constant and real-time collection of all relevant farm information. Researchers and decision-makers will be able to access these outputs in real-time to guide their decisions. In particular, the platform will provide metrics such as lists of cows with high probability to develop a metabolic disorder, information about the type of health problem that is most likely to occur, an indication of the number of cows in negative energy balance, a list of cows in severe negative energy balance, and so on. In the end, human decisions will be guided by these outputs, and we are positive these decisions will promote improved health, welfare, and profitability.Aim 3: We will integrate the computer vision-based outcomes into a centralized database. Such database will contain the real-time prediction of body weight, shape and biometrics, and animal identification. The automation of data collection, integration and harmonization will be performed by CHTC. For the predictive analyses of health problems associated with NEB, data regarding individual cows' health events such as ketosis, endometritis, abomasum displacement, and milk fever will be used. Staff-recorded reports of health problems will be validated as a true health event using proof of diagnosis or treatment. The longitudinal data from all sources will be tracked until the health event occurred. Thereafter, different lengths of time series leading up to the health event will be tested, in order to determine the optimum data to carry relevant signal and to assess how early health problems can be detected accurately. We will explore the use of RNN (Choi et al., 2016) and Logistic Regression (Chandler et al., 2017) for early prediction of health events, such as ketosis, abomasum displacement, clinical mastitis, and metritis.