DEVELOPMENT AND IMPLEMENTATION OF AN ECONOMICALLY VIABLE COMPUTER VISION SYSTEM TO MONITOR AND CONTROL METABOLIC DISORDERS IN DAIRY COWS

<b>Project Methods</b><br> Aim1.A RFID-camera systems will be installed at the UW-Madison Research Farms (Arlington-WI and Marshfield-WI) in three barn locations: (1) above the water tank in the dry cow pens; (2) above the water tank in the close-up pens; and (3) at the exits of the milking parlor. The CVS will be developed to extract 3D image-features related to cows' body shape. Our objective is to monitor a total of 1,000 dairy cows from -60 to +60 DRTC. This sample size will provide enough statistical power to detect even low correlations between image-based features and BCS and animal performance variables. The CVS will have depth cameras from Intel® RealSense™ Depth Camera D455, already tested by our research group, that will acquire a top-down view infrared and depth images from the cows' dorsal area. We will collect BCS from all cows to train a deep neural network to perform BCS classification that will be used as a predictor of metabolic diseases associated with negative energy balance and will serve as a control measurement of body shape changes. Subsequently, two important steps will be performed: feature extraction, and model development. 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, dorsal width, body volume, eccentricity, and Fourier shape descriptors. The computational features will be extracted using the feature maps of the pre-trained CNN for animal body segmentation. Additionally, we will use the features of the last dense layer of a pre-trained CNN used to classify body condition score. The output of the proposed deep neural networks will be used in combination with the biological features as image-based predictors, and in combination with other covariates to early detect health issues associated with negative energy balance.Aim 2.Two groups of covariate data sets will be evaluated.First is the PREPARTUM GROUP:Set 0) use of BCS generated using the CVS;Set 1)variables from body shape (biological and computational) obtained from CVS during the prepartum period;Set 2) variables from feeding behavior obtained from CVS during the prepartum period;Set 3) Set 1 + Set 2;Set 4)Set 1 + Set 2 + cow records from herd management software. Second is the POSTPARTUM GROUP:Set 5)Set 1 +variables from body shape obtained from CVS in the postpartum until the diagnostic of health problem; andSet 6)Set 3 + variables from body shape and feeding behavior obtained from CVS in the postpartum until the diagnostic of health problem. 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, and the time series (body shape features, feeding behavior and cow records) measured before the event will be used to create predictive models. Different lengths of time series and varying intervals from the last predictor measurement to the health event will be evaluated to determine the optimum data to carry relevant signal and to assess how early health problems can be detected accurately. 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. We will explore the use of Recurrent Neural Networks and Logistic Regression for early prediction of health events, such as ketosis, abomasum displacement, clinical mastitis, and metritis. Lastly, we will use individual cow image sequences acquired during the prepartum period (-60 to 0 DRTC) to predict potential health problems after calving.Aim 3.We will populate an independent and unbiased web portal, openly and freely available, within the University of Wisconsin-Madison Dairy Management to list all the digital technologies available for dairy farming. We will tap on Dr. Dorea's class materials of AN/DYSCI 875 that has already a long list of technologies including activity sensors, boluses, computer vision, sound recognition, etc. and build upon it to maintain the world's most comprehensive and the most up-to-date list of technologies available in the dairy industry. For each technology, there will be a systematic description of what the technology does, how it does it, how much it cost, pictures and videos (as available), and contact information. Once the site is launched, we will publish an article in the Hoard's Dairyman magazine about the existence of the web portal. Our outreach specialist will be in charge to maintain the list up-to-date, clearly organized, and completely searchable. We will take every contact opportunity with farmers and other stakeholders (e.g., extension meetings) to highlight and point to the audience to the web portal. As part of our awareness efforts, we will launch a monthly podcast, "The Digital Livestock," to discuss technologies in the livestock world.Based on our experience of creating AI tools for outreach and participating on K-12 demonstrations, open-science events, and UW-Madison Science Expedition, we will expand and increase our participation to dairy farming stakeholders including farm managers, employees, consultants, extension educators, vendors, and others. We will participate in all possible hands-on events such as yearly WI state fair and the World Dairy Expo in Madison-WI with a booth to demonstrate precision livestock technologies.We will develop online training modules addressed to farmers, extension educators, and farm consultants that will deal with three aspects of farm knowledge, all of them related to improve their readiness to adopt new technologies in the realm of digital agriculture and precision dairy farming: 1) installations, 2) personnel, and 3) data management. The training module about installations will deal on how new or improved installations planned at the farm could be made more technology friendly by including connectivity hubs, outlets, or strategic lights. Similarly, the personnel module will deal on how the farmer could support technological education of existing or new employees and will include not only resources available online and everywhere else, but also contact of knowledgeable professionals who could respond specific farm questions.Aim 4.We will develop a dynamic optimization and simulation model of a dairy herd. The model will be constructed using a hierarchical Markov chain process in monthly steps/stages. Each cow will be assigned a state based on production characteristics, and the model's aim will be to maximize the net present value (NPV) of a cow in each state-stage. During the transition period, cows will have a risk for transition diseases based on the reported incidences from the participating farms. Transition diseases will be classified as uterine (UTD) and non-uterine diseases (NUTD). We will evaluate how disease risk and losses would change if changes in management decisions due to the use of the CVS and early identification of high-risk cows were implemented.To assess the impact of the use of CVS and its potential impacts, we will apply the Becker-DeGroot-Marschak auction (BDM) approach to elicit willingness-to-pay (WTP) for the four lettuce categories.Under the BDM auction procedure, subjects individually submit sealed bids for a good. Next, a random number or price is drawn from a pre-specified distribution. Individuals whose bid is greater than the randomly drawn price "win" the auction and can purchase the good at the randomly drawn price. We will conduct separate experiments for beverage milk products and cheese to determine whether WTP is affected differently based on the product made from farm milk.Econometric models including participant demographic characteristics will then be used to assess correlates of WTP estimates.