Recipient Organization
UNIVERSITY OF MISSOURI
(N/A)
COLUMBIA,MO 65211
Performing Department
Animal Science
Non Technical Summary
Computer vision and data-driven approaches are used to enhance existing DNA and growth beef industry datasets (>150,592 cattle with 50 million DNA variants) with dense data from three-dimensional (3D) cameras (~5,000 cows, ~7 million points per cow) to answer 170-year-old questions in ecology and evolution (Bergmann's Rule). Using yield, environmental stress, and pathogen resistance, the project creates DNA predictions to help farmers select cattle that thrive in the environment at their ranch, ultimately improving the environmental efficiency of beef production. Identifying adapted, energy efficient cattle allows family farms and ranches across the USA to be more profitable and improves animal welfare, an important societal issue. In Objective 1, 3D computer vision is used to measure the surface area and volume of cattle to test if surface-area-to-volume ratio (SA:V) influences genetics for growth under cold stress or growth under high feed resources. Larger animals have a lower SA:V, allowing them to retain heat more efficiently in winter. One hypothesis suggests larger animals live at higher latitudes (further north in this study) because they use the lower SA:V to better deal with cold stress. An alternative hypothesis suggests that animals are larger at higher latitudes because more food is available when the animals are growing. This research creates two DNA predictions, one tailored to cold stress and the second tailored to high feed resources. Genetic merits of animals within each of these production contexts will be directly compared to their SA:V ratios. In Objective 2, this project tests the hypothesis that random models versus machine learning methods better predict various cattle traits (feed intake, fat reserves, body weight, etc.) using processed 3D point clouds and environmental data. The project uses random models typically used in DNA predictions, but rather than DNA variants as the predictors, the analysis uses the large 3D camera data as predictors.Besides advances in computer vision algorithms (e.g. Structure from Motion) and new approaches to apply data-driven machine learning to agriculture problems, this project provides a definitive test of James's Rule (intraspecific Bergmann's Rule) that animal size increases with latitude due to an increased ability to maintain body temperature with a smaller SA:V ratio. This is contrasted with Geist's Rule that body size increases due to advantages of larger animals to utilize seasonal burst of food at higher latitudes. Explicitly comparing direct SA:V measurements with precise climate-by-growth and resource-by-growth predictions, this project will answer which force drives evolution in body size across latitudes. This will be the first definitive test of Bergmann's rule. By comparing random models versus deep learning to predict a variety of beef cow production traits using 3D camera data, this project indicates whether random models can replace or clarify artificial intelligence in many contexts. This project creates new data recording and selection tools based on Internet of Things and 3D computer vision algorithms to address the pressing need to measure and improve efficiency in pasture, cow-calf production (a $64 billion industry). This enables convenient and affordable measurement of beef cow efficiency (a vital aspect of beef production that is not currently predicted).Partnering with the University of Missouri School of Journalism Strategic Communication capstone course provides strategic plans to educate a variety of audiences on the true environmental impact of cattle and how technology adoption decreases this impact. The three-pronged approach 1) educates lay audiences 2) delivers online educational outreach to farmers/ranchers and 3) provides traditional extension programming to farmers/ranchers through regional and national in-person presentations. The first prong will reach a national audience through cable television. The second and third prong will reach approximately 7,000 people per year. With the National Center for Applied Reproduction and Genomics, a "Data Science Best Practices" module is created to give veterinary and graduate students a primer on data-driven agriculture. Partnerships with industry create applied genetic evaluations to predict genetic, environment, and management interactions, which allow precision selection not currently possible with one-size-fits-all national evaluations. The cow scanning system will be commercialized and added onto existing commercial platforms that collect feed and water intake. The 3D data is used to predict cow efficiency. Collectively, these impacts improve the financial security of rural farmers.
Animal Health Component
33%
Research Effort Categories
Basic
34%
Applied
33%
Developmental
33%
Goals / Objectives
This project has two main goals.Use high-throughput phenotyping and "big data" to better predict production efficiency.Use high-throughput phenotyping and "big data" to test rules of evolution and ecology.For big data to be effective in livestock industries, we do not need more data, we need more useful data. Livestock industries need information-rich data to predict economically important traits that are not currently measured. Livestock production also needs cost effective, accurate ways to predict traits that are expensive to measure. However, electronic, high-dimensional records of beef cattle traits do not currently exist. We will address these needs by creating a cattle body scanner using from Red-Green-Blue, thermal and time-of-flight (ToF) cameras to create three-dimensional data with heat and color measures (4D-RGB); extract phenotypes from 4D-RGB data; and apply machine learning to predict traits.ObjectivesObjective 1)The first objective is to test the hypothesis that surface-area-to-volume ratio (SA:V) influences genotype-by-environment effects in Bos taurus beef cattle production using phenotypes based on 4D-RGB data generated with novel cattle scanners. An alternative hypothesis is that SA:V influences genotype-by-management effects in beef production.In 1847, Carl Bergmann proposed that body size increases at higher latitudes. It is believed that Bergmann's Rule was originally intended for comparisons across species. However, James's Rule has been suggested as an appropriate name for comparisons of body size and latitude within a species. There are two possible mechanisms for the effect of surface area (SA) and volume (V) on production in domesticated cattle. First is the Heat Conservation Hypothesis. Larger animals have lower SA:V compared to smaller animals. This allows larger animals to more efficiently conserve heat. Interestingly, it may actually be adaptation to heat and humidity, rather than to cold, that is responsible for the relationship between body size and latitude. Second is the Resource Availability Hypothesis (also referred to as Geist's Rule or eNPP Rule). This hypothesis states that primary plant productivity during the growing season (referred to as eNPP by Huston and Wolverton) influences the growth and reproduction of animals. This seasonal surge of resource availability favors larger animals.Beef cattle represent a unique opportunity to test these two hypotheses.Beef cattle are raised across the U.S. with essentially no management interventions to limit the effects of heat and cold stress, allowing the maximum influence of genotype × environment effects.Resources available to cattle, especially in the growing phase, are controlled by farmers and ranchers. These tend to fall into two groups: high-inputs (high resource availability) and low-inputs (low resource availability).There are large phenotypic and genomic databases available for beef cattle.Further, we are creating a system to measure the SA and V of cattle in an inexpensive, high-throughput manner. We will collect ~6,475 of SA and V phenotypes, which has not previously been done in any animal species.Restated, Objective 1 will test two hypotheses of forces that drive body size ecology and evolution in animals. We will test if heat retention as a function of larger body size altering SA:V affects adaptation to local climates. This is James's Rule (intraspecific Bergmann's Rule) that body size increases with latitude. We also test if resource availability (genotype × management interaction) is associated with body size. This is Geist's Rule, that pulses in primary plant productivity direct growth in animals.Objective 2)In Objective 2 we will test if multimodal data extracted from RGB, thermal and ToF cameras can be used to predict several production traits (feed intake, body condition score [level of fat reserves], mature weight, hair shedding, foot and skeletal scores, udder scores, etc.), comparing random regression models and a novel deep learning approach using XAI techniques.The second objective is to test the hypothesis that random models produce the same prediction accuracy as artificial intelligence (AI) methods when a large number of features are used to predict quantitative traits. Random regression models, such as Ridge Regression/GBLUP, BayesB, and BSLMM, are often used in genomic prediction to estimate the genetic merit of animals through a supervised process involving thousands to millions of DNA features. The effect of each feature is reported from these statistical models, and there are many software implementations that can handle very large data sets. However, these random regressions do not take advantage of the added discriminant power that neighboring/correlated data points, complementary features, and other underlying, nonlinear properties that a large number of features can provide with the use of many pattern recognition and machine learning approaches, e.g. deep learning (DL). Moreover, through the use of explainable AI techniques such as those derived from topological data analysis (TDA), we will test when and why random regression can/cannot compete with machine learning, and using what set of high dimensional image data. Indeed, eXplainable AI (XAI) is a promising research area, and this proposed objective has the potential to provide never-seen parallels, insights and tools to be used in random regression, in current PR/ML approaches, or both, in order to better understand predictions.As a related pursuit of better understanding AI and random regression methods, we will also test if the SA:V predicts an animal's appetite and feed intake. Perhaps the greatest ecological niche of cattle and other ruminants is their capacity to utilize cellulose and other plant materials indigestible to mammalian enzymes. The ruminant digestive tract evolved from non-ruminant species. Thus, similar to non-ruminant species, intake in cattle is regulated, in part, by chemostatic mechanisms that sense oxidation of metabolic fuels. However, different from non-ruminants, intake of cattle can also be regulated by ruminal volume. Ruminal volume is closely related to body size in ruminants, and when rates of ruminal fermentation of cellulosic materials are limited then intake of cattle is more closely described by ruminal fill than by caloric density of diets. Cattle that are more spherical in shape, have lower SA:V, as spheres have the lowest SA:V. We hypothesize that more spherical shaped cattle have more digestive system capacity and will eat more fermentable feed before their appetite is satiated.
Project Methods
Because this is an interdisciplinary project combining robotic vision, computer science, nutrition, metabolism, genetics and genomics, a multitude of methods will be used.First, Time of Flight, infrared, and Red-Green-Blue cameras will be used to scan cattle. Robotic vision algorithms (such as Structure from Motion, Bundle Adjustment, etc.) will be used to create a three-dimensional point cloud of the animal, from which surface area and volume will be measured.Surface area is a key parameter to measure an animal's basal metabolic rate.We will use multivariate genomic predictions models (such as genomic BLUP) to create genotype-by-environment, genotype-by-management, and genotype-by-environment-by-management genetic predictions of yearling growth. We will also create genomic predictions of surface-area-to-volume ratio. We will compare surface-area-to-volume to these context-specific genomic predictions of growth and body size.We will use various machine learning and statistical measures to predict cattle traits from both the raw and converted three-dimensional camera data. Predicted traits will include feed intake, body condition score, hair shedding score, and foot structure.