AI & Computer Vision in Animal Breeding

Livestock science has become a data science. Milking robots, wearable sensors, cameras, genotyping chips and weather stations now generate streams of data on every animal, every day. “Big data” is usually described by the 5 Vs: Volume (how much), Velocity (how fast it arrives), Variety (tabular, images, sound, sequence), Veracity (how trustworthy) and Value (what decision it supports).

Not all big data is big in the same way. Tall data has many observations and relatively few variables, classical statistics works well and significance is easy to reach. Wide data is the “big p, small n” world: far more variables than animals, as when each animal carries hundreds of thousands of genetic markers. Here ordinary regression breaks down (collinearity, multiple testing, overfitting), so we turn to penalised / regularised regression and dimension-reduction techniques.

Animal breeding's workhorse is the mixed model, y = Xβ + Zu + e, separating fixed effects (e.g. herd-year-season) from random genetic merit (breeding values). Genomics made this model “wide”; multi-omics (transcriptomics, proteomics, microbiomics) is widening it further. AI does not replace this framework, it extends our ability to use messy, high-dimensional data within it.

Digital agriculture combines sensors, communication networks, drones (UAVs), robotic machinery, data analytics and AI to manage farms more precisely. Applied to animals, this is precision livestock farming (PLF): continuous, automated, individual-level monitoring.

The value chain is a pipeline: collect data (wearables, cameras, milk meters), store it (structured tables and unstructured images/sound), analyse it, visualise it (dashboards, alerts), and turn it into smart decisions. Among all sensing technologies, wearable accelerometers are the most common, and animal behaviour (activity, rumination, oestrus) is the most-collected phenotype, but cameras are rising fast because a single image can be extraordinarily informative.

A charming starting point is Francis Galton's 1906 observation that the average guess of 800 fair-goers estimated an ox's weight almost perfectly, the “wisdom of crowds.” Modern ensemble methods use the same idea: many weak models combined beat any single one.

Supervised learning predicts a labelled outcome (e.g. body weight) from inputs; unsupervised learning finds structure without labels (e.g. clustering animals). Three families dominate:

Artificial neural networks (ANNs) are flexible non-linear regressions inspired by the brain: inputs feed weighted hidden units, an activation function (sigmoid, tanh, ReLU) adds non-linearity, and back-propagation tunes the weights. ANNs easily overfit, so we use early stopping and weight decay (regularisation).

Support vector machines (SVMs) find the boundary with the widest margin between classes; the “kernel trick” lets them draw curved boundaries for data that aren't linearly separable. Decision trees split data with nested if-then rules; on their own they're unstable, so we combine many: bagging, random forests (each tree sees a random subset of predictors) and boosting (each model fixes the previous one's mistakes).

Computer vision works because an organism's status, health, growth, behaviour, leaves visually distinguishable cues. Light is electromagnetic radiation; cameras (CCD or CMOS sensors with a Bayer filter) turn it into numbers. The eye sees colour through three cone types (sensitive to red, green, blue), and digital images mimic this with RGB channels.

An image is simply a matrix of numbers: a black-and-white image is one binary matrix, a grayscale image one matrix of intensities, and a colour image three stacked matrices (R, G, B). Bit depth sets how many levels each pixel can take, 8-bit gives 2⁸ = 256 levels, 16-bit gives 65,536.

Processing an image means applying operators. A point operator changes each pixel on its own (e.g. brightness scaling). A linear filter combines each pixel with its neighbours using a small grid of weights called a kernel, this operation is convolution, and it can blur (mean filter), sharpen, or detect edges depending on the kernel. Hold onto this idea: it is the heart of the deep-learning module.

A perceptron is a single artificial neuron: it weights its inputs, sums them, and passes the result through an activation function (sigmoid, tanh, ReLU) that introduces non-linearity. Stack many neurons in many layers and you have a deep neural network.

Networks learn by minimising a loss function, mean squared error for continuous traits, cross-entropy for classification. Gradient descent nudges the weights downhill on the loss surface, and back-propagation efficiently computes those gradients layer by layer. Because deep models overfit readily, we apply dropout (randomly switching off connections during training) and early stopping (halting when validation performance stops improving).

For images we use convolutional neural networks (CNNs). Instead of hand-designing filters (Module 4), a CNN learns the kernels from data. Convolutional layers detect local patterns (edges → textures → parts → objects); pooling layers downsample to add robustness; a final dense layer turns the learned features into a prediction.

The phenotype equation is simple to state and hard to satisfy: Phenotype = Genetics + Environment. Genomics gave us cheap genotypes; the bottleneck moved to phenotypes. High-throughput phenotyping uses cameras, wearable sensors, robots and drones to measure many animals, often and cheaply. A good phenotype is accurate and precise, measured non-invasively, and scalable to thousands of animals at low cost.

Computer vision now delivers traits that were once subjective or labour-intensive:

• Body weight & growth, continuous monitoring of cattle weight and energy balance from images.
• Body condition score, from 3-D images at >90% accuracy, a key health, welfare and feed-efficiency indicator.
• Black soldier fly larvae, automated weight and sex prediction (YOLOv8 detection) for insect production.
• Ribeye area & carcass shape in live calves from 3-D body shape.
• Parasite detection, reading the ocular conjunctiva (FAMACHA) to flag parasitised sheep, removing human subjectivity.
• Identification & behaviour, recognising individuals and their feeding behaviour without tags.

Sensors also describe the environment. Enviromics uses farm- and region-level environmental descriptors to study genotype-by-environment interaction (G×E), why the best animals in one system are not the best in another, a central question for breeding programmes across the diverse environments of East Africa and Denmark.

Key takeaways

• AI and machine learning let us use big data for prediction, interpretation and causal inference in breeding.
• Computer vision enables non-invasive, high-throughput phenotyping at large scale.
• Neural networks, SVMs and tree ensembles are the core ML tools; CNNs are the workhorse for images.
• The payoff for breeding is more, better and cheaper phenotypes, and therefore faster, fairer genetic progress.

Glossary

Further reading & resources

• Gianola, D. & Rosa, G. J. M. (2015). One hundred years of statistical developments in animal breeding. Annual Review of Animal Biosciences 3:19–56.
• Kuhn, M. & Johnson, K. Applied Predictive Modeling. Springer.
• Free, certified courses on data and sustainable agriculture at the FAO e-learning Academy.