Exploring a siamese neural network as a novel individual cow identification technique for daily monitoring free water intake.

Abstract

A dairy cow needs enough water to produce good quality milk. Over the last decades, the amount of milk produced per cow has increased a lot. This means that the amount of water needed per cow has increased as well. It is unclear whether cows nowadays can drink enough water at the dairy farm, because individual water intake is currently not being monitored. An easy way of measuring the amount of water a cow drinks needs to be developed and artificial intelligence can be used to achieve just that. Using one camera per drinking trough, we want to survey the drinking behaviour. A cow can be detected in an image using algorithms like YOLOv5. We then need to detect whether it is drinking. By locating its head using an algorithm called DeepLabCut, we can predict a cow is drinking when its head is close to the drinking trough. Next, the cow needs to be identified in order to track its water consumption. We developed a model with the aim of doing this, called a siamese neural network. It can learn to 1) distinguish images of different cows and at the same time 2) learn to recognise different images of the same cow. These steps combined can let us analyse a video and track when a cow was drinking and which cow it was. The siamese neural network seems to be promising but needs to be improved by removing the background from an image, such that only the cow is visible. When this is achieved, we can easily measure the time a cow spends drinking. Combined with data from water flow meters, we can monitor how much water a cow drinks.