Prediction of Lameness in Cows in the Cattle Farming

Introduction

A significant problem the cattle farming is trying to solve, is the identification of lameness in cattle. The aim of the research is to present a unique computational analysis for lameness prediction based on machine learning and deep learning methods. Lameness is one of the most critical issues with respect to the health of farm animals. Problems created in production derived from lameness are catastrophic for the farmer. It is important to detect the lameness in time and with reliability [1], in order to reduce the cost but also to ensure the health of the animal. It has been observed that animals which suffer from lameness presents various symptoms, such as difficulty in walking [3], they lie down more compared to healthy animals [3], [5] and [7], stand less [3] and graze less [6].

Aim

The aim of the research project was to use the capability of an innovative real- time location sensor and combined accelerometer to measure potential differ- ences in behaviour over time for lame and non-lame cows within a structured environment using computational methods.

Data

The data is collected from a commercial dairy farm in Essex, UK. The data has data time attribute capturing individual cow’s position. The cows are fitted with a wireless sensors to track the spatial location. The sensors were mounted on cows using a neck collar that incorporates a counterweight to keep the sensor in a stable position at the top of the neck [4]. In the below table, the description of the data is mentioned. Columns Description Cow-id Unique identifier of the cow Day Day representing the count of days Time Represents the specific time when the sensor readings are captured X-coordinate X-coordinate position of the sensor in the barn Y-coordinate Y-coordinate position of the sensor in the barn.

Approach

Initially the research project requires a study of similar approaches that has been done in the past focusing a particular question of whether there are any behavioural differences between lame and non-lame cows. It further requires to study the technological algorithms and advancements that has been imple- mented and used in this field of study.

Secondly, understanding how illness may affect animal movement behaviour, space-use, and interactions with the local landscape could provide useful insights and indicators for monitoring and managing a range of animal species. Inves- tigation of other diseases that could affect feeding and lying behaviour also becomes necessary and in this context, the use of cow-mounted accelerometers to measure cow behaviour would serve well to identify behaviours in our dataset [2] A time series analysis to identify specific behaviours of lameness cows so that to identify initial stage of symptoms a cow can experience when becoming lame. This could be helpful for farmers in identifying specific cows earlier and can provide appropriate treatment to avoid lameness. Analyzing the data to generate heat maps and other visualizations over time for individual cows and either up sampling or down sampling the date- time data would help gain solid understanding to get effective insights in order to distinguish the lame and non-lame cows. Once gained a effective understanding of the data set, the research work would focus on implementing a machine learning model (Boosting) and a deep learning model (CNN over heat maps) to classify lame cows from non-lame cows.

Methods

The research work is keen on implementing a boosting algorithm on the tabular dataset for classification. The term ‘Boosting’ refers to a family of algorithms which converts weak learner to strong learners. It is an ensemble algorithm for improving the model. The idea of boosting is to train weak learners sequentially to correct its previous errors where the goal is to ”teach” a model F F to predict.

At each stage m m, 1 ≤ m ≤M1 ≤ m ≤ M, of gradient boosting, it may be assumed that there is some imperfect model FmFm (at the outset, a very weak model that just predicts the mean y in the training set could be used). The gradient boosting algorithm improves on FmFm by constructing a new model that adds an estimator h to provide a better model (2): Fm+1(x) = Fm(x) + h(x)Fm+1(x) = Fm(x) + h(x). (2) To find h, the gradient boosting solution starts with the observation that a perfect h would imply (3): h(x) = y − Fm(x) (3)

Therefore, gradient boosting will fit h to the residual y − Fm(x)y − Fm(x). As in other boosting variants, each Fm+1Fm+1 attempts to correct the error of its predecessor Fm. Feature engineering will be carried out in adding more variables like the time duration the cows spent on feeding, resting, milking and other activities. The parameters involving in the boosting algorithm will be tuned to find its global minimum.

Based on the cow’s behaviour and data analysis, heat maps will be generated capturing the amount of time cows spent on the barn in different areas as shown in the figure above, which will be passed as an input to a Convolutional neural network to classify the lame and non-lame cows based on their behaviours cap- tured at various times of the day or week inside the barn. However, whether to choose the timeframe for day or week depends on research and data exploration of the data. The algorithm will be evaluated on classification accuracy and if it deals with unbalanced dataset it will be evaluated based on precision, recall and f1 score.