Lameness is a significant concern in dairy cattle management, affecting both animal welfare and farm productivity. Despite efforts to mitigate its impact, traditional methods of lameness detection often overlook early signs, leading to delayed intervention and prolonged suffering for affected cows. This challenge underscores the need for more effective and proactive approaches to identifying and managing lameness. This study seeks to create an objective lameness detection methodology using sensor data from cattle limbs. Inertial Measurement Units (IMUs) were attached to cattle in Eastern Macedonia and Thrace, Greece, to record movement data. The collected data were preprocessed to address missing values, and features from both the time and frequency domains were extracted. Key gyroscope and accelerometer features were selected through Neighborhood Components Analysis. These features were then used to train LeastSquares Support Vector Machine (LS-SVM) and Multiclass Random Forest (MRF) models to classify lameness severity into healthy, mild, and severe categories, achieving an overall accuracy of more than 0.90 for both models. MRF has shown a better performance than LS-SVM.
Lameness in dairy cattle is a major welfare concern and economic burden, leading to decreased
milk production, increased treatment costs, and early culling [
Techniques such as human observation and pressure-sensitive walkways have historically been
used to detect lameness, though these methods are labor-intensive and costly [
Machine learning (ML) algorithms, particularly support vector machines (SVM) and random
forests (RF) are increasingly applied to analyze IMU data for lameness detection [
Data collection was carried out using the Blue Trident IMU device from (Figure 1). Four devices were placed on the limbs of each animal. The device features three types of sensors: an accelerometer, a gyroscope, and a magnetometer, all capturing data across three axes. In this study, only the data from the accelerometer and gyroscope were used, with a sampling rate of 500 Hz. Consequently, the dataset consists of six columns for sensor readings, one for timestamps, and another for the animal's class or state at the time of measurement, which was determined by expert visual observation and used as the ground truth. Each animal was observed freely for 10 minutes while the IMU sensors were active.
Feature extraction is widely used in the analysis of locomotion monitoring data. This approach
involves deriving new features from the raw data to improve the interpretation of the captured
information and to enable more precise analysis compared to using raw data alone. The extracted
features can be categorized into several groups, such as statistical (e.g., mean, standard deviation),
time domain (e.g., minimum and maximum values, signal energy), and frequency domain (e.g.,
dominant frequency, spectral area) [
Neighborhood Components Analysis (NCA) was employed to identify the most relevant features for lameness detection. NCA is a feature selection method that seeks to maximize the accuracy of classification by learning a linear transformation of the input data. It selects features that improve the performance of the model by emphasizing those that contribute most to class separation. In this study, a threshold of 0.5 was applied to select the top-performing features for further analysis.
In this study two machine learning algorithms have been employed for the detection of lameness in the dairy cattle; the Multiclass Random Forest (MRF) and the Least-Squares Support Vector Machine (LS-SVM).
A Multiclass Random Forest model extends the traditional Random Forest algorithm to address
classification tasks involving more than two classes. This model builds multiple decision trees, each
trained on a random subset of the dataset. Each tree independently classifies the input into one of
the available classes, and the final prediction is made based on the majority vote from all the trees
involved [
The Least-Squares Support Vector Machine (LS-SVM) is a modified version of the traditional
SVM, designed to solve classification tasks more efficiently by transforming the problem into a set
of linear equations [
For the evaluation of classification algorithms, the methods used were accuracy, precision and recall, as presented in Equations (1), (2) and (3). where TP, TN, FP and FN represent the True Positive, True Negative, False Positive and False Negative samples in the confusion matrix, respectively.
The accuracy estimation method is one of the most common, but it suffers due to its sensitivity
to imbalanced data. Another issue is that two classification algorithms may have the same accuracy
but different performance in terms of the correctness of the decisions they make [
The features selection as it was mentioned, was based on the NCA algorithm, as is shown in Figure 2.
The selected variables, as they have been decided by NCA algorithm were the mean, the standard deviation and the maximum value across the y axis of the accelerometer, the zero crossing across the x axis of the gyroscope and the signal area, across the z axis of the accelerometer. The mean, standard deviation, and max in the y-axis of the accelerometer measure changes in vertical limb movement, providing insights into stride length and height. The zero crossing in the x-axis of the gyroscope may reflect shifts in limb orientation and balance, essential for detecting irregular gait patterns. The signal area in the z-axis of the accelerometer tracks overall limb movement energy, which is critical for identifying reduced or exaggerated movement, both of which can indicate lameness. These features together offer a comprehensive view of movement abnormalities.
Hyperparameter selection is a crucial step in the machine learning pipeline, affecting model performance, training time, and generalization. Techniques like grid search are commonly used for hyperparameter tuning. The grid search method systematically tests combinations of hyperparameters, using cross-validation to ensure results aren’t dependent on data splits. The best hyperparameter set is determined based on performance metrics from a validation set. The list of the possible hyperparameters for each ML algorithm and their values range is given in Table 2.
The selected hyperparameters after the grid search method were the following: for the LS-SVM model, a C of 10, a Gamma of 0.05 and an Epsilon of 0.01 and for the MRF model a number of trees of 100, a Gini criterion, a maximum depth of 10 and a minimum samples to split a node of 2.
The performance of the classification models was evaluated using a confusion matrix, which provides insights into the model's accuracy and ability to distinguish between the different lameness LS-SVM
MRF
C Gamma (sigma)
Epsilon Number of Trees
Criterion Maximum Depth
Minimum Samples to Split a
Node
Short Description trade-off between model complexity
and training error. width of the Gaussian kernel
tolerance for errors number of decision trees in the model function to measure quality of a split longest path from the root to a leaf
node the minimum number of data samples required for a node to be split into child nodes
0.5, 1, 10, 20, 50 0.01, 0.05, 0.1, 0.5 0.1, 0.05, 0.01, 0.001 50, 100, 200, 500
Gini, Entropy None, 10,20, 50 2, 5, 10, 20 status of the cattle. The data split followed a 70-30% scheme, for training and test, respectively, with the test set containing 168 values per feature for each class. The following confusion matrices of Table 3 and performance metrics summarize the results.
The results from both confusion matrices demonstrate a strong overall ability of the models to detect varying levels of lameness in cattle, with notable differences in performance between the LSSVM and Random Forest classifiers. Comparatively, Multiclass Random Forest outperforms LS-SVM in both precision and recall across each lameness category. For precision, Random Forest achieved slightly higher values across all categories (0.97, 0.97, and 0.92 for healthy, mild, and severe lameness, respectively) compared to LS-SVM (0.96, 0.92, and 0.89). In terms of recall, Random Forest also displays superior results (0.96, 0.90, and 0.99) in comparison to LS-SVM (0.90, 0.87, and 0.98), underscoring its enhanced effectiveness in identifying actual cases across all categories, especially in the "healthy" and "mild lameness" classes.
Overall, Random Forest demonstrates a more balanced and reliable classification performance in
this scenario, effectively capturing a greater number of true instances with fewer false alarms. This
improvement can likely be attributed to Random Forest’s ability to capture complex, non-linear
relationships within the data, a characteristic that is often limited in LS-SVM models due to their
reliance on a fixed kernel function [
• • • •
The use of Inertial Measurement Units (IMUs) combined with machine learning models (MRF and LS-SVM) was effective in detecting lameness in cattle, achieving over 90% accuracy. MRF outperformed LS-SVM across all categories. This is likely due to MRF's ability to handle complex, non-linear relationships.
Neighborhood Components Analysis (NCA) successfully identified key gyroscope and accelerometer features, enhancing model performance.
Overall, sensor-based detection offers an objective and efficient alternative to traditional visual assessments, with the potential for early intervention and improved animal welfare.
This research was funded by the Action «Rural Development Programme of Greece 2014-2020» under the call Measure 16 “Co-operation”, Sub-Measure 16.1 – 16.2» that is co-funded by the European Regional Development Fund and Region of Eastern Macedonia and Thrace, project «COWLAM - Lameness identification system at milk- producing cow units in the Region of Eastern Macedonia and Thrace» (Project code: Μ16ΣΥΝ2-00031).