INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue IX September 2025
Page 4748
www.rsisinternational.org
Machine Learning Image Classification model to Identify Cattle in
Kenya
Benard Onyango1, Obadiah Musau2, Kennedy Ondimu3
Technical university of Mombasa, Kenya
DOI: https://dx.doi.org/10.51244/IJRSI.2025.1208004131
Received: 12 Sep 2024; Accepted: 19 Sep 2025; Published: 25 October 2025
ABSTRACT
Classifying cattle using muzzle images is an emerging technology in livestock management for recognition
and classification. This study used Convolutional Neural Networks (CNN) algorithm to uniquely identify
cattle by using their muzzle patterns which are unique to every single cattle. The study used a dataset of 4,923
muzzle images of different cattle breeds which were pre-processed to improve the dataset’s performance and
reduce overfitting. The Convolution Neural Network used several convolutional layers to capture muzzle
patterns, pooling and dense layers to differentiate breeds. Adam optimizer and categorical cross-entropy loss
were employed for model training. The results revealed high accuracy, verifying muzzle images as an
effective biometric method for cattle identification. Transfer learning via pre-trained models positively
impacted model accuracy and generalization. The technology can be integrated into livestock management
and breeding programs, as well as agricultural and farming systems.
Keywords: Biometric Identification, Muzzle Images, Convolutional Neural Networks, Keras Framework.
INTRODUCTION
Cattle rustling refers to the unlawful theft of cattle, often involving violent and destructive methods. This
practice not only result in loss of livestock but also leads to so many deaths, destruction of property, and
displacement of families. It poses a serious threat to peace and security along transnational borders,
undermining governance and stability. The root causes of cattle rustling are deeply rooted in social, cultural,
economic, political, and historical disputes. According to a 2024 report by the Kenya Institute of Public Policy
Research and Analysis (KIPPRA), Kenya’s Parliamentary Committee on Administration and National
Security highlights several drivers of banditry. These include competition for scarce resources such as water
and pasture, widespread access to firearms, cycles of retaliatory attacks, political mobilization, land and
border disputes, as well as neglect and underdevelopment in vulnerable regions. According to the report, areas
most affected by cattle rustling include Turkana, Baringo, West Pokot, Elgeyo Marakwet, Samburu, and
Laikipia. Data from the National Crime Research Centre (2020) revealed alarming rates of cattle rustling, with
Laikipia at 82%, Turkana at 70%, Samburu at 46%, and Elgeyo Marakwet at 45.5%, compared to a national
average of 37.2%.
Several Machine learning studies aimed at solving animal identification problems have been conducted. A
study by Hang and Basanta (2019) on developing a model for identifying bird species from their images. The
model processed bird images as input and classified them into specific species using convolutional neural
networks (CNNs) alongside other machine learning techniques.
While the study successfully demonstrated the effectiveness of CNNs for image-based bird classification, its
approach encountered limitations when distinguishing between highly similar or nearly identical objects.
R.W. Bello et al. (2020) developed a machine-learning model for identifying Cattle using their unique body
patterns. The model accepts image inputs and output classifications by categorizing the cattle into specific
groups using convolutional neural networks. While the study demonstrated the power of machine learning
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue IX September 2025
Page 4749
www.rsisinternational.org
models for classifying cattle using their unique body patterns, its approach suffered limitations in scenarios
where objects are highly similar or nearly identical.
In the study by Alharbi et al. (2019), the researchers aimed to classify animals as either predators or pests
based on their biometric features. The model accepts biometric inputs and categorizes inputs accordingly,
utilizing machine learning techniques such as Support Vector Machines (SVM) and Multi-Layer Perceptron
(MLP). However, this approach faced limitations in cases where objects were highly similar or nearly
identical, and situations where biometric inputs can be open to manipulation.
According to a study by Crall et al. (2013), a machine-learning model was developed to classify animal
species using a labelled dataset, with coat colour as the primary input. The model applied the hot-porter
algorithm. While this model proved effective in species identification based on color patterns, it was however
unable to accurately classify animal species in cases where coat color was altered by scratches from
vegetation or injuries from fights.
In a study by Parham and Stewart (2016) to assist in tracking animals in the wild, the researcher developed a
model that identifies zebras from images. The model employed Naive Bayes nearest neighbor algorithm.
However, the model faced limitations in cases where the zebra's coat color has been altered by scratches from
vegetation, injuries sustained during fights, or deliberate tampering.
Many existing studies on the use of machine learning algorithms demonstrate the power of machine learning
models for classifying cattle using their unique body patterns. However, most of these studies highlight
challenges in accurately classifying objects that are highly similar or nearly identical especially where the
researcher relies heavily on biometric features that are open to manipulation. Again, most studies have shown
limited or no use of muzzle images for cattle identification despite their uniqueness thus representing a
significant research gap.
Studies have also shown that external factors such as vegetation scratches, injuries, or deliberate tampering
can modify animal’s distinguishing features, further limiting the reliability of these methods. These gaps
underscore the need for alternative identification techniques that enhance classification accuracy, minimize
susceptibility to manipulation, and remain effective under varying environmental conditions.
METHODS
Dataset Collection
A dataset of cattle muzzle images was assembled, including various breeds to ensure diversity. To maintain
uniformity, all images went through pre-processing, standardizing resolution and aspect ratio. Additionally,
data augmentation techniques including image rotation, flipping, and contrast adjustments were implemented
to strengthen the model’s robustness. The dataset used in this study was sourced from Kaggle.com and it is
made up of 4,923 muzzle images of 268 beef cattle. Images were captured using a mirrorless digital camera
equipped with a 70–300 mm F4-5.6 lens. The dataset spans three cattle breeds: Angus, Angus x Hereford, and
Continental x British cross. All images were obtained under a Creative Commons 4.0 license and were taken
at a right angle from a distance of one meter, minimizing disturbance to the animals. Only clear, properly
cropped muzzle images were included in the final dataset.
Data Pre-processing
Data pre-processing techniques were applied on the raw data to transform it into a clean, usable format to
enhance model performance, it included the following:
Resizing: Images were resized to 244 by 244 pixels using Keras's tf.image.resize() function to ensure
compatibility with CNN input dimensions.
Normalization and Augmentation: Image pre-processing involved normalization, where pixel values were
scaled to a consistent range to stabilize and accelerate training, and augmentation, where transformations
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue IX September 2025
Page 4750
www.rsisinternational.org
such as rotation and flipping were applied to increase data variability and enhance the model’s ability to
generalize.
Model Development
A CNN architecture using Keras was used. The model consists of convolutional layers for feature extraction,
pooling layers for dimension reduction, and fully connected layers for classification. Transfer learning with
pre-trained models was utilized to improve model performance.
Figure 1: Basic Architecture of a Convolutional Neural Network (Adapted from Research Gate, 2025).
Training and Validation
Dataset was split into training, validation, and testing subsets. The developed model was trained using
categorical cross-entropy loss and the Adam optimizer. Early stopping and learning rate scheduling was
employed to prevent overfitting. 80% of the images were used for training, while the remaining 20% was split
equally into validation and test sets. To split a dataset in Keras (or TensorFlow) train_test_split script was
used from the sklearn. model selection module. 10% of the images were utilized for model validation during
training, and the final 10% were employed as an independent test set to evaluate the model’s performance in
real-world situations
Deployment
The trained model was then converted to TFLite format so that it could be used on the mobile device. An
Android application was created in Java and Kotlin and integrated with the TensorFlow Lite model for real-
time classification. Heroku cloud platform was used to provide computing services that enable users to build,
deploy, manage, and scale applications and services via the internet.
Figure 2: Model Deployment Diagram
Android Mobile
Flash API
Horeku Cloud
Flask Server
Download Load Images
Upload
Download
Classifier
Model
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue IX September 2025
Page 4751
www.rsisinternational.org
Model Evaluation
Model evaluation involves using different metrics to check a model's performance. This helps to identify its
strengths and weaknesses. This step is important for figuring out how effective a model is during the early
research stages and for ongoing monitoring. To evaluate the performance and validity of the classifier model,
800-labelled images were used for training. The output was then organized into a table, and the results were
examined to evaluate its classification performance. The evaluation matrix compared the actual target values
with the model's predictions. The model showed strong performance in accurately identifying the images, as
shown below.
Figure 3: Model Evaluation Diagram
The model was evaluated based on accuracy, precision, recall, and F1-score. A comparative analysis was
conducted with existing identification methods. The application’s performance was the assessed in real-world
situations, including changing lighting conditions and image quality. The model achieved high accuracy and
showed strong performance in classifying muzzle images. The study used detailed performance metrics and
visual results, including graphs and confusion matrices, to confirm its effectiveness. The model's prototype
was tested through a series of experiments to measure its accuracy in identifying individual cattle. The test
dataset represented real-world conditions and was used to evaluate the model’s performance on data it had
never seen before. The model achieved a test accuracy of 98%, with 98% of test images correctly labelled.
This highlights its strength and reliability in identifying cattle based on their muzzle patterns.
Figure 4: Diagram of Confusion Matrix
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue IX September 2025
Page 4752
www.rsisinternational.org
Figure 5: Cross Entropy Loss Diagram
RESULTS
The prototype model achieved a test accuracy of 98 % in identifying individual cattle from muzzle images,
demonstrating strong robustness and effectiveness. The use of data augmentation, normalization, and transfer
learning likely contributed to this high performance by enabling the model to generalize effectively across
unseen images.
DISCUSSION
This accuracy aligns with findings in related research where convolutional neural networks (CNNs) have
shown superior performance in visual classification tasks. The integration of the model into an Android
platform enhances its practicality and accessibility, particularly in regions affected by cattle rustling, offering
significant value for livestock management and security. The results further suggest that muzzle patterns
provide a highly reliable biometric trait for cattle identification, comparable to human fingerprint recognition.
However, further testing under real-world farm and market conditions is necessary to validate the system’s
reliability beyond controlled datasets. Beyond technical performance, the deployment of such systems in rural
communities carries important ethical, economic, and practical implications. Ethically, questions of data
privacy and ownership must be addressed, as farmers may be concerned about how biometric records of their
cattle are stored, who controls access, and whether such data could be misused for surveillance or commercial
exploitation. Economically, while the technology promises to reduce financial losses from cattle theft and
disputes over ownership, the costs of smartphones, internet access, and system maintenance may be
prohibitive for smallholder farmers unless subsidized or supported through cooperative models. Practically,
farmer adoption hinges on the usability of the application, the availability of training, and the degree of trust
in digital tools, especially in communities where technological literacy may be limited. Ensuring equitable
access, transparent governance of data, and affordable implementation strategies will therefore be critical to
realizing the potential of this system for improving livestock security and management in rural settings.
CONCLUSION
This study successfully developed a cattle identification model using muzzle images and convolutional neural
networks (CNNs), effectively addressing the challenge of reliably distinguishing individual animals in regions
prone to cattle rustling. The system achieved high prediction accuracy, making it practical even in situations
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue IX September 2025
Page 4753
www.rsisinternational.org
where other models fail to classify animals accurately due to altered features such as coat color changes from
vegetation scratches or injuries. The findings affirm CNNs as a powerful tool for biometric-based cattle
identification.
RECOMMENDATIONS
Future research should focus on expanding the dataset to capture greater variability, improving error detection
mechanisms to minimize misclassification, and exploring additional biometric traits to further enhance the
model’s reliability and applicability in diverse real-world environments.
REFERENCES
1. KIPPRA. (2024, January 8). Banditry and lawlessness in arid and semi-arid lands of Kenya: Which
way out? Retrieved from https://kippra.or.ke/banditry-and-lawlessness-in-arid-and-semi-arid-lands-
of-kenya-which-way-out/
2. Alsaadi, E. M. T. A., & El Abbadi, N. K. (2020). An Automated Classification of Mammals and
Reptiles Animal Classes Using Deep Learning. Iraqi Journal of Science, 2361–2370.
3. Bello, R., & Abubakar, S. (2019). Development of a software package for cattle identification in
Nigeria. Journal of Applied Sciences and Environmental Management, 23(10), 1825–1828.
4. Crall, J. P., Stewart, C. V., Berger-Wolf, T. Y., Rubenstein, D. I., & Sundaresan, S. R. (2013).
Hotspotter—Patterned species instance recognition, 230–237.
5. Huang, Y.-P., & Basanta, H. (2019). Bird Image Retrieval and Recognition Using a Deep Learning
Platform. IEEE Access, 7, 66980–66989. https://doi.org/10.1109/ACCESS.2019.2918274
6. Parham, J., & Stewart, C. (2016). Detecting plains and Grevy’s Zebras in the real world. 1–9.