In the context of cattle farming, understanding the feeding behavior of animals is essential to ensure their welfare and maximize productivity. However, monitoring and interpret- ing the acoustic signals associated with grazing, particularly the sound events related to grass intake, pose a significant challenge. This study proposes an innovative method based on 1D convolutional neural networks to automatically classify such sound events during grazing. The approach was developed using a balanced dataset composed of 322 prehension samples and 1000 non-prehension samples, extracted from audio recordings of grazing cattle in real conditions. The results obtained show a high accuracy of 100% during the testing and validation phases of the model. However, there is concern about overfitting of the model due to the limited size of the dataset used. Consequently, future expansion of the dataset is suggested by collecting a larger and more diverse number of audio recordings to improve the generalization and robustness of the model in real cattle farming contexts.
Recent advances in automated monitoring systems have
opened up new possibilities for precision breeding,
including the radio signal power [
The feeding behaviour of grazing animals is an aspect whose knowledge represents an added value in understanding the methods of direct use of the herbage by the animals and the consequent implications in terms of pasture response and production (milk, growth) and qualitative characteristics of products (nutritional, nutraceutical and sensorial aspects). Parameters such as the herbage intake, the number of prehensions, the rumination activity are characterized by objective detection dificulties, taking into account the complete freedom of behaviour of grazing animals, especially in extensive conditions. The traditional method consists of direct observations of the herbage prehension behaviour but also, with greater executive dificulty, of the number of bites made per minute, the time dedicated to eating and the time dedicated to ruminating.
To overcome the executive dificulties that characterize this method, several techniques for the automatic detection of some grazing behavioural parameters have been developed.
Among the first automatic detection systems were those
based on the 24h-recording of chewing movements by
pressure sensors nose bands [
In more recent years the use of accelerometers is probably
the most adopted precision farming practice, being able to
detect changes in the position of the neck, head and mouth
[
This allows detecting activities such as prehension,
chewing, rumination, searching, lying, but also urination and
defecation if accelerometers are attached to various points
along the back of the spine or in the tail as illustrated by
the studies of Marsden and Shorten et al. [
However, [
In [
Determining the feeding behaviors of dairy cows is crucial for assessing their productivity and health status. Various research contributes to progress in livestock management, providing a basis for the development of more precise and reliable decision support tools.
Audio systems represent an additional mean available
for the detection of grazing behavior. The recording of
sounds, suitably codified through direct observations or
video recordings, can be an efective mean of characterizing
eating behavior, as the activities of bite prehension, chewing,
rumination can be recognized on the basis of the sound
frequencies recorded over the day at pasture [
The studies conducted by Chelotti et al. presented an
analysis system called Real-Time Chews-Boluses
Recognition Algorithm (CBRTA) [
Milone et al.’s work [
In the research conducted by Li et al. [
In this work, the authors propose a study aimed at the identification of a suitable method to classify and quantify the sounds detected during grazing in an extensive system, representing the key aspect for its use in real farm conditions. The objective of the research is to quantify the number of herbage prehensions made by 20 h grazing cows through the development of an audio classifier.
In Subsection 1.1, we will focus on the development and implementation of the proposed methodology for audio event classification in the context of cattle farming. In particular, we present a detailed analysis of the application context and the provided data, as well as describe the development process of the 1D Convolutional Neural Network (CNN) and the creation of the dataset for model training and testing.
In Section 2, we will describe the results obtained following the training and testing phase of the neural network.
Finally, there is the concluding section that gathers the obtained results, study limitations, and future work.
The choice to use a 1D convolutional neural network (CNN
1D), the same as used in the study in [
The data analysis phase included evaluating the quality,
variability, and distribution of the provided audio samples.
A detailed analysis was conducted to identify the presence
of background noise, variation in signal quality among
different samples, and consistency in labeling. This thorough
examination allowed for establishing a solid foundation for
creating a balanced and representative dataset, necessary
for the efective training of the network [
The approach chosen for audio classification in the context of cattle farming not only addresses the immediate needs of the project but also lays the groundwork for a scalable and adaptable platform for future research and development needs.
To facilitate the training process, a dedicated Python script was developed for creating a structured dataset. This script eficiently organized audio samples in preparation for subsequent training and validation phases. The dataset creation phase was crucial to ensure the efectiveness and accuracy of the audio classification network. It required careful planning and execution, particularly in terms of data preparation and verification.
Initially, the Python script developed for this phase began by reading the CSV file containing the audio data and their corresponding labels. This file served as the backbone of the dataset, providing essential information needed for the subsequent classification process.
A fundamental aspect was verifying the semantic correctness of the timestamps associated with each audio segment. This step was essential to ensure that each label precisely matched the desired audio segment.
To maximize process eficiency and minimize errors, an automated procedure was introduced to pre-examine each label. This verification mechanism was tasked with identifying and flagging any inconsistencies or errors in the labels, such as invalid timestamps or incorrect formats.
In case an error is detected, the procedure specifically reports the row in the CSV file that requires revision. This approach aims to facilitate prompt corrective action, reducing the risk of introducing incorrect data into the final dataset.
Once the dataset has been verified and cleaned, the script proceeds to calculate the maximum and average lengths of the audio segments. This step is crucial for understanding the variation in audio segment sizes and for setting appropriate parameters during the network training phase.
Knowledge of the maximum and average lengths of the audio segments plays a key role in configuring the convolutional network. It influences crucial aspects such as input size, layer structure, and output management, ensuring that the network is optimized to efectively handle the variety of audio segments in the dataset.
For the creation of the audio classification dataset focused on distinguishing between prehension and non-prehension, specific procedures were followed to ensure the balance and quality of the dataset.
Given the limited nature of available samples, a total of 322 samples were obtained for the prehension class. Considering the abundance of data for the non-prehension class, it was chosen to limit this category to 1000 segments, randomly selected to avoid bias in the dataset.
Due to the discrepancy in the number of samples between the two classes, a strategy was adopted to balance the dataset. This is essential to prevent overfitting or bias towards the more represented class. Therefore, techniques such as downsampling of the most represented class (nonprehension) were employed.
The dataset, in turn, was divided into training, validation, and test sets. This division allows for efective model training, validation to prevent overfitting, and testing its performance on unseen data.
These stages represent a systematic and balanced approach to creating a dataset for an audio classification project, ensuring that the data is representative, balanced, and of high quality.
For the training phase of the machine learning model, we follow the following approach, keeping in mind the specific context of the project, where we faced the challenge of an imbalanced dataset and a limited number of samples.
Training is conducted for 100 epochs. It was noticed that already in few epochs (Fig. 1), the network had achieved high accuracy. In fact, it is observed that the network reaches 100% accuracy in just 20 epochs, indicating a possible overfitting due to the limited number of samples.
Despite the high accuracy, it is important to emphasize that this result may not be indicative of the actual performance of the model on unseen data, due to the limited training dataset.
The final phase included thorough testing and validation of the network, using confusion matrices (see Fig. 2) to assess performance. Despite achieving 100% accuracy, this result suggests the possibility of overfitting.
The neural network model demonstrated high accuracy, reaching 100% in both testing and validation phases. However, this exceptionally high result raises concerns regarding the phenomenon of overfitting, where a model overly adapts to the training data, thereby losing the ability to generalize to new data.
Despite the promising performance of the classifier, the limited availability of data makes it dificult to determine with certainty whether the network is actually capable of efectively functioning in real-world scenarios or if it is sufering from overfitting. Therefore, acquiring a larger and more diverse dataset is recommended for further testing and to improve the model’s generalization. This step is crucial to confirm the classifier’s reliability in the real context of cattle farming.
Despite the limitations, the results obtained provide a preliminary indication of the problem’s feasibility. To further improve the model, expanding the dataset, utilizing data augmentation techniques, or exploring more complex models and regularization techniques to reduce overfitting are considered.
It is essential to increase the size and variety of the dataset to ensure a more robust evaluation of the model.
Testing various neural network architectures and training
parameters to find the optimal configuration is suggested
[
In conclusion, the training phase has provided significant insights into the feasibility of the project, despite the limitations imposed by the dataset.
Future developments include a significant increase in the dataset size and also the classes to be identified.
This study was carried out within the Agritech National Research Center and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) – MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.4 – D.D. 1032 17/06/2022, CN00000022). This manuscript reflects only the authors’ views and opinions, neither the European Union nor the European Commission can be considered responsible for them.