Machine Learning for Precision Broiler Weight Estimation

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Nov 26 2023

In a paper published in the journal Agriculture, researchers introduced an innovative method utilizing machine learning algorithms for non-invasive broiler weight estimation in large-scale production. Employing Gaussian mixture models, Isolation Forest, and the Ordering Points To Identify the Clustering Structure (OPTICS) algorithm in a two-stage clustering process achieved accurate predictions of individual broiler weights.

Their approach combined polynomial fitting and gray models for real-time predictions, continuously refining parameters based on accuracy feedback. Evaluation using the Symmetric Mean Absolute Percentage Error (SMAPE) metric on 111 datasets revealed considerable accuracy, presenting a promising solution for cost-effective and precise broiler weight monitoring in large-scale farming setups.

Background

Broiler weight monitoring is crucial for large-scale farming; companies seek high qualification rates to maximize profits, while farmers aim for incentives by meeting weight standards. Traditional methods involve growth models, image processing, audio analysis, and sensor-based predictions, each facing limitations in accuracy and scalability.

Previous studies explored growth curve models for broilers, image-based weight estimation, audio analysis for weight correlation, and direct weight prediction using sensor data. Challenges include model specificity, noise sensitivity in image and audio analysis, and limited evaluation on smaller datasets, necessitating more robust, scalable solutions for broiler weight.

Automated Broiler Weight Monitoring Methodology

The primary aim of this research was to devise an automated method leveraging machine learning algorithms for precise broiler weight estimation using weight sensors as the singular data source. Researchers integrated diverse machine learning clustering algorithms into three stages to derive the average weight of an individual chicken from noisy and outlier-infused data.

Machine learning encompasses algorithms that discern patterns and features from substantial datasets, categorized into supervised, unsupervised, and reinforcement learning. In this context, unsupervised learning methods played a pivotal role in the data analysis and mining processes, offering avenues for comprehensive insights.

The methodology commenced by replacing evident outliers and invalid data—such as negatives or 0s—using a multiple interpolation machine learning algorithm in the initial phase. Subsequent stages segmented the entire dataset into distinct monitoring units at fixed intervals, enabling the calculation of mean body weights through multiple clustering and anomaly detection algorithms. The third phase focused on predicting mean weights for subsequent intervals, building upon the compiled list of individual weights.

The method encompassed 36 farms housing Cobb 500 broilers for data collection, with weight measurements facilitated by KOKOFARM electronic scales. The breeding cycle spanned 28 to 31 days, encompassing up to 30,000 broilers per batch. Raw data, collected every second and aggregated into Comma-Separated Values (CSV) files every minute, underwent rigorous preprocessing using Python 3.9 and critical libraries such as sci-kit-learn and SciPy.

Researchers conducted experiments on a high-performance server with an Intel Xeon Gold Central Processing Unit (CPU), an NVIDIA A100 80 GB Peripheral Component Interconnect Express (PCIe), 251 GB  Random Access Memory (RAM), and an Ubuntu Linux operating system. The preprocessing stage involved critical-value-based noise reduction and a MissForest-based approach to null-filling. These methods aimed to eliminate noise and replace it with null values, employing imputation techniques to fill the missing data. It prepared the dataset for further analysis by addressing noise and outliers within the time-series data.

Segmentation and Gaussian Mixture Model (GMM)-based data modeling were pivotal in the subsequent stages. GMM facilitated data segmentation, breaking it into Gaussian distribution components to discern patterns within the dataset. Subsequently, isolation forest-based outlier sieving, and OPTICS-based multiple clustering processes refined the dataset, enabling the determination of average broiler body weights through algorithmic model analysis.

The adaptive forecasting phase combined multinomial regression with gray models to forecast broiler weights. Polynomial regression, mainly the Gompertz function, was employed for fitting mathematical curves to the data, while gray models compensated for the late starting point of the polynomial-appropriate prediction models, offering short-term forecasting with limited data points. This integration facilitated automatic real-time monitoring and prediction of average broiler body weight, representing a robust methodology for industry application.

This comprehensive methodology, amalgamating diverse machine learning techniques, sensor-based data collection, and predictive modeling, holds promise for precise and automated broiler weight monitoring, critical in large-scale farming setups.

Broiler Weight Estimation: Data Journey

This comprehensive study navigates through the processes of data preprocessing, analysis, and estimation of broiler weight. Researchers managed extensive CSV files by filtering timestamps and broiler weight values. The team implemented noise reduction and MissForest techniques to tackle missing and noisy data, ensuring a more robust and accurate analysis. The analysis phase navigated complexities in handling missing broiler weight values, emphasizing challenges posed by 0s and diverse noise.

Utilizing MissForest and validating populated data showcased the algorithm's efficiency, highlighting the necessity for data padding and the effectiveness of the imputation technique. Additionally, the approach highlighted difficulties in accurately weighing numerous chickens, revealing insights into the proposed algorithm's effectiveness despite human-induced challenges during transportation. Validating across 111 datasets showcased practical feasibility, showing errors below 0.2, emphasizing its potential real-world applications in farming contexts.

Conclusion

To sum up, this study revolutionizes broiler weight management by leveraging machine learning algorithms with simple electronic scales. Its innovative approach simulates natural weighing processes, ensuring accurate predictions without complex technology. While future research could expand this system's capabilities by integrating additional farm factors, this methodology promises efficient decision-making in broiler farming. It holds potential for broader application in livestock weight estimation.