TLDR: A study developed and compared machine learning models (Random Forest, LSTM, SVR) to forecast one-month advanced weight gain in mob-based cattle. Using historical data, weather, and age factors from 108 Angus cattle, the Random Forest model consistently showed superior accuracy (96.70% with R² of 0.970) when both weather and age were included. The research highlights the importance of these factors for accurate predictions and introduces an automated data pre-processing tool to aid future research.
Farmers managing large livestock operations constantly seek ways to improve efficiency and reduce risks. One critical area is accurately predicting how much weight their cattle will gain. This information can help them fine-tune feeding plans, make smarter breeding decisions, and better prepare for market shifts or unpredictable weather. A recent study introduces a new method for forecasting mob-based cattle weight gain (MB-CWG), offering a significant step forward for the agricultural industry.
The research, titled “Mob-based cattle weight gain forecasting using ML models,” was conducted by Muhammad Riaz Hasib Hossain, Md Rafiqul Islam, S.R. McGrath, Md Zahidul Islam, and David Lamb. Their work focuses on predicting the one-month advanced weight gain of entire herds, rather than individual animals, using historical data.
To achieve this, the team explored three prominent machine learning models: Random Forest (RF), Support Vector Regression (SVR), and Long Short-Term Memory (LSTM). They gathered extensive data from the Charles Sturt University Farm, specifically from 108 Angus cattle between February and October 2022. This dataset included not only the cattle’s background information but also crucial environmental factors like daily rainfall and temperature, which are known to influence cattle growth.
A significant part of the study involved meticulous data preparation. The researchers tackled inconsistencies in daily weight measurements, as not all animals visited the weighing unit every day. They processed this raw data, removed outliers, and calculated monthly average weights. Weather data was also aggregated into monthly averages to align with the cattle weight records. To make the data more useful for the machine learning models, they engineered new features, such as the cattle’s current age in months, and rainfall and temperature data from the current month, one month prior, and two months prior.
To thoroughly evaluate the models, four different datasets were created, each with a unique combination of features. These ranged from a comprehensive dataset including both weather and age factors, to a baseline dataset that excluded both. This allowed the researchers to understand how different types of information impacted prediction accuracy.
The results were compelling. The Random Forest (RF) model consistently outperformed the SVR and LSTM models across all scenarios. When both weather and age factors were included, the RF model achieved remarkable accuracy, with an R² value of 0.970 and an overall accuracy of 96.70% in testing. Even when some factors were excluded, RF maintained strong predictive capabilities, demonstrating its robustness. The LSTM model also performed well, especially with a full set of features, highlighting its ability to capture patterns over time. However, it showed more sensitivity to missing data compared to RF. The SVR model provided acceptable results in simpler conditions but struggled more with complex or reduced feature sets.
A key takeaway from the study is the significant improvement in prediction accuracy when both weather conditions and the age of the cattle are considered. This underscores their critical role in understanding and forecasting cattle growth trends.
Beyond the predictive models, the researchers also developed an innovative automated data pre-processing tool. This software helps to standardize inconsistent raw data formats and generate benchmark datasets, making it easier for other researchers and farmers to prepare their data for similar analytical tasks. This tool is publicly available on GitHub, fostering further research and practical application.
While the study yielded promising results, the authors acknowledge some limitations, such as the data being sourced from a single farm, which might affect the generalizability of the findings. Future work aims to validate the models across diverse farms, incorporate more variables like feed intake and genetics, and move towards daily prediction intervals for more real-time insights. The potential for hybrid models combining the strengths of RF and LSTM is also being explored.
Also Read:
- Automated Pig Behavior Monitoring: A New Computer Vision System for Farm Welfare
- Enhancing Precipitation Forecasts with Adaptive Mixture of Experts and Multimodal Climate Data
This research offers valuable tools and insights for livestock managers, enabling them to make more informed decisions, optimize resource allocation, and ultimately enhance the productivity and economic viability of cattle farming operations. For more details, you can read the full paper here.
