Challenges
The customer worked on implementing a specialized livestock management system (LMS, or cattle management system) to automate handling of their animals. One of the features required by the customer is the ability to estimate whether animals were suitable for milk production, or they should be removed from the herd. Wrong decisions led to selection of inappropriate cows who got sick or died before their first lactation. Through that, the company was losing time and money on efforts regarding this selection.
They asked for a solution that could optimize the process of predictive analytics helping to decide whether the animal should grow till lactation or be removed.
About the Project
Our client is a large agriculture holding with a few livestock farms (around 20 000 cows and calves in total) specialized in producing and selling meat and dairy products.
Region
Europe
Time
5 months
Team
3 specialists
Solution by Steps
This case is an example of usage of AI and predictive analytics in agriculture, so to understand how it helps solving the problem of predicting animal death probability, we highlight main steps of development.
- 1
Data collection
We collaborated with developers of the custom livestock management system and implemented a lightweight extract-transform-load (ETL) pipeline to collect and prepare data from LMS. The resulting dataset on the health characteristics and life events of more than 17 000 animals. This dataset has become a basis for further development.
- 2
Feature engineering
The main task, crucial for providing a comprehensive predictive analytics solution, was to extract the most significant data parameters (features) from the livestock life history reports.
We converted the histories of life events into a feature table which included the most important indicators, such as cow’s weight, the dates and number of previous diseases, its birth month and blood parameters. The data covered the first 6 and 12 months of an animal’s life.
It is worth noticing that during the feature engineering process we faced a certain amount of erroneous data (outliers) contained in the holding reports. Through that we included special data cleaning steps and provided recommendations for future data collecting on the customer side.
The example of outliers included unusually low or unusually hard weight of an animal, or incorrect event codes. For algorithm stability, such values were replaced by empty values at PoC phase. Besides, our team added recommendations to avoid outliers into the livestock management system.
- 3
ML model training and further analysis
The prepared data was used for the creation of a training and testing datasets which were then used to train an ML. The model learned to identify crucial data entities and find certain patterns inside the data. The training usually takes about half an hour for one run, including data preparation time and accuracy estimation.
The created AI model was able to perceive data regarding animals' health condition and disease history, analyze and provide advanced predictions on whether death or lactation possibility of each cow.
To further verify the solution, the final report also contained the information on how different features affect prediction. Such a chart helps the experts on the customer side to additionally verify the results.
During the whole project, our Data Science team was in contact with veterinary experts on the customer side to verify the assumptions, data processing pipeline, feature importance and the results: both intermediate and final.
According to our experience, proper communication with subject matter experts (SMEs) and stakeholders is one of key factors to the project success.
Wish to see how the AI can help your business through advanced predictive analytics?
Share your ideas and problems, and we’ll connect you with the right expert for a free consultation.
Key Results
Our solution has demonstrated that combined usage of available animal live events data and technologies of artificial intelligence in agriculture can predict the survival probability with significant accuracy (about 70% in terms of ROC AUC value).
Through that, the agriculture company has improved its decision-making processes regarding cows’ usage for lactation or meat production.
Moreover, our team has formulated and presented several approaches for improvement of prediction quality along with recommendations to get better data (such as having specific rules and limits for data input on farms).
The ML prediction can run in offline mode: periodic runs, for example once per week. The ML results are further available in livestock management system to provide information for veterinary specialists and farm managers to take appropriate actions with animals.
Our Expert
"This project is a powerful example of a huge perspective of AI usage in farming. Dedicating decades of work to the industry, I also can say that predictive analytics is only one of the many ways of applying ML technologies to agriculture. AI technologies already can enhance supply management by predicting market demand, optimizing logistics and reducing food waste."
In further perspective usage of such a technology in agriculture can be represented by irrigation automation, soil and crops condition monitoring and waste management.
Yuri Svirid, PhD. — CEO Silk Data
Frequently Asked Questions
Various AI tools can transform animal care, automizing and optimizing lots of processes. Automated health monitoring (through wearables or image recognition), advanced behavioral analysis (through animal activity tracking), physiological and psychological research – these are only small part of AI capabilities.
With the right approach, the prospects of further AI development for agriculture are promising. They can be represented by autonomous farming equipment, such as self-driving trucks, harvesters and drones, genetic data analysis for proper animal selection, advanced climate modeling and much more.
Yes, provided the farm has sufficient animals and a livestock management system covering the important events about animals’ life and health. Note however, that the model trained on data for cows cannot be used for pigs or poultry, because of different life timespans, types of life events and diverse diseases.
