In the evolving field of agriculture, the dairy industry has embraced new techniques and technologies to enhance the efficiency, productivity, and management of the herd. Specifically, dairy managers use software embedded with precision dairy technologies to manage individual cows in a herd setting or to check if the herd is on target for performance. This concept of taking information from sensors and making informed decisions to manage livestock is called precision livestock farming (PLF). For example, producers use milk capture technology to track milk production in each cow. If the herd deviates from their milk production by 20% on average, a PLF producer would use this information to inform their decisions, such as troubleshooting the feed bunk, calling their veterinarian, or checking with parlor staff. Producers use PLF to make informed management decisions because of the capability behind machine learning algorithms (ML) and artificial intelligence (AI). This article explains ML, AI, and the importance of identifying the farm’s goals for a technology before purchase.
Understanding Machine Learning in Dairy Farming:
One subset of AI is machine learning (ML). Computer engineers use ML algorithms and statistical models to train ML models and test the algorithm against different scenarios, including making data predictions. There are two main types of ML: supervised, where a computer scientist adjusts the algorithm based on ML feedback, and unsupervised, where the computer algorithm adapts to new information automatically [1]. Generally, an engineer will not embed an ML algorithm into an AI software platform until a certain threshold for accuracy, precision, and error across multiple iterations is met. Unsupervised ML, or AI software can make predictions about patterns in new datasets without direct input from human users. This allows for automated predictions about individual cows in a herd setting.
What is artificial intelligence?
Artificial intelligence is software that mimics the human thinking process and adapts to additional information [2]. Dairy producers use PLF systems that are embedded with AI to analyze and interpret predictions about their cattle. An easy way to understand AI is by thinking of the human brain. As a human brain learns and practices something, it becomes more efficient when completing a task or solving a problem. Essentially, AI is like the brain of a computer, the more information it receives, the better the answers and responses it generates. However, AI is not perfect, and only works as well as the quality of the data and the “robustness” of the software [3]. Producers should only select precision technologies that are validated for the metric of interest because the primary role of AI is to provide an extra set of eyes and ears for the producer [3]. Validation is important because AI-embedded software requires “robustness,” or an ability to generalize the predictions of the algorithm to different contexts and situations. After all, no dairy is the same. The goal of PLF is to save time, improve cattle performance, and provide data for more informed decision-making for the herd. Therefore, as a dairy producer, it is fundamental to investigate if the PLF system of interest is validated scientifically and within the company to perform the specific task of interest.
Precision livestock farming: what should I look for before making the purchase?
One example of PLF is tracking the health or reproductive status of individual cows in the herd and using that information to intervene. There are many types of precision technologies: robotics, external sensors, and wearable sensors that attach to the cow in some way to track feeding behavior, rumination, temperature, or activity status, (aspreviously described).
One common PLF system that producers use involves wearable technologies, which are sensors attached to the cow. Specific information is collected from a cow and locally stored on the sensor until a cow is near a base station. The cow’s tag will be triggered to download to the station and transmit to a cloud server, where an AI platform will interpret the data and make predictions about individual cows. Cows who deviate from their normal behavior will have an alert generated for review through the software interface. There are two types of wearable PLF systems to consider for the farm:
- Saving labor: High emphasis on specificity, or animals that are truly negative for the condition of interest.
- Replacing skilled labor: High emphasis on sensitivity, or animals that are truly positive for the condition of interest.
- Saving labor: An alternative to health exams on each transition cow.
Many dairy producers screen each transition cow with intensive health exams for the first 10 days in milk because metabolic diseases could negatively compromise her entire lactation [5]. Recently, Cornell researchers observed that using rumination monitoring systems daily to decide who to screen for a health exam allows for a less labor-intensive strategy than locking up each cow [6,7]. Specifically, veterinarians performed health exams on each transition cow in one group of cows for the first 10 days in milk (farm standard protocol). For the second group of cows, the veterinarians performed health exams only on cows who had PLF-generated alerts from the rumination system. Both protocols required that staff walk the fresh pen daily as well to safeguard any cows who were extremely sick and not identified. Researchers observed that there was no difference in disease detection rates, or disease treatment rates between the two protocols, saving the farm $$ in labor costs when they adapted the PLF system. For the labor reduction system to work well, the PLF system should be validated with very high specificity > 90% meaning that 90/100 cows that the system says are healthy are healthy. We want the system to rarely mislabel healthy cows who do not need exams to save labor. Identifying which cows do not require exams saves the farm labor and allows healthy transition cows more time at the feed bunk.
Transition cow health alert systems can often be incorporated into a milking parlor to sort cows based on alerts for further clinical examination. Photo courtesy of Shelby Felder
2. Replacing skilled labor: Using robotics to identify scouring calves
In a different scenario, perhaps a farm is limited in their labor to observe cattle for the disease, but the mortality rate for the disease is high. This is the case for dairy calves on most farms, where complications from diarrhea such as dehydration are the leading causes of death in preweaned calves [8]. The only correct way to diagnose diarrhea in calves is by observing the fecal consistency, or fluidity of the diarrhea which is labor intensive [9]. For this scenario, Penn State and U. Guelph researchers used robotic milk feeder data to design an algorithm to flag calves at risk for diarrhea from the day before to the day after the calf had diarrhea [10]. Calves were offered at least 15 L/d milk volume, and the alert was generated based on changes in the previous 2 d milk intake or drinking speed. This algorithm was diagnostically accurate, which means that there was a high sensitivity of > 80% meaning that 80/100 calves that the system says were sick had diarrhea. This is important because early intervention for a calf to recover from diarrhea is fundamental for getting ahead of dehydration. That is why when selecting a PLF system, it is very important to make sure that the system is selected based on what you want it to do: save labor on a task performed on everyone (removing animals from the checklist) or using the PLF system as skilled labor (using the system to screen for sick cattle).
Finding sick calves is challenging in group housing. Researchers from Penn State and U. Guelph observed that an alert was diagnostically accurate for flagging calves at risk for scours using data from an automated milk feeder.
We do not have the labor: Reproductive management
It is well known that replacement heifers who calve later than 23-24 months of age can impact economic success for a dairy [11]. Heifers are a large economic investment, and each additional day that she is on feed without milking she is costing the dairy money [12]. Missing just one heat cycle can easily put dairy producers behind schedule. However, producers can place wearable sensors on their heifers to passively observe for estrus behaviors. Estrus behaviors can include evidence of mounting another heifer or standing to be mounted (recorded as increased head or neck movements by the sensor), or an increased overall activity index relative to that heifer’s behavioral baseline [13]. This type of system may be preferred for dairies over more labor-intensive methods such as CIDR, Kamar strips, tail chalk, and observing for heats, or producers may use a PLF system in conjunction with a synch protocol to improve their conception rates. Economists suggest that for a PLF system to improve pregnancy rates on a dairy, the system should last 5+ years, and the dairy should not already be in the top 10% for reproductive performance for conception rate compared to their peers [14]. There are many sensor systems available, and each varies regarding how well it classifies heifers with estrus [15]. It is important to check that the PLF system you are purchasing has at least 80% sensitivity, meaning that of 100 heifers that the system labels as heifers in heat, 80 are in heat. Furthermore, consider evaluating the heifer before insemination for signs of estrus behavior prior to breeding off the alert. Does the heifer seem restless, or extremely friendly? This is important to avoid breeding heifers that are not in estrus.
In summary, make sure that the system of interest is scientifically validated, and that you select a system with the sensitivity, or specificity that meets your needs.
Source: extension.psu.edu
