Can custom heat detectors benefit your farm?
Activity sensors likely have a large potential to serve your farm well. Customizability is important in all farming operations and this example of heat detection sensors shows how it is possible. The goal is to have a better understanding of when generic algorithms are supposed to work well and when they are not. This is important for manufacturers if they are to strive for the equity of treatment of all farmers. For farmers who may benefit from the use of custom algorithms, the industry should encourage more active discussions on which data should be made available to the farmers.
The sensor technology works in three stages:
- The mechanical device collects technical data on the animal’s motion, sound, and temperature etc.
- The data is interpreted into the levels of rumination, eating, and activity behavior.
- Those results are then translated into actionable information for herd management like alert systems for detecting heat, sickness/injuries, and any sign of abnormal behavior. In this last component, there is room for customization that could better serve individual farmers’ needs.
Customizability in action
Here is an example of sensor technology based on a custom heat detection model being developed for the West Central Research and Outreach Center (WCROC) in Morris, MN. The dairy herd has been equipped with two popular activity sensor systems: SCR by Allfex Inc. and CowManager by Agis Automatisering BV Inc.
The herdsman combines the information from SCR, CowManager, and standing patches to ensure accurate heat detection during breeding seasons. They are researching how well these sensors can help detect heat in a herd of mixed breeds of animals that are grazed during summer months. They are developing a custom alert system because the generic programs are not working for the setup at WCROC and farms like it.
Figure 1 is an example of the preliminary results. The cow being monitored is on her third lactation and calved on March 11, 2017, she likely came in heat (estrus) five times from early April to mid-July at intervals of roughly 21 days. The model was able to predict those five dates, including two actual insemination dates, along with a couple of false positive alerts in early April. This particular case was closely monitored during the model development process, which partly explains why it shows our model in a favorable light.
Our model is based on the breeding records and SCR data from April 2014 to August 2017, including a total of 892 lactations and 4,338 cow-days. The model can be used to predict how likely a given cow on a given day is in heat, which are then mapped into discrete heat state predictions with an arbitrary threshold.
Figure 2 presents how the distribution of those probabilities by the actual heat state relate to the user-defined threshold shown as a vertical dashed line.
The tradeoff of an alert system is that the lower the threshold, the more the alert produces both true positives and false positives. This sensors performance was also better in the non-grazing season where there is less daily activity.
Results of the model:
- Accuracy of 84% - the rate of correct prediction for heat state over all cow-dates.
- Sensitivity of 78% - the rate of successful heat detection out of all heat incidents.
- Specificity of 86% - this meaning the remainder, 14%, is the rate of false positives out of all heat alerts for the threshold of 40% probability.
This model is developed with rumination and activity data from SCR, combined with information on animal’s breed, organic vs conventional herd type management, calving dates, and insemination dates. The key ingredient here is insemination records, which serve as a good approximation to the true heat incidents.
The plan is to turn the model into a supplemental alert system in the form of an app meant for the Morris dairy. In the future, the Morris model may be shared with others, particularly those who manage similar grazing herds as the Morris farm.
October 14, 2017