Cainthus
Nominated Award: Best Use of AI in Sector
Website of Company: https://www.cainthus.com/
Cainthus is a computer vision and artificial intelligence AgTech startup founded with a vision to provide affordable measurement within agriculture to increase the global sustainability of the industry. The founders realized that the agricultural industry lacked accurate digital measurement tools available at the commercial farm level and that it would be very difficult to fix a lot of global agricultural problems without having such measurement. After getting exposed to the latest computer vision technology in 2013, it became apparent that computer vision could be the tool that enabled affordable digital measurement for farmers. Today, we are focused on the dairy industry.
Our mission is to have a positive impact on global sustainability on two levels. First, we aim to contribute to an economically sustainable agriculture industry in which farmers can sustain thriving and profitable businesses while providing food that is affordable to the most at risk in the economy. Second, we envision a more environmentally sustainable global agriculture system, a complex but increasingly important and urgent objective.
We believe that we can achieve a positive impact on economic and environmental sustainability by providing measurement tools that enable farmers to operate more efficiently at scale. Cainthus was founded with the purpose of enabling further intensification of the farming industry to enable us to produce more food with less waste and land use.
For example, if every dairy cow was as productive as a US dairy cow, we could reduce the global dairy herd from approximately 275 million cows to 69 million cows without losing a drop of milk. Any meaningful reduction in the number of cows in the dairy industry will yield a significant reduction in carbon emissions and land use for crops that support the industry.
We have found the best product-market fit and technical fit for developing a first-of-its-kind application of computer vision to provide digital measurement and artificial intelligence on the farm in the dairy industry. We have built a solution for 24/7 intelligent monitoring of large-scale dairy farms and using computer vision and artificial intelligence, we translate visual information into actionable insights that enable dairy farmers to make data-driven decisions to improve farm operations, animal health, and welfare and, ultimately, increase efficiency and profitability.
We are the first to bring computer vision onto dairy farms, a challenging feat on its own. Furthermore, with our first solutions, Cainthus has demonstrated the ability to increase
milk production per cow by 4.5%, and a reduction of carbon footprint per cow of 2 – 4% depending on the model used. As more insights are generated relating to welfare and health related concerns, we estimate that Cainthus solutions can decrease the Global Warming Potential (GWP) and GWP per pound or litre of milk by up to 10%.
Reason for Nomination:
Our first goal was to prove that we could develop and operationalize a commercial product using computer vision and artificial intelligence to provide observation-based measurements on dairy farms. Based on early experience, we knew that this was a very complex and challenging problem to solve. Therefore, we started by identifying the simplest technical problem that we thought we could solve that would also prove to be a commercially viable product.
The first problems that we identified were focused on monitoring the feeding operations on the farm. Optimal feeding operations are critical to the dairy business — cows must have high feed availability to support maximum milk production and feed represents more than 50 percent of the costs on the farm — therefore, farms must balance high feed availability with minimal waste on a daily basis for each pen (a group of cows).
One of the biggest challenges that we face in general is access to farms and data for research and development. There are no publicly available datasets representing dairy farm feeding operations! We partnered with several farms that generously allowed us to experiment with different cameras and various camera configurations in order to collect data and design the camera solution for the product.
Because we wanted to minimize any disruption on real farms, we also developed a 3D barn model complete with 3D cows and with which we could simulate placing different cameras with different lenses into various locations to determine the field-of-view and resolution of each setup. We selected a camera with a 180 degree fisheye lens that provides a large field-of-view along the feeding alley as well as into the pens because this would allow us to cover the largest area within the barn with sufficient resolution with the smallest number of cameras.
Next, we needed a hardware solution for collecting image data for research and development that would also support delivering real-time measurements and insights on the farm in a web application. Limited internet connectivity in rural areas and the large data size of images ruled out any option to stream images directly from the farm over a broadband connection. After several trials and experiments, our hardware solution consists of small computing devices containing GPUs and robust to the farm environment distributed throughout the farm connected to cameras with PoE cables and connected to the internet with a cellular connection.
With access to image data from farms, we could begin research and development on observing feeding operations using computer vision and artificial intelligence. We have developed solutions to observe the feed availability and the feeding events on the farm. Our feed availability solution observes the feed in each pen and monitors when the feed is running low. We have solved this with image segmentation to locate the feed in the image followed by classification of low and high feed.
Some of the challenges we faced when developing this solution were the subjectivity of the task, the amorphous nature of the feed shape and texture, variability in feed colour and appearance, challenging lighting conditions, shadows, and eating cows obscuring the feed.
Our feed events monitoring solution observes the times when new feed is delivered to each pen, when that feed is pushed closer to the feeding cows (the feed gets pushed out of their reach as they eat), and when feed is cleaned away in preparation for the next day’s routine. Our feed events monitoring solution is an end-to-end deep-learning solution that takes as input two images from subsequent timestamps and uses a state-of-the-art attention mechanism to focus on areas of the feed that have changed or moved in the two images and detect feed deliveries, push-ups, and clean-outs. Some of the challenges we faced when developing this solution include the variability of operations and practices on different farms and the subtle and elusive differences in the
appearance of feed as feeding events are performed.
Both of these solutions are implemented in our first product that was released commercially in 2020 and is now deployed on farms across the US and monitoring tens of thousands of cows. We have since developed a second product that observes cow behaviour and welfare and milking operations.
In both of our products, we have developed analytics on top of the core observations described above to provide actionable insights, trends, and recommendations to farmers that enable them to detect and respond to problems early and maintain efficient feeding operations while guaranteeing cow comfort and welfare.
Additional Information:
The description above only touches the surface of challenges that we’ve faced and innovative problem-solving that we’ve engaged in to overcome those challenges! Another area that I’d like to highlight is the operationalization of artificial intelligence on dairy farms so that it works not only in a curated dataset but in a continuous real-world application.
As mentioned above, we have limited access to data to develop our solutions and have employed techniques to develop generalized models and solutions that work in different environments. We have also implemented ongoing monitoring to measure the performance of the solutions in operation and detect any model or data drift. We have implemented pipelines for continuous training and maintenance of our solutions with new data continuously collected from farms so that we can keep our models updated with changing real-world conditions and guarantee the quality of our artificial intelligence to customers.
At Cainthus, we have had tremendous achievements so far, bringing a revolutionary idea to life on dairy farms with a never-ending series of challenges and complexities. The number of surprises we’ve faced along the way — cows unplugging our equipment on numerous occasions as only one example — has made it that much more interesting and rewarding!