Digital Goat Farmer's Aid

Designing intelligence in interaction

Abstract

In livestock agriculture, earmarks are widely used to identify individual animals. Goat farmers are legally obligated to attach two earmarks to the goats' ears within the first six months of their lives. These earmarks have embedded identification numbers, which are used to report the goats' residence and migration information to the Dutch government's "Identificatie & Registratiesysteem (I&R)" (RVO, n.d.). Several problems are surrounding the earmarking of goats. A part of the problem focuses on animal welfare. Next to this, the business standard proves to be time- and labor-intensive. Lastly, the regulations for goat identification are sharpened, meaning that the farmer is at risk of receiving costly fines. However, no alternative is easily accessible. During this project, we explore the viability of the Digital Goat Farmer's Aid (DGFA), a computer vision-based goat identification as an alternative to earmarks.

To create a goat identification system, a normalized dataset of facial pictures of goats is required. As such a dataset is not publicly available, we created a dataset of goats ourselves. This dataset was preprocessed, after which features were extracted using Matlab. This processed data was used in a multilayer perceptron neural network to train a goat identification system using Neuroph Studio. Various types of data was extracted and compared with each other to get the best result.

The research resulted in an 83.3% predicted accuracy proving the feasibility of goat identification. This is not robust enough for implementation in the agricultural industry, but it does show that the identification of goats via image recognition is feasible. Discussion of future approaches towards face recognition on goats suggested the implementation of a deep learning convolutional neural network to better define specific head features of individual goats.

Responsibilities

Collaborating with goat farmer and potential stakeholders

Development business opportunity

Pre-processing data

Feature extraction

Development multilayer perceptron neural network

Identification of goats based on facial features

Identification of goats based on facial features

Identification of goats based on facial features

Identification of goats based on facial features

Identification of goats based on facial features

Data collection

Technology & Realisation

I believe, as a designer, you should be capable to recognize the different possibilities of applying intelligence in your concepts. This means that you should also be able to implement this intelligence into your prototypes to validate (a part of) your concept. Therefore, I wanted to focus on the realization of prototype.

After the concept became clear, I looked into related work among human facial recognition and tried to find out what fits best within our concept. I ended up with a conventional neural network. However, I learned that this type of deep learning requires a lot of computation power we did not have. Although I still believe this is the best neural network for our concept, I think we made the right move by focusing on image recognition to find out if we actually could identify different goats.

After the obtained feature extraction data, we uploaded the data to Neuroph Studio. In Neuroph Studio, I created a multilayer perceptron neural network. I tried several combinations of the collected variables in our network, to see which combinations would result in the best goat identification. Because I worked mainly on the data and neural network, it made me realize what the impact of the quality of data is on the results. It showed me the importance of a good qualitative data collection and preprocessing phase. In order to avoid under- and overfitting problems these things should be kept in mind.

Math, Data & Computing

One of my responsibilities was creating the dataset, by taking photos of 24 randomly selected goats while they were positioned in a milking carousel. From these photos, I created a set of training-photos and test-photos. Thereafter, I could distinguish image processing methods for goats. For every feature extraction method, I created a Matlab code. The code first loads the data, after which it performs several feature extraction methods.

As part of my job, I focused on extracting different features from the goats head in Matlab. I had no experience with Matlab, but I can see how it might become useful for me in the future, since it is easy to clean, plot and analyze your data. After extracting data, I tried making sense of the extracted features by plotting the data. By doing this, I could understand why certain data would work or would not work for our neural network. Because I worked mainly on the data and neural network, it made me realize what the impact of the quality of data is on the results. It showed me the importance of a good qualitative data collection and preprocessing phase.

Business & Entrepreneurship

During one of our first brainstorms, we talked about the opportunities for machine learning in the field of agriculture, since there is a lot of data and no strict rules around privacy regulation. During an interview with a goat farmer, we found an opportunity for among goat farms. The solution, face recognition among goats, was straight forward and involved stakeholders seemed really interested. Therefore, I got in contact with multiple companies in order to find out if we were able to create a new value proposition. It showed me what the financial benefits of our concept were in an efficient way.

Finally, one of the companies I approached is currently looking further into face recognition among goats and shows there is value in our concept.

References

I&R wordt strenger | Geitenhouderij. (2019). Retrieved from https://www.vakbladgeitenhouderij.nl/ir-wordt-strenger/

RVO. (2015, July 15) Oormerken schapen en geiten. Retrieved from https://www.rvo.nl/onderwerpen/agrarisch-ondernemen/dieren/dieren-registreren/schapen-en-geiten/oormerken-voor-schapen-en-geiten

© 2020 Jesper van Bentum