One of the major critical barriers to using AI and image analysis in the agriculture and food industry is the trust in its effectiveness.

Today, manual visual inspection of grains remains one of the crucial quality assurance procedures throughout the value chain, ensuring the journey of grains from the field to the table and guaranteeing that farmers receive the right price for their crops.

At the Danish-owned family company FOSS, high-tech analytical instruments are developed for the agriculture and food industry, as well as the chemical and pharmaceutical industries.

Since its founding in 1956 by engineer Nils Foss, development and innovation have been high priorities. As a global producer of niche products, staying ahead of competitors is essential.

Hence, collaboration with researchers from the country’s universities is a crucial part of the company’s digital journey. In a project at the National Research Centre for Digital Technologies (DIREC), the company, along with researchers from Technical University of Denmark and University of Copenhagen, aims to map how AI and image analysis can replace the subjective manual inspection of grains with an automated solution based on image processing. The goal is to develop a method using deep learning neural networks to monitor the quality of seeds and grains using multispectral image data. This method has the potential to provide the grain industry with a disruptive tool to ensure quality and optimize the value of agricultural commodities.

The agricultural and food industry is generally a very conservative industry, and building trust in digital technologies is necessary, explains senior researcher Erik Schou Dreier from FOSS. The development of AI, therefore, cannot stand alone. To encourage farmers to adopt the technology, it is crucial to instill confidence in how it works. In this process, researchers use explainable AI to elucidate how the algorithms function.

Today, grain is assessed manually in many places, and replacing manual work with a machine requires trust. Because the work is performed by humans, it is a fairly subjective reference method used today. Humans may not necessarily perform the work the same way every time and can arrive at different results. Therefore, there will be some uncertainty about the outcome.

Mapping and explaining algorithms

– The result is more precise when using AI and image analysis in the process. However, for these new technologies to gain widespread acceptance globally, a model is needed to explain how AI works and arrives at a given result, says Erik Schou Dreier.

Many people have inherent skepticism toward self-driving cars. Self-driving cars need to be even better and safer at driving than us humans before we trust them. Similarly, the AI analysis models we work with must be significantly better than the manual processes they replace for people to trust them. To build that trust, we must first be able to explain how AI analyzes an image and arrives at a given result. That is the goal of the project—to interpret the way AI works, so people can understand how it reads an image.

We typically accept a higher error rate among humans than machines. For us humans to trust the algorithms, they need to be explainable.
Erik Schou Dreier, senior researcher

PhD student Lenka Tetková from Technical University of Denmark is part of the project and spends some days at FOSS’ office. Here, she works with images of grains in two different ways, partly to improve image qualification and partly to better understand how classifications work so they can be enhanced.

– I sometimes use the example of a zebra and a deer to explain how image classification works. Imagine you have a classification that can recognize zebras and deer. Now, you get a new image of an animal with a body like a deer, but the legs resemble those of a zebra. A standard model will not be able to recognize this animal if it hasn’t seen the animal during training. But if you provide it with additional information (metadata) – in this case, a description of all kinds of animals – it will be able to infer that the image corresponds to an okapi, based on its knowledge of zebras, deer, and the description of an okapi. That is, the model will be able to use information not present in the images to achieve better results, explains Lenka Tetková and continues:

– In this project, we want to use metadata about the grains, such as information about the place of origin, weather conditions, pesticide use, and storage conditions, to improve the classification of grains.