top of page

How BASF is using Deep learning to create the wheat of the future


BASF's broad portfolio ranges from chemicals, plastics, performance production products to oil and gas.

Through research and innovation, BASF supports customers in nearly every industry in meeting current and future needs of society.

Their innovation Center in Ghent aims to deliver solutions for BASF's seed business by developing valuable traits for various crops through biotechnology solutions and modern breeding techniques.

Their activities cover all aspects involved with the R&D process, from the initial stages of research to life cycle management, from the lab to the field.

The Opportunity

In order to decrease the time-to-market for new wheat variations BASF looked into numerous ways to improve the breeding process in testing new genetic lines.

Research activities at BASF have led to the creation of large amounts of wheat datasets.

Manually evaluating images from the field and greenhouse is a substantial task. This creates a need to accurately detect certain metrics (phenotypical markers) involved in quality of the wheat plants using automated qualification.

The Solution

A strong team of data-scientists (members…) at BASF recently started using Deep learning for phenotypical quantification of biological images.

These metrics included: wheat stomata count and morphology, vascular bundles in wheat peduncle microscopy images, spike detection, spike detection in both the field and the greenhouse …

One important challenge in these kinds of projects is the preparation of your data. The images contain a lot of noise, circumstances can vary such as lighting, there is a need for annotation, …

Provisional lessons learned:

  • Focus on data quality and quantity: for example, rotate your images so you can increase the number of images in your training dataset.

  • Efficiently tackle data annotation: for example, by outsourcing it (Samasource, Crowdflower).

  • Track model experiments: develop a systematic approach for storing code, parameters, datasets,

  • On-premise v/s cloud infrastructure: for example, investing in on-premise systems or working together with VSC lets you iterate and try different experiments.


bottom of page