Using machine learning in phenotyping and seed testing

Digital phenotyping is well known as non-invasive imaging-based method to obtain data on plant phenotypic traits. Likewise, seed testing methods are available to measure germination processes, or determine seed and seedling quality.

After recording, images are processed, however classical image processing frequently reaches its limits, because plant material is complex, or parts that need differentiation are highly similar. With color-based thresholds and filtering procedures, some interesting plant traits are not accessible. To tackle this challenge, LemnaTec provides a toolbox for machine learning, where our customers can label exactly these traits in recorded images that are of interest to them, and we feed these labelled images into our machine learning procedure. Doing this, our analytical software can be trained – like our customers’ analysis specialists – to look for the analysis-specific properties according to the current demand. Such training can target e.g., plant quality traits, disease symptoms, cultivar-specific features, or it can aim to identify foreign objects and weed seeds in seed batches. Of course, this list is not limited to the given examples, virtually any feature that is detectable at the outer surface of plants, seeds, fruits, pests, or pathogen in visible light, but also NIR/SWIR or fluorescence images can be trained as analysis target.