b'FIGURE 1: FINAL F1-SCORE ASSESSMENTS OF CROP VARIETY CLASSIFICATION AND SENSITIVITY ANALYSISSource: Geo4Adataset is performed. This emulates data augmentation which added additional variance to the existing dataset. Furthermore, a random initialization is used prior to training of the machine learning model to probe the model stability. The figure above shows an example of the sensitivity analysis.IF SATELLITES CAN RECOGNIZE After including three observations, all 3-performance met- POTATO VARIETIES, THEN THIS rics are more than 75 per cent within a 1-confidence interval. It should be added that for the variety Bellini, only a limitedCOULD BE A VERY HELPFUL number of fields are being tested. ENFORCEMENT INSTRUMENT ANSWERS NEEDEDIt was foreseen that in this first pilot year not all questions whichIN THE FIGHT AGAINST ILLEGAL we posed ourselves are being answered, Geert Staring explains.MULTIPLICATION AND TO ENFORCE For instance, they now have some experience with five com- PLANT BREEDERS RIGHTS.pletely different varieties, but what if two varieties are geneti-cally close to each other? Can remote sensing techniques in that case distinguish those two varieties from each other? What is the influence of the soil type on the obtained results? And how about the legal status of those data, can this be seen as legally obtained evidence when going to court? Therefore, more infor- need to be further analysed. Nevertheless, this feasibility study mation needs to be gathered. But the first years experiences byexceeded expectations and results published in scientific peer using this technique for recognizing varieties in the field lookreviewed papers. It marks a most promising starting point for very promising and they also found satisfying answers on theany further development. The HyperF-Tensors feature engi-legal use of retrieving those data.neering, as well as the machine learning model can be applied and tested with additional potato varieties, if more in-situ data FOLLOW UP are made available. Due to the promising results obtained from Given the results from the sensitivity analysis within the feasibil- available high resolution (HR) data of the Sentinel-2 satellite, ity study, the developed machine learning method outperformsvery high resolution (VHR) data was not yet included in the Breeders Trust metric threshold of an accuracy greater than 75analysis. Meanwhile, Breeders Trust has decided to start phase per cent. It is important to emphasize that the dataset provided2 before moving on to implementation in practice, which is is biased by restricting the area of interest to the observationplanned for 2023. points provided by Breeders Trust. According to Mangnus, itWe first want to test this instrument extensively before remains an open question how the classification quality of thewe use it in our enforcement toolbox, said Staring. We are not developed machine learning method is altered when additionalunder pressure: once the program is running in practice, we can potato varieties are included into the used dataset. Additionally,retrieve data up to 2017. the effect of planting strategies like planting distance, soil type,Grass seed breeders and vegetable seed breeders already fertilization and irrigation on the classification performanceshowed interest in using this technique.50IEUROPEAN SEEDIEUROPEAN-SEED.COM'