Suppose one day it is possible to distinguish the most important Plant Breeders’ Rights protected potato varieties (or a variety of any other species) anywhere in the world by using remote sensing technology, and that it even is possible to distinguish seed potato fields from ware potato fields of a certain variety. If it turns out to be possible — that remote sensing techniques can recognize potato varieties — then this could be a very helpful enforcement instrument in the fight against illegal multiplication and to enforce Plant Breeders’ Rights, for instance, by monitoring the farm saved seed acreages in a certain part of the world.
That is the dream of Geert Staring, general director of Breeders Trust (www.breederstrust.eu), seated with his head office in Brussels and representing the PBR interests of the 11 most prominent West European seed potato breeders and nine grass seed breeders.
The Remote Sensing project
The project aims to support Breeders Trust to enforce plant breeders’ rights and eliminate illicit propagation and marketing of seed potatoes. The project is performed by GEO4A, situated with its headquarters in Emmeloord in the Netherlands with a specialty in potatoes. In June 2020, Breeders Trust signed a contract with GEO4A to conduct a feasibility check in order to find an answer if potato variety recognition by using remote sensing technology, anywhere in the world, is possible. Breeders Trust and GEO4A jointly started the feasibility pilot project with five potato varieties. The North-East Polder in the Netherlands was chosen because of the high number of potato fields of many different varieties situated close to each other.
Guido Mangnus, managing director GEO4A, mentions that due to the technical challenges, a step-by-step approach was chosen in close coordination with Breeders Trust. “The first phase focusses specifically on the feasibility of variety classification by means of Earth Observation (EO) data in general. For the development of the potato variety recognition, Breeders Trust provided sample locations and crop variety at those locations as in-situ data. These sample points are supplemented by GeoVille (Austria) with earth observation data (Sentinel 2) as a basis for any further actions on a potato variety methodology”, Mangnus explains. Now, one year later, the results of this feasibility check are known and look very promising.”
Methodology
In the first step, data exploration is used to analyse the data provided by Breeders Trust. Therefore, the spectral characteristics of each of the five selected varieties are visually checked. These in-situ data and Sentinel-2 observations are combined for the potato variety recognition, using an in-house developed feature engineered dataset called HyperF-Tensors. Given the underlying dataset, these HyperF-Tensors represent unique fingerprints for the five potato varieties and are used as an input for the machine learning model classification.
Results and performance metrics
For each number of observations included into the training of the machine learning model, a random split of the train and test dataset 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 below shows an example of the sensitivity analysis.
After including three observations, all 3-performance metrics are more than 75 per cent within a 1σ-confidence interval. It should be added that for the variety ‘Bellini’, only a limited number of fields are being tested.
Figure 1: Final F1-score assessments of crop variety classification and sensitivity analysis
Answers needed
It was foreseen that in this first pilot year not all questions which we posed ourselves are being answered, Geert Staring explains. For instance, they now have some experience with five completely different varieties, but what if two varieties are genetically 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 information needs to be gathered. But the first years’ experiences by using this technique for recognizing varieties in the field look very promising and they also found satisfying answers on the legal use of retrieving those data.
[tweetshare tweet=”Can remote sensing techniques with satellites distinguish two varieties that are genetically close to each other?” username=”eyJtZDUiOnRydWV9″]
Follow up
Given the results from the sensitivity analysis within the feasibility study, the developed machine learning method outperforms Breeders Trust metric threshold of an accuracy greater than 75 per cent. It is important to emphasize that the dataset provided is biased by restricting the area of interest to the observation points provided by Breeders Trust. According to Mangnus, it remains an open question how the classification quality of the developed machine learning method is altered when additional potato varieties are included into the used dataset. Additionally, the effect of planting strategies like planting distance, soil type, fertilization and irrigation on the classification performance need to be further analysed. Nevertheless, this feasibility study exceeded expectations and results published in scientific peer reviewed papers. It marks a most promising starting point for any further development. The HyperF-Tensors feature engineering, as well as the machine learning model can be applied and tested with additional potato varieties, if more in-situ data are made available. Due to the promising results obtained from available high resolution (HR) data of the Sentinel-2 satellite, very high resolution (VHR) data was not yet included in the analysis. Meanwhile, Breeders Trust has decided to start phase 2 before moving on to implementation in practice, which is planned for 2023.
“We first want to test this instrument extensively before we use it in our enforcement toolbox,” said Staring. “We are not under pressure: once the program is running in practice, we can retrieve data up to 2017.”
Grass seed breeders and vegetable seed breeders already showed interest in using this technique.
