Method for assigning quality classes to seeds of a population

Abstract

The present invention relates to a method for assigning quality classes to seeds of a population, the method comprising, in a first analysis step: acquiring three-dimensional (3D) tomography image data relating to seeds of one or more samples of the population by means of an imaging three-dimensional tomography method, determining quality properties of the sample seeds and / or of young plants developing therefrom and classifying them into one or more quality classes on the basis of the determined quality properties and assigning quality classes determined for each sample seed to the corresponding 3D tomography image data relating to said sample seed; and in a second step: generating synthetic data regarding the phenotype from the associated 3D tomography image data relating to a sample seed for at least one spatial position of the sample seed, and annotating the synthetic data regarding the phenotype with the quality classes assigned to the associated 3D tomography image data. In a further step, one or more quality classes can be assigned to any seeds of the population for which real data regarding the phenotype is captured or measured.

Description

[0001] DG2.DD16911 1 / 01

[0002] September 24, 2025

[0003] 1

[0004] Method for assigning quality classes to seeds of a basic population Description

[0005] The present invention relates to a method for assigning quality classes to seeds from samples of a population (e.g., a number of seeds of a fruit type, a variety, a seed lot) in order to subsequently classify further, unknown seeds of this population into these quality classes based on their similarity to the pre-classified seeds, in particular to then be able to sort them into corresponding sorting / quality fractions.

[0006] Technical background

[0007] The goal of every seed processing operation is to select from a batch those seeds that will produce healthy, vigorous, and fast-growing seedlings that germinate quickly and uniformly. Such seedlings form the basis for maximum yields in food, feed, or raw material production, while maximizing the use of finite resources like soil, water, and nutrients. Optimized seeds of this kind are therefore also the foundation for the most sustainable plant production possible. This applies equally to the processing of ornamental plant or forestry seeds, or, for example, the production of high-quality wild plant seeds for storage in gene banks.

[0008] Externally damaged seeds have a high probability of not germinating or of producing abnormal, damaged seedlings. Various methods are therefore used to remove broken or externally damaged seeds from the seed stock intended for further processing. In contrast, it is generally impossible to tell from the outside of an externally intact fruit or undamaged seed what its germination rate will be and what plant quality it will produce. Experience shows that seeds that are clearly too small have a higher probability of not germinating or of producing only small plants, as the seeds are generally immature and contain an underdeveloped embryo. Therefore, excessively small seeds are usually sorted out by sieving.It is also known from experience that lighter seeds or fruits often have reduced germination capacity and vigor, and may also produce abnormal or undersized seedlings because they contain either no embryo at all, an underdeveloped embryo, or too little endosperm. For example, gravity processing is used to sort out the lighter seeds from the seed mixture and select only the DG2.DD16911 1 / 01.

[0009] September 24, 2025

[0010] 2 specifically heavier seeds with a presumably well-developed embryo and sufficient nutritive tissue to be used as commercial seed.

[0011] However, there are no universally valid objective limits regarding the seed size, specific gravity, or color at which germination will be rapid and complete, or when healthy and vigorous seedlings can be expected. Whether sorting by size, specific gravity, or color has actually led to the desired result, and whether the processed seed consistently produces uniformly fast-germinating and vigorous seedlings while the rejected seed fails to achieve this, can therefore only be analyzed after the fact. Generally, this is validated using visual germination tests on the sorted fractions to draw conclusions about the sorting quality. However, these conclusions lead to considerable time delays and often necessitate at least one further processing step.In particular, these quality tests do not provide direct measurements of the internal quality of seeds. Instead, germination tests can only draw indirect conclusions about the quality of the seeds in the sample, but not what kind of quality defects are present and how the seeds can be sorted specifically for them.

[0012] From European patent EP 2 525641 B1, "Method and apparatus for assessing the germination properties of plant seeds," a method used in industry is known that enables the internal and external (three-dimensional) 3D phenotyping of seeds and the resulting seedlings using 3D X-ray computed tomography (3D-CT). Seeds and the resulting seedlings are examined using 3D-CT over a predetermined period, particularly at specific intervals, so that the quality of the individual seedlings can be quantified, for example, by measuring the lengths, volumes, and growth directions of all plant organs during germination and early growth. A classification algorithm classifies these seedlings according to their quality, for example, into usable, conditionally usable, and unusable plants, and annotates the seeds or other characteristics from which they originate.their 3D CT image data accordingly.

[0013] This method (hereinafter also referred to as the "phenoTest") therefore provides information about seed characteristics, such as a specific shape, size, or seed filling level (the volume of the embryo and its endosperm compared to the volume of the seed cavity or the entire seed), in order to produce usable seedlings. DG2.DD16911 1 / 01

[0014] September 24, 2025

[0015] 3

[0016] The correlation between seed properties and seedling properties derived from the phenoTest, and the resulting quality criteria, cannot yet be used on an industrial scale with high throughput for lot-specific seed processing. 3D X-ray computed tomography (e.g., gantry CT) is not yet technically feasible for processing very small objects like individual seeds at high throughput rates of hundreds of kilograms of seed per hour and would be extremely complex.

[0017] European patent EP 2 588255 B1, "Method for classifying objects contained in seed lots and corresponding use for seed production," describes a seed sorting method suitable for the high throughput and resolutions described above. This method uses a laser light sectioning technique (optionally in combination with other optical inspection methods) for the 3D volume, shape, and surface determination of individual seeds, as well as for the detection of other unwanted objects contained in the seeds. Seeds and unwanted objects are each classified and then sorted into at least two corresponding sorting fractions, depending on the automatic classification performed.The classification is based on the recorded morphological information (surface structure, volume, shape and dimensions of the seeds) as well as optionally on further information determined via optional additional examination methods and sensors (such as color and / or (two-dimensional) 2D x-ray images).

[0018] By scanning the surface of seeds and other objects on a high-speed conveyor belt with a laser line, recording the deflection of this laser line as the objects pass by, and reconstructing the respective 3D surface, the external 3D shape and surface structure of the visible side of the seed, the height of the seeds (maximum and average height), as well as the position and any tilting of the seeds at high transport speeds can be determined. A classification algorithm categorizes the objects passing under the laser line on the conveyor belt, for example, according to pure seeds, foreign matter, stubble, fragments, stones, clods of earth, etc., as well as according to the desired external shape and size of the seed, and in one configuration, controls the pneumatic sorting of these objects.The blowing out of unwanted objects after they have left the conveyor belt at the end of the belt (or their removal by other means) so that at least two fractions of the seed can be obtained (usable seeds on the one hand and unusable seeds and unwanted objects on the other). DG2.DD16911 1 / 01.

[0019] September 24, 2025

[0020] 4

[0021] To additionally determine internal structures, such as the extent and surface area of ​​the embryo in 2D projection or defects like embryo fractures or necrosis, within the same flow-through process, the aforementioned European patent, EP 2 588 255 B1, proposes combining this with 2D flow-through X-ray imaging, in which the seeds are also subjected to 2D X-ray imaging. This is intended to lead to improved classification and sorting into quality classes by combining the sorting of seeds according to external phenotypic parameters, such as seed size and shape, with simultaneous sorting according to internal properties, such as embryo size or an approximation of the grain filling degree. Volume information from external and internal phenotypic parameters can also be combined with 2D X-ray projections, thus achieving a more reliable classification, for example, of the grain filling degree, into these quality classes.

[0022] However, generating the necessary training data presents several problems: Depending on its position on the conveyor belt, the 2D X-ray projections and / or 3D height profiles of the same seed can differ considerably. Furthermore, generating both 2D X-ray projections and / or 3D height profiles for different possible positions of a seed, especially in combination with subsequent seed germination, is extremely complex. Generating training data in this way is not realistically suitable for producing sufficient data points for robust classification models, particularly using artificial intelligence and machine learning, and especially for refinement within a population (e.g., a variety or seed lot).

[0023] This invention is the subject of determining which seed properties, in which combination, are actually responsible for good, rapid germination and vigorous seedlings for each individual seed, and how these properties can be determined on a training sample, calculated as a model and used as training data.

[0024] Summary of the invention

[0025] Investigations within the scope of the present invention have shown that 3D X-ray tomography for sorting seeds with very high throughput and high resolution is currently not technically feasible, meaning that only 2D flow-through X-ray imaging can be used for this purpose. However, a 2D scan results in a significant loss of information, as the third dimension—i.e., the volume, the external 3D shape, the thickness or height of the seed, the seminal organs, or the seminal cavity—is not captured.

[0026] September 24, 2025

[0027] 5 can be detected and therefore provides only insufficient discrimination. Furthermore, investigations within the scope of the present invention have shown that the combination of the various quality-relevant seed parameters differs in practically every seed lot due to growing conditions, genetics, and other partly morphological or biochemical factors, so that a specific classification and sorting for each population (e.g., seeds of the same variety, lot, growing conditions) can be advantageous. The investigations have further shown that even with good seed filling, germination capacity and seedling quality can be insufficient because, depending on other parameters, for example, the seed coat may be too thick, the seed shape unfavorable, or the embryo damaged.On the other hand, relatively small seeds, which are otherwise usually considered inferior, can under certain circumstances exhibit very good and rapid germination and produce very good seedlings due to the early start of growth.

[0028] The present invention according to claim 1 relates to a method for assigning quality classes to seeds of a population comprising in a first analysis step: acquiring three-dimensional, 3D, tomography image data by means of an imaging three-dimensional tomography method of seeds of one or more samples of the population, determining quality characteristics of the sample seeds and / or of seedlings developing therefrom and classifying them into one or more quality classes based on the determined quality characteristics and assigning quality classes determined for the respective sample seed to the corresponding 3D tomography image data of this sample seed;and in a second step: generating synthetic phenotypic data from the associated 3D tomography image data of a sample semen for at least one spatial position of the sample semen and annotating the synthetic phenotypic data with the quality classes assigned to the associated 3D tomography image data.

[0029] Based on this, a classical image analysis algorithm and / or a neural network (especially artificial intelligence, AI) can be trained with these synthetically generated, quality-class-annotated phenotypic data for simulated seed layers from a sample to create a classification model. In a subsequent step, one or more quality classes, or derived classification or sorting classes, can be assigned to any seeds in the population from which real phenotypic data corresponding to the synthetically generated data are collected, as explained in detail below. DG2.DD16911 1 / 01

[0030] September 24, 2025

[0031] 6

[0032] For a better understanding of the invention and the following description, the following definitions and explanations of terms are provided: "Phenotype data" of a seed can be image data and / or measurement data relating to the phenotype. Phenotype refers to all observable characteristics of an organism, in this case a seed, including both its external appearance and internal properties. Therefore, the following will also refer to "external phenotype data" (characteristics of external appearance) and "internal phenotype data" (characteristics of internal properties). External phenotype data will also be referred to as "3D morphology data," where morphology refers to the external shape and structure of a seed; thus, 3D morphology data describes the shape, appearance, and external structure."3D morphology image data" can therefore be image data depicting the morphology of the seed, such as a height profile image, while "3D morphology measurement data" can be measurement data relating to the morphology, such as the volume or height, length, width, etc., of the seed, or the presence of external damage. Data relating to the internal phenotype are also referred to as 2D projection data. "2D projection image data" can therefore be image data representing a projection, particularly an X-ray projection, of the seed interior onto a plane, while "2D projection measurement data" generally refers to measurement data that can be attributed to the seed interior or a projection of the seed interior, such as measurements of the areal extent of the embryo, the areal extent of the seed cavity, the density of various seed organs, the presence of internal fractures or damage, etc.

[0033] Generating "synthetic" phenotypic data from the corresponding 3D tomography image data—in contrast to actually measuring or capturing image data—means generating virtual measurement data or virtual image data from 3D tomography image data. For example, (synthetic) 2D X-ray projection images can be calculated or generated from a 3D X-ray tomography image using image processing techniques; similarly, (synthetic) elevation images can also be calculated or generated from a 3D X-ray tomography image using image processing techniques. Furthermore, synthetic measurement data, in particular the aforementioned 3D morphology measurement data and / or 2D projection measurement data, can be derived from the 3D tomography image data or from the synthetic image data derived from it (such as 2D X-ray projection images or elevation images). DG2.DD16911 1 / 01

[0034] September 24, 2025

[0035] 7

[0036] The present invention therefore proposes a method for assigning quality classes to seeds from one or more samples of a population. The method according to the invention includes a first, preliminary experimental analysis step aimed at assigning one or more quality classes to a seed from which 3D image and / or measurement data have been acquired. These quality classes are derived from specific measurements of quality properties or characteristics, such as weight, shape, appearance, color, etc., of the same seed and / or from quality properties such as the seedling quality (e.g., germination and shoot growth rate, shoot size and shape, etc.) of the resulting seedling. The at least one resulting quality class is then assigned to the 3D image and / or measurement data of this seed.While this experimental analysis step relates to seeds of one or more samples of a population, subsequent classification steps of the inventive method relate to any seeds of initially the same population, whereby corresponding classification models can be optimized or generalized to specific species or seed lots.

[0037] First step of the analysis:

[0038] In a first, preliminary experimental analysis step, three-dimensional (3D) X-ray computed tomography image data, hereinafter abbreviated as 3D-CT (X-ray) image data, are acquired from seeds of one or more samples of a population using a three-dimensional X-ray tomography imaging technique. Other three-dimensional tomography imaging techniques, such as magnetic resonance imaging (MRI), are also suitable. To obtain statistically robust data, especially training data for neural networks or classical image processing, a sufficiently large sample size is required. The size of this sample depends on various factors, such as sample quality and heterogeneity, the type and frequency of the quality features to be distinguished, and the desired classification accuracy.The "capture" of image data should in particular include the recording of corresponding images and the generation of corresponding 3D image data, as is generally known from the state of the art.

[0039] The 3D tomography image data essentially already includes all 3D information about each examined seed, such as information on external structure and shape, i.e., the morphology of a seed ("external parameters"), as well as information on embryo size, grain filling level, possible damage, etc., i.e., the internal phenotype ("internal parameters"). For example, a classification based on measured values ​​for DG2.DD16911 1 / 01

[0040] September 24, 2025

[0041] 8. These internal and / or external parameters of the seed are determined using 3D X-ray CT or other measurement methods. Image data can be advantageously evaluated using an image processing algorithm that, through an artificial intelligence approach or defined threshold values ​​for measured values ​​from classical image processing, can perform an automatic classification into quality classes. However, visual or other algorithmic classification and annotation of the seed based on visual observation or evaluated (image) data from other sensors is also possible. Thus, quality parameters, as a basis for classification into quality classes, can often be read directly from the seed or with the aid of additional sensors.

[0042] Based on the three-dimensional tomography image data acquired for the seeds, and especially on measurements from other applied measurement methods, quality characteristics or parameters can be assigned to the seeds, enabling their classification into one or more quality classes (e.g., underfilled, broken, small, etc.). Visually determined or image-processed quality characteristics, such as shape, size, specific gravity, etc., can thus be incorporated into the determination of the seeds' quality classes. The definition and delimitation of quality classes depend on the purpose and use of the seeds. For example, a class of "small seeds" may be defined differently depending on its intended use, for instance, by using threshold values.

[0043] Preferably, these seeds from the sample are subsequently germinated, and the seedling quality of each germinated seed is determined. This is expediently done using the aforementioned phenoTest method, in which the development of the seedling during germination can also be visualized three-dimensionally using 3D X-ray computed tomography, and the phenotype can be automatically measured. However, an additional or alternative form of seed or seedling evaluation (e.g., visual or optical measurement methods) can also be performed.

[0044] In particular, seedling quality can be determined based on the completeness of germination, its speed, and / or the speed of growth (biomass production per unit of time) and / or the size of the shoot. Quality classes are also derived from seedling qualities determined in this way and / or from other measured quality characteristics. These quality classes of seedlings are DG2.DD16911 1 / 01

[0045] September 24, 2025

[0046] 9 subsequently assigned to the 3D tomography image data of the respective corresponding seed from which the seedling emerged.

[0047] In the well-known ISTA classification, for example, there are three quality classes: "germinated," "ungerminated," and "abnormal." In the present invention, it is possible in principle to define at least one, but preferably more, quality classes.

[0048] These seed-seedling data pairs will subsequently serve as training data for predicting germination in unknown seeds and for use in the classification of seeds in a later sorting process according to these quality classes.

[0049] It should be emphasized that the acquisition of three-dimensional tomography image data (such as 3D CT image data) of the seeds can be performed using various experimental setups. In practice, for example, the seeds of the training sample can be X-rayed in isolation in appropriate holders, each with its own position identifier (here, too, it is advisable to separate the seeds from the substrate and the holder using image processing). With this method, the original "numbering" of the seeds must be strictly maintained during subsequent germination, as data sets from different acquisition methods are obtained, which then need to be correctly matched.

[0050] In the second method of CT imaging in the future germination medium, e.g., using the phenoTest procedure, the seed can be X-rayed in the same medium and at the same location where germination is to occur (e.g., in filter paper). Here, it is advisable to apply a segmentation algorithm that separates or segments the CT images of individual seeds from their surroundings (filter paper, container wall, other seeds in the X-ray path). This method is very reliable because the seeds can be X-rayed directly on site, initially dry, and later the seeds and seedlings can be X-rayed in a moist state during germination. This type of 1:1 analysis can be applied equally well in soil or another germination medium (instead of filter paper). In the latter case, the vigor and tolerance to physical / mechanical stress (covering soil layer) or biochemical stress (chemical soil composition) could also be tested simultaneously.DG2.DD16911 1 / 01.

[0051] September 24, 2025

[0052] 10

[0053] This preliminary first analysis step therefore assigns a quality class, or preferably several quality classes, to each seed of one or more samples of a population and the corresponding 3D tomography image data of this seed.

[0054] Since 3D imaging tomography methods (e.g., 3D X-ray CT or MRI) are currently not technically feasible with the necessary throughput and resolution for measurements or classifications in flow mode, the 3D X-ray data and the associated quality classes cannot be directly translated into classification and / or sorting decisions in flow mode. The invention therefore proposes at least one further step to enable the use of the quality classes for classification or sorting with existing high-throughput measurement methods (e.g., laser light sectioning and / or 2D flow X-ray).

[0055] In the following, a “first embodiment” of the further step for assigning quality classes refers to the data on the external phenotype explained above, while a “second embodiment” of this further step refers to the data on the internal phenotype explained above.

[0056] First embodiment of the further step:

[0057] According to the invention, at least one further step follows the first analysis step. In a first embodiment, three-dimensional (3D) morphology images and / or measurement data (generally data on the external phenotype) are first synthetically generated for seeds of the sample from the corresponding 3D tomography image data of the seed for at least one predetermined view. Measurement or image data of the "morphology" or "external phenotype" refers to the surface structure and / or the geometry of the seed, i.e., externally visible measured values ​​such as the volume of the seed or its dimensions in all spatial directions, in particular, for example, the height profile of the seed at a specific spatial position, e.g., on a planar surface.

[0058] It should be noted again that the corresponding 3D tomography image data acquired generally also contain all image information regarding the overall context surrounding the individual seed in the direction of transmission (e.g., container, filter paper, and other seeds in front of or behind the seed in question, or, as described above, the substrate and holding device for the seed). The image data of the individual seed must therefore be expediently separated from its respective environment. DG2.DD16911 1 / 01

[0059] September 24, 2025

[0060] 11. This will be achieved, for which image processing methods can be used in particular. "Generating" 3D morphology data here refers specifically to calculating and / or extracting such measurement data, especially using image processing methods, from the associated 3D tomography image data.

[0061] Unlike a seed "floating" freely in space, such 3D morphology data, or rather its elevation profiles such as elevation images, can depend on the spatial orientation of the seed as it is detected by a sensor (also referred to here as the "view" of the seed) if the seed is lying in different positions on a flat surface (e.g., on a conveyor belt). This is particularly relevant for seeds that could assume a number of physically stable or probable positions on a planar surface or in free fall, meaning that the views of this seed from the perspective of a (hypothetical at this stage) sensor can differ.According to the invention, 3D morphology data are generated for at least one, but in particular for several, such views of the seed, wherein each such view advantageously corresponds to a possible physically stable position of the seed on a planar surface or a physically probable view in free fall. For this purpose, it is advantageous to simulate or calculate physically stable or probable positions of the seed that it can assume on a planar surface or in free fall for the 3D volume of the seed isolated from its surroundings. As explained below, it is advantageous if this view corresponds to, or includes, the view from which a subsequent (real) 3D image of the seed is acquired for a high-throughput sorting process (for example, laser light sectioning).

[0062] The terms "synthetic" or "virtual" images or data are used below when they are generated for a seed isolated from its background for a specific view or position of the seed in three-dimensional space, preferably corresponding to a physically stable or probable position. Therefore, in this further step, the synthetically or virtually generated views of the 3D X-ray volumes of the seeds for different views / positions are transposed, in particular, into synthetic / virtual elevation images, such as those generated by laser triangulation.

[0063] The generated 3D morphology and external phenotypic measurement data are now assigned the same quality classes as the corresponding quality classes of the original 3D tomography image data. This means, for example, that 3 DG2.DD16911 1 / 01

[0064] September 24, 2025

[0065] 12 different views (elevation images) of the same seed are assigned the same quality classes as the original 3D image volume.

[0066] Second embodiment of the further step:

[0067] The present invention provides a second embodiment comprising a further step, wherein the second embodiment can be carried out in addition to or as an alternative to the first embodiment described above. General statements made regarding the first embodiment shall also apply to the second embodiment, unless otherwise described or unless this leads to contradictions. In any case, the second embodiment, like the first embodiment, follows the first upstream empirical analysis step as detailed above.

[0068] According to the second embodiment, synthetic / virtual two-dimensional (2D) projection image data and / or 2D internal phenotypic measurement data (generally, data on the internal phenotype), such as the area, shape, and density of the embryo, are generated for the seeds in the sample (instead of or in addition to the external phenotypic data discussed above, i.e., virtual 3D morphology data such as elevation images). The "generation" of this 2D image data includes, in particular, image processing methods, whereby in this case, artificial or synthetic 2D projection image data can be generated or "created"—that is, calculated and / or extracted—from the corresponding 3D tomography image data of the seed (from the first analysis step), at least in a predetermined projection direction. Again, it is advantageous to isolate or separate the 3D tomography image data of the individual seed from its respective environment beforehand, as explained above.Subsequently, these virtual 2D projection image data of the seed, which contain image information from the entire transmission path of a seed, are assigned the same quality classes as those of the original 3D tomography image data of this seed (from the first analysis step). The same applies to measurement data on the internal phenotype calculated from these 3D tomography image data, such as grain filling degree, the presence of necrosis, infections, or damage from mechanical impact or insects. These internal phenotypic measurement data are also assigned the same quality classes as those of the original 3D tomography image data of this seed (from the first analysis step). DG2.DD16911 1 / 01.

[0069] September 24, 2025

[0070] 13

[0071] It is also advantageous if physically stable or probable positions of the seed are simulated or calculated from the freestanding 3D tomography data and the synthetic projections are generated or synthesized as image or measurement data for these positions. As explained below, it is useful if the aforementioned projection direction coincides with that in which the subsequent (real) "2D X-ray images" of the seeds are acquired or measured for analysis in a stationary state and / or a sorting process in a flow, e.g., using flow-through X-ray imaging. The (synthetically) generated 2D projection image data typically differ for a given external shape of a seed depending on the projection direction, i.e., depending on the spatial position of the seed in which it is detected or irradiated by the sensor.It is therefore advantageous to choose the projection direction such that it is perpendicular to the seed, which is in one of the possible physically stable positions that the seed can assume on a flat surface, e.g., on a conveyor belt, depending on its shape. For sorting in free fall, this applies accordingly to the physically most probable positions. These spatial positions can be determined experimentally, but can also be simulated or calculated using data modeling. Depending on the possible position, a different 2D projection image is obtained.

[0072] All 2D projections of a seed in its possible different positions, created on the basis of the same 3D CT original volume, are then assigned the quality classes of the original volume (as above, for example, with the elevation images).

[0073] Use of synthetic data for the classification of unknown seeds in the population

[0074] In a subsequent classification step, which will be explained later, any seeds from the population can be captured using appropriate optical imaging methods. The resulting real 2D or 3D image data can then be automatically compared with the previously generated synthetic 2D projection image data or synthetic 3D morphology data (elevation image data). Based on the highest degree of agreement, the seeds can be assigned corresponding quality classes, enabling classification for a sorting process (or simply for other seed descriptions). The same applies to real 2D or 3D measurement data, which are automatically compared with the previously generated internal or external phenotypic measurement data and assigned corresponding quality classes based on the highest degree of agreement.The method according to the invention can thus, for the first time, perform the analysis in the first step, which in particular can correspond to the “phenoTest” discussed at the outset, DG2.DD16911 1 / 01.

[0075] September 24, 2025

[0076] 14 empirically derived, objectively measured quality parameters of the seed and / or seedling (“germination quality classes”) are used to determine seed quality classes, and these quality classes are subsequently used as annotations of the 3D tomography image data and the derived internal and external phenotypic image data or measurement data of the dry seed – particularly for training neural networks – for the automatic classification of seeds in the high-throughput sorting process.

[0077] Creation of classification models

[0078] The invention further proposes an assignment of one or more quality classes to any seeds of the population, wherein, firstly, synthetic data on the external phenotype, i.e., 3D morphology image and / or measurement data and associated quality classes for sample seeds are determined in a method described above, and wherein, in a subsequent step, real optical 3D image data are acquired from any seed of the population by means of an optical imaging method, wherein the actually acquired 3D image data are compared with the synthetic data on the external phenotype, and, depending on the degree of agreement, the corresponding quality classes are assigned to the seed.Alternatively or additionally, data on the internal phenotype and corresponding quality classes for sample seeds are first determined using a procedure described above. In a subsequent step, real 2D image data are acquired from an arbitrary seed from the population using an optical imaging method. The acquired 2D image data are then compared with the synthetic data on the internal phenotype, and the corresponding quality classes are assigned to the seed based on the degree of agreement. In particular, depending on the degree of agreement, the corresponding quality classes, along with their respective probability of agreement, can be assigned to the seed.

[0079] The invention further relates to a method for assigning quality classes to (any) seeds of a population for the purpose of subsequent description, classification, and / or sorting. In a further step, the synthetic data generated in this way, each annotated with the corresponding quality classes for simulated stable and / or probable positions of a large number of seeds from a sample of a population, can now be incorporated into a [DG2.DD16911 1 / 01]

[0080] September 24, 2025

[0081] 15 several classification models for unknown seeds of this population are implemented.

[0082] For this purpose, as described above, synthetic 3D external and / or 2D internal phenotypic image and / or measurement data are extracted from the respective 3D CT image data and annotated with the quality class(es) of the corresponding seed. It is advantageous to use this synthetically generated 3D external and / or 2D internal phenotypic image and / or measurement data as training data for a classical image analysis algorithm and / or a neural network (Kl). Specifically, this would involve training with the external and internal phenotypic measurement data and / or with the synthetically generated elevation images and 2D projection images in the possible orientations of the seeds. From this, a classification model can be calculated, which can then be applied to real-world 3D external and / or 2D internal phenotypic image and / or measurement data from any seeds in the population, particularly in high-throughput applications.

[0083] Such a classification model, for example based on artificial intelligence, could be trained as follows: All synthetically generated image and / or measurement data of seeds annotated with a specific quality class (e.g., "ungerminated seed") could be trained into a neural network under this class or label. Based on this trained data, the neural network would then classify measurement and image data of unknown seeds from the same or a different population into one or more quality classes according to the "best fit" principle, as is well known in the state of the art.

[0084] Real classification step

[0085] In a subsequent real-world classification step, in the first embodiment of the invention, optical 3D image data of any seeds from the population are acquired using an optical imaging method (here, for example, laser light sectioning) while they are flowing on a conveyor belt, a stationary surface, or in free fall. The acquired "real" 3D image data of a seed are then assigned the corresponding quality classes based on the best-fit 3D morphology data from the training data. These quality classes form the basis for subsequent classification and / or sorting. "Acquisition" of optical 3D image data here refers specifically to the recording of corresponding 3D images and the generation of corresponding 3D image data from the recorded 3D images. Any form of optical imaging method is suitable. DG2.DD16911 1 / 01

[0086] September 24, 2025

[0087] 16

[0088] 3D imaging, in particular laser light sectioning or stereoscopic techniques, is used. With such an optical 3D imaging method, corresponding 3D image data is acquired from any seeds in the population. This 3D image data contains information about the external structure and / or shape of the seeds, i.e., their morphology. Therefore, the term "3D image data" should explicitly include the various dimensional values ​​of a seed (length, height, width, volume, etc.). Accordingly, these acquired real 3D image data can be assigned quality classes, analogous to the seeds in the sample. This is done by comparing the acquired 3D image data with the synthetically generated 3D morphology data (i.e., the 3D image and measurement data) for different positions and identifying the 3D morphology data with the highest degree of agreement, to which quality classes have already been assigned in the previous analysis step.

[0089] If seeds need to be sorted into different fractions, sorting rules or instructions can be defined based on the quality classes of unknown seeds. For example, in a simple case, the quality classes "ungerminated" and "abnormal" can be assigned to the "reject" fraction, and the "germinated" fraction to the "utilization" fraction. In practice, it is often useful to have more quality classes to create more fractions or to refine the sorting fractions. For sorting purposes, these quality classes can then be translated into sorting instructions, allowing different quality classes to be grouped into a single fraction and sorted accordingly. Rules can also be defined to determine which quality classes take priority for sorting decisions in case of conflict.For example, if sorting is to be done in one step for seed size and germination capacity, a seed that has been assigned to the quality classes "large" (i.e. "good") and "not germinated" (i.e. "bad") by the classification model could be classified in the sorting fraction "reject" because germination capacity was given priority over seed size.

[0090] According to the second embodiment, which can be used alternatively or additionally to the first embodiment, quality classes are first determined for seeds from a sample of the population according to the method described above. In a subsequent step, real 2D image data, in particular 2D X-ray image data, are acquired from any seeds in the population using an optical imaging method. The acquired real 2D image data of a seed are then assigned corresponding quality classes based on the best-fit synthetic 2D projection image data from the previously examined sample, and a DG2.DD16911 1 / 01

[0091] September 24, 2025

[0092] 17. Subsequent classification or sorting decision is derived. In the subsequent classification step of the second embodiment, true 2D image data are thus acquired in flow or on a stationary planar surface of any seeds from the population using an optical imaging method (e.g., 2D flow X-ray). The "acquisition" of image data again includes, in particular, the recording of images and their evaluation or processing. This specifically includes the recording of 2D X-ray images of the seeds. This optical imaging method has the advantage that not only the external form of the 2D projection is recorded, but also information about the interior of the seeds (areal size and shape of the embryo, etc.) can be obtained. This thus allows an automatic comparison of the generated synthetic 2D projection image data with the actually generated 2D image data.Measurement data of any seed and its corresponding assignment to quality classes. Based on this assignment, the captured real 2D image data of a seed are also assigned one or more quality classes of the original 3D seed volume from which the virtual 2D projections were generated. Subsequently, in an analogous manner to the first embodiment, each seed is classified into a class or sorting fraction according to predefined rules based on its assigned quality classes.

[0093] As explained above, the two embodiments of the invention can be used independently of each other to assign quality classes to seeds within a population. However, a combination of both embodiments can increase the accuracy of the classification. For example, the acquired 2D image data, in particular 2D X-ray image data, of a fully filled seed of ellipsoidal shape and a partially filled seed of spherical or otherwise ellipsoidal shape may be approximately the same if their 2D projections result in (approximately) identical images. In this respect, additional information on the external shape, i.e., 3D morphology data (e.g., the height of the seeds), can provide further valuable information that can be used in addition to the classification. For example, if one of the two quality classes indicates "poor germination quality," the seed in question can already be classified as "reject."As already shown for the first embodiment, rules and priorities of the quality classes can be defined, which are applied in case of possible contradictions between quality classes and define the assignment to a class and / or sorting fraction.

[0094] Overview of an exemplary sequence of possible steps in an embodiment of the method according to the invention: DG2.DD16911 1 / 01

[0095] September 24, 2025

[0096] 18

[0097] 1. Generating 3D X-ray volume image data of individual seeds using computed tomography or other methods (e.g., MRI imaging)

[0098] 2. Separation and isolation of seed image data from image data of the environment (filter paper, holder, other seeds, etc.)

[0099] 3. Optional segmentation, extraction and calculation of measured values ​​for external and internal phenotypic properties of the isolated seed for classification into quality classes based on measured values.

[0100] 4. Assignment of one or more quality classes to the 3D image and / or measurement data of each individual seed.

[0101] • via the measurements from 3D X-ray CT scans of the semen and / or

[0102] • via measurements from other procedures on the seed and / or

[0103] • via visual or automatic annotation using image data on the seed (e.g. classification as “broken”, “necrotic”, “contamination”, etc.) and / or

[0104] • about the germination of individual seeds and measurement / determination of their germination quality and seedling characteristics and annotation of this empirically determined seedling data for the individual seeds

[0105] 5. Generation of various virtual views of the 3D CT volume image data, corresponding to the most probable physically stable positions on a plane (e.g., a conveyor belt) or physically probable positions in free fall.

[0106] 6. Generation of synthetic (virtual) 3D height profiles of these different X-ray volume views related to these (virtual) layers

[0107] 7. and / or generation of synthetic (virtual) 2D X-ray projections for each of these positions / views

[0108] 8. Assignment of the same quality classes (see point 4) to all synthetic 3D elevation image and / or 2D X-ray image data generated from the same 3D semen volume and / or the derived measurements from classical image processing for the different positions for the external and / or internal phenotype.

[0109] 9. Creation of classification models in which internal and external phenotypic parameters, either as synthetic images or as measured values, are used as a basis for specific positions of a seed annotated with one or more quality classes. The classification model (in one implementation, an artificial intelligence model) is intended to use this training data to identify unknown DG2.DD16911 1 / 01

[0110] September 24, 2025

[0111] 19

[0112] Classifying seeds from a population into corresponding quality classes with a certain probability

[0113] 10. and / or creation of classification models for predicting, for example, germination behavior based on the external and internal phenotypic properties of the seed (via training of Kl or with the help of threshold values ​​for measured values)

[0114] 11. Conveying unknown seeds from the same population onto a plane, for example a conveyor belt, or by free fall, and measuring the seeds with appropriate measuring devices (here preferably laser light sectioning and / or 2D flow X-ray) and comparing image data and / or measured values ​​of real height profiles generated in the flow and real 2D X-ray projections of unknown seeds with the aim of finding the greatest possible similarity to trained seeds and their quality classes in near real time and classifying the unknown seeds accordingly into these quality classes, for example for sorting purposes.

[0115] 12. and / or calculation of probabilities for the correctness of the classification into one or more quality classes and use as an additional parameter for the classification and / or sorting.

[0116] 13. or sorting with other conventional sorting / cleaning machines based on measured values ​​(e.g. size, specific weight) for quality classes, for example to control processing with sieves and / or gravity.

[0117] Details:

[0118] If, according to the first embodiment, the 3D tomography image data or the synthetic 3D morphology measurement data extracted therefrom, such as synthetic elevation images, are to be used as training data for the external appearance of the seeds, for example, in a subsequent laser triangulation as an optical imaging method for any seeds to be sorted from the seed lot, the problem can arise, particularly when conveying the seeds on a conveyor belt, that the position of the seeds on the conveyor belt can be random and does not necessarily correspond to the view of the "trained seed," i.e., the corresponding measured seed from the sample of the population (e.g., seed lot). Therefore, in one embodiment of the method according to the invention, several relevant 3D views of the dry seed, annotated with quality classes, e.g., from germination testing, are used as training data for the sorting process.These spatial positions advantageously correspond to the possible stable or most probable positions of each seed. In other words, for different views of a seed, a 3D morphology image and / or - DG2.DD16911 1 / 01.

[0119] September 24, 2025

[0120] 20 measurement data sets were generated and assigned the same quality classes as the original 3D volume.

[0121] In the second embodiment of the invention described above, direct use of the anatomical, i.e., internal, 3D tomography image data, for example from the preceding computed tomography scan of the seeds, as training data for subsequent sorting, for example according to the 2D flow X-ray principle, is generally not readily possible. While the computed tomography principle does involve generating 2D X-ray images, for example of the container with the seeds, from hundreds of angles to create the 3D reconstructions, i.e., the 3D tomography image data, each of these 2D views contains all the information of the X-ray path. Even if a specific seed can be located within the 3D image volume, "cutting out" this image area from the 2D image of the overall context would include not only that seed, but also all structures in front of and behind it (i.e., any other seeds, filter paper, and the container wall).Therefore, in such cases, the inventive method proposes the following for generating synthetic 2D projection image data from the associated 3D tomography image data of a seed: Instead of the original 2D data of the entire vessel containing the seed, which are initially generated during 3D computed tomography, only the 3D image volume data of the individual seeds isolated from the vessel and the surrounding medium are to be used, and synthetic 2D projections are to be generated from these 3D image volume data.

[0122] In a suitable embodiment, synthetic / artificial / virtual 2D projection image data for a seed is generated by isolating the seed from its environment using image processing within the 3D tomography image data. Subsequently, at least one 2D projection image is calculated for the isolated seed, initially from a predetermined spatial direction. This is particularly advantageous when using the aforementioned phenoTest method, in which a large number of seeds are germinated in a germination vessel, separated from one another by filter paper, in order to determine the germination quality of each seed.

[0123] It should be noted that the computational generation of synthetic 2D projections from 3D models or 3D tomography image data is generally state of the art. For example, Algebraic Reconstruction Technique (ART) should be mentioned here. An exhaustive discussion of this technique would exceed the scope of this document. DG2.DD16911 1 / 01

[0124] September 24, 2025

[0125] 21

[0126] The patent application is too extensive; therefore, for further information on this technique, please refer to the literature.

[0127] As already explained above, the seeds of the seed lot are expediently conveyed on a conveyor belt or by free fall, particularly at high throughput (see definition above), while they are assigned quality classes that can be translated into a sorting decision for each seed according to defined rules. It is advantageous to sort the seeds of the seed lot into different sorting fractions according to their assigned quality classes. Seed sorting can be carried out, for example, by pneumatic sorting, in which suitably positioned air nozzles subject the seeds to air pulses as they leave the conveyor belt in free flight or free fall. Depending on the sorting instructions, this either prevents the seeds from drifting or deflects them in such a way as to enable sorting into different fractions.For further information on this technique, please refer to the literature.

[0128] It is possible and advantageous to synthetically generate 2D projection image and / or measurement data from the corresponding 3D tomography image data of the sample seed for at least one predetermined projection direction onto the sample seed, which is located in at least one physically stable or probable position. In particular, the at least one predetermined projection direction for generating synthetic 2D projection image and / or measurement data can be a spatial direction that corresponds to the one in which a subsequently used sensor (for example, 2D flow X-ray and / or laser light sectioning) would detect this seed in its respective spatial position. For several such possible stable or probable positions, it is advantageous to generate several corresponding synthetic 2D projection image and / or measurement data sets and use them as training data.

[0129] As previously explained, it is advantageous if the generated synthetic 3D morphology image and / or measurement data of the sample seeds are created in such a way that they correspond to real 3D morphology image and / or measurement data of any other seeds in the population acquired using a high-throughput classification procedure, in order to be compared with these data and assign these seeds to the corresponding quality classes. Similarly, it is advantageous if the generated synthetic 2D projection image and / or measurement data of the sample seeds are created in such a way that they correspond to real 2D projection image data acquired using a high-throughput classification procedure. DG2.DD16911 1 / 01

[0130] September 24, 2025

[0131] 22 and / or measurement data from any other seeds in the population, in order to be compared with these, with the aim of assigning these seeds to the corresponding quality classes.

[0132] It is advantageous to assign any seed of the population to a classification class and / or a sorting class based on its assigned quality classes, in particular together with the respective probability of agreement, wherein these seeds of the population are conveyed in particular on a conveyor belt or in free fall, especially in high-performance throughput, while quality classes and / or classification and / or sorting classes are assigned to them in particular in real time.

[0133] As already explained above, it is advantageous to use artificial intelligence or neural networks in the implementation of the invention. In particular, artificial intelligence or neural networks can be used to assign quality classes to any seeds in the population. Specifically, 3D morphology data generated from different views, annotated with the corresponding quality classes of the 3D seeds, can be used as training data. Based on such a classification model, real 3D image data of a seed (e.g., acquired using a laser light sectioning method) can then be compared with the synthetically generated or virtual 3D morphology data (from the 3D tomography data) and, in the case of a "best match," assigned to the corresponding quality class(es).Furthermore, the virtual 2D projection image data generated for several predetermined projection directions can be used as training data for the assignment of quality classes by comparing the real 2D image data of a seed (e.g., acquired using flow-through X-ray imaging) with the synthetic 2D projection image data generated from the 3D tomography data and classifying the seeds into one or more corresponding quality classes if there is a "best match".

[0134] Since this comparison generally cannot result in a 100% match between the training data and the data generated in the flow of the arbitrary seeds, probabilities for the correctness of the assignment to quality classes can be determined using appropriate statistical methods.

[0135] In a further embodiment of the invention, a seed and / or seedling analysis is subsequently carried out again for at least a part of a sorting fraction obtained after sorting according to the present invention, in order to determine the DG2.DD16911 1 / 01

[0136] September 24, 2025

[0137] The system allows for the review and, if necessary, modification of the 23 originally assigned quality classes. This enables corrections to the seed annotations and / or sorting rules, which can then be automatically used to further optimize the analysis and classification algorithm.

[0138] In a further embodiment of the method according to the invention, which has already been discussed above, other or additional measurement parameters determined for the seeds of the sample from the seed lot are acquired using appropriate measurement methods, to which the quality classes of the corresponding 3D seed can then be assigned. In this embodiment, synthetic 3D morphology measurement data can first be generated directly or indirectly from the associated 3D tomography image data of a sample seed for at least one predetermined position of the sample seed, and the synthetic 3D morphology measurement data can be annotated with the quality class(es) assigned to the associated 3D tomography image data.Alternatively or additionally, synthetic 2D projection measurement data can be generated directly or indirectly from the associated 3D tomography image data of a sample seed for at least one predetermined position of the sample seed, and the synthetic 2D projection measurement data can be annotated with the quality classes assigned to the associated 3D tomography image data.Subsequently, real 3D morphology measurement data can be acquired from a seed in the population using appropriate measurement methods and compared with the synthetic 3D morphology measurement data. Depending on the degree of agreement, the seed is then assigned the corresponding quality classes. Alternatively, real 2D projection measurement data can be acquired from a seed in the population using appropriate measurement methods and compared with the synthetic 2D projection measurement data. Depending on the degree of agreement, the seed is again assigned the corresponding quality classes. In particular, the quality classes assigned in this way are also used to determine the aforementioned classification class and / or sorting class. These measurement parameters can also be used as training data, provided that suitable sensors are available for classifying any seeds in the population.Such measurement methods can include, for example, the determination of color spectra or magnetic resonance spectra. The correlation of such additional measurement parameters with the quality classes generated by 3D tomography can further improve the classification accuracy in subsequent classification steps, provided suitable sensors are available. DG2.DD16911 1 / 01.

[0139] September 24, 2025

[0140] 24

[0141] Furthermore, the invention relates to a device for carrying out the method according to the invention. The properties and configurations of this device are described analogously to the description of the method according to the invention.

[0142] Examples of implementation

[0143] The present invention and its embodiments will be explained in more detail below using an exemplary embodiment in conjunction with the figure.

[0144] Figure 1 shows a possible flow diagram of the process proposed according to the invention using a beech seed as an example.

[0145] In detail:

[0146] First experimental analysis step according to Fig. 1A

[0147] Step 1.1 (Figure 1A): a) 3D-CT X-ray image acquisition of a beech seed from a sample of a population of beech seeds in a container; b) Separation of this seed from the container and freeing it in space to obtain the 3D-CT image data (“3D image volume”) of the seed and / or, in particular, additional measurement of the external and / or internal phenotype, in particular using image analysis methods, especially of the acquired 3D-CT image data.

[0148] Optional germination of the seed, for example in filter paper for the aforementioned phenoTest procedure or in another medium, e.g. soil.

[0149] Step 1.2 (Figure 1A):

[0150] Determination, in the present example, of a quality class "Fill level = well filled" based on the quality characteristics determined in step 1.1b), for example by classifying the measurement data of the internal phenotype via threshold values. Further exemplary quality classes for quality determination based on measurement data of the internal / external phenotype could be: "Seed size = large" or, in the case of a poor seed, "Damage = broken", "Pre-germination = pre-germinated", etc. DG2.DD16911 1 / 01

[0151] September 24, 2025

[0152] 25

[0153] Optionally, from step 1.1c), the seedling quality can be determined and, in this example, classified into a germination quality class of "normally germinated." This classification can be done visually, but preferably using image processing with the phenoTest method mentioned earlier. In the case of indirect quality determination via the seedling in step 1.1c), the seedling quality class (in the example, "germination quality = normally germinated") is additionally assigned to the seed. Other exemplary seedling quality classes could be "growth = fast," "vigor = high," "seedling size = large," as well as the corresponding lower quality classes "growth = slow," "vigor = low," "seedling size = small," etc.

[0154] The 3D volume of the seed in the present embodiment is thus now annotated with the quality class 1 "filling level = well filled" and the quality class 2 "germination quality = normally germinated", whereby according to the invention only one quality class or a plurality of quality classes can be assigned to a seed based on the direct seed quality properties and / or the indirect seedling quality properties.

[0155] The following second step involves the generation of synthetic 3D morphology data and / or 2D projection data according to Figure 1B.

[0156] Step 1.3 (Figure 1B):

[0157] First, a simulation or calculation is performed, in this example for a three-sided beech seed, to determine the possible, physically stable spatial positions for the isolated 3D image volume that the seed can assume relative to a sensor (e.g., on a conveyor belt). Three possible spatial positions are identified for the three-sided beech seed.

[0158] Step 1.4 First embodiment (Figure 1B): a. Generation of synthetic (virtual) elevation images (3D morphology image data) based on the calculated / simulated different positions of the 3D image volume (from step 1.3), as is known per se from the prior art; b. and / or or, in particular, additionally determination of 3D morphology measurement data, i.e., the external phenotype of the seed for the calculated / simulated different positions via measurement data (length, width, height, volume, etc.) using image processing methods.

[0159] Step 1.5 Second embodiment (Figure 1B): DG2.DD16911 1 / 01

[0160] September 24, 2025

[0161] 26 a) Alternatively or additionally, generation of synthetic (virtual) 2D X-ray projections (2D projection image data) based on the calculated / simulated different positions of the 3D image volume (from step 1.3); b) and / or or, in particular, additionally, determination of 2D projection measurement data, i.e., the internal phenotype for the calculated / simulated positions via measurement data (fill level, fractures, damage) using image processing methods.

[0162] Step 1.6 (Figure 1C): a) Assigning the quality classes of the 3D image volume of the seed to the synthetically generated height images and / or 2D X-ray projections and / or to the external and / or internal phenotypic measurement data from steps 1.4 and / or 1.5. b) Training a classical image analysis algorithm or a neural network (Kl) with the synthetic data annotated with quality classes from step 1.6a) to generate a classification model that can classify unknown seeds into corresponding quality classes based on their similarity to the trained, annotated synthetic data.Generation of a classification model for classifying unknown seeds into the corresponding quality classes based on the principle of "Best Fit" with a probability of agreement per class, using measurements of the external and / or internal phenotype recorded with real sensors and / or the actually generated 2D X-ray images and / or the actually measured 3D height profiles.

[0163] Step 1.7 (Figure 1C):

[0164] Alternatively or additionally, rules are defined for classifying quality grades into higher-level classification classes or – in the case of subsequent seed sorting – into sorting classes, for example, "well-filled" + "normally germinated" = "good fraction" for later sorting into fractions. In case of conflicts, clear rules or priorities can be defined here, as is already the case in practice with state-of-the-art sorting algorithms for seed sorters. For example, "well-filled" and "not germinated" could fall into the sorting class "rejects" because the quality class "germination" has been assigned a priority above the quality class "fill level".

[0165] Real classification step for any seeds of the population according to Figure 1 D:

[0166] Step 2.1 (Figure 1 D) DG2.DD16911 1 / 01

[0167] September 24, 2025

[0168] 27 a) Detection of unknown seeds with real sensors (in the example “Seed 2”), for example, in flow on a conveyor belt, to generate a height image of the position in which the seed is detected by the sensor (in the example “Position 1”). b) and / or measurement of the external phenotype of Seed 2 based on the height image. c) and / or detection of unknown seeds with real 2D X-ray sensors (in the example “Seed 2”), for example, by flow-through X-ray imaging on a conveyor belt, to generate a 2D X-ray image of the position in which the seed is detected by the sensor (in the example “Position 1”). d) and / or measurement of the internal phenotype of Seed 2 based on the 2D X-ray image.

[0169] Step 2.2

[0170] Classify the unknown seed (here "Seed 2") using the classification model from step 1.6 into the quality classes that most closely match the trained synthetic data ("Best Fit") with a probability <=100% that the seed actually falls into these quality classes. In the example, a correct assignment using the classification model would be, for example: "Seed 2 = well-filled + normally germinated" with assigned match probabilities.

[0171] Step 2.3 (Figure 1 D)

[0172] Alternatively or additionally, the seed (in the example, "Seed 2") can be classified into a sorting class according to the rules defined in step 1.7, based on which a sorting decision is made. In the example, according to the defined sorting rules, Seed 2 falls into the "Good Fraction" sorting class, which means it could be sorted into this fraction on a conveyor belt sorter, for example, by pneumatic blowing (e.g., using a "blow-out map"). Another seed, which would be assigned to the "Reject Fraction" based on its quality class, could then be sorted into a "Reject" fraction.

Claims

DG2.DD16911 1 / 01 September 24, 2025 28 Claims 1. Procedure for assigning quality classes to seeds of a population, comprising: in a first analysis step: Acquisition of three-dimensional, 3D, tomography image data using a three-dimensional imaging tomography method of seeds from one or more samples of the population, Determining quality characteristics of the sample seeds and / or of developing seedlings and classifying them into one or more quality classes based on these characteristics, and assigning the corresponding 3D tomography image data of the sample seed to the respective sample seed; and in a second step: Generating synthetic phenotype data from the associated 3D tomography image data of a sample semen for at least one spatial position of the sample semen and annotating the synthetic phenotype data with the quality classes assigned to the associated 3D tomography image data.

2. Method according to claim 1, wherein the one or more quality classes of a sample seed are determined by germination of the seed, wherein one or more seedling qualities of the germinated sample seed are determined and these are assigned to one or more quality classes.

3. Method according to claim 2, wherein the seedling quality is determined depending on one or more of the parameters: rate of germination, rate of shoot growth and size of the shoot, size and shape as well as direction of growth of the individual plant organs.

4. Method according to one of the preceding claims, wherein the acquisition of the 3D tomography image data of one of the sample seeds comprises the step of isolating the sample seed in question from the image data of its environment.

5. Method according to any one of the preceding claims, wherein the generation of synthetic data on the phenotype is the generation of synthetic data on the DG2.DD16911 1 / 01 September 24, 2025 29 internal phenotype in the form of synthetic 2D projection image and / or measurement data, in particular 2D x-ray projection image and / or measurement data.

6. Method according to claim 5, wherein, to generate synthetic 2D projection image and / or measurement data for at least one spatial position of the sample seed, 2D projection image and / or measurement data are synthetically generated from the associated 3D tomography image data of this sample seed for at least one predetermined projection direction onto the sample seed, which is located in at least one physically stable or probable position.

7. Method according to any of the preceding claims, wherein the generation of synthetic phenotype data comprises the generation of synthetic external phenotype data in the form of synthetic 3D morphology image and / or measurement data.

8. Method according to claim 7, wherein, to generate synthetic 3D morphology image and / or measurement data for at least one spatial position of the sample seed, 3D morphology image and / or measurement data are synthetically generated from the associated 3D tomography image data of this sample seed for one or more predetermined spatial directions, each corresponding to a view of the sample seed in which it is located in a physically stable or probable position.

9. Method according to one of the preceding claims, wherein the generated synthetic data on the phenotype of the sample seeds are generated in such a way that they correspond to real data on the phenotype of any other seeds of the population obtained in a high-throughput classification procedure, in order to be able to be compared with these with the aim of assigning these seeds to the corresponding quality classes.

10. Method for assigning one or more quality classes to any seeds of the population, wherein, first, synthetic data on the external phenotype and associated quality classes for sample seeds are determined in a method according to one of the preceding claims, and wherein, in a subsequent step, an arbitrary seed of the population is used to determine the quality classes of the sample seeds. DG2.DD16911 1 / 01 September 24, 2025 30 optical 3D image data are acquired using an optical imaging method, wherein the actually acquired 3D image data are compared with the synthetic data on the external phenotype, and depending on the degree of agreement, the corresponding quality classes are assigned to the seed; and / or wherein, first, synthetic data on the internal phenotype and corresponding quality classes for sample seeds are determined in a method according to one of the preceding claims, and in a subsequent step, 2D image data are acquired from any seed of the population using an optical imaging method, wherein the actually acquired 2D image data are compared with the synthetic data on the internal phenotype, and depending on the degree of agreement, the corresponding quality classes are assigned to the seed.

11. Method according to claim 10, wherein, depending on the degree of similarity, the corresponding quality classes are assigned to the seed together with the respective probabilities of similarity.

12. Method according to one of claims 10 to 11, wherein a laser light sectioning method is used as an optical imaging method for capturing optical 3D image data.

13. Method according to one of claims 10 to 12, wherein an X-ray imaging method is used as an optical imaging method to acquire 2D image data.

14. Method according to any one of claims 10 to 13, wherein any seed of the population is assigned to a classification class and / or a sorting class based on the quality classes assigned to it, in particular together with the respective probability of conformity according to claim 11.

15. Method according to claim 14, wherein any seeds of the population are conveyed on a conveyor belt or in free fall, in particular at high throughput, while quality classes and / or classification and / or sorting classes are assigned to them based thereon.

16. Method according to one of claims 10 to 15 in conjunction with claim 6, wherein the at least one predetermined projection direction onto the sample seed, for DG2.DD16911 1 / 01 September 24, 2025 31 the synthetic 2D projection image and / or measurement data are generated, corresponding to at least one projection direction in which a sensor for capturing the real 2D image data captures a seed of the population in a physically stable or probable position, in particular if it is located on the plane of the conveyor belt according to claim 15.

17. Method according to one of claims 10 to 16 in conjunction with claim 8, wherein the one or more predetermined spatial directions from which the sample seed is viewed or captured for generating the synthetic 3D morphology image and / or measurement data are selected as spatial directions from which a sensor for capturing the real 3D image data captures a seed of the population in a physically stable or probable position, in particular when it is located on the plane of the conveyor belt according to claim 15.

18. Method according to any one of claims 10 to 17, wherein artificial intelligence is used to assign one or more quality classes and any classification and / or sorting classes derived therefrom according to claim 14 to the arbitrary seeds of the population.

19. Method according to one of claims 10 to 18 in conjunction with claim 14, wherein the seeds of the population to which classification and / or sorting classes are assigned are sorted into corresponding sorting fractions.

20. Method according to claim 19, wherein for at least a part of a sorting fraction of the seeds of the population sorted according to the corresponding classification and / or sorting classes, at least one quality class is subsequently determined again for each of these seeds according to the first analysis step of claim 1 in order to check the originally assigned quality classes and, if necessary, to change them.

21. A method according to any one of claims 10 to 20, wherein real 3D morphology measurement data from a seed of the population are acquired using appropriate measurement methods and these are compared with the synthetic 3D morphology measurement data according to claim 7, wherein the corresponding quality classes are assigned to the seed depending on the degree of similarity, and / or wherein real 2D morphology measurement data are acquired using appropriate measurement methods. DG2.DD16911 1 / 01 September 24, 2025 32 Projection measurement data from a seed of the population are acquired and compared with the synthetic 2D projection measurement data according to claim 5, wherein the corresponding quality classes are assigned to the seed depending on the degree of agreement, wherein in particular the quality classes thus assigned are additionally used for determining the classification class and / or the Sorting class according to claim 14 may be used.

22. Device for carrying out a method according to one of claims 1 to 21.

Images