Evaluation of optical spectral data by means of artificial intelligence

Abstract

For analysis of the state of a plant (2) or of another eukaryote, an evaluation device or an evaluation system (30) has an evaluation circuit (34) which is designed to carry out an evaluation (36) of spectral data. The spectral data or data generated on the basis thereof by preprocessing are divided into multiple parts and analyzed using an analysis logic. The analysis logic comprises at least one attention mechanism.

Description

[0001] EVALUATION OF OPTICAL SPECTRAL DATA USING ARTIFICIAL INTELLIGENCE

[0002] TECHNICAL FIELD

[0003] The invention relates to methods and evaluation systems that can be used in connection with a spectral analysis of a plant or other eukaryote. The invention particularly relates to such methods and evaluation systems in which a wavelength-resolved optical parameter is evaluated. The invention particularly relates to such devices, systems, and methods that can be used with crops or cultivated plants.

[0004] BACKGROUND

[0005] The analysis of eukaryotes, such as plants or plant parts, as well as food products derived from plant or animal substances, is of great importance. It can be used, for example, to ensure adequate nutrient supply and / or to detect potential diseases.

[0006] Conventional methods for analyzing plant health have traditionally relied on human expert assessment. Techniques that use measuring devices to objectively and quantitatively analyze plant health are becoming increasingly important. For example, such techniques can be used to detect over- or under-supply of certain nutrients.

[0007] WO 2019 / 169434 A1, US 7 804 588 B2, US 11 320 307 B2 and DE 10 2018 103 509 B3 disclose exemplary techniques for this purpose.

[0008] To understand the invention disclosed here, features of specific artificial intelligence (AI) models are helpful, namely the so-called attention mechanism (or attention module) and the so-called transformer model. Fundamentals are disclosed, for example, in A. Vaswani et al., "Attention Is All You Need," 2017, Advances in Neural Information Processing Systems 30 (NIPS 2017), pp. 6000-6010 (https: / / arxiv.org / abs / 1706.03762). Transformer models with their attention mechanisms are frequently used in speech recognition. Examples are disclosed in US 10789427 B2, US 11494561 B2, US 2022 / 0005465 A1, and US 2022 / 0108689 A1. Another application of transformer models with their attention mechanisms in image processing is disclosed in A. Dosovitskiy et al., "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale", International Conference on Learning Representations (IPCLR) 2021.These techniques will be referred to below, even though they have not yet been discussed in the context of plant condition analysis by analyzing spectral data recorded on the plant.

[0009] The application DE 10 2023 113 716.6, which was not yet published on the priority date of the present application, discusses the application of a ViT (Vision Transformer) to image data of a plant.

[0010] There is still a need for methods, devices and systems that offer mechanical support, optionally even further automation, in the evaluation for the condition analysis of plants, other eukaryotes or food products derived from them.

[0011] SUMMARY

[0012] The invention is based on the object of providing improved methods, devices, and systems with which information about the condition of plants, other eukaryotes, or food products derived therefrom can be obtained. In particular, the invention is based on the object of providing methods, devices, and systems that provide improvements with regard to condition analysis based on acquired spectra. The invention is also based on the object of providing such methods, devices, and systems that can be used for the evaluation of spectra obtained from crops or cultivated plants.

[0013] According to the invention, a method, an evaluation device or an evaluation system, and a system are specified as defined in the independent claims. The dependent claims define preferred and advantageous embodiments.

[0014] A method for evaluating spectral data according to one aspect of the invention comprises the following: receiving the spectral data by at least one evaluation circuit, wherein the spectral data comprises spectral data acquired from a eukaryote, a part of a eukaryote, or a food product, and analyzing the spectral data by the at least one evaluation circuit. Analyzing the spectral data by the at least one evaluation circuit comprises: dividing the spectral data or data generated therefrom by preprocessing into a plurality of parts, analyzing the plurality of parts using analysis logic, wherein an input of the analysis logic receives the plurality of parts, and wherein an output of the analysis logic provides codings of the plurality of parts, wherein the analysis logic has at least one attention mechanism, and determining at least one evaluation result based on the codings.

[0015] The method provides various technical effects and advantages. Spectral data can be evaluated using at least one attention mechanism. This allows a condition analysis of the eukaryote (e.g., a plant), a part of a eukaryote (e.g., a plant leaf), or a food product to be determined using the attention mechanism. The evaluation can thus be performed using a data-driven analysis logic. Subjective assessment criteria and potentially resulting evaluation errors can be reduced.

[0016] The method may be or comprise a method for determining a condition of a plant or a part of a plant by analyzing the spectral data, wherein the evaluation result is an evaluation result representing the condition.

[0017] This allows the state of a plant or plant part to be determined using analysis logic that can be generated data-driven. By using analysis logic that includes at least one attention mechanism, relative relationships between different parts of the spectral data (e.g., characteristic sequences of local maxima and / or minima that are characteristic of a state) can be efficiently recognized during evaluation and used for state analysis.

[0018] The at least one evaluation result may comprise at least one quantitative analysis result.

[0019] This makes it possible to determine evaluation techniques not only for determining qualitative results (e.g., identification of present and absent pathological conditions or situations), but also for quantitative evaluation.

[0020] The at least one quantitative analysis result may have a continuously variable value from a range of values.

[0021] This allows continuously variable characteristics (e.g. nutrient concentrations) to be determined and output objectively from the spectral data.

[0022] The at least one quantitative analysis result may have a value selected from an ordinal scale.

[0023] This can be used, for example, to express membership in different states represented by the ordinal scale. The ordinal scale can be determined and output for different nutrient concentration ranges (e.g., nutrient deficiency / acceptable concentration / oversupply, or a value from another scale that maps the nutrient concentration to values ​​on an ordinal scale).

[0024] The at least one quantitative analysis result may comprise a concentration of at least one chemical element or at least one chemical compound.

[0025] This allows analysis techniques to be used to determine concentrations based on spectral data. Such concentration determination is of significant importance, for example, in agricultural engineering. The at least one analysis result can include at least one nutrient concentration.

[0026] This allows analysis techniques to be used to determine nutrient concentrations based on spectral data. This type of nutrient concentration determination is of crucial importance, for example, in agricultural engineering.

[0027] The at least one evaluation result may have a value on a nominal scale. The nominal scale may, for example, represent the presence or absence of different types of biotic and / or abiotic stress.

[0028] This means that the analysis logic can be used, for example, for a qualitative analysis.

[0029] The analysis logic may include an encoder. The encoder may include a stack of multiple attention blocks, each of which includes at least one attention mechanism.

[0030] This allows artificial intelligence (AI) models to be used to evaluate the spectral data to determine the codings known, for example, from text analysis or image data analysis. The encoder provides the codings.

[0031] The at least one attention mechanism may comprise at least one self-attention mechanism.

[0032] This allows KL models to be used to evaluate the spectral data in order to determine the codings known, for example, from text analysis or the analysis of image data.

[0033] The at least one attention mechanism may comprise at least one multi-head attention mechanism.

[0034] This allows KL models to be used to evaluate spectral data to determine codings known from text analysis or image data analysis, for example. The use of a multi-head attention mechanism can provide particularly reliable results.

[0035] The method may include wavelength-dependent position coding. This wavelength-dependent position coding preserves relative position information of various features of the spectral data for analysis, for example, information about the wavelength-dependent sequence in which data points of an optical characteristic of the spectral data were acquired.

[0036] This allows the spectral data to be evaluated taking into account the relative position of the various features (such as local maxima and / or local minima) of the spectral data. The method can include preprocessing of the spectral data prior to the analysis logic. During preprocessing, various data can be used in addition to the spectral data, for example, spectra of standards and / or calibration measurements.

[0037] Through preprocessing, "features" can be determined from the spectral data. These features can then be further processed using an encoder that has at least one attention mechanism.

[0038] Preprocessing may include filtering of the spectral data.

[0039] This allows the spectral data to be prepared in such a way that it can be further processed particularly efficiently by the analysis logic with at least one attention mechanism.

[0040] Filtering parameters can be determined data-driven.

[0041] This makes it possible to carry out preprocessing in a data-driven and therefore objective manner.

[0042] Preprocessing can be done using another Kl model.

[0043] This makes it possible to carry out preprocessing in a data-driven and therefore objective manner.

[0044] The method may comprise training the analysis logic or at least a portion of the analysis logic using training data.

[0045] This makes it possible to carry out the evaluation in a data-driven and therefore objective manner.

[0046] The training may comprise training a transformer encoder having at least one self-attention mechanism.

[0047] This makes it possible to carry out the evaluation in a data-driven and therefore objective manner.

[0048] The training may involve joint training of the transformer encoder and preprocessing prior to the transformer encoder.

[0049] This makes it possible to carry out the evaluation in a data-driven and thus objective manner, whereby the functionality of the transformer encoder and the preprocessing are inherently coordinated through training.

[0050] The training data can contain multiple spectra and associated annotations.

[0051] This allows an analysis logic to be trained whose input receives the spectral data.

[0052] The training data can contain multiple spectrally derived data and associated annotations.

[0053] This allows for training analysis logic whose input receives data derived from the spectral data. The training data can include training data derived from the spectra through augmentation.

[0054] This allows a training dataset to be generated that meets specific requirements. This is particularly desirable when, for example, the analysis logic provides results in the form of levels of an ordinal scale, and augmentation ensures that the different levels of the ordinal scale are sufficiently represented in the training data. Similarly, augmentation can be used advantageously when, for example, the analysis logic provides values ​​from at least one continuous value range to ensure that the at least one continuous value range is sufficiently represented in the training data (for example, with a sufficient density per value interval).

[0055] The spectral data may include spectral data collected from a plant, a part of a plant, a plant-like eukaryote, or a plant or animal food.

[0056] This enables a spectrum-based determination of the condition of plants, plant parts, plant-like eukaryotes (such as algae) or a food, whereby the evaluation is enabled using a data-driven analysis logic.

[0057] The method may comprise outputting the evaluation result via a human-machine interface.

[0058] This makes it possible to provide a result of the evaluation, for example, status information concerning a plant.

[0059] The method may comprise outputting the evaluation result or data derived therefrom via a data interface.

[0060] This allows a result of the evaluation, for example status information concerning a plant, to be provided in a form that is particularly suitable for storage or further automatic use.

[0061] The method may include executing a control function based on the evaluation result.

[0062] This allows the evaluation result, for example, information about the condition of a plant, to be used for automatic control functions. This can be of particular interest in agricultural engineering.

[0063] The control function can be executed automatically.

[0064] This means that a result of the evaluation, for example status information concerning a plant, can be used for automatic control functions.

[0065] According to a further aspect of the invention, a method is provided for generating analysis logic of an evaluation system or an evaluation device for evaluating spectral data. The analysis logic comprises at least one KI model having at least one attention mechanism. The method comprises training the KI model using training data by a computing device or a computing system and providing parameters of the trained KI model by the computing device or the computing system to the evaluation system or the evaluation device for application to the spectral data, wherein the spectral data are acquired from a eukaryote (for example a plant or a plant part) or a food product (for example a plant or animal food product).

[0066] This allows the analysis logic to be generated in a data-driven manner, improving the objectivity of analysis results.

[0067] The training may comprise training a transformer encoder having at least one self-attention mechanism.

[0068] This makes it possible to provide the analysis logic for evaluation in a data-driven and therefore objective manner.

[0069] The training may involve joint training of the transformer encoder and preprocessing prior to the transformer encoder.

[0070] This makes it possible to provide the analysis logic for evaluation in a data-driven and thus objective manner, whereby the functionality of the transformer encoder and the preprocessing are inherently coordinated through training.

[0071] The training data can contain multiple spectra and associated annotations. This allows for training of an analysis logic whose input receives the spectral data.

[0072] The training data can contain multiple spectrally derived data and associated annotations.

[0073] This allows an analysis logic to be trained whose input receives data derived from the spectral data.

[0074] The training data may include training data derived from the spectra by augmentation.

[0075] This allows a training dataset to be generated that meets specific requirements. This is particularly desirable when the analysis logic, for example, provides classification results, and augmentation ensures that the different classes are adequately represented in the training data.

[0076] The spectral data may include spectral data collected from a plant, a part of a plant, a plant-like eukaryote, or a plant or animal food.

[0077] This enables a spectrum-based status determination of plants, plant parts, plant-like eukaryotes (such as algae), or a food product, with the evaluation being enabled using data-driven analysis logic. The provision of parameters of the trained AI model for use by the evaluation system or device can include an output via a data interface and / or via a data network.

[0078] This allows the parameters that enable the application of the analysis logic by the evaluation device or the evaluation system to be provided, even if the evaluation device or the evaluation system is located remotely from the computing device performing the training.

[0079] The method may comprise the use of the parameters of the trained Kl model by the evaluation device.

[0080] This allows the data-based trained analysis logic to be applied to spectral data, for example to perform a condition analysis of plants or plant-like organisms (e.g. algae).

[0081] According to a further aspect of the invention, machine-readable data are provided which are determined by the method for training the analysis logic and which can be used by the evaluation device or the evaluation system in order to use the analysis logic to evaluate spectral data which were acquired on a eukaryote, a part of a eukaryote or a food product.

[0082] This makes it possible to achieve the effects described with reference to the methods.

[0083] According to a further aspect of the invention, an evaluation system or an evaluation device for evaluating spectral data is provided, comprising: an interface for receiving the spectral data, wherein the spectral data comprises spectral data acquired on a eukaryote, a part of a eukaryote, or a food product; and at least one evaluation circuit configured to perform at least the following processing steps for analyzing the spectral data: dividing the spectral data or data generated therefrom by preprocessing into a plurality of parts, analyzing the plurality of parts using analysis logic, wherein an input of the analysis logic receives the plurality of parts, and wherein an output of the analysis logic provides codings of the plurality of parts, wherein the analysis logic comprises at least one attention mechanism, and determining at least one evaluation result based on the codings.

[0084] The evaluation system or device achieves various technical effects. Spectral data can be evaluated using at least one attention mechanism. This allows a status analysis of the eukaryote (e.g., a plant), part of a eukaryote (e.g., a plant leaf), or a food product to be determined using the attention mechanism. The evaluation can thus be performed using data-driven analysis logic. Subjective assessment criteria and potentially resulting evaluation errors can be reduced.

[0085] The at least one evaluation circuit, the evaluation system, or the evaluation device can be configured to carry out the method according to one of the exemplary embodiments. The effects achieved thereby correspond to the effects explained with reference to optional method features.

[0086] According to a further aspect of the invention, a system or a device is provided which comprises the evaluation system or the evaluation device according to the invention and a spectral analytical detection device for detecting the spectral data.

[0087] The spectral analysis detection device can comprise one or more of the following configurations: a spectrometer, a hyperspectral camera, a controllable multispectral illumination source in combination with an optoelectric converter (e.g., a camera chip), or a multispectral camera. The spectral analysis detection device can comprise a measurement system that enables the recording of multiple wavelengths. For this purpose, the spectral analysis detection device can comprise, for example, a spectrometer (in which light is split depending on the wavelength), filter-based systems (by blocking light components), single-wavelength diodes, or other configurations. For example, a detector with multi-sensor channels that are sensitive to different spectral ranges can be used.

[0088] This enables the application of the analysis logic, which includes at least one attention mechanism, to spectral data acquired with such modalities.

[0089] The spectral analytical detection device can be configured to detect the spectral data in a reflection arrangement (in which the illumination and wavelength-resolved detection device are arranged on the same side relative to the object) or a transmission arrangement (in which the illumination and wavelength-resolved detection device are arranged on different sides relative to the object) or a transflectance measurement (in which radiation from the illumination penetrates the sample under investigation, is reflected at the back, penetrates the sample a second time and is detected on the same side as the illumination by a wavelength-resolved detection device).

[0090] This enables the application of the analysis logic, which includes at least one attention mechanism, to spectral data acquired in a reflection, transflection, or transmission geometry.

[0091] The spectral analysis detection device may comprise a positioning system and may be configured to determine and transmit position data relating to a position and, optionally, an orientation of the spectral analysis detection device in association with the acquired spectral data. The positioning system may comprise, for example, a navigation satellite system (GNSS) and / or acceleration sensors.

[0092] This makes it possible to assign evaluation results obtained using the analysis logic to different positions in a cultivation area (e.g. a field).

[0093] In the method and system, the spectral analysis detection device can be mounted on a vehicle, an aircraft, or a robot. The method can include controlling at least one actuator for positioning the spectral analysis detection device.

[0094] This allows spectral data to be collected efficiently and partially or fully automatically, and a status analysis associated with the respective acquisition position can be carried out.

[0095] In the method and system, the spectral analysis detection device can be mounted on a sample conveying system, for example, a conveyor belt. The method can include controlling at least one actuator for positioning the sample conveying system and / or the sample on the sample conveying system.

[0096] This allows spectral data to be collected efficiently and partially or fully automatically, and a status analysis associated with the respective acquisition position can be carried out.

[0097] According to a further aspect of the invention, a use of the methods, evaluation devices, evaluation systems, or systems according to embodiments is provided for determining cultivation measures. Using a method, an evaluation device, an evaluation system, or a system according to an embodiment, a condition analysis for plants or plant-like organisms (e.g., algae) can be carried out, wherein, depending on the condition analysis, it is determined which measures need to be taken to improve plant health.

[0098] As a result, the methods, evaluation devices, evaluation systems or systems disclosed here can be used to improve cultivation measures.

[0099] The methods, evaluation devices, evaluation systems, systems, and system components according to embodiments of the invention achieve various effects. In particular, the acquisition and / or evaluation of spectral analytical information can be supported using one or more images acquired with a near-field camera.

[0100] The devices, methods, systems and system components are available in various

[0101] This includes, but is not limited to, the condition analysis of crops or plants in agricultural engineering. BRIEF DESCRIPTION OF THE FIGURES

[0102] Embodiments of the invention are described with reference to the figures. In the figures, similar or identical reference numerals designate elements with similar or identical design and / or function.

[0103] Figure 1 is a schematic representation of a system having an evaluation device according to an embodiment.

[0104] Figure 2 shows a schematic representation of an analysis logic of the evaluation device.

[0105] Figure 3 shows a schematic representation of another analysis logic of the evaluation device.

[0106] Figure 4 shows a schematic representation of another analysis logic of the evaluation device.

[0107] Figure 5 shows a schematic representation of another analysis logic of the evaluation device.

[0108] Figure 6 shows a schematic representation of another analysis logic of the evaluation device.

[0109] Figure 7 is a flowchart of a process.

[0110] Figure 8 is a flowchart of a process.

[0111] Figure 9 shows results of a quantitative nutrient concentration determination, showing both a result determined according to the invention and a result determined using a conventional technique.

[0112] Figure 10 shows results of a quantitative nutrient concentration determination, showing both a result determined according to the invention and a result determined using a conventional technique.

[0113] Figure 11 shows results of a quantitative nutrient concentration determination, showing both a result determined according to the invention and a result determined using a conventional technique.

[0114] Figure 12 is a schematic representation of a system according to an embodiment.

[0115] Figure 13 is a flowchart of a process.

[0116] Figure 14 is a flowchart of a process.

[0117] Figure 15 is a block diagram of an apparatus with a spectral analytical detection device.

[0118] Figure 16 is a block diagram of an apparatus with a spectral analytical detection device.

[0119] Figure 17 is a block diagram of a device with a spectral analysis detection device. Figure 18 is a block diagram of a device with a spectral analysis detection device.

[0120] Figure 19 is a block diagram of an apparatus with a spectral analytical detection device.

[0121] Figure 20 is a flowchart of a process.

[0122] Figure 21 shows schematically an intermediate result of the method of Figure 20.

[0123] Figure 22 illustrates further aspects of the method of Figure 20.

[0124] Figure 23 shows a schematic representation of a system.

[0125] Figure 24 is a flowchart of a process.

[0126] Figure 25 shows an embodiment of a spectral analytical detection device in systems and methods according to embodiments.

[0127] Figure 26 shows a further embodiment of a spectral analytical detection device in systems and methods according to embodiments.

[0128] Figure 27 shows a further embodiment of a spectral analytical detection device in systems and methods according to embodiments.

[0129] Figure 28 shows a further embodiment of a spectral analytical detection device in systems and methods according to embodiments.

[0130] Figure 29 is a flowchart of a method according to an embodiment.

[0131] DETAILED DESCRIPTION OF EMBODIMENTS

[0132] Embodiments of the invention are described with reference to the figures. In the figures, similar or identical reference numerals designate elements with similar or identical design and / or function.

[0133] While embodiments are described in connection with a nutrient analysis or other condition analysis of a crop or plant, the embodiments are not limited thereto.

[0134] The features of the embodiments can be combined with each other unless this is expressly excluded in the following description.

[0135] Methods, evaluation devices, evaluation systems, and systems according to embodiments of the invention are configured to evaluate at least one spectral data acquired from a eukaryote (in particular a plant or a plant-like organism such as an algae), a part of a eukaryote (for example, a leaf or another plant part), or a food product. An evaluation result can, for example, be a state variable of the sample being spectrally analyzed, for example, a nutrient concentration. The terms spectral data and spectrum, as used here, refer to an optical variable acquired for a plurality of wavelengths (for example, reflection, transflection, transmission, etc.), which are present in a form that enables assignment of the respectively acquired optical variable to the associated wavelength.The spectral data and spectra may comprise the optical quantity for wavelengths in at least one wavelength range, wherein the wavelengths for which the optical quantity was recorded may differ from one another by a constant or wavelength-dependently variable wavelength spacing.

[0136] The spectral data and spectra can be acquired using a spectral analysis acquisition unit. The term "spectral analysis acquisition device" as used here encompasses a device that is capable of and configured to acquire an optical measurement variable for multiple wavelengths at the same measurement area of ​​the plant. The spectral analysis acquisition device may comprise a spectrometer. However, this is not necessarily required. For example, as an alternative or in addition to using a spectrometer, the spectral analysis acquisition device may be configured to acquire the optical measurement variable for the multiple wavelengths sequentially over time, for example, by actively irradiating with different wavelengths and determining the intensity of the scattered or reflected light detected in each case.

[0137] At least one attention mechanism is used to evaluate the spectral data. The attention mechanism is known per se, but has so far been used primarily in applications such as text analysis and image analysis. Further information on the attention mechanism and its implementation can be found, for example, in the documents in A. Vaswani et al., "Attention Is All You Need," 2017, Advances in Neural Information Processing Systems 30 (NIPS 2017), pp. 6000–6010 (https: / / arxiv.org / abs / 1706.03762) and A. Dosovitskiy et al., "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale," International Conference on Learning Representations (IPCLR) 2021.

[0138] The applicant has found that these techniques, which have not previously been discussed in connection with a plant condition analysis by analyzing spectral data recorded on the plant, can surprisingly also be applied to the plant condition analysis, in particular to the analysis of spectral data recorded on the plant, whereby the measures disclosed in detail here are taken to adapt the conventional methods.

[0139] In general, the attention mechanism is a computational method for analyzing data in an artificial intelligence (AI) model that, after training, is capable of recognizing relationships in the input data (relevant for a task defined by training data). The attention mechanism transforms individual parts of the input data (also called tokens) or representations of these parts (also called embedded tokens) into a new representation, whereby this transformation can incorporate information from the entire input data. To transform a token, the token is compared with one, several, or optionally all other tokens of the input data by applying a comparison function. The tokens are then weighted using an aggregation function and transferred into the new representation, whereby the weighting is determined from the result of applying the comparison function.An attention mechanism therefore generally has an aggregation function and a comparison function.

[0140] An attention mechanism can process the tokens before applying the comparison function and / or the aggregation function. The processing of the tokens before applying the comparison function and / or the aggregation function can have trainable parameters.

[0141] One possible form is what is now known as "scaled dot-product attention" according to A. Vaswani et al., "Attention Is All You Need", 2017, Advances in Neural Information Processing Systems 30 (NIPS 2017), pp. 6000-6010 (https: / / arxiv.org / abs / 1706.03762), which determines at least one query (Q) for each token, at least one key (K) for each token, and, based on the pairwise comparison of the at least one query and the keys, determines weights for aggregating the values ​​(V) obtained from each token. The determination of query, key, and value can involve trainable parameters. The trainable parameters, which are trained during training, can define how input data is processed to obtain query, key, and value. This processing can be implemented, for example, by linear projection matrices Wq, W K and W v be defined, whereby these projection matrices can be learned during training.

[0142] When used in the context considered here for evaluating spectral data, an analysis logic comprising at least one attention mechanism is configured to process multiple pieces of spectral data (which may, for example, cover different wavelength ranges) or input data determined therefrom through preprocessing. The at least one attention mechanism can be at least one self-attention mechanism, in particular a multi-head self-attention mechanism.

[0143] The use of analysis logic that features at least one attention mechanism enables the detection of extensive pairwise relationships in spectral data. Parameter values ​​of the comparison operations are data-dependent. This has surprisingly proven particularly advantageous for the analysis of spectral data. In particular, after training, a local or global context in the spectral data is taken into account during processing by the analysis logic, which features at least one attention mechanism. This contrasts, for example, with conventional filters in CNNs ("convolutional neural networks"), which have fixed weights after training and are limited to detecting local relationships depending on the size of the learned filters.

[0144] Other processing may be performed before and / or after the attention mechanism (for example, normalization and an FF NN ("Feed Forward Neural Network", hereinafter and in this field of technology also referred to as a "feed forward network" or "feed forward layer") and residual processing). A block that may have additional processing in addition to the attention mechanism is also referred to as an attention block. Thus, as shown in Fig. 3 and Fig. 4, the analysis logic may have several sequentially executed attention blocks, each of which may have at least one attention mechanism and optionally further processing before and / or after the attention mechanism. The attention mechanism may have a self-attention mechanism, in particular a multi-head self-attention mechanism.

[0145] The analysis logic comprising the attention mechanism can comprise multiple attention blocks. In particular, the analysis logic can comprise an encoder with a stack of multiple attention blocks. The attention blocks can each comprise further processing before and / or after the attention mechanism, for example, a feedforward network and / or a normalization layer such as layer normalization and / or other normalization functions and / or residual calculation.

[0146] The analysis logic can be trained using training data such that the output of the analysis logic provides a state variable associated with the state of the eukaryote, part of the eukaryote, or food product. In particular, the analysis logic can be trained such that its output provides a quantitative state variable (e.g., a nutrient concentration) from a continuous value range, an ordinal value scale, or a nominal scale.

[0147] The term "eukaryotes" as used here includes plants and plant-like organisms, in particular algae. The devices, systems, and methods disclosed here can be configured, in particular, for analyzing the condition of plants or plant parts.

[0148] Figure 1 shows a system 10 configured to evaluate spectral data acquired from a plant 2. The system 10 comprises a device 20 with a spectral analysis acquisition device for acquiring spectral data from the plant 2 and an evaluation device 30. The functions of the evaluation device 30 can also be implemented by a system, for example, a server system.

[0149] The evaluation device 30 is configured to evaluate the spectral data acquired by the device 20. The evaluation device 30 has one or more interfaces 31 for receiving the spectral data. The evaluation device 30 has a memory system 33 in which parameters (in particular the parameters trained during training) of an analysis logic for performing an evaluation 36 of the spectral data can be stored non-volatilely. The evaluation device has at least one evaluation circuit 34 configured to perform the evaluation 36 of the spectral data. Based on an evaluation result, an interface controller 35 can generate an output and provide it via a data interface and / or a human-machine interface 32.

[0150] The at least one evaluation circuit 34 used to perform the various control and processing functions may comprise one or more integrated circuits to control the interface(s) 31 and the human-machine interface 32, as well as to evaluate the spectral data. The one or more integrated circuits may, for example, comprise any one or any combination of the following circuits or circuit components: an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a processor (e.g., a central processing unit (CPU), graphic processor unit (GPU), or tensor processor unit (TPU), a controller, one or more quantum gates, a quantum information processing circuit, and other integrated circuits.

[0151] The memory system 33 can store machine-readable instruction code which, when executed by the evaluation circuit 34, causes the functions and steps disclosed herein to be executed. The memory system 33 also stores parameters that enable an evaluation of the spectral data, for example, for a state analysis. The parameters stored in the memory system 25 for evaluating the spectral data can, for example, include one or more of the following parameters: parameters for processing tokens of an attention block or of multiple attention blocks of a stack of attention blocks (for example, parameters that define the dependence of query, key, and value on the input data); parameters of at least one further trainable part of the KI model, for example, neural networks, normalization layers, and / or feedforward layers that can be arranged upstream and / or downstream of the attention mechanism or mechanisms.

[0152] Although the evaluation system 30 is shown separately from the device 20 in Figure 1, the acquisition of the spectral data and its evaluation can be carried out in an integrated device. In particular, the evaluation circuit 34 can be arranged in the same housing as the spectral analysis acquisition device.

[0153] Figure 2, Figure 3, Figure 4, Figure 5, and Figure 6 each show embodiments of an analysis logic that can be used for evaluating 36 the spectral data. The analysis logic can have an input for receiving spectral data 40 or data derived therefrom by preprocessing. The analysis logic in each case has at least one attention mechanism. The analysis logic can, in particular, have at least one self-attention mechanism. The analysis logic can have at least one encoder of a transformer, which can have at least one self-attention mechanism and, in particular, a stack of attention blocks, each with at least one (multi-head) self-attention mechanism.

[0154] In the embodiments of the spectrum evaluation 36 shown in Figures 1, 3, 4, and 5, the spectral data 40 or input data obtained therefrom through preprocessing are divided into several parts. The parts can cover different wavelength ranges. The parts can cover interleaved or overlapping wavelength ranges. The evaluation circuit 34 is configured to generate 42 the parts.

[0155] In the embodiments of the spectrum evaluation 36 shown in Figures 2, 3, 4, and 5, position coding 43 is performed. This occurs in such a way that an optical measurement variable recorded in the spectral data for a specific wavelength or wavelength range (which can be recorded, for example, in a reflection, transflection, or transmission arrangement) remains assignable to the corresponding wavelength or wavelength range. The position coding 43 is particularly technically advantageous because common variants of attention mechanisms are invariant with respect to the order of tokens in the input. In order to make the information about relative positions available to the Kl model, position information is therefore added to the tokens by position coding.

[0156] For example, an unlearned position encoding can be added by summing the (optionally processed) tokens with oscillating functions, such as sine functions, which may have different frequencies. For example, techniques based on page 6, section 3.5 of the publication A. Vaswani et al., "Attention Is All You Need," 2017, Advances in Neural Information Processing Systems 30 (NIPS 2017), pp. 6000-6010 (https: / / arxiv.org / abs / 1706.03762) can be used. The position encoding 43 can be learned or predefined. When learning the position encoding, the position encoding can be trained together with the transformer.

[0157] In the embodiments of the spectrum evaluation 36 shown in Figures 2, 3, 4, and 5, the portions of the spectral data 40 or the portions of input data obtained from the spectral data by preprocessing are further processed by a transformer encoder 44 and / or transformer decoder 44' (Figure 5). The transformer encoder 44 and / or transformer decoder 44' has at least one attention mechanism 45, 55. An output of the transformer encoder 44 and / or transformer decoder 44' provides a code for each of the portions of the spectral data 40 or input data obtained therefrom by preprocessing. In the embodiments of the spectrum evaluation 36 shown in Figures 2, 3, 4, and 5, the codes are mapped to a target value by a subsequent further processing 49. This can be done, for example, by a linear or non-linear combination of the codings.More complex pre- and / or post-processing steps are possible and are described in more detail below.

[0158] Figure 3 shows an embodiment in which the transformer encoder 44 has a stack 46 of several attention blocks 47. Several of the attention blocks 47 of the stack 46 can have identical configurations. The attention blocks 47 can thus be run through one after the other. Each attention block 47 has at least one attention mechanism 45. Optionally, before and / or after each execution of the attention mechanism 45 in the respective attention block 47, further operations 48, 48' can be performed and / or further

[0159] Processing layers are executed, for example one, several or all of: a feed forward network 48', a normalization layer 48, a residual calculation.

[0160] Figure 4 shows an embodiment in which the transformer encoder has a multi-head attention mechanism 55. In the embodiment shown, the analysis logic has several attention blocks 57 in a stack 56. The attention blocks 57 can be run through sequentially. Each attention block 57 has a multi-head attention mechanism 55. The multi-head attention mechanisms 55 are each designed as multi-head self-attention mechanisms in Figure 4. Optionally, before and / or after each execution of the multi-head attention mechanism 55 in the respective attention block 57, further operations 58, 58' can be performed and / or further processing layers can be executed, for example one, several or all of: a feed-forward network 58', a normalization layer 58, a residual calculation.

[0161] Figure 5 shows an embodiment in which the spectral data 40 is preprocessed. Preprocessing can be performed, for example, by convolution with one or more filters 50. Other preprocessing methods are possible that generate features from the spectral data. The features advantageously remain assigned to wavelengths or wavelength ranges. The filtering 50 is accordingly advantageously performed with a local filter kernel. The preprocessing, for example, convolution with a filter kernel, can be performed using a further KI model that is applied to the spectral data 40 and whose output can be further processed as already described above, for example by a transformer decoder 44', which can have at least one attention block 47, 57 and advantageously a stack of attention blocks 47, 57. The parameters of the preprocessing can be learnable.

[0162] Preprocessing may involve the use of additional data that is different from the spectral data. The additional data may, for example, include standards (e.g., measurements on a white standard) and / or calibration measurements. The additional data may be specific to the spectral analysis acquisition device used and may be stored in a non-volatile manner for use in the preprocessing to be performed for the respective spectral analysis acquisition device.

[0163] Figure 6 illustrates that the spectrum evaluation 36 can have multiple KL models. For example, preprocessing 50, if provided, can be performed using a first KL model 61, while a second KL model 62 has at least one attention block 47, 57, which in turn has at least one attention mechanism 45, 55. The second KL model 62 can be configured as a transformer encoder and can have any of the configurations already explained (in particular, any of the configurations explained with reference to Figures 2, 3, 4, and 5) or can be configured as a transformer decoder.

[0164] In the spectrum evaluation 36 explained with reference to Figures 2, 3, 4, 5, and 6, the spectral data is considered as a one-dimensional vector, which is then further processed using at least one Kl model 44, 44', 61, 62. The at least one Kl model has at least one attention mechanism. Advantageously, the at least one Kl model has a stack of attention blocks, each of which has an attention mechanism to recognize relations within the spectral data or the input data obtained therefrom through preprocessing and to use them, for example, for a condition analysis of a plant.

[0165] Trained parameters of the AI ​​models can be stored non-volatilely in a memory system of the evaluation device or the evaluation system. The trained parameters can vary depending on the implementation of the attention mechanisms. Further features of the training process are described with reference to Figure 8, Figure 12, Figure 13, and Figure 14.

[0166] Figure 7 is a flowchart of a method 70. The method may be performed by the system 10 according to one embodiment.

[0167] At 71, a wavelength-resolved detection of at least one optical measurement parameter is performed on a sample, which may be a eukaryote, a part of a eukaryote, or a food product. The detection may, in particular, comprise a detection of an optical measurement parameter on a plant or a part of a plant. The detection may also comprise a detection of an optical measurement parameter on a plant-like organism (e.g., an algae). The detection may be performed in a reflection arrangement, a transflection arrangement, or a transmission arrangement.

[0168] At 72, the spectral data acquired at 71 is evaluated. The evaluation can include a condition analysis of the plant based on the acquired spectral data. During the evaluation, the spectral data is processed using analysis logic that has at least one attention mechanism. The analysis logic can be configured to provide a result of a quantitative condition analysis at its output. The at least one attention mechanism is used to recognize relationships within the spectral data and to use them for the quantitative condition analysis. The quantitative condition analysis can include a quantitative determination of at least one nutrient concentration. A result of the quantitative determination can represent the nutrient concentration(s) on a continuous or ordinal value scale.

[0169] At 73, an evaluation result is provided. The result can be output via a data interface (e.g., for executing a control function) or via a human-machine interface.

[0170] The methods, evaluation devices, evaluation systems, and systems according to embodiments can be configured such that several different trained AI models are provided and used. The several different trained AI models can be assigned to one, several, or all of the following: a plant variety, a plant species, at least one plant stage, at least one leaf stage, a position of the leaf within the plant, a region (e.g., the USA or Europe), a cultivation type (e.g., outdoors or indoors), a cultivation substrate type (e.g., soil-based or soilless, such as coco substrate or hydroponics).

[0171] The methods, evaluation devices, evaluation systems, and systems can be designed such that a selection of the trained AI model to be used depends on a user input that specifies one, several, or all of the following: a plant variety, a plant species, a plant stage, a leaf stage, a position of the leaf within the plant, a region, a cultivation type, a cultivation substrate type. The methods, evaluation devices, evaluation systems, and systems can comprise a control of a human-machine interface to enable an input that specifies the plant variety, the plant species, the plant stage, the leaf stage, the position of the leaf within the plant, the region, the cultivation type, and / or the cultivation substrate type.The methods, evaluation devices, evaluation systems, and systems can be configured so that user input can be made via a human-machine interface on the device. The methods, evaluation devices, evaluation systems, and systems can also be configured so that user input can be received via an interface from a storage system. The latter is useful, for example, if the user input is not made in the field environment where the data acquisition is performed.The methods, evaluation devices, evaluation systems and systems can be designed such that a selection of the trained Kl model to be used is made depending on an image evaluation of an image that was acquired from the sample to be analyzed (for example a plant or another eukaryote) and that is analyzed by the evaluation circuit in order to automatically determine the plant variety, the plant species, the plant stage and / or the leaf stage and, based thereon, to retrieve and use the Kl model to be used to analyze the spectral data.

[0172] Figure 8 is a flowchart of a method 80 by which the evaluation circuit 34 can be configured for the evaluation of the spectral data.

[0173] At 81, training data is generated. Generating the training data may include capturing training spectra. The training spectra are acquired using a spectral analysis acquisition device, which may be identical in design to the spectral analysis acquisition device used to acquire the spectral data to be analyzed. The training spectra or data derived therefrom (e.g., by convolution with a filter or by preprocessing with a first class model) may be annotated.

[0174] The generation of annotations that are part of the training data can, for example, involve determining nutrient concentrations. The determination of nutrient concentrations can be done in the laboratory (e.g., using chemical and / or physical methods) or expert-based. The training spectra can be annotated with the nutrient concentrations. This can also be done implicitly by grouping training spectra into groups with similar nutrient concentrations and then annotating the groups.

[0175] Depending on the desired application, the annotations can also be generated in such a way that the training spectra are provided with annotations different from a nutrient concentration.

[0176] The training data may include spectra from at least 100 plants at different plant and / or leaf stages. The training data may include spectra from at least 1000 leaves at different plant and / or leaf stages. As explained in Figures 9 to 11, good results are achieved with a sufficiently large number of training spectra.

[0177] At 82, analysis logic is trained. This may include training with the training data. Part of the training data may be used for validation and another part of the training data may be used for testing. The training may include techniques known to those skilled in the art, such as gradient-based optimization methods, e.g., gradient descent. The training optimizes an objective function (e.g., an L2 loss) over a training set (typically by minimizing the objective function). The training may be terminated on a validation set (determined for the respective training stage) depending on the value of the objective function.

[0178] The training at 82 comprises training at least one attention block that has an attention mechanism. During the training of the at least one attention block, parameters for processing tokens can be determined, for example, parameters of a mapping that determine how K, Q, and V depend on the input data.

[0179] At 83, the trained analysis logic is stored in a memory system for use by the evaluation circuit 34 in evaluating the spectral data. The evaluation system or device can use the trained analysis logic to perform a state analysis of a eukaryote based on the spectral data.

[0180] Figure 9, Figure 10, and Figure 11 show results of the inventive technique in quantitative status analysis (in this specific case, quantitative nutrient concentration determination) for different nutrients, namely nitrogen (N) and phosphorus (P). Bars 91, 93, and 95 show the results for a technique according to the invention. Bars 92, 94, and 96 show the results for a conventional technique in which a PLS ("partial least squares") method was applied to spectral data. Figure 9 shows an R2 measure of agreement, i.e., a coefficient of determination where 100% represents perfect agreement and 0% no agreement. Figures 10 and 11 show a mean absolute error (MAE) of agreement, i.e., a deviation measure where 0 represents perfect agreement and larger values ​​represent poorer agreement.The comparison was made with laboratory-determined nutrient concentrations. The uncertainties are also shown as error bars at the top of the result bars.

[0181] As can be seen, the nutrient concentrations determined using the techniques of the invention are better than the results obtained using conventional techniques in all cases presented.

[0182] The nutrients N and P, which are shown representatively in Figure 9, Figure 10 and Figure 11, are particularly important nutrients in agricultural technology.

[0183] The following describes details of the training of the analysis logic, the design of the analysis logic and its use for the results shown in Figure 9, Figure 10 and Figure 11.

[0184] Collection of training data

[0185] Creating a dataset: Studies were conducted on 587 groups of strawberry plants. For each group, several individual measurements were recorded from leaves of several plants in close proximity, resulting in a total of 17,686 individual measurements in the dataset. The leaves were removed and analyzed in the laboratory to determine nutrient concentrations. These laboratory values ​​serve as target values ​​for training.

[0186] Filtering the dataset: Training data of poor quality were excluded. Filtering applied both to the quality of the spectra and the quality of the target values ​​(i.e., nutrient concentrations).

[0187] Division into training, validation, and test data: The data was divided into training data (70%), validation data (15%), and test data (15%). While the portion of validation data is optional, it is helpful for finding good hyperparameters for the AI ​​model and training procedure.

[0188] Cross-folds: The data sets were divided into training, validation, and test data sets multiple times (6 times) into so-called folds. This allows for more statistically robust conclusions to be drawn.

[0189] Machine learning model

[0190] Two different machine learning models were trained and tested. In a first preferred embodiment, the spectral data is divided into several spectral ranges (i.e., converted into tokens), processed linearly ("embedded"), the processed tokens are position-encoded, and then processed in a transformer encoder that has a sequence of attention blocks with a (self-)attention mechanism. In a second embodiment, the processing by the sequence of attention blocks occurs downstream of processing by a deep learning (DL) architecture.

[0191] Specifically, the Transformer Encoder has a sequence of attention blocks that can be implemented as follows:

[0192] Layer standardization (“layernorm”)

[0193] Multi-head self-attention (with scaled dot-product attention and with linear projections for processing the inputs to obtain Q, K, V)

[0194] Feed forward network consisting of layer normalization, linear NN ("linear neural network", also known as "Linear Model" in technology), GELU ("Gaussian Error linear unit") as non-linear activation function, linear NN

[0195] The above-mentioned attention block is executed several times in succession.

[0196] The mapping to the target value (Block 49 in Figure 3 and Figure 4) was realized by

[0197] - Averaging of the transformed tokens,

[0198] - Layer normalization, linear NN. In the second variant of the machine learning model, in which processing is performed by a sequence of attention blocks following a DL architecture, the DL architecture was implemented as follows:

[0199] - Convolution

[0200] - Dropout (during training)

[0201] - Rectified Linear Unit (ReLU) as a non-linear activation function

[0202] - Convolution

[0203] - Dropout (during training)

[0204] - Rectified Linear Unit (ReLU) as a non-linear activation function where this sequence was run through several times.

[0205] Both variants have hyperparameters, which include model hyperparameters and training hyperparameters. These were first optimized separately for each fold on the respective validation set using hyperparameter optimization. These fold-specific hyperparameters were then combined to form a common hyperparameter set for all folds, with each value within it roughly corresponding to the median of the fold-specific parameters. It should be noted that these parameter sets differ between the predictions of different nutrients.

[0206] (A) Using the spectral data to generate the input of the transformer encoder

[0207] Such a configuration is shown in Figure 2, Figure 3 and Figure 4. The spectral data are interpreted as a sequence of several spectral ranges, and these spectral ranges (i.e., parts of the spectral data) are analyzed together with attention mechanisms or attention blocks.

[0208] The spectral data was divided into several vectors of equal length. These are also referred to here as parts of the spectral data or partial spectra. In technology, the term "patches" or "tokens" is also commonly used. Optionally, the spectral data was transformed to a specified length before being divided, for example, by truncating the last and / or first entries or by appending additional entries at the beginning or end, for example, through periodic continuation.

[0209] These parts of the spectral data (i.e., the multiple vectors of equal length) serve directly as input to a transformer encoder. There, they are first processed ("embedded"). To avoid losing information about the position of a part of the spectrum relative to the overall spectrum, position coding is used. The thus processed and position-coded tokens are transformed according to the definition of the attention mechanism, whereby the resulting coding (also referred to in technology as "transformed tokens" or "transformed inputs") of each individual sub-spectrum can simultaneously encode relevant information about all other sub-spectra. Thus, the local and global context of the spectral data is taken into account during evaluation.The final mapping to a target value (in this case, a nutrient content), as performed in Figures 2, 3, and 4 during mapping operation 49, can be achieved, for example, by a linear mapping of the coding of the first partial spectrum to that target value. Other configurations are possible, for example, by averaging all token codings followed by a linear mapping.

[0210] As a concrete implementation used for the data generated in Figures 9, 10, and 11, several multi-head attention mechanisms were executed sequentially. Specifically, this was implemented as follows:

[0211] Since the spectrum consists of 636 individual values, it is first divided into three vectors of length 212, each of which serves as a subspectrum of the input. Each token is, as already explained, processed ("embedded") and position-encoded.

[0212] The depth of the transformer is set to 5, so that the subspectra are each transformed 5-fold by an attention block that features a multi-head attention mechanism (with the attention blocks having additional operations, as explained above). The internal dimensionality was 1024.

[0213] When analyzing with an attention block that has a multi-head attention mechanism, each spectrum is simultaneously considered multiple times in parallel, with the outputs being concatenated and serving as input to an MLP with hidden dimension 512 in each block.

[0214] The resulting codings of the parallel processed partial spectra are finally averaged and normalized, and the resulting vector serves as input for a linear regressor, which maps the output to a specific nutrient concentration value.

[0215] (B) Use of a transformer to process data generated from the spectral data by another Kl model

[0216] Here, at least one attention block is applied to the output data of another AI model, which may, for example, comprise a DL model. The at least one attention block may be part of a transformer decoder.

[0217] The spectral data is first convolved with one or more filters to detect essential, primarily locally observable features. The feature space can be extended by one dimension by applying a filter bank to obtain a plurality of convolution results for an ID spectrum.

[0218] The results of the convolutions can then serve as input for an attention mechanism, which finally maps them to a target value.

[0219] In concrete terms, this can be implemented as follows: First, a convolutional encoder (here: CNN = convolutional neural network) is run through. This consists of several layers, which in turn are formed by a sequence of convolutional, dropout, and nonlinearity layers (e.g., ReLU = rectified linear unit). Optionally, the dimensionality can be reduced using pooling layers. After the encoder, a transformer-based decoder is run through. This has at least one attention mechanism. In this specific case, the decoder was constructed from exactly one decoder layer, which is composed as follows: o Multi-head self-attention o Normalization o Multi-head cross-attention with learned query o Normalization o Feed-forward network consisting of linear NN, ReLU, dropout during training, linear NN, dropout during training o Normalization

[0220] The resulting vector serves as input to a linear regressor, which maps the output to a specific nutrient concentration value.

[0221] The use of such a decoder can be particularly useful for the simultaneous quantitative determination of the nutrient concentration of several nutrients.

[0222] Training the machine learning model

[0223] The machine learning model was trained by optimizing an objective function that minimizes a deviation (e.g., the mean square deviation (i.e., an L2 loss)) between the laboratory target values ​​and the respective baseline of the machine learning model over a training set. Training progress was monitored on a validation set.

[0224] The training was carried out on a training set from the groups of training spectra determined as described above.

[0225] Application phase of the trained machine learning model

[0226] Newly recorded spectra from strawberry plant leaves were evaluated using the trained machine learning model. The leaves from which the newly recorded spectra were acquired were analyzed in the laboratory to determine the R2 and MAE values ​​shown in Figures 9 to 11.

[0227] The results of the transformer-based architecture are superior to those of PLS ​​for all nutrients studied. Transformer-based architectures represent an attractive alternative to existing chemometric models for the analysis of spectral data. TI

[0228] Figure 12 illustrates the procedure for providing the trained analysis logic. In a particularly simple case, a set of training spectra is acquired using a device 20' that has the same configuration of a spectral analysis acquisition device as device 20. The training spectra can be acquired, for example, from plants 101 with different nutrient supplies. The method can involve the targeted induction of different nutrient supplies by applying different fertilizers or by planting the plants 101 in substrates with different nutrient supply levels.

[0229] The leaves 102 (or other plant areas) from which the training spectra were acquired can be analyzed. In particular, a chemical, physical, or physico-chemical analysis 103 can be performed.

[0230] The training spectra 104 and the analysis results for the corresponding leaves 102 (or other plant areas) are used by a computer or server system 105 to train the analysis logic. Training can involve training at least one attention block, which has at least one attention mechanism, or a Cl model with a stack of attention blocks. Different analysis logics can be trained for different state analyses (e.g., different nutrients). For the simultaneous analysis of multiple nutrients, the last attention block can advantageously be executed as a cross-attention block, and a query can be trained for each nutrient, and these can be processed together.

[0231] Parameters and optionally also hyperparameters of the trained analysis logic can be stored for use by the evaluation circuit 34, for example, in the memory system 33 of the evaluation system 30. Spectral data acquired by the device 20 can be analyzed using this analysis logic. A result can be output via a data interface and / or provided via a human-machine interface, for example, via a human-machine interface of the device 20.

[0232] Figure 13 is a flowchart of a method 110 that may be used to implement step 81 of the method 80.

[0233] At 111, training spectra are acquired from leaves. The training spectra can be recorded in central leaf regions, at leaf tips, or at other areas of leaves that can be reliably identified in both the training and application phases.

[0234] At 112, the leaves are analyzed. This can involve determining nutrient concentrations using chemical, physical, or physicochemical techniques. Alternatively or additionally, an expert-based analysis can be performed. At 113, the training spectra can be grouped based on the nutrient concentrations determined at 112. The different groups can be annotated with the respective target laboratory values ​​from step 112.

[0235] Figure 14 is a flowchart of a procedure 120 that may be used to implement step 82 of the method 80.

[0236] At 121, the training data can optionally be augmented. Augmentation of training data can be done depending on random parameters.

[0237] Training occurs at 122. As mentioned, training can be performed using techniques known to those skilled in the art, such as gradient-based optimization methods, such as gradient descent. Training can optimize (typically minimize) an objective function (e.g., an L2 loss) over a training set. During training, outliers can optionally be filtered out. The optional filtering out of outliers can also be performed before training.

[0238] With reference to Figures 15, 16, 17, 18, and 19, exemplary embodiments of the device 20 with the spectral analysis detection device are described. These devices can be used alone or in combination with the evaluation system or the evaluation device 30.

[0239] The devices each comprise a spectral analysis detection device, which can be configured differently. The spectral analysis detection device is configured to detect spectral data from a eukaryote (in particular a plant), a part of a eukaryote, or a food product.

[0240] The device 20 can have a storage system 24 for temporarily storing the spectral data. The device 20 can have a data interface 22 for transmitting acquired spectral data or data derived therefrom through preprocessing to the evaluation system 30 or the evaluation device 30. The device 20 can have a human-machine interface 23. A result of the evaluation of the spectral data can be output via the human-machine interface 23. This result can either be determined locally in the device 20 or received by the evaluation device or the evaluation system 30.

[0241] The device 20 has at least one processing circuit 25. The at least one processing circuit 25 can perform optional preprocessing 26 of acquired spectral data. The optional preprocessing can include, for example, compression and / or removal of low-quality spectra. The at least one processing circuit 25 can execute an interface controller 27 to store the spectral data acquired by the spectral analysis acquisition device in the storage system 24 and / or to control the data interface 22 for transmitting the spectral data and / or to control the human-machine interface 23 for outputting an evaluation result.

[0242] The device 20 of Figure 15 has a spectral analytical detection device which has a spectrometer 21.

[0243] The device 20 of Figure 16 comprises a spectral analysis detection device comprising a hyperspectral camera 21'. The spectral data acquired at a specific point on the test object can be transmitted and analyzed as spectral data.

[0244] The device 20 of Figure 17 comprises a spectral analysis detection device comprising a controllable multispectral light source 29 and a camera 21" or another optoelectric sensor. A light source controller 28 can control the multispectral light source 29 such that light with different wavelengths can be emitted sequentially over time, and a corresponding optical parameter can be detected with the optoelectric sensor or the camera 21". Alternatively or additionally, the sample can be illuminated with multispectral mixed light, with detection being performed by a detector with a filter arrangement. In this case, a filter arrangement can be used that transmits only a specific wavelength to the underlying pixel, or a filter arrangement that randomly transmits several light wavelengths to increase sensitivity.

[0245] As shown in Figure 18, the device 20 can have a positioning system 121. The positioning system 121 can have a satellite-based positioning system (in particular a GNSS system). The positioning system 121 can have an acceleration sensor-based positioning system. Results of a position determined by the positioning system 121 can be stored in association with acquired spectral data. This enables the results of the condition analysis to be associated with geopositions. In this way, it is possible to determine, for example, in which areas of a cultivation area (for example, a field) there is an undersupply or oversupply of plants with nutrients. Such an embodiment can also be used if the spectral analysis detection device has a different configuration (for example, the configuration explained with reference to Figure 16 or Figure 17).

[0246] As shown in Figure 19, the device 20 can have a camera 122. The camera 122 can be designed as a near-field camera configured to capture images when the device 20 is in direct contact with a leaf. The images can be processed and used, for example, using the techniques disclosed in DE patent application 102023 113704.2 of Carl Zeiss AG. In particular, the captured images can be used to determine at which positions on a plant the spectral analysis data acquisition should take place. In each of the disclosed embodiments, the spectrum evaluation 36 or at least part of the spectrum evaluation 36 can be carried out locally in the device 20. In this case, it is not necessarily required (but still possible) to output the spectral data for evaluation in a computer system separate from the device 20 via a data interface.This is shown by way of example in Figure 19, but also applies to the other disclosed embodiments of the device 20. The parameters required for executing the analysis logic can then be stored in a memory system 24, from which the processing circuit 25 can retrieve them.

[0247] The processing circuit may comprise one or more circuits to perform the explained functions. The one or more integrated circuits may, for example, comprise any one or any combination of the following circuits or circuit components: an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a processor, a CPU, a GPU, a TPU, a controller, one or more quantum gates, a quantum information processing circuit, or other integrated circuits.

[0248] Figure 20 shows a flowchart of a method 130. The method 130 can be executed automatically by the device 20 and / or an evaluation device or an evaluation system 30 coupled to the device 20.

[0249] At 131, a near-field image is received. At least one near-field image can be received from a plant section directly adjacent to device 20.

[0250] At 132, the at least one near-field image is evaluated to identify potential sources of error, such as foreign matter or mechanical damage. The evaluation may include determining a suitability assessment that indicates whether plant-inherent or measurement-related sources of error are present. The suitability assessment may quantify the suitability of the at least one image based on the presence or absence of potential sources of error in a binary manner (suitable / unsuitable), on an ordinal suitability scale, or on a continuous suitability scale. The evaluation may include determining a pixel-by-pixel or region-by-region segmentation of the at least one image to identify which regions of the at least one image are free of sources of error and thus suitable for further analysis based on spectral data.

[0251] At 133, a result of the suitability assessment is used for further analysis based on spectral data. This can be done by selectively acquiring spectral data depending on the suitability assessment and / or by generating metadata representing the suitability.

[0252] Figure 21 shows an example of a leaf 3. A leaf section of the leaf 3 is in use with the device 20 in contact with the device 20. The leaf 3 may have regions 141, 142 in which, due to foreign substances such as dirt, pollen or liquids or living organisms such as insects, a nutrient concentration determination based on spectral data entails an increased risk of falsifying the analysis results. The leaf 3 may have regions 143 in which, due to mechanical damage, necrosis, chlorosis, anthocyanosis, or spoiled leaf areas, a nutrient concentration determination based on spectral data entails a risk of falsifying the analysis results. The suitability assessment detects whether such regions exist and / or whether sufficiently large regions exist that are free of the verified error sources and are sufficiently large for further analysis based on spectral data.

[0253] Figure 22 is a schematic diagram 150 for further illustrating the processing of the at least one near-field image. The at least one near-field image can be processed to evaluate the suitability of a plant section adjacent to the device 20 for further analysis based on spectral data.

[0254] The suitability assessment 152 can be determined based on the absence of various possible error sources. The plant section can be assessed as suitable if all verified error sources are absent. More complex techniques can be used. For example, the plant section can be assessed as suitable if the area fraction of the plant section surrounded by the contact surface of the device 20 that is free of all verified error sources is greater than a given threshold.

[0255] The suitability assessment 152 can influence the spectrum-based analysis 151. This can be done in various ways, as already explained in detail. For example, the suitability assessment 152 can be used to ensure that a spectrum-based analysis is only performed on plant sections assessed as suitable. Alternatively or additionally, the suitability assessment 152 can be used to mark the result of the spectrum-based analysis (e.g., as reliable / limited reliability) or to evaluate it (e.g., to weight it).

[0256] To determine the suitability assessment, one, several or all of the following checks can be carried out based on at least one near-field image:

[0257] Foreign matter detection 153: Detection of liquid and / or solid foreign matter (such as dust or pollen) in the plant section depicted in at least one image;

[0258] Necrosis detection 154: Detection of necrotic areas in the plant section depicted in at least one image;

[0259] Chlorosis detection and / or anthocyanosis detection: Detection of areas with chlorosis or anthocyanin abnormalities in the plant section depicted in at least one image; Detection of spoiled areas 155 in the plant section depicted in at least one image;

[0260] Pest detection 156: Detection of pests or insects in the plant section depicted in at least one image;

[0261] Position verification 157: Detection of incorrect positioning of the device on the plant. For this purpose, for example, automatic detection of the plant section can be used to determine whether the device is positioned on a plant section of the correct type (e.g., a leaf at a desired leaf stage);

[0262] Crack / break detection 158: Detection of mechanical damage in the plant section depicted in at least one image.

[0263] Fungal detection 159: Detection of an area affected by fungi (e.g. mildew) in the plant section shown in at least one image.

[0264] The image-based verification of suitability for further analysis, the implementation of the evaluations performed for this purpose, and / or the use of the result (i.e., suitability) can be performed using the techniques disclosed in German patent application 10 2023 113 704.2 filed by Carl Zeiss AG on May 25, 2023. Reference is made to this application for further details.

[0265] Alternatively or additionally, automatic control or user guidance with regard to the acquisition of spectral data can also be achieved using all techniques disclosed in German patent application 10 2023 113 709.3 filed by Carl Zeiss AG on May 25, 2023. Reference is also made to this application for further details. An exemplary implementation is described with reference to Figure 23.

[0266] Figure 23 is a schematic representation of a system 10 comprising the device 20, the evaluation system 30, and a camera 160. The camera 160 can be provided separately from the device 20 or structurally installed in the same housing as the spectral analysis detection device of the device 20. The various system components can communicate with each other via data connections, in particular a communications network 4. The communications network 4 can comprise a wireless network, a cellular network, and / or other wired or wireless communication paths.

[0267] In operation, the camera 160 can capture an image of a plant or plant part. The evaluation system 30 can evaluate the image to determine whether the plant and / or which parts thereof are suitable for acquiring the spectral data. The determination of suitability can be made using the techniques disclosed in German patent application 10 2023 113 709.3 of Carl Zeiss AG, filed on May 25, 2023. A measurement recommendation can be provided via a human-machine interface of the device 20 to assist the user in acquiring the spectral data. The measurement recommendation can indicate which plant and / or plant part is particularly suitable for acquiring the spectral data. The measurement recommendation can depend on conditions to be analyzed (for example, nutrients to be analyzed), which can be user-defined as described in German patent application 10 2023 113 709.3 of Carl Zeiss AG.

[0268] In a further embodiment, the system 10 comprises the device 20 and the evaluation system 30, which can be communicatively connected via a communications network or other communications connection, as explained with reference to Figure 1. The camera 160 is thus optional.

[0269] Figure 24 is a flowchart of a method 170 that may be performed by the system 10.

[0270] At 171, an image-based measurement recommendation is generated that indicates where spectral data should be acquired. This can be done as described in German patent application 10 2023 113 709.3 of Carl Zeiss AG.

[0271] At 172, a wavelength-resolved detection of at least one optical measurement variable takes place using the device 20.

[0272] At 173, the spectral data acquired at 172 are analyzed. At least one attention mechanism is applied for this purpose. The analysis of the spectral data can be performed as already described in detail here.

[0273] Below, various additional features are described that can be used optionally to support the technical effects already described.

[0274] Augmentation to enlarge the training dataset: To enlarge the training data and cover possible variations, the training spectra can be augmented for training.

[0275] The augmentation may consider or use one, several or all of the following modifications:

[0276] Simulating variations that are expected during the manufacturing of the spectral analysis acquisition device. This includes, for example, a shift in the pixels of the image sensor between different image sensors of the same design, where the shift can be in the nanometer or sub-nanometer range. This pixel shift leads to shifts in the acquired spectrum.

[0277] Simulating expected variations of the light source (e.g., degradation over time or temperature-dependent variation of intensity and spectral signature)

[0278] Multiplicative and / or additive variations of amplitudes Shifting the spectrum towards shorter or longer wavelengths Adding noise to the spectrum Simulating variations in the distance between sample and sensor Simulating variations in the distance between sample and light source Selective smoothing of the spectrum

[0279] Optimization of hyperparameters during training: Optimization of hyperparameters of the AI ​​model (e.g., a depth of an encoder or decoder) and / or the training procedure (e.g., a learning rate) is beneficial.

[0280] Use of specific hardware in the training phase: Training is computationally intensive. Therefore, the use of suitable hardware such as GPUs, TPUs, etc. is advantageous.

[0281] Use of specific hardware in the analysis phase: The application of the trained analysis logic can be significantly accelerated with specialized hardware. Examples include GPUs or special integrated circuits optimized for executing neural networks or other ML models. Accordingly, it is advantageous if the evaluation circuit 34 has at least one such hardware component that is particularly suitable for executing the AI ​​model(s).

[0282] Use of the methods, evaluation systems, evaluation devices, and / or systems for various types of condition analyses: In the context of exemplary embodiments, reference was made to the determination of nutrient concentrations by evaluating spectral data, with the spectral data being acquired from plant leaves. The disclosed techniques are particularly suitable for this application, since the relationship between spectral data and nutrient concentrations can be complex, particularly for micronutrients. Therefore, nonlinear techniques, in particular those using at least one attention mechanism, are particularly suitable.

[0283] However, the disclosed techniques have numerous other possible applications, such as the investigation of nutrient content in water or other applications of analytical technology. Even in such applications, the use of Cl models comprising transformer-based architectures is possible and useful for identifying complex relationships.

[0284] Further applications include further condition analyses of plants, plant parts or plant-like organisms, which are described below.

[0285] Modification of the analysis logic architecture: Several specific examples of architectures with at least one attention mechanism have been explained. Further modifications are possible, for example, by varying the various model hyperparameters (number of attention blocks, inner dimension, activation function, etc.). Furthermore, various variants of the implementation of the attention mechanisms and / or their approximation are also possible, for which the person skilled in the art can obtain further information, for example, from the specialist literature cited here.

[0286] Output (e.g. visualization) of uncertainties: In addition to spectrum-based state analysis, for example to determine nutrient concentration, an (uncertainty of the analysis) can also be determined and used. The uncertainties can be determined in various ways and output as qualitative or quantitative uncertainty measures. For example, the uncertainties can be determined based on initial values ​​of classifiers, based on uncertainties determined during training, based on a scatter of spectral data results with multiple trained machine learning models, based on a scatter of the nutrient concentrations determined spectrum-based on different leaves or different plants, or in other ways. The uncertainties can be output via a human-machine interface and / or used to automatically control the spectral analysis acquisition device.

[0287] The results of the spectrum-based state analysis can be filtered or weighted based on the associated confidence or uncertainty. For example, only those results whose confidence exceeds a certain threshold can be used further (e.g., for visualization). Alternatively, only the k most certain results can be output. The threshold or the number k can be user-defined.

[0288] Design and mounting of the device 20: The device 20 can be designed as a portable, in particular as a manually held device 20.

[0289] Figure 25 schematically shows an embodiment of the device 20 as a manually held device having a structure 180 for holding the device 20.

[0290] Figure 26 shows a further embodiment in which a vehicle 182 according to one embodiment has one or more devices 20 mounted thereon. Accordingly, the device 20 or the devices 20 can be configured with a support structure for attachment to an agricultural vehicle.

[0291] Figure 27 shows a further embodiment in which a robot 184 has the device 20. A base 186 of the robot 184 can be stationary or movable, for example, along a rail system. A controller 188 can control one or more actuators for positioning the device 20.

[0292] A relative movement between the device 20 and the material to be examined (eukaryotes or food products) can also be achieved by moving the material to be examined relative to the device 20 by a conveyor or other device. The device 20, and in particular the spectral analysis detection device, can be mounted in a fixed location. Figure 28 shows a further embodiment in which a flying object 190 comprises the device 20. The flying object 190 can be remotely controllable and / or configured for autonomous flight operation.

[0293] In yet another embodiment, the device 20 can be configured as a stationary system. The device 20 can be mounted with a stationary support near a conveyor on which plants or other eukaryotes or food products are transported past the device 20.

[0294] Areas of application: The disclosed devices, evaluation devices, evaluation systems, systems, and methods can be used for various applications. In particular, the devices, evaluation devices, evaluation systems, systems, and methods can be used on crops or cultivated plants. The plants on which the techniques can be used can, in particular, also include forestry plants (e.g., trees), hybrids, and / or genetically modified plants.

[0295] Examples of such plants can be:

[0296] • Crops

[0297] • Fodder plants

[0298] • Fiber plants

[0299] • Oil plants

[0300] • Ornamental plants

[0301] • Industrial crops, such as o Oilseeds o Tobacco o Hemp o Hops o Aromatic, culinary and medicinal plants o Seeds for herbaceous oil plants o Seeds for linseed (and consequently fibre flax) o Energy crops o Plants used for the production of feedstocks for renewable energy production

[0302] The analysis of the spectral data may include, but is not limited to, the qualitative and / or quantitative determination of nutrient concentrations. For example, the condition analysis may include analyses regarding one, several, or all of the following plant conditions:

[0303] Nutrient supply problems, for example o nutrient deficiency o toxicity biotic stress, for example due to o nematodes o insects o arachnids o fungi o bacteria o viruses abiotic stress, for example due to o drought o excessive water supply and / or waterlogging o salinity o temperature-related stress (cold, frost, heat) o UV light o metal toxicity o mechanical stress (e.g. pressure, squeezing, pressing).

[0304] The results of the spectral data analysis can be used to improve plant conditions. This can be achieved through automatic or semi-automatic adjustment of artificially added substances (such as approved fertilizers or approved insecticides) and artificially added water. Numerous other applications are possible, for example, for monitoring plant health with regard to environmental and regulatory issues.

[0305] Further areas of application: Even if embodiments have been described in the context of a condition analysis of plants, the devices, evaluation devices and evaluation systems and methods can also be used elsewhere, for example for the spectrum-based analysis of an agricultural product.

[0306] Figure 29 is a flowchart of a method 200 that can be carried out using the apparatus 20 and / or the evaluation device 30 or the evaluation system 30.

[0307] In 201, spectral data is acquired and analyzed from a plant or other eukaryote. The analysis involves the use of at least one attention mechanism.

[0308] At 202, a result of the evaluation is used to determine measures to improve conditions for the plants. The measures can then be implemented to improve the conditions. The disclosed methods, devices, evaluation devices, and systems provide various technical effects. In particular, they offer improvements with regard to the detection of complex non-linear relationships based on spectral data, for example, for analyzing the condition of plants. The methods, devices, and systems can be used to improve the evaluation of spectral data on crop plants or cultivated plants.

[0309] Compared to techniques in which plant parts must first be separated, then sent to a laboratory and examined there, the techniques disclosed here offer the advantage that the evaluation can be carried out quickly, for example while the user with the device 20 is still on site at the plant.

[0310] Embodiments of the invention can be used in particular in connection with plant cultivation in agriculture, which is a central economic and industrial sector of humanity. Global trends such as a growing population and a simultaneous decrease in the availability of agricultural land lead to a constantly growing demand for crop yields per area. At the same time, the unsustainable exploitation of soils leads to an increased need for targeted fertilization in order to maintain or even increase previous yields. Furthermore, a growing awareness of sustainable management and increasing regulatory pressure are creating new requirements for precisely tailored fertilization, in particular to avoid over-fertilization.

[0311] Of particular importance for the growth of any crop is the appropriate supply of nitrogen (N), as it is the most important element of chlorophyll and thus essential for optimal metabolism. In addition to N, other macronutrients are also relevant, such as phosphorus (P) and potassium (K), which can often be added to the soil together with N as so-called NPK fertilizers, as well as calcium (Ca), magnesium (Mg), and sulfur (S). Mg is the central element in the chlorophyll ring.

[0312] Additionally, depending on the plant type, other nutrients may also be relevant for optimal growth and yield, although these are often required in significantly smaller amounts and are therefore also referred to as micronutrients. These include, for example, boron (B), molybdenum (Mo), copper (Cu), manganese (Mn), zinc (Zn), iron (Fe), and chlorine (Cl).

[0313] Embodiments of the invention can contribute to ensuring optimal conditions with regard to nutrient supply throughout the entire growth phase of a plant. Only in this way can nutrient deficiencies or nutrient oversupply be appropriately responded to through adapted fertilizer applications (through a specific fertigation program in the substrate or soil growth, or as additional soil and / or foliar fertilization). Embodiments of the invention allow for regular monitoring of plants. Compared to laboratory-based leaf analysis, the condition analysis can be performed quickly. This is particularly desirable for fast-growing plants and plant species with rapid fruit ripening (e.g., strawberries and lettuce).Embodiments of the invention thus also address the objective that there should not be a delay of several days between measurement and availability of the evaluation result, as otherwise the measurement result would already be invalid and would not allow for a meaningful adjustment of the nutrient application. Rapid evaluation is also desirable when environmental factors change rapidly (e.g., temperature, precipitation, etc.), as otherwise the conditions would change significantly between the measurement and the time of analysis. Rapid evaluation is also advantageous in hydroponic cultivation, where it enables rapid adjustment of applied fertilizers.

[0314] Embodiments of the invention enable spectral data acquisition at the intended application site without the need to transmit physical samples. Embodiments of the invention can thus serve to provide a non-destructive, automated, and simultaneously rapid (near real-time) solution for detecting deficiencies and / or oversupplies of multiple nutrients in plants.

[0315] While exemplary embodiments have been described with reference to the figures, modifications can be implemented in further exemplary embodiments. For example, the modifications already explained can be used cumulatively or alternatively. The evaluation device 30 does not necessarily have to be provided separately from the device 20, but can also be structurally connected to it or integrated into the device 20.

[0316] While embodiments have been described that can be used in crops or cultivated plants, the disclosed techniques can also be used in other fields of application.

[0317] The present disclosure also encompasses embodiments with any combination of features mentioned or shown for different embodiments. It also encompasses individual features in the figures, even if they are shown there in connection with other features and / or are not mentioned above or below. Furthermore, the alternative embodiments described in the figures and the description and individual alternative features thereof may be excluded from the subject matter of the invention or from the disclosed subject matter.

[0318] The terms "comprise" and "comprise" and derivatives thereof indicate a non-exhaustive relationship and do not exclude the presence of other elements or steps. The indefinite article "a" or "an" and derivatives thereof do not exclude the presence of a plurality of the corresponding elements. The functions of several features listed in the claims can be fulfilled by one unit or one step, respectively. A machine-readable instruction code that can be executed by a programmable circuit to carry out methods according to embodiments can be stored and / or distributed on a suitable medium, such as an optical storage medium or a solid-state medium provided together with or as part of other hardware. The instruction code can also be distributed in other forms, such as a modulated data signal sequence.

[0319] Embodiments of the invention provide improved techniques for evaluating spectral data and are particularly applicable to the condition analysis of plants.

Claims

CLAIMS 1. A method for evaluating spectral data (40), comprising: Receiving the spectral data (40) by at least one evaluation circuit (34), wherein the spectral data (4) comprise spectral data (40) acquired on a eukaryote (2), a part of a eukaryote (3) or a food product; Analyzing the spectral data (40) by the at least one evaluation circuit (34), comprising: Dividing the spectral data or data generated from it by preprocessing into several parts, Analyzing the plurality of parts using an analysis logic (44), wherein an input of the analysis logic (44) receives the plurality of parts, and wherein an output of the analysis logic (44) provides encodings of the plurality of parts, wherein the analysis logic comprises at least one attention mechanism (45, 55), and Determining at least one evaluation result based on the codings.

2. The method according to claim 1, wherein the at least one evaluation result comprises at least one quantitative analysis result.

3. The method according to claim 2, wherein the at least one quantitative analysis result comprises a continuously variable value from a range of values.

4. The method of claim 2 or claim 3, wherein the at least one quantitative analysis result comprises a value selected from an ordinal scale.

5. The method according to any one of claims 2 to 4, wherein the at least one quantitative analysis result comprises a concentration of at least one chemical element or at least one chemical compound.

6. Method according to one of claims 2 to 5, wherein the at least one evaluation result comprises at least one nutrient concentration.

7. Method according to one of the preceding claims, wherein the analysis logic comprises: an encoder (44) comprising a stack of a plurality of attention blocks, wherein the plurality of attention blocks each comprise an attention mechanism (45, 55).

8. The method according to any one of the preceding claims, wherein the at least one attention mechanism (45, 55) comprises at least one multi-head attention mechanism (55).

9. Method according to one of the preceding claims, further comprising a wavelength-dependent position coding (43).

10. Method according to one of the preceding claims, wherein the at least one attention mechanism (45, 55) has at least one learnable parameter.

11. The method according to claim 10, wherein the at least one learnable parameter defines a processing of an input of the respective attention mechanism (45, 55).

12. Method according to one of the preceding claims, further comprising preprocessing (50, 61) preceding the analysis logic.

13. The method according to claim 12, wherein the preprocessing is carried out using a further artificial intelligence, Kl, model (61).

14. The method according to any one of the preceding claims, further comprising training the analysis logic or at least a portion of the analysis logic using training data, wherein the training data comprises: a plurality of spectra (104) and annotations associated therewith or a plurality of data derived from spectra (104) and annotations associated therewith.

15. Method according to one of the preceding claims, wherein the spectral data (40) comprise spectral data (40) acquired on a plant (2), a plant part (3), a plant-like eukaryote or a plant or animal food.

16. Evaluation system or evaluation device for evaluating spectral data (40), comprising: an interface (31) for receiving the spectral data (40), wherein the spectral data (40) comprises spectral data (40) acquired on a eukaryote (2), a part (3) of a eukaryote, or a food product; at least one evaluation circuit (34) configured to perform at least the following processing steps for analyzing the spectral data (40): Dividing the spectral data (40) or data generated therefrom by preprocessing into several parts, Analyzing the plurality of parts using an analysis logic (44), wherein an input of the analysis logic (44) receives the plurality of parts, and wherein an output of the analysis logic (44) provides encodings of the plurality of parts, wherein the analysis logic (44) has at least one attention mechanism (45, 55), and Determining at least one evaluation result based on the codings.

17. System (1) or device (20), comprising: the evaluation system (30) or the evaluation device according to claim 16 and a spectral analytical detection device (21; 21'; 21", 29) for detecting the spectral data (40).

Images