Crop / weed / pest identification from images using soil adapted image augmentations

Abstract

Computer implemented method for generating target domain data, a computer implemented method for training a neural network, and a system, an apparatus, a use, and a computer program element adapted therefor, including providing to the neural network source domain data relating to a first agricultural domain; providing to the neural network target domain data relating to a second agricultural domain; determining a transformation relation between the source domain image data and target domain image data; and generating predicted target domain data with the target domain image data including the target domain background data and the target domain item data, and the target domain labeling data relating to a parameter of at least one target domain item identified in the target domain image of the second agricultural domain.

Description

[0001] CROP / WEED / PEST IDENTIFICATION FROM IMAGES USING SOIL ADAPTED IMAGE AUGMENTATIONS

[0002] TECHNICAL FIELD

[0003] The present disclosure relates to a computer-implemented method for providing generated labeled image data as training data for a machine learning procedure for image classification a plant, weed and / or pest in a target domain, an apparatus for providing generated labeled image data as training data for a machine learning procedure for image classification a plant, weed and / or pest in a target domain, a use of generated labeled image data of a plant or a plant part infested with a pathogen as training data, a system for image classification a plant, weed and / or pest in a target domain, and a computer program element.

[0004] TECHNICAL BACKGROUND

[0005] The general background of this disclosure is the providing of training data for a machine-learning procedure for image classification a plant, weed and / or pest in a domain of an agricultural field.

[0006] The goal of digital farming is to provide farmers with useful tools and solutions to understand the problems, to plan and to treat their crop protection measures. For instance, carefully monitoring and determining the appearance, type, and infestation of fields with weeds, pest and diseases is essential for digital farming products and solutions to provide farmers with the best recommendation for a treatment. It also facilitates fully automated digital solutions in the future. Therefore, it is desired to tools and products to achieve above mentioned goals.

[0007] The quality of artificial intelligence solutions for the identification of weed, pest and disease infestation of agricultural fields relies highly on the underlying data, in particular data sets. Especially in image recognition, in particular recognition of subject matters in images, and modern machine learning procedures require a high amount of data to train models with acceptable quality. In agriculture, labeled data is scarce and high efforts and costs are required to create high quality massive data with annotations and labels. However, it is not only necessary to have large datasets, but also versatile datasets representing best variations and generalizations well to any condition that comes with images from users of digital farming products.

[0008] In common agricultural practice, when scaling solutions of a trained model to a new domain or region, often a decrease in quality of the model solution occurs, as the training data may have not the same information contend or information density, or particular information is missing in data for new regions. Especially in cases, where a model must predict based on data that differ from what was used in training, and in cases where information is missing for agricultural domains it is expected that the quality of the learning process deteriorates. It is a so-called domain shift where the distribution of the data in the training set is different from unseen data.

[0009] There is a need for predicted information to provide high quality training data when a neural network is trained, and models are deployed. Especially in image recognition for digital farming relevant use cases, such as weed, pest and disease identification, environmental factors play a big role and large datasets with labels are required to cover them. However, labeled image data are not available for some regions.

[0010] Labeled image data is an important requirement for receiving accurate model solutions, but generating of labeled image data is a time and money consuming part and therefore very costly, especially if large image collections need to be created with annotations.

[0011] It has been found that a need exists for an alternative and smarter way for generating training data for a machine-learning procedure for image classification a plant, weed and / or pest in a less time and moneyconsuming manner.

[0012] SUMMARY OF THE INVENTION

[0013] The present invention provides a computer implemented method for generating target domain data, a computer implemented method for training a neural network, and a system, an apparatus, a use, and a computer program element adapted therefor according to the subject matters of the independent claims, wherein further embodiments are incorporated in the dependent claims.

[0014] According to an aspect of the invention, there is provided a computer-implemented method for generating training data for a neural network to be trained, comprising: providing to a training data generation neural network source domain data relating to a first agricultural domain, wherein the source domain data include source domain image data relating to a source domain image and source domain labeling data, wherein the source domain image data include source domain item data relating to at least one source domain item of the first agricultural domain and wherein the source domain image data further include source domain background data relating to a source domain background of the first agricultural domain, wherein the source domain labeling data relate to a parameter of the at least one source domain item; providing to the training data generation neural network target domain data relating to a second agricultural domain, wherein the target domain data include target domain image data relating to a target domain image, wherein the target domain image data include target domain item data relating to at least one target domain item of the second agricultural domain and wherein the target domain image data further include target domain background data relating to a target domain background of the second agricultural domain; determining by the training data generation neural network a transformation relation between the source domain image data and target domain image data based on a relation of the source domain background data and the target domain background data; identifying the target domain item data relating to at least one target domain item in the target domain image data relating to the target domain image; determining based on the transformation relation, based on the source domain image data and based on the source domain labeling data target domain labeling data relating to a parameter of the at least one target domain item of the second agricultural domain; allocating the determined target domain labeling data to the identified target domain item data relating to the at the least one target domain item identified in the target domain image data; and generating predicted target domain data comprising the target domain image data including the target domain background data and the target domain item data, and the determined target domain labeling data.

[0015] Thus, a computer implemented method can be provided which may allow generation of training data even for regions for which training data are not available in detail. In particular it is possible to generate image data with labeling or classification data with respect to, e.g., a plant, weed and / or pest in a target domain, for which such labeling or classification data is not available or only labeling or classification data are available which are limited over labeling or classification data of a source domain. The source domain data, the target domain data and the predicted target domain data may comprise data relating to a plurality of images of the respective domain.

[0016] In an example the source domain background data may relate to soil background data and the target domain background data may relate to soil background data.

[0017] The term “source domain” is to be understood broadly in the present case and represents an area of a whole agricultural field, domain or a part respectively sub-area of the whole agricultural field or domain being different to the whole agricultural field or a part respectively sub-area of the whole agricultural field or domain. A source domain may be a domain, i.e., agricultural field or a part respectively sub-area of the whole agricultural field, for which detailed information, in particular plant, weed and / or pest information, as well as soil information are provided or known. A source domain in particular for the present invention may have allocated information for a background, which may be soil for plants, which may beneficial plants or malicious plants, for diseases, pests and other relevant information for that agricultural field or domain. For a source domain the information may be provided as a strong correlated information, which may not only include images represented by image data, but also information on recognized items on images and additional information to the recognized items. Also referred to as labeling. Labeling may be provided for whatever is included in the source domain, be it soil as background, plants or diseases or the like. This may include labelling of identified or recognized items with additional information, which may be provided by manually adding of information and / or a reference to identified or recognized items, by which reference information can be added from an external database, an internal database, or a self-generated database, e.g., by a machine learning process and the like, but is not limited thereto.

[0018] The term “target domain” is to be understood broadly in the present case and represents an area of a whole agricultural field or a part respectively sub-area of the whole agricultural field being different to the whole agricultural field or a part respectively sub-area of the whole agricultural field. For this invention the target domain can be understood as a domain for which less information may be available compared to the source domain, but for which target domain labeling information should be provided based on information of the source domain. The target domain may have allocated information for a background. The information for a background may be soil for plants. The information for a background may be beneficial plants or malicious plants. The information for a background may be diseases, pests and other relevant information for that agricultural field or domain. The information for a background has a lower information content compared to that of the source domain. A lower information content may be, e.g., a lower grade of labeling. For a target domain the information may be available only as not or low correlated information. For a target domain, e.g., labelling information may be missing. Missing labeling for a target domain may be for whatever is included in the target domain be it soil as background, plants or diseases or the like. A target domain may lack a labelling of identified or recognized items with additional information. A target domain, e.g., a whole agricultural field or a part respectively sub-area of a whole agricultural field of a target domain may be located in a climate zone being different to the climate zone of the whole agricultural field or a part respectively sub-area of the whole agricultural field of the source domain. Therefore, e.g., the soil color and / or structure of the whole agricultural field or the part respectively sub-area of the whole agricultural field of the target domain may differ from, e.g., the soil color and / or structure of the whole agricultural field or the part respectively sub-area of the whole agricultural field of the source domain. Nevertheless, a certain relation exists between image data of a source domain and image data of a target domain. Based on the information for the source domain and the relation between the source domain and the target domain, information can be deduced for the target domain. This may include information, which was implemented manually in the source domain, like, e.g., identification of particular information identifying a plant, a crop, a weed, a disease, or any other item. In other words, the target domain may be a domain, i.e., an agricultural field or a part respectively sub-area of the whole agricultural field, which has not been evaluated so far in detail, and for which, e.g., no plant, weed and / or pest information have been identified or labeled. The target domain may differ in its soil types, soil colors, soil structure, soil moisture, plant density and foreign objects from the source domain. An aspect of the target domain may be that it has elements that are different from that of the source domain in terms of, e.g., background information (e.g., its soil types, soil colors, soil structure, soil moisture, plant density and foreign objects), whereas other elements may correspond, e.g., for plants, weeds, pests etc., where the actual plant, weed and / or pest information stays the same or more similar than, e.g., the background information.

[0019] A domain may also be an abstract space having different properties, wherein a transformation relation may be indicated that transforms or converts one domain to the other. A first domain having a first characteristic may be converted into a second domain having a second characteristic. The first and second domain may be a first and second image domain having different characteristic. In this way the first domain may comprise an original image and the second domain may comprise a converted image of the original image.

[0020] The term “image” is to be understood broadly in the present case and represents any two-dimensional or three-dimensional visual representation of a subject or item. An image may be an artifact, such as a photograph or other two-dimensional picture, which resembles a subject or item. In the context of signal processing, an image is a distributed amplitude of color(s). An image does not have to use the entire visual system to be a visual representation. The image may be a greyscale image, which uses the visual system's sensitivity to brightness across all wavelengths, without taking into account different colors. A black and white visual representation of a subject or item may be also interpreted to be an image, even though it does not make full use of the visual system's capabilities. Images can be still but can be moving or animated. Further, images may be captured by optical devices - such as cameras, mirrors, lenses, telescopes, microscopes, etc. An image may be understood as a composition of one or more single images, wherein an image is to be understood as a two- or higher dimensional pixel, voxel or vector or the like array.

[0021] The term “source domain image” means an image of the source domain or the first agricultural domain. Likewise, the term “target domain image” means an image of the target domain or the second agricultural domain.

[0022] The term “labelling” or “labeled image data” is to be understood broadly in the present case and comprises any data indication information with respect to plants, weeds and / or pests etc. in a domain. For instance, the plant, weed and / or pest information may include information about the species, genus, family, category, texture, shape, growth stage, coloring, visual appearance and / or density of plants, in particular crop plants, weeds, e.g., weed plants, and / or pest, but is not limited thereto. In other words, the labeled image data may be images comprising additional information with respect to the plants, weeds and / or pests in the source domain. The labeled image data may be, e.g., one of the following types of labels for images: labels describing the whole image, boxes describing locations and individual features of objects on the image (e.g., a box for a particular plant with coordinates and the taxonomy and growth stage of that plant), and segmentations being similar to the boxes but more precise, since the object is not marked by a box, but a polygon instead. In this context, individual objects or whole regions can be marked and described with attributes and / or metadata. The labeling of the image data can be provided by an Al. For example, if there is a model and / or NN, that can identify the locations of individual plants on an image, this model and / or NN can be applied to that image and the output, e.g., the locations of plants, can be used as a label to train another neural network NN.

[0023] However, using labels from an Al can only be as good as the Al and / or neural network. Accurate labeling may result in a good training. But, if the Al and / or NN makes mistakes when labelling or is not performing well in a particular scenario, those mistakes may be passed on in the output labels. Additionally or alternatively, the labeling of the image data may be a manual labeling by human annotators. Alternatively, the labeling of the image data may be a combination of labels provided by Al and the manual inspection and correction by human annotators.

[0024] The term “unlabeled image data” is to be understood broadly in the present case and comprises no or less data indication information with respect to plants, weeds and / or pests in, e.g., the target domain compared to a source domain. In other words, the unlabeled image data may be solely images without any further information than the pictorial representation of the target domain or may be images of the target domain with less information than images of the source domain. Unlabeled image data may be individual images, which do not have labeled information. However, there can be information about an “unlabeled” dataset in terms of location, time, use case, what plants are on that field, etc., like e.g., general information about a whole dataset. Unlabeled image data may be image data with empty attribute fields and / or empty metadata fields in an image data structure.

[0025] The term “background” is to be understood broadly in the present case and comprises any information with respect to the background of the respective domain. This background may be, e.g., soil, in particular the soil in the background of the image data, e.g., the visible soil surface behind respectively under, the plants on the soil surface, in the image data. For instance, the soil background may have properties which may be evaluated and used for processing, e.g., soil types, soil colors, soil structure, soil moisture, plant density and foreign objects, but is not limited thereto.

[0026] Parameters of a background, also referred to as background parameters, may be employed in the method of the present disclosure, for example a soil parameter.

[0027] According to the present disclosure, the parameter of the background, e.g. soil parameter, may be a characteristic of background, or, in other words may characterize the background. A characteristic of the background and / or an attribute of the background may define, describe and / or classify the background.

[0028] The term “cluster” may be a combination of same and related aspects and differentiation of different aspects. E.g., all data relating to a particular soil color may be clustered. The clustering may render the entire procedure more robust, as related data may be clustered. In case a data pool includes domain data relating to, e.g., 90% red soils and 10% grey soils, and a clustering is carried out on red soils and grey soil, deterioration of the red soil data by grey soil data can be avoided and vice versa. By clustering the labeled image data and / or the unlabeled image data with respect to the soil background information, the variance in the images can be reduced so that a higher quality of soil-adapted image augmentations can be provided and produced, respectively. This in turn may, e.g., improve the quality of a machine learning process and an image recognition model when trained on original images of a source domain and soil adapted augmented images of a target domain, wherein the augmented images increase the number of training data.

[0029] The term “agricultural field” or “agricultural domain” as used herein refers to an agricultural field to be treated. The agricultural field may be any plant, in particular crop, cultivation area, such as a farming field, a greenhouse filed, or the like. An agricultural field may comprise crops, weeds, volunteer plants, crop from a previous growing seasons, beneficial plants or any other plants present on the agricultural field. The agricultural field may be identified through its geographical location or geo-referenced location data. A reference coordinate, a size and / or a shape may be used to further specify the agricultural field. The agricultural field may be separated in areas, respectively domains.

[0030] The term “data” as used herein is to be understood broadly in the present case and represents any kind of data. Data may be single numbers / numerical values, a plurality of a numbers / numerical values, a plurality of a numbers / numerical values being arranged within a list, two-dimensional maps or three-dimensional maps, but are not limited thereto. As far it is referred to element data representing an element, the data are to be understood as data being able to characterize the respective element, i.e., include the information to distinguish the respective element from other elements. This does not mandatorily mean that the element data include all data of that element. However, the element data may extend over the amount of data which is required to distinguish the element from other elements but may include information on element details.

[0031] In an example data may be related to and / or associated with physical objects, physical states, physical conditions, physical parameter and / or measurement values. Data may represent objects from the physical domain in a digital domain. Data may be generated by a measurement device, a camera, a digitizer and / or a sensor. Data may be a digital representation of the physical environment. Data may be stored in a database. Whilst physical objects may have an analogous characteristic, data may summarize predefined ranges to a digital value by quantifying the analogous value. In another example data may be referred to as digital values that may be used to describe an object and / or a quantity. In a further example data may be used to represent an object and / or to describe the object.

[0032] Data may define a whole category, set and / or group of digital information. Data may be referred to by their use such as image data may be used to represent an image, e.g. as pixels, and labeling data may be used to label and / or annotate items. A specific value, item and / or physical property may be represented by a parameter. A label may, for example, establish an association between an item and an appropriate parameter. In an example, weather data may describe substantially all data relating to weather conditions and a parameter such as temperature, humidity and / or pressure may refer to a specific value to be examined. In another example, soil data may comprise conductivity, soil color and / or soil humidity as soil parameter. Which parameter may be used to describe an overall phenomenon such as whether, soil etc may depend on the purpose of a specific embodiment. In an example data may comprise one or more parameter and / or data may be derived from one or more parameter. In a further example an image may be seen as data, and specific elements and / or items identified on the image may be described by a parameter. In other words, a parameter may be seen as a specific value identified for an element, an item and / or object. Allocating a value to a parameter may help representing a specific local situation. For example, fields may have crops, plants and / or weeds. However, the specific type of crop, plants and / or weeds and their specific properties such as grow density, grow state and / or distribution may be very specific for each field to be examined. In this way parameter may provide the individual and / or unique character of a physical situation to the digital representation and / or to the data. Parameter may be described by different names, properties, metadata and / or values. These names, properties, metadata and / or values may be provided as labels. Labels may be allocated to an item and / or parameter during a labelling process and may remain related to the respective item and / or parameter in order to describe the item and / or parameter in more detail. Labels and / or labeling data may allow to allocate a certain meaning to an item and to make a parameter understandable for a machine. The labels and / or annotations may follow a predefined standard by employing a predefined ontology and / or a predefined hierarchy in order to provide a common understanding of the items. In this way a processor, a system, a neural network and / or a computer may be enabled to understand the meaning of the items and / or parameter and to may relate them to other data and / or other dataspaces. A parameter may be a property of the physical domain and be represented by labeling data in the digital and / or data domain. The labeling data may describe and / or carry the parameter.

[0033] The term “neural network” as used herein is to be understood broadly in the present case and represents any black box (especially for computer vision tasks), transforming device and / or mapping device. Provision of domain data to a neural network can be understood as using the domain data as input for a neural network. The neural network may carry out different actions, e.g., image recognition, data conversion and / or data generation. When providing, e.g., domain data and / or domain image data, and / or domain background data and / or domain item data and / or labeling data to a neural network, the data, e.g., in case of image data, is a data set representing a matrix of pixels, which is provided as an input to the neural network. Likewise, domain item data and domain labeling data include data representing an item and include data, which establish an allocation or information considered as a label to an element to be labeled. The neural network then may extract the most relevant features and information, to solve a certain task of determining a transformation relation or determining allocation of information to further elements as a labeling procedure or generate predicted target domain image data representing a predicted target domain image being labeled, which labeling was carried out by the neural network. From the results, it can be estimated and predicted what the neural network used as relevant features.

[0034] Relevant features may be the soil type, the soil color, the soil structure, the soil moisture, the plant density and the foreign objects, but is not limited thereto. Additionally or alternatively, there could be other useful patterns respectively parameters in the image data which are not meaningful to the human, but can be exploited by the Al algorithm to solve the task better (e.g., a watermark which correlates to certain fields). The images of a particular plant type can be taken at a specific time, e.g., in the morning or in the evening, and the neural network learns the brightness respectively shadow of the image as a relevant feature. The term "transformation relation” may be considered as generating a relation between particular aspects of, e.g., a source domain and a target domain and to conclude based thereon further aspects, which are included in one domain, e.g., the source domain, but not included in the other domain, e.g., the target domain. A transformation relation may be generated, e.g., by using or applying a generative adversarial algorithm of a GAN (Generative Adversarial Network). A transformation relation may be a complex relation, which extends over a simple one-to-one relation, but may establish a transformation relation also for items which have a different appearance in an image, however are identified to be equivalent items. Items may be for example plants or plant portions, wherein equivalent items may also an equivalence which considers different growing states of a plant, different nutrition conditions and so on.

[0035] The term “allocation” may be considered as establishing a relation between particular data to a particular element or item. The item or element then “carries” this information represented by the data and can be used, in case an action is to be carried out with respect to that item or element, for which the allocated data or information is required. In other words, the allocation may establish the availability of related data or parameter with respect to an element or item.

[0036] According to the present disclosure, the parameter of the at least one source domain item may be a characteristic (in other words, an attribute) of the source domain item, or, in other words may characterize the source domain item. The parameter of the at least one target domain item may be a characteristic (in other words, an attribute) of the target domain item, or, in other words may characterize the target domain item. Characteristics of an item, particularly the source domain item and / or the target domain item, may define, describe and / or classify said item. Labeling data may relate to a parameter of an item, such as the source domain item or the target domain item. For example, labeling data may comprise one or more labels, a label being representative of a parameter of an item, e.g. a source domain item or a target domain item.

[0037] According to an embodiment, the source domain background data relate to soil background data and the target domain background data relate to soil background data, wherein the soil background data in particular are soil color data.

[0038] Thus, the soil parameter may be used to determine a transformation relation. The source background data relate to soil background data typical for the source and the target background data relate to soil background data typical for the target domain. The neural network may, based on the source domain data, extract item data and related labeling data and based on the transformation relation and the extracted item and labeling data merge this with soil data of the target domain to correlate the items in the target domain with labeling data derived from the source domain. The soil color may serve as a significant parameter, as it can also be used for a later clustering. The color of soil also may represent an integral parameter, which allows conclusions not only on the soil composition, but also on humidity, etc. The soil parameter, in particular the color may be used as training data for a machine learning procedure, e.g., for image classification and labeling items, in particular plantation items. However, this can also be applied for weed and / or pest labeling and classification in a target domain. Beside the color, also soil types, soil structure, soil moisture, plant density and / or foreign objects may be used for this purpose.

[0039] According to an embodiment, the source domain item includes a beneficial source domain item and the target domain item includes a beneficial target domain item, and the source domain labeling data relate to a parameter of the beneficial source domain item and the target domain labeling data relate to a parameter of the beneficial target domain item, in particular the plantation, includes at least one plant parameter.

[0040] Thus, a focus can be set for the computer implemented invention for generating predicted target domain data and also the training of a neural network on beneficial items, like crop and plant portions. Beneficial items are not limited thereto and may also be, e.g., beneficial insects. Beneficial items may also be parts of a beneficial plant and may be classified or labeled with data relating to species, genus, family, category, texture, etc. for generating prediction target domain data and generating training data for a machinelearning procedure.

[0041] A beneficial item and / or beneficial parameter may be an item and / or parameter labeled as being beneficial and / or having a desired property.

[0042] According to an embodiment, the source domain item includes a malicious source domain item and target domain item include a malicious target domain item, and the source domain labeling data relate to a parameter of the malicious source domain item and the target domain labeling data relate to a parameter of the malicious target domain item, in particular the at least one of a weed and pest, includes at least one of a weed parameter and pest parameter.

[0043] Thus, a focus can be set for the computer implemented invention for generating predicted target domain data and also the training of a neural network on malicious items, like weeds, pests or diseases or parts thereof. Malicious items are not limited thereto and may also be other items. Malicious items may also be parts of a malicious weed, pest or disease and may be classified or labeled with data relating to species, genus, family, category, texture, etc. for generating prediction target domain data and generating training data for a machine-learning procedure.

[0044] A malicious item and / or malicious parameter may be an item and / or parameter labeled as being malicious and / or having an undesired property.

[0045] Which items and / or parameter may be seen to be desired and which items and / or parameter may be seen to be undesired may be predefined. In an example a list of beneficial properties may be provided comprising desired properties, parameter and / or items, and / or a list of malicious properties may be provided comprising undesired properties, parameter and / or items.

[0046] It should be noted that an item may also be identified as being beneficial and malicious at the same time for different reasons. In this case, it may be determined whether the beneficial part or the malicious part is more relevant, and it can be selected to act accordingly for the beneficial and malicious procedure, respectively. It should also be noted that a relation can be determined for beneficial and malicious share and a weighted action can be applied reflecting the weight of the beneficial and malicious procedures, respectively.

[0047] According to an embodiment, at least one plant parameter, be it beneficial or malicious, is selected out of a group, the group consisting of a plant portion or plant infested, a plant portion or plant infested with a pathogen a biotic plant disease symptom and an abiotic plant disease symptom. A plant portion may be selected out of a leave, stem, and / or panicle. A biotic plant disease may be selected out of a fungal disease, viral disease and / or bacterial disease.

[0048] According to an embodiment, the method further comprises generating a cluster by clustering of images of the first agricultural domain and generating a cluster by clustering of images of the second agricultural domain, respectively, and selecting a set of images of the cluster of the first agricultural domain and the cluster of the second agricultural domain, respectively, and use the respective selected set images as the source domain image data and target domain image data, respectively.

[0049] Thus, it is possible to reduce the effort and required computational power. The clustering may render the entire procedure more robust, as related data may be clustered. As an example, clustering may combine same and differentiate different aspects, e.g., in case all data relating to a particular soil color are clustered. In case a data pool includes domain data relating to, e.g., 90% red soils and 10% grey soils, and a clustering is carried out on red soils and grey soil, deterioration of the red soil data by grey soil data can be avoided and vice versa. By clustering the labeled image data and / or the unlabeled image data with respect to the soil background information, the variance in the images can be reduced so that a higher quality of soil-adapted image augmentations can be provided and produced, respectively. This in turn may, e.g., improve the quality of a machine learning process and an image recognition model when trained on original images of a source domain and soil adapted augmented images of a target domain.

[0050] According to an embodiment, the method further comprises providing background adapted image augmentations where items of the target domain are augmented with a background which is derived from the source domain and the transformation relation. The background adapted image augmentations may provide images where items of the target domain are augmented with a background which is derived from the source domain and the transformation relation.

[0051] Thus, the aspects and items which may be identified on a target domain image can be augmented with aspects and items derived from the source domain data and the transformation relation.

[0052] According to an embodiment, determining a transformation relation between the source domain image data and target domain image data is carried out using a generative adversarial network GAN, in particular a cycleGAN.

[0053] Thus, the use of the generative adversarial network (GAN) may support determination of a transformation relation to generate predicted image data including labeled image data based on soil background information, such that the labeled image data can be easily and time and cost efficiently generated. As an example, a GAN generator may firstly create a fake image from real images of fields (e.g., predicted images labeled with plant identifiers relating to the target domain). For this purpose, the GAN may, e.g., process labeled images from a source domain in Brazil and unlabeled images from a target domain in Argentina. A GAN discriminator than may try to find out whether the image is a fake image or real image from target domain. This comparison may be used to improve and consolidate the transformation relation. The GAN then may generate corresponding predicted images of the target domain from the images of the source domain. The GAN may recognize the matching aspects of the images of the source domain and the target domain and "overlay" them with the delta aspect of the target domain. This may happen on the layer of the unlabeled images. According to an embodiment, then the metadata of the labels may be transferred either one-to-one or according to a predetermined relation. Thus, the process of generating training data also related to a target domain can be improved with regard to the determination of the transformation relation by using a generative adversarial network GAN accelerating the scalability of machine learning models to new unseen regions, i.e., target domain, for which, e.g., unlabeled images exist. For this purpose, as described above, e.g., soil colors may be transferred from a source domain to target domain using cycleGANs.

[0054] Images of the source domain and images of the target domain, which are not provided with labeling data, or provided with a reduced amount of labeling data, are clustered according to particular aspects, which may be, e.g., the aspect of background information, in particular according to soil information. Thus, the variance in the images is reduced and high quality soil adapted image augmentations are produced. This, in turn, improves the quality of a machine learning image recognition model when trained on original and soil adapted augmented images. Thus, the scalability of machine learning models to new unseen regions, i.e., a target domain, for which, e.g., unlabeled images exists can be accelerated.

[0055] The general application of cycleGANs is, e.g., described e.g. in Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. "Unpaired Image-to-lmage Translation using Cycle-Consistent Adversarial Networks", in IEEE International Conference on Computer Vision (ICCV), 2017; Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, Luc van Gool, "SCAN: Learning to Classify Images without Labels", European Conference on Computer Vision (ECCV), 2020; D. Gogoll, P. Lottes, J. Weyler, N. Petrinic, and C. Stachniss, "Unsupervised Domain Adaptation for Transferring Plant Classification Systems to New Field Environments, Crops, and Robots," in IROS , 2020; and Armin Mehri, Angel D. Sappa; “Colorizing Near Infrared Images Through a Cyclic Adversarial Approach of Unpaired Samples”, CVPR Workshops, 2019.

[0056] According to an embodiment, determining of the transformation relation includes applying the generative adversarial network GAN, wherein the generative adversarial network GAN comprises a generator being adapted for generating provisional predicted images for the second agricultural domain with parameters from source domain items based on the transformation relation, and a discriminator being adapted for evaluating the generated provisional predicted images for the second agricultural domain.

[0057] Thus, a more reliable and stable transformation relation may be achieved, as described above.

[0058] The term “applying” is to be understood broadly in the present case and comprises any method for providing or executing, e.g., a generative adversarial network GAN with the labeled image data from the source domain and the unlabeled image data from the target domain. The labeled image data from the source domain and the unlabeled image data from the target domain may be input data for the generative adversarial algorithm of the GAN. In other words, applying is the step of using the model and using the generation or predicted target domain image data after it has been successful trained. As mentioned above, input data, i.e., the labeled image data from the source domain and the unlabeled image data from the target domain, are passed to the neural network, in particular a GAN, which may determine a transforming relation of the visual appearance of the labeled source domain data to the target domain (e.g., while maintaining the individual labels). Therefore, the output may be images with a visual appearance of the target domain and having individual labels (e.g., boxes, segmentations, etc.) of the input image. This output data can be used to train new neural network(s) with more variant data without the need of providing manually labeled data from the target domain, e.g., a CNN may be trained for image recognition with more variant data, i.e., target domain data in addition to the source domain data.

[0059] According to an aspect of the invention, there is provided a computer implemented method for training a neural network to be trained, the method comprising: providing to the neural network to be trained source domain data relating to a first agricultural domain, wherein the source domain data include source domain image data relating to a source domain image and source domain labeling data, wherein the source domain image data including source domain item data relating to at least one source domain item of the first agricultural domain and source domain background data relating to a source domain background of the first agricultural domain, wherein the source domain labeling data relating to a parameter of the at least one source domain item; providing to the neural network to be trained predicted target domain data generated according to the computer implemented method of the above described aspect; and training the neural network to be trained based on the source domain data and the predicted target domain data.

[0060] Thus, more and better training data can be obtained to train a neural network, Further, it is not required to prepare all labeled data and images of the source domain again manually for the target domain. The neural network and the model can be trained with data which again are generated inter alia by a neural network. Thus, a kind of two-step approach may be carried out with a neural network supported data generation and a training of a further neural network with such generated data. It should be noted that the neural network of the first step may be different from the neural network of the second step, and may be chosen or designed adapted to the requirement of the respective step. However, both neural networks may also be implemented within a single neural network, if desired and required.

[0061] The term “machine learning procedure” is to be understood broadly in the present case and may comprise any process including decision trees, Naive Bayes classifications, nearest neighbors, neural networks, convolutional or recurrent neural networks, generative adversarial networks GAN, support vector machines, linear regression, logistic regression, random forest and / or gradient boosting algorithms, but is not limited thereto. The machine-learning procedure may be organized to process an input having a high dimensionality into an output of a much lower dimensionality. Such a machine-learning algorithm is termed “intelligent” because it is capable of being “trained.” The algorithm may be trained using records of training data. A record of training data comprises training input data and corresponding training output data. The training output data of a record of training data may be the result that is expected to be produced by the machine-learning algorithm when being given the training input data of the same record of training data as input. The deviation between this expected result and the actual result produced by the algorithm may be observed and rated by means of a “loss function”. This loss function may be used as feedback for adjusting the parameters of the internal processing chain of the machine-learning algorithm. For example, the parameters may be adjusted with the optimization goal of minimizing the values of the loss function that result when all training input data is fed into the machine-learning algorithm and the outcome may be compared with the corresponding training output data. The result of this training may be that given a relatively small number of records of training data as “ground truth”, the machine learning algorithm is enabled to perform its purpose well for a number of records of input data that higher by many orders of magnitude.

[0062] The term “training data” is to be understood broadly in the present case and comprises any data set of examples used during the learning process of a machine-learning algorithm or machine-learning model and is used to fit the parameters (e.g., weights) for a machine learning algorithm for image classification a plant, weed and / or pest in the target domain, but is not limited thereto. The transformed labeled image data from the source domain and the unlabeled image data from the target domain may be used as training data for a machine-learning process for image classification a plant, weed and / or pest in the target domain, but is not limited thereto. Additionally, the training data may be data being specifically adapted for corresponding specific machine-learning process or machine-learning models but is not limited thereto.

[0063] According to an embodiment, training of the neural network to be trained is based on the source domain image data as labeled source domain image data with source domain labelling data and predicted labeled target domain image data of the predicted target domain data.

[0064] According to an embodiment, the neural network to be trained is a convolutional neural network (CNN) and comprises an image recognition model. According to an embodiment, the method of training a neural network comprises deploying a model based on the training of the neural network based on the source domain image data and the predicted target domain image data.

[0065] According to an aspect of the invention, there is provided a use of source domain data and predicted target domain data as defined according to the computer implemented method as described above for training of a neural network.

[0066] According to an aspect of the invention, there is provided a system for generating training data for a neural network to be trained, the system comprises: a source domain data input interface being adapted for receiving source domain data relating to a first agricultural domain, and including source domain image data relating to a source domain image and source domain labeling data, wherein the source domain image data include source domain item data relating to at least one source domain item of the first agricultural domain and source domain background data relating to a source domain background of the first agricultural domain, wherein the source domain labeling data relate to a parameter of the at least one source domain item; a target domain data input interface being adapted for receiving target domain data relating to a second agricultural domain, wherein the target domain data include target domain image data relating to a target domain image, wherein the target domain image data including target domain item data relating to at least one target domain item of the second agricultural domain and target domain background data relating to a target domain background of the second agricultural domain; an image recognition unit being adapted for identifying at least one target domain item in the target domain image; a training data generation neural network, which is adapted for determining a transformation relation between the received source domain image data and the received target domain image data based on a relation of the source domain background data and the target domain background data, determining based on the transformation relation, based on the source domain image data and based on the source domain labeling data target domain labeling data relating to a parameter of the at least one target domain item of the second agricultural domain, allocating the determined target domain labeling data to the identified target domain item data relating to the at the least one target domain item identified in the target domain image data; and generating predicted target domain data with the target domain image data including the target domain background data and the target domain item data, and the determined target domain labeling data.

[0067] In an example the system for generating training may also comprise the neural network to be trained. In this way, the predicted target domain data and / or source domain data may be forwarded to the network to be trained. The output interface of the system for generating training data for the neural network may be connected to an input interface of the neural network to be trained.

[0068] According to an aspect of the invention, there is provided an apparatus for generating training data for a neural network, the apparatus comprising one or more computing nodes and one or more computer- readable media having thereon computer-executable instructions that are structured such that, when executed by the one or more computing nodes, cause the apparatus to perform the computer implemented method as described above.

[0069] According to aspect of the invention, there is provided a computer program element which when executed by a processor is configured to carry out the computer implemented method as described above.

[0070] Any disclosure and embodiments described herein relate to a computer implemented method, an apparatus, a system, a model, a neural network, a computer program element, and a use thereof as outlined out above and vice versa. The technical effect provided by any of the embodiments and examples equally apply to all other embodiments and examples and vice versa.

[0071] BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In the following, the present disclosure is further described with reference to the enclosed figures:

[0073] Fig. 1 illustrates a schematic diagram of a computer-implemented method for providing predicted target source image data according to an exemplary embodiment of the invention.

[0074] Fig. 2 illustrates a schematic diagram of a computer-implemented method for providing predicted target source image data and training data for a machine learning process according to an exemplary embodiment of the invention.

[0075] Fig. 3 illustrates a schematic sketch of a source domain / first agricultural domain with labeled image data and a target domain / second agricultural domain with un-labeled image data for providing a neural network therewith according to an exemplary embodiment of the invention. Fig. 4 illustrates a schematic sketch of a source domain with labeled image data analogue to Fig. 3 and a target domain with labeled image data generated by a neural network according to an exemplary embodiment of the invention.

[0076] Fig. 5 illustrates a schematic sketch of a data processing of data of a source domain with labeled image data and a target domain with un-labeled image data for providing a neural network therewith according to an exemplary embodiment of the invention.

[0077] Fig. 6 illustrates a schematic sketch of a data processing of data of a source domain with labeled image data and a target domain with labeled image data generated by a neural network and providing the same to a neural network to be trained according to an exemplary embodiment of the invention.

[0078] Fig. 7 illustrates a schematic sketch of a system for data processing of data of a source domain and a target domain based on application of a neural network according to an exemplary embodiment of the invention.

[0079] DETAILED DESCRIPTION OF EMBODIMENT

[0080] The following embodiments are mere examples for implementing the method, the apparatus, the system or application device disclosed herein and shall not be considered limiting.

[0081] The invention provides a computer implemented method for generating target domain data, a computer implemented method for training a neural network, and a system, an apparatus, a use, and a computer program element adapted therefor, including providing to the neural network source domain data relating to a first agricultural domain; providing to the neural network target domain data relating to a second agricultural domain; determining a transformation relation between the source domain image data and target domain image data; and generating predicted target domain data with the target domain image data including the target domain background data and the target domain item data, and the target domain labeling data relating to a parameter of at least one target domain item identified in the target domain image of the second agricultural domain.

[0082] The references used in the following relating to elements and items to which, e.g., data relate, refer to the illustrations in Fig. 3 and 4. The references used in the following relating to data refer to the data flow structure as illustrated in Fig. 5 and Fig. 6. The references used in the following relating to structure refer to the structure as illustrated in Fig. 7.

[0083] Fig. 1 illustrates a schematic diagram of a computer-implemented method for providing predicted target source image data 206 according to an exemplary embodiment of the invention. For the purpose of the method, source domain data 105 relating to a first agricultural domain 100 are provided to a neural network (NN) 300. The source domain may be an agricultural field, e.g., in a particular region of the world, e.g., in Brazil. The agricultural field may be a field for which a large amount of labeled images are available. The source domain data 105 include source domain image data 115 relating to a source domain image 110 and source domain labeling data 155. The source domain image 110 may be an image taken by a camera, which camera by be a camera on a field device or a remote camera like a satellite camera. The source domain image data 115 includes source domain item data 135 relating to at least one source domain item 131, 132 of the first agricultural domain 100 and source domain background data 145 relating to a source domain background 140 of the first agricultural domain 100. The source domain item 131 , 132 may be an item on the source domain image 110, e.g., a plant or a weed, but is not limited thereto. The background 140 may be, e.g., soil, wherein the image may allow recognizing structures and color of the soil. The source domain labeling data 155 relate to a parameter 161 , 162 of the at least one source domain item 131, 132. The labeling data may be, e.g., information on a particular plant as an item 131 , 132, which information may be, e.g., a species, growth stage, specific abnormalities, nutrition and watering level and saturation and the like. The labeling in the past was done manually, this means that a person had analyzed an image and had allocated particular information to an identified item on the image. This procedure was very time consuming and it is desired to not need to repeat this for other regions, but to rather transfer this knowledge for other regions.

[0084] For this purpose, a training data generation neural network 300 is provided not only with the source domain data, which includes labeled image information, but also with target domain data 205 which relates to a second agricultural domain 200, which may also a region somewhere in the world, e.g., in Argentina, for which however no or less labeled image data is available. The target domain data 205 include target domain image data 215 relating to a target domain image 210. The target domain image data 215 include target domain item data 235 relating to at least one target domain item 231 , 232, which may be plant or a weed growing in the second agricultural domain 200. The target domain image data 215 also include target domain background data 245 relating to a target domain background 240 of the second agricultural domain 200, which is, e.g., the soil of the second region. The target domain image data 215 however has no labeling or less labeling than the source domain image data, wherein the labeling knowledge of the of source domain should be made available also for the target domain. The training data generation neural network 300 therefore determines a transformation relation 305 between the source domain image data 105 and target domain image data 205, which may allow using the labeling knowledge of the source domain also in the target domain. As the labeling relates to particular items in the respective domain, target domain items 231 , 232 are identified in the target domain image 210, and based on the transformation relation 305, the source domain image data 115 and the source domain labeling data 155 target domain labeling data 255 are determined relating to the respective target domain item 231, 232 of the second agricultural domain 200. These target domain labeling data 255 are allocated to the respective identified target domain item 231 , 232 identified in the target domain image data 215. The transformation relation allows using the labeling data of the source domain for labeling of items of the target domain, although the background of the target domain may be different to the background of the source domain. Based on the gained information, in particular the labeling, predicted target domain data 206 are generated with the target domain image data 215 including the target domain background data 245 and the target domain item data 235. The predicted target domain data 206 also include the target domain labeling data 255 relating to a parameter 261 , 262 of the respective target domain item 231 , 232 identified in the target domain image 210. Thus, the images of the target domain may be labeled based on the transformation relation between the source and target domain, and the labeling of the source domain, without the need for a manual labeling of the target domain.

[0085] Fig. 2 illustrates a schematic diagram of a computer-implemented method for providing predicted target source image data and training data for a machine learning process according to an exemplary embodiment of the invention. As above described with respect to Fig. 1 , labeled images 110, or to be exact data of labeled images 115 of a first domain, here the source domain are used as input for generating labeled images for the target domain, or to be exact, for generating labeled image data. In the following providing, e.g., an image should be understood as providing image data for the purpose of data processing. Further, unlabeled or less labeled images 210 of a second domain, which is the target domain are provided. Both images, those of the source domain and those of the target domain can be clustered in order to increase the data quality. For the further data processing sets of images can be selected from the respective cluster of images of the respective domain. These selected set of images of the source domain and the target domain are applied to a training data generation neural network 300. While applying the selected set of images to the training data generation neural network 300, the neural network can also be trained. The training data generation neural network 300 predicts translated images from the target domain. These images of the target domain then are labeled and may be provided to get a soil adapted image augmentation, based on the soil background of the target domain, which usually is different from the soil background of the source domain. The predicted soil adapted augmented labeled images for the target domain and the labeled images of the source domain then are used for training an image recognition model to deploy a new model or to adapt or improve the present model.

[0086] Fig. 3 illustrates a schematic sketch of a source domain with labeled image data and a target domain with un-labeled image data for providing a neural network therewith according to an exemplary embodiment of the invention. Fig. 3 illustrates on the left an image 110 of the first or source domain for a first agricultural domain 100, which has a background 140, which is formed, e.g., by a soil on the agricultural field 100. The image 110 also includes items 131 , 132, which may be beneficial plants 131 like crop, or may be malicious plants 132, like weeds. The image also includes a labeling, which is here illustrated with boxes and only for illustrational purposes for a few items. Both, the beneficial plants 131 and the malicious plants 132 can be labeled. The beneficial plant 131 may for example include as labeling a crop species or a designation of plant parts, a growth stage, a nutrition or watering level or saturation or the like. The malicious plant 132 may for example include as labeling 162 a weed species or a disease identifier. It should be noted, that although not illustrated here, the beneficial plant 131 may also have a malicious labeling 162, in case this labeling identifies a disease on that plant.

[0087] Fig. 3 illustrates on the right an image 210 of the second or target domain for a second agricultural domain 200, which has a background 240, which is formed, e.g., by a soil on the agricultural field 200. This background usually is different from the background 140 of the first or source domain. The image 210 of the target domain also includes items 231 , 232, which may beneficial plants 231 like crop, or may malicious plants 232, like weeds. The image 210 of the target domain in this example however does not include a labeling. Both images 110, 210 of Fig. 3 can be provided to the training data generation neural network 300 for processing, as illustrated in Fig. 5.

[0088] Fig. 4 illustrates a schematic sketch of a source domain with labeled image data and a target domain with labeled image data generated by a neural network according to an exemplary embodiment of the invention. The left image 110 of the source domain is the same as that of Fig. 3 to which it is referred here. The training data generation neural network 300 has generated a labeling also for the image 210 of the second, here target domain. Although the items 231 , 232 of the target domain have different appearance in the image, the training data generation neural network 300 can derive the labeling from the source domain and apply the respective labeling to the items 231 , 232 of the target domain doe to a machine learning process. The right image is a generated predicted image for which no manual labeling was required, as the labeling was provided by the training data generation neural network 300, as illustrated in Fig. 5. Both images 110, 210 of Fig. 4 can be provided to a further neural network 400 to be trained for training this neural network, as illustrated in Fig. 6.

[0089] Fig. 5 illustrates a schematic sketch of a flow chart for processing data of a source domain comprising labeled image data and a target domain comprising un-labeled image data for providing a neural network therewith according to an exemplary embodiment of the invention. The training data generation neural network 300 is provided with source domain data 105 and with target domain data 205. The source domain data include source domain image data 115 and source domain labeling data 155. The source domain image data 115 include source domain item data 135 relating to at least one source domain item 131, 132 of the first agricultural domain 100 and source domain background data 145 relating to a source domain background 140 of the first agricultural domain 100, as described above.

[0090] The target domain data 205 include target domain image data 215 and the target domain image data 215. The target domain image data 215 include target domain item data 235 and target domain background data 245. Based on the source domain data 105 and the target domain data 205 the training data generation neural network 300 determines a transformation relation 305 for transforming data from the source domain to the target domain. Based thereon predicted target domain data 206 are generated with the target domain image data 215 including the target domain background data 245 and the target domain item data 235. The predicted target domain data 206 also includes the target domain labeling data 255 relating to parameters 261, 262 of the respective target domain item 231 , 232 identified in the target domain image 210.

[0091] Fig. 6 illustrates a schematic sketch of a flow chart for processing data of a source domain comprising labeled image data and data of a target domain comprising labeled image data generated by a neural network and providing the same to a neural network to be trained according to an exemplary embodiment of the invention. The neural network to be trained 400 is provided with source domain data 105 and with predicted target domain data 206. The source domain data include source domain image data 115 and source domain labeling data 155. The source domain image data 115 include source domain item data 135 relating to at least one source domain item 131, 132 of the first agricultural domain 100 and source domain background data 145 relating to a source domain background 140 of the first agricultural domain 100, as described above. The predicted target domain data 206 are generated as described above with the target domain image data 215 including the target domain background data 245 and the target domain item data 235. The predicted target domain data 206 also include the target domain labeling data 255 relating to parameters 261, 262 of the respective target domain item 231 , 232 identified in the target domain image 210, as described above. As the predicted target domain data 206 are generated by applying a training data generation neural network 300, as described in Fig. 5, and these generated predicted target domain image data 206 are again used for training a further neural network 400, the entire process can be considered as a kind of stepped procedure. It should be noted that both steps can be performed directly in sequence. Alternatively, the second step of training the neural network 400 can be performed later, e.g., after having generated a larger data pool of a larger number of generated predicted target domain image data 206. It should be noted that the respective domain data 105, 205, 206 may include domain image data 115, 215 for a plurality of images 110, 210 with different items thereon in order to provide a broad data source for a high quality machine learning process.

[0092] Fig. 7 illustrates a schematic sketch of a system 10 for data processing of data of a source domain 100 and a target domain 200 based on application of a neural network according to an exemplary embodiment of the invention. The system 10 comprises a source domain data input interface 11 . The source domain data input interface 11 is adapted for receiving source domain data 105 relating to a first agricultural domain 100, as described with respect to Fig. 1 and Fig. 5. The system 10 further comprises a target domain data input interface 12. The target domain data input interface 12 is adapted for receiving target domain data 205 relating to a second agricultural domain 200, as described with respect to Fig. 1 and Fig. 5.

[0093] The system 10 further comprises an image recognition unit 13 being adapted for identifying target domain items 231, 232 in the target domain image 210. The system 10 further comprises a training data generation neural network 300, which is adapted for determining a transformation relation 305 between the received source domain image data 105 and the received target domain image data 205, as described above with respect to Fig. 1 and Fig. 5. Further, the training data generation neural network 300 is adapted for determining based on the transformation relation 305, the source domain image data 115 and the source domain labeling data 155 target domain labeling data 255 relating to at least one target domain item 231 , 232 of the second agricultural domain 200; allocating the determined target domain labeling data 255 to the at least one target domain item 231 , 232 identified in the target domain image data 215; and generating predicted target domain data 206 as described above with respect to Fig. 1, Fig. 4 and Fig. 5. The system 10 further comprises a predicted target domain data output interface 14 being adapted for output of the predicted target domain data 206. The present disclosure has been described in conjunction with exemplary embodiments and modifications thereof. However, other variations can be understood and effected by those persons skilled in the art and practicing the claimed invention, from the studies of the drawings, this disclosure and the claims. Notably, in particular, the any steps presented can be performed in any order, i.e. the present invention is not limited to a specific order of these steps. Moreover, it is also not required that the different steps are performed at a certain place or at one node of a distributed system, i.e. each of the steps may be performed at a different nodes using different equipment / data processing units. In the claims as well as in the description the word “comprising” does not exclude other elements or steps and the indefinite article “a” or “a” does not exclude a plurality. A single element or other unit may fulfill the functions of several entities or items recited in the claims. The mere fact that certain measures are recited in the mutual different dependent claims does not indicate that a combination of these measures cannot be used in an advantageous implementation.

Claims

Claims1 . Computer-implemented method for generating training data for a neural network to be trained (400), comprising: providing to a training data generation neural network (300) source domain data (105) relating to a first agricultural domain (100), wherein the source domain data (105) include source domain image data (115) relating to a source domain image (110) and source domain labeling data (155), wherein the source domain image data (115) include source domain item data (135) relating to at least one source domain item (131 , 132) of the first agricultural domain (100) and wherein the source domain image data (115) further include source domain background data (145) relating to a source domain background (140) of the first agricultural domain (100), wherein the source domain labeling data (155) relate to a parameter (161 , 162) of the at least one source domain item (131 , 132); providing to the training data generation neural network (300) target domain data (205) relating to a second agricultural domain (200), wherein the target domain data (205) include target domain image data (215) relating to a target domain image (210), wherein the target domain image data (215) include target domain item data (235) relating to at least one target domain item (231 , 232) of the second agricultural domain (200) and wherein the target domain image data (215) further include target domain background data (245) relating to a target domain background (240) of the second agricultural domain (200); determining by the training data generation neural network (300) a transformation relation (305) between the source domain image data (115) and target domain image data (215) based on a relation of the source domain background data (145) and the target domain background data (245); identifying the target domain item data (235) relating to at least one target domain item (231 , 232) in the target domain image data (215) relating to target domain image (210); determining based on the transformation relation (305), based on the source domain image data (115) and based on the source domain labeling data (155) target domain labeling data (255) relating to a parameter (261 , 262) of the at least one target domain item (231 , 232) of the second agricultural domain (200); allocating the determined target domain labeling data (255) to the identified target domain item data (235) relating to the at least one target domain item (231 , 232) identified in the target domain image data (215); andgenerating predicted target domain data (206) comprising the target domain image data (215) including the target domain background data (245) and the target domain item data (235), and the determined target domain labeling data (255).

2. Method according to claim 1, wherein the source domain background data (145) relate to soil background data and the target domain background data (245) relate to soil background data, wherein the soil background data in particular are soil color data.

3. Method according to any one of the preceding claims, wherein the source domain item (131 , 132) includes a beneficial source domain item (131) and the target domain item (231 , 232) includes a beneficial target domain item (231), and the source domain labeling data (155) relate to a parameter(161) of the beneficial source domain item (131) and the target domain labeling data (255) relate to a parameter (261) of the beneficial target domain item (231).

4. Method according to any one of the preceding claims, wherein the source domain item (131 , 132) includes a malicious source domain item (132) and target domain item (231 , 232) includes a malicious target domain item (232), and the source domain labeling data (155) relate to a parameter(162) of the malicious source domain item (132) and the target domain labeling data (255) relate to a parameter (262) of the malicious target domain item (232).

5. Method according to any one of the preceding claims, wherein the method further comprises generating a cluster by clustering of images of the first agricultural domain (100) and generating a cluster by clustering of images of the second agricultural domain (200), respectively, and selecting a set of images of the cluster of the first agricultural domain (100) and the cluster of the second agricultural domain (200), respectively, and use the respective selected set images as the source domain image data (115) and target domain image data (215), respectively.

6. Method according to any one of the preceding claims, wherein the method further comprises providing background adapted image augmentations where items of the target domain are augmented with a background which is derived from the source domain and the transformation relation.

7. Method according to any one of the preceding claims, wherein determining a transformation relation (305) between the source domain image data and target domain image data is carried out using a generative adversarial network GAN.

8. Method according to claim 7, wherein the determining of the transformation relation (305) includes applying the generative adversarial network GAN wherein the generative adversarial network GAN comprises a generator being adapted for generating provisional predicted images for the second agricultural domain (200) with parameters (161 , 162) from source domain items (131 , 132) based on the transformation relation (305), and a discriminator being adapted for evaluating the generated provisional predicted images for the second agricultural domain (200).

9. Computer implemented method for training a neural network (400) to be trained, the method comprising: providing to the neural network to be trained (400) source domain data (105) relating to a first agricultural domain (100), wherein the source domain data (105) include source domain image data (115) relating to a source domain image (110) and source domain labeling data (155), wherein the source domain image data (115) including source domain item data (135) relating to at least one source domain item (131 , 132) of the first agricultural domain (100) and source domain background data (145) relating to a source domain background (140) of the first agricultural domain (100), wherein the source domain labeling data (155) relating to a parameter (161 , 162) of the at least one source domain item (131, 132); providing to the neural network to be trained (400) predicted target domain data (206) generated according to the computer implemented method according to any of the preceding claims; and training the neural network to be trained (400) based on the source domain data (105) and the predicted target domain data (206).

10. Method according to claim 9, wherein training of the neural network to be trained (400) is based on the source domain image data (115) as labeled source domain image data (116) with source domain labelling data (155) and predicted labeled target domain image data (216) of the predicted target domain data (206).11 . Method according to any one of claims 9 and 10, wherein the method of training a neural network comprises deploying a model based on the training of the neural network based on the source domain image data and the predicted target domain image data.

12. Use of source domain data (105) and predicted target domain data (206) as defined according to any of claims 1 to 8 for training of a neural network.

13. System for generating training data for a neural network to be trained (400), the system (10) comprises: source domain data input interface (11) being adapted for receiving source domain data (105) relating to a first agricultural domain (100), and including source domain image data (115) relating to a source domain image (110) and source domain labeling data (155), wherein the source domain image data (115) include source domain item data (135) relating to at least one source domain item (131 , 132) of the first agricultural domain (100) and source domain background data (145) relating to a source domain background (140) of the first agricultural domain (100), wherein the source domain labeling data (155) relate to a parameter (161 , 162) of the at least one source domain item (131, 132); target domain data input interface (12) being adapted for receiving target domain data (205) relating to a second agricultural domain (200), wherein the target domain data (205) include target domain image data (215) relating to a target domain image (210), wherein the target domain image data (215) include target domain item data (235) relating to at least one target domain item (231 , 232) of the second agricultural domain (200) and target domain background data (245) relating to a target domain background (240) of the second agricultural domain (200), image recognition unit (13) being adapted for identifying at least one target domain item (231, 232) in the target domain image (210); a training data generation neural network (300), which is adapted for determining a transformation relation (305) between the received source domain image data (115) and the received target domain image data (215) based on a relation of the source domain background data (145) and the target domain background data (245); determining based on the transformation relation (305), based on the source domain image data (115) and based on the source domain labeling data (155) target domain labeling data (255) relating to a parameter (261 , 262) of the at least one target domain item (231, 232) of the second agricultural domain (200);allocating the determined target domain labeling data (255) to the identified target domain item data (235) relating to the at least one target domain item (231 , 232) identified in the target domain image data (215); and generating predicted target domain data (206) with the target domain image data (215) including the target domain background data (245) and the target domain item data (235), and the determined target domain labeling data (255); and predicted target domain data output interface (14) being adapted for output of the predicted target domain data (206).

14. An apparatus for generating training data for a neural network, the apparatus comprising one or more computing nodes and one or more computer-readable media having thereon computerexecutable instructions that are structured such that, when executed by the one or more computing nodes, cause the apparatus to perform the computer implemented method according to any one of claims 1 to 11.

15. A computer program element which when executed by a processor is configured to carry out the method of any one of claims 1 to 11 .

Images