Computer implemented quantification of seed moisture content and / or thousand kernel weight based on color images of a seed sample

Abstract

The invention relates to an easy-to-handle tool for non-invasive indirect measurement of seed parameter values based on acquisition and computer implemented analysis of an image of a small seed sample collected in the field. The solution can be implemented as portable device or as system comprising one or more portable image acquisition tools connected to a remote computer system for image analysis.

Description

Analyzing the part of the image pixels assigned to seeds, by way of one or more of: Determining, using a moisture content model trained to solve a regression problem of correlating image features with seed moisture content value, a seed moisture content value for the seed sample based on feature analysis of the image pixels assigned to seeds; and / or Determining, using a TKW model trained to solve a regression problem of correlating image features with TKW value, a TKW value for the seed sample based on feature analysis of the image pixels assigned to seeds Causing output of one or more of the determined value(s).

[0130] Further object of the application is a computer-implemented method comprising: Causing receiving image data from a scene comprising a seed sample of a plurality of seeds, said image data comprising quantized values representing color of image pixels; Assigning each image pixel of the image data to seeds or background using a segmentation algorithm; Analyzing the part of the image pixels assigned to seeds, by way of one or more of: Determining, using a moisture content model trained to solve a regression problem of correlating image features with seed moisture content value, a seed moisture content value for the seed sample based on feature analysis of the image pixels assigned to seeds; and / or Determining, using a TKW model trained to solve a regression problem of correlating image features with TKW value, a TKW value for the seed sample based on feature analysis of the image pixels assigned to seeds; Causing output of one or more of the determined value(s).

[0131] Image data can be received from the image acquisition tool described above or from a data repository comprising image data acquired with the imaging device.

[0132] In an embodiment, the one or more processors are part of a remote unit for image analysis. Alternatively, the one or more processors of the image acquisition tool described above can comprise the one or more processors configured to conduct image analysis, that is the one or more processors can be embedded in the imaging device. For example, a mobile phone or a smartphone with imaging and computing capacity can be used. In an embodiment, the solution can be configured for offline / remote use with limited computing capacity. Embodiment for “light” configuration is further described below.

[0133] Image pretreatment

[0134] In an embodiment, the one or more processors can be configured to conduct pretreatment of the image data. The one or more processors can be configured to cause display of the results of said pretreatment of image data and / or introduce the pretreated image data into one or more algorithms for analysis of image data.

[0135] In an embodiment, image pretreatment comprises quality check of image data. In an embodiment, the solution can be configured to generate an alarm in case inadequate condition for image acquisition are identified, for example insufficient contrast, inadequate sample (too much, too little seeds), inadequate seed positioning in the image (seed cumulated in one location), default in lighting, lack of definition (blur), etc.

[0136] Processors configured for image pretreatment may be integrated in the imaging device, in which case treated image data can be transmitted or stored in an image data repository for later analysis. Alternatively, or additionally processors configured for pretreatment of image data can be integrated in a remote computer system for image analysis.

[0137] As the imaging device of the invention allows controlled reproducible imaging conditions, minimal pretreatment is sufficient.

[0138] In an embodiment, image pretreatment comprises automated check for enough seeds in the image (generically referred as seed count value).

[0139] In an embodiment, the automated check of seed number is based on the ratio of the surface of the seed collecting area covered by the seed sample to the whole surface of the seed collecting area. As a matter of example, for best computer-implemented model-based analysis of a OSR seed sample, it was found advantageous that a sample comprising around 300 seeds is imaged. 300 OSR seeds as a monolayer statistically roughly cover one quarter of the blue seed collecting area of the image acquisition device described above. In an embodiment, adequate number of seeds can be checked by visual inspection of the surface covered with seeds or blue background. Alternatively, an algorithm can be implemented which automatically assign pixels to seeds (not blue pixels) or background (blue pixels) and evaluate the surface of the image covered by pixels assigned to seeds in the image compared to the number of pixels attributed to background / blue seed collecting area. In an embodiment, seed count can be determined by a segmentation algorithm as further described below. In an embodiment, a warning can be implemented to be displayed in the application in case there are not enough seeds in the image. Warning may or may not prevent the user from saving the image. The system may be configured to warn the user before or after image analysis, that the seed sample was outside the optimal range for seed count. In an embodiment, the system can be configured to disregard images, wherein the number of seeds is outside the required range or below a minimum threshold for seed count.

[0140] In an embodiment, image pretreatment may comprise cropping and / or resizing the image to focus on the seed collecting region, preferably the whole surface of the collecting region for further analysis.

[0141] Cropping of image as used herein refers to dividing received image into smaller patches (x, y) to remove unwanted areas from an image, typically outer areas are cropped to improve image framing, change the aspect ratio, or accentuate or isolate the subject matter from its background In an embodiment of the present solution, cropping of image is done to remove the unwanted surrounding and not relevant areas from the image and to select the region of interest (ROI).

[0142] In an embodiment, image pretreatment may comprise image resizing. It is preferred that image resizing is conducted after image cropping. In an embodiment, image resizing comprises down scaling the image.

[0143] In an embodiment, it is preferred to fit the image to the requirements / characteristics of utilized trained model(s) for image analysis, in particular neural networks further described below. In an embodiment, cropping is conducted with cropping parameters used for cropping the training data set for model training. In an embodiment, resizing is conducted with resizing parameters used for resizing the training data set for model training.

[0144] In an embodiment, image pretreatment comprises one or more of blurriness check or blur filtering. For this purpose, an algorithm is used that checks the blurriness of the image with a threshold. In an embodiment, the image is converted to a gray scale image and a Laplacian function is used to calculate a blurriness coefficient. In an embodiment, the system is configured to display a warning if the image blurriness coefficient is below a defined threshold. It is preferred that the blurriness coefficient is saved in the image metadata and is available in the image repository. The system can be configured to warn the user before or after image analysis, that the blurriness coefficient is above a certain threshold. Alternatively, the system may be configured to automatically disregard (filter out, also referred to as “blur filtering”) images with insufficient sharpness. Known method for computer-implemented blurriness check can be used for blurriness check (e.g. https: / / ieeexplore.ieee.org / document / 7894491).

[0145] Thanks to reproducible imaging conditions provided by using the image device described above, normalization of environmental variations, in particular illumination effects is not needed. Considering variations related to the camera, detector sensitivity and / or embedded camera automatic image pretreatment algorithms, it is preferred that training data for model training and new images are acquired with the same camera / image acquisition unit.

[0146] Pretreated images are saved in an image repository and / or transmitted for image analysis by the one or more processors.

[0147] Seed count values

[0148] As already mentioned, a quality check for imaging relates to evaluating seed count.

[0149] In an embodiment, a seed count in seed sample can be determined based on the number of centroids identified by the segmentation algorithm further described below (output of the segmentation algorithm) from the image / heatmap of the seed sample.

[0150] In an embodiment, seed count is determined during performance of the ground truth methods for generation of a training data set. In an embodiment, ground truth seed count is stored in the image metadata. Ground truth seed count can be used to validate the segmentation algorithm.

[0151] Exact seed count is usually not needed for proper running of a model-based image analysis for a new sample. In most cases, a ratio of the surface of the seed collecting area covered by the seed sample to the whole surface of the seed collecting area is used as “seed count value” to evaluate adequate seed count criteria being fulfilled.

[0152] Determining best range for seed count values

[0153] As already mentioned, it was shown that the solution described above is improved when the number of seeds in a seed sample for image acquisition comprises enough separable seeds. For example, adequate number of OSR seeds is chosen within a range from 300 to 1000 seeds.

[0154] The skilled person will appreciate that the solution is not limited to OSR; it can be applied to seed types, which seed maturity can be recognized by color change, for example but not limited to brassica and canola.

[0155] In an embodiment, the solution comprises determining a best seed count range for which the solution is implemented.

[0156] In an embodiment, criteria for choosing the seed count range are one or more of: Seed size, seed form; The seed count range can be chosen to provide statistically relevant seed information for a location; and / or Measurement of ground truth values may require a particular seed count range. In an embodiment, seed count range for image-based measurement is like or close to seed count range for ground truth measurement for improved validity of model with ground truth values.

[0157] In an embodiment, the surface of the seed collecting region in the image acquisition device is chosen, so the seed sample with adequate seed count nicely spreads as a monolayer, and so it can be effectively reorganized on the seed collecting region by wobbling the imaging device (means only part of the seed collecting region is covered by the seed sample).

[0158] In an embodiment, best range for seed count is identified once by experiment for best performance of the model-based value determination. For this purpose, a training data set comprising image data and respective ground truth value with different seed counts is acquired, a model for value determination (moisture content model, TKW model or moisture dependent TKW model further described below) is trained with said training data set; validity of imagebased value determination in relation to seed count is determined in a validation test. Best seed count range showing best validity score is selected for robust method. A training data set comprising image data of a seed sample and respective ground truth value for the sample within best seed count range is prepared and the model is trained with said training data set.

[0159] In an example, seed samples containing from 50 to 1000 OSR seeds are prepared, with increments of 50 seeds using a seed counting machine (such as https: / / www.pfeuffer.com / de / produkt / contador). The number of seeds is stored as metadata for each sample as ground-truth number of seeds. Color images are acquired for each sample. The sample is submitted to one of the ground truth measurement methods below to obtain training data sets with different seed counts.

[0160] Sub-processes of image analysis

[0161] In an embodiment, the computer system comprises one or more processors configured to conduct image analysis in an automated manner using one or more algorithms by running a computer (sub-)program with corresponding instructions in a working memory of the one or more processors.

[0162] In an embodiment, the computer system comprises one or more method repository / library, wherein sub-programs comprising instructions for one or more of image pretreatment, model training, image analysis, data processing, etc., can be stored to be retrieved by the one or more processors and run as needed.

[0163] In an embodiment, trained models are stored for later use in a model repository of the computer system.

[0164] In an embodiment, one or more of the data, method and / or model repositories can be connected to the computer system using a transmitter unit.

[0165] Many algorithms for image analysis are known in the art; algorithms adapted to the solution described herein are detailed below.

[0166] As already mentioned, computer-implemented image analysis comprises: Assigning each image pixel of the image data to seeds or background; Selecting image part comprising the pixels assigned to seeds; Analyzing the image part comprising the pixels assigned to seeds, using one or more models / algorithms for image-based determination of values for the sample.

[0167] Image segmentation

[0168] In a first subprocess (also referred to as image segmentation), each pixel of the image data is assigned to seed or background using a segmentation algorithm.

[0169] Unsupervised assigning of pixels to seeds or background can be achieved by performing e.g. k-mean, k-nearest neighbours, DBSCAN methods, unsupervised trained classification networks, autoencoders, variational autoencoders or one or more trained classification networks, e. g. GAN networks.

[0170] Each seed comprises a group of pixels with similar spectral signatures; it can be characterized by its cluster centroid, TWCV and assigned a class label (e. g. seed / background or seed yes / no). Image pixel assigned to seeds are further referred to as “seed pixel”.

[0171] In an embodiment, assignment of pixel to seed or background can be achieved using a deep convolutional neural network (CNN) trained for image segmentation and configured to output a B&W segmentation map with two classes seed or background. Examples of deep convolutional neural network (CNN) and training thereof are further described below.

[0172] In an embodiment, an Encoder-Decoder Neural Network can be used to assign pixels to seed or background.

[0173] In an embodiment, a U-Net CNN model is used as a powerful encoder-decoder model. Alternatively, a SegNet deep CNN model can be used to perform image segmentation (ref. https: / / ieeexplore.ieee.org / abstract / document / 7803544).

[0174] In an embodiment, a customized version of the U-Net convolutional network / model (U-Net CNN) described by Ronneberger et al. is implemented to segment the seed region [Ronneberger et al., (2015): U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab, N., Hornegger, J., Wells, W., Frangi, A. (eds) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. MICCAI 2015. Lecture Notes in Computer Science, vol 9351. Springer, Cham, https: / / doi.org / 10.1007 / 978-3-319-24574-4 281], This algorithm was chosen for its exceptional performances in several semantic segmentation tasks in biological and biomedical applications [Ali et al., (2020): Predictive Modeling of Antibiotic Susceptibility in E. Coli Strains Using the U-Net Network and One-Class Classification, in IEEE Access, vol. 8, pp. 167711-167720, 2020, doi: 10.1109 / ACCESS.2020.30228292; Sodjinou et al., (2022): A deep semantic segmentationbased algorithm to segment crops and weeds in agronomic color images, information Processing in Agriculture, Volume 9, Issue 3, Pages 355-364],

[0175] In an embodiment, the customized version of the U-Net CNN model was trained from scratch on a training set of segmented images of seed samples.

[0176] In an embodiment, the training set of segmented images is obtained by way of manually segmenting images of seed samples using an image processing tool. Manual segmentation typically comprises cropping seed region by (automatic) thresholding, then performing morphological features (protrusion and concavity of each seed, curvature of the seed boundary, roughness, and / or non-smooth surface, etc....) to determine the best seeds region(s). In an example FIJI-ImageJ tool was used for manual segmentation of training images of seed samples. Then these seed sample images with the corresponding segmented images were used to train the U-Net model from scratch, wherein training the CNN from scratch comprises varying one or more of input image size, number and order of CNN layers, training options, etc.... [Training A Model From Scratch - MATLAB & Simulink (mathworks.com); https: / / de.mathworks.com / solutions / deep-learning / examples / training-a-model-from-scratch.html],

[0177] In an embodiment, the proposed version of U-Net CNN model was trained once on 25 images of OSR seed samples; the obtained trained U-Net model, can be used to select the seed region and create a heatmap of the seed region for new images within 1 second.

[0178] In an embodiment, the solution is configured to create a spatial heat map for the segmented seed region. A heat map (or heatmap) is a data visualization technique that shows magnitude of a phenomenon, for example color, in two dimensions. The variation in color may be by hue or intensity, giving obvious visual cues to the user about how the phenomenon is clustered or varies over space. Generating a heatmap of the segmented seed region images allows better visualization of the seed edges of segmented images. In an embodiment, a continuous heatmap is created to avoid very pixelized image. Figure 8c shows an example of a spatial continuous heatmap image generated from an image segmented by the trained U-Net CNN model.

[0179] In an embodiment, the solution is configured for causing display of the heatmap(s).

[0180] Seed size

[0181] In an embodiment, average seed sizes can be determined by way of dividing the number of pixels assigned to seeds, by the number of centroids identified by the segmentation algorithm. In an embodiment, the segmentation algorithm is configured for counting the pixels assigned to each identified centroid; number of pixels for each centroid is seed size. Average seed size, seed size range and / or seed size distribution can be computed for a sample based on determined singular seed size.

[0182] Narrow distribution of seed size for a seed sample for an area or sub-area is a sign of homogeneous crop maturation. Abnormalities in seed size distribution for a sample, an area or sub-area may be used to warn user that attention is needed and / or disregard image for analysis.

[0183] In other words, size analysis can be used as quality check to improve reliability of modelbased measurement by image analysis described below.

[0184] Color analysis

[0185] In an embodiment, the one or more processors are configured for estimating average color, color range and / or color distribution for a seed sample, or for a seed batch. Narrow color distribution for an area or sub-area is a sign of homogeneous crop maturation.

[0186] Narrow color distribution for a seed sample / for a seed batch is also advantageous for the robustness of the method. In an embodiment, abnormalities in color distribution for a sample, an area or sub-area can be identified by thresholding and used to warn user that the validity of the color-based measurements is at stake and / or attention is needed. In an embodiment, thresholding can be used to sort out images with abnormal color distribution.

[0187] In other words, color analysis can be used as quality check to improve reliability of model-based measurement by image analysis described below.

[0188] Color-based moisture estimation

[0189] For crops wherein the seed moisture content correlates with seed color, the seed color information from the color image of a seed sample can be used for computer-implemented estimation of seed moisture content value of the seeds. It is preferable that color images are captured with the imaging device described above for reproducible imaging conditions in the color range of interest.

[0190] In an embodiment, a sub-program comprising instructions is used for computer-implemented indirect estimation of a moisture content value for a seed sample based on analysis of seed color values in a color image of the seed sample.

[0191] In an embodiment, said color-based estimation of moisture content value comprises inputting part of image assigned to seeds, and optionally image metadata as needed, into an algorithm / a model configured for estimation of moisture content value based on image feature (seed color) values of seed pixels. This algorithm is further referred to as “moisture content model”.

[0192] In an embodiment, the moisture content model can be one or more machine learning algorithms, in particular neural networks (NN). In an embodiment, one or more machine learning algorithms trained as regressor. Two different types of NNs were tested: Convolutional NNs (CNN) and transformer-based NNs. Specifically, examples of tested CNNs are mobile net, vanilla network, inception v3, and ResNet50. Among the transformer-based NNs, the vision transformer method was tested. In an embodiment, ResNet50 convolutional neural network model. In the example described below for OSR seeds, ResNet50 convolutional neural network model resulted in the best mean squared error (MSE). The tested vision transformer method was found to require more seed images for training, which may be a problem when limited image data is available (only one harvest season for example).

[0193] In an embodiment, supervised correlation between image feature (seed color) values and moisture content value is achieved using trained CNN, e.g. ResNet50, as moisture content model. Training of moisture content model is achieved by self-learning CNN with a training algorithm for feature discovery in a training data set. Trained models can be stored for later use in a model repository as described above.

[0194] In an embodiment, the moisture content model is trained using a training algorithm and a training data set to learn extracting features for correlation between seed color values (image feature values) with moisture content value (measurement of interest).

[0195] A training data set for training of the moisture content model comprises color image data for several seed samples and at least ground truth for moisture content value for each seed sample as metadata, wherein the ground truth moisture content value for each imaged seed sample is obtained separately using established method of measurement for said value. Ground truth moisture content value or moisture content value of ground truth refers to a value measured by a ground truth measurement method for seed moisture content established in the field; ground truth measurement method can be invasive or non-invasive. Typically, it is a method conducted in a laboratory requiring more or less efforts, equipment and expertise.

[0196] In an embodiment, a training data set for training the moisture content model comprises color images of seed samples at different ripeness / maturity period (time point) comprising an adequate number of seeds and respective ground truth moisture content values, so training data cover the whole color range for the seeds of interest.

[0197] In an embodiment, samples with ground truth moisture content values in increments of around 5-10 % are considered for building the training data set.

[0198] In another embodiment, samples for training data set are collected on regular basis during one or more crop seasons (e. g. daily) in the area(s) of interest, typically starting when the seeds are separable.

[0199] In an embodiment, ground truth for moisture content value is obtained by thermogravimetric measurement with a determined number of seeds. Gravimetric measurement method typically comprises machine-counting the number of seeds in the sample, weighing the seed sample before and after drying the seed sample in an oven under defined conditions, and determining a seed moisture value, wherein the seed moisture content value equals seed weight loss after drying.

[0200] In an embodiment, a training data set is prepared, said training data set comprising color images of seed samples and respective ground truth moisture content value as measured by a ground truth measurement method. The seed samples for the training data set are selected within a range from moisture content value from high moisture content values (exhibited by humid, unmature, separable seeds) to low (harvestable seeds) and respective ground truth moisture values. In other words, training data comprises images for seed samples from the time point at which seeds can be separated by user for good quality image acquisition.

[0201] In an embodiment, it is preferred that a moisture content model is trained for each seed genotype (inbreeds). In this embodiment, image data comprises seed genotype, typically introduced by user, and saved as metadata. For new seed sample images, seed genotype information is used by for selection of moisture content model out the model repository; selected model is used for image analysis and model-based determination of moisture value.

[0202] In an embodiment, genotype metadata of the seed sample is considered as influencing factor for the model in the training data set. In other words, training data set comprises color image data and ground truth moisture content value for seeds samples for different genotypes. In an embodiment 20-30 genotypes are considered for training.

[0203] In an embodiment, the training data set is designed to comprise several samples for each condition, one condition being a combination of a given moisture content for a given genotype. For each condition, several batches of seeds / seed pods of each genotype are harvested at the same time at the same location. Samples for the same condition are together referred to as batch, typically a batch comprises two samples.

[0204] In an embodiment, further influencing factors for training of the moisture content model can be one or more of the locations of the seed sample acquisition, year to year variability, and weather conditions, etc.

[0205] In an embodiment, a training data set can be selected to comprise one or more representative locations for an area, field, growing region and / or macrozone.

[0206] In an embodiment, a training data set can be selected to comprise image data and ground truth data from different years to check the year-to-year variability and germination rate.

[0207] In an embodiment, a ResNet50 CNN is trained from scratch to perform as an image regressor on color images of the seed sample as acquired and evaluate a moisture content value for each seeds image. In an embodiment, the pretrained ResNet50 CNN model is fine-tuned by adding batch normalization layer, average pooling layer, dense layer, and dropout layer to improve the CNN learning and overcome the model overfitting.

[0208] According to good practice, performance evaluation of the trained model is tested after training using a data set comprising color images of seed samples with known ground truth for moisture content value of the seed samples. In an embodiment, leave-one-batch-out crossvalidation concept (LOBO-CV) is used to evaluate the performance of the trained models, in which one ground truth moisture content value for each batch is determined and used once as validation set. Color images of the validation data set are introduced to the trained model; average of determined color-based moisture values for a batch is compared with known ground truth of moisture content value for a batch. In an alternative embodiment, the collected imaging dataset is divided into two splits: training set and test set. Therein, the model is trained in the training set. Thereafter, the ground truth and the moisture determination based on the trained model are compared to evaluate the performance of the test set. This method is usually known as training-test split validation technique.

[0209] In an embodiment, for offline implementation of the solution on a device with limited computing ability, e. g. in a mobile application without cloud access to a computer system for image analysis, a ResNet50 model, is trained based on cropped color images of the seeds with the advantage to provide a light model to simplify the developments of the mobile app. Alternatively or additionally, a light customized version of a moisture content model can be used for running the solution in offline mode. For this purpose, a smaller model size is used to enable the offline app function. For example, TensorFlow Lite can be used instead of regular TensorFlow.

[0210] In an embodiment, the image data set for training the moisture content model comprises color images of similar size, wherein similar size is obtained by resizing the acquired images (pretreatment). It is preferred that the resizing parameters for the training data set are applied to new color image data for determination of moisture content value using the trained model.

[0211] Similarly, it is preferred to run a moisture content model trained with cropped color images on new color image data previously submitted to cropping pretreatment.

[0212] Model information for storage in the model repository therefore comprises pretreatment information applicable at model use. Model information may also comprise value on offline and / or cloud usability.

[0213] Size-based TKW estimation

[0214] In an embodiment, the seed size information can be used for computer-implemented estimation of TKW value for dried seeds from color images, preferably captured with the imaging device described above for reproducible imaging conditions in color range of interest.

[0215] In an embodiment, said size-based TKW value estimation for a seed sample comprises preprocessing image data of said seed sample (pretreatment described above; mainly the heatmap generation and images cropping and resizing) and inputting preprocessed image data, and image metadata as needed, into an algorithm / a model configured for estimation of TKW value based on analysis of the seed assigned pixels in the image. This algorithm is further referred to as “TKW model”.

[0216] In an embodiment, the TKW model can be one or more machine learning algorithms, in particular neural networks (NN), or convolutional networks (CNN). In an embodiment, one or more machine learning algorithms trained as regressor / trained for solving regression problem are used.

[0217] In an embodiment, supervised correlation between seed size and TKW value is achieved using trained CNN, e.g. ResNet50, as TKW model. Training of TKW model is achieved by selflearning CNN with a training algorithm for feature discovery in a training data set. Obtained trained models can be stored for later use in a model repository of the computer system or connected to the computer system.

[0218] In an embodiment, training is achieved using a model, a training data set, and a training algorithm for feature correlation. In an embodiment, a ResNet50 CNN model is finetuned. Its performance is evaluated based on training-test split validation (as described for color-based moisture estimation), wherein two image batches for each TKW value are used to train the ResNet50 and a third batch is used to evaluate the obtained trained model.

[0219] In an embodiment, a training data set for training the TKW model comprises images of seed samples at different ripeness / maturity period (time point) comprising an adequate number of seeds, and TKW value obtained for imaged seed sample by a ground truth measurement method. In other words, training data comprises images for seed samples from the time point at which seeds can be separated by user for good quality image acquisition. “Ground truth TKW value” or “ground truth for TKW value” refer to the value measured by ground truth measurement method for TKW.

[0220] In an embodiment, ground truth for TKW value is obtained by optimetric measurement using Opto-Agri von OptoMachines: https: / / optomachines.fr / home / optomachines-equipment-for-seeds-analysis / opto-agri-tsw-seed-biometry / . Said method comprises machine-counting the seeds of a sample. Optimetric measurement method using Opto-Agri from OptoMachines is established for TKW measurement on dry seeds. For ground truth TKW measurement of seeds, in particular immature / humid seeds, plants or plant pods are collected and first let dry in open air (simulating field swathing conditions) until that the seeds reach around 8% moisture content value (regular moisture content value for harvest time). Then the seeds are collected / cleaned and the ground-truth TKW is obtained using the Opto-Agri machine according to standard apparatus instructions. The skilled person will appreciate that other measurement method can be used for measurement of ground truth TKW for a seed sample; a measurement method established in the technical field is preferred.

[0221] In an embodiment, it is preferred that a TKW model is trained for each seed genotype. In this embodiment, image data comprises seed genotype, typically introduced by user, and saved as metadata. For new seed sample images, seed genotype information is used by for selection of TKW model out a model repository; selected model is used for image analysis and model-based determination of TKW value.

[0222] In an embodiment, genotype metadata of the seed sample is considered as influencing factor in the training data set. The training data set comprises images and ground truth TKW value for seeds samples for different genotypes. In an embodiment, 20 to 30 inbreeds are considered.

[0223] In an embodiment, further influencing factors can be considered for training of the TKW model, examples are mentioned above for the moisture content model.

[0224] In an embodiment, model information for storage of a trained TKW model in the model repository comprises pretreatment information applicable at model use.

[0225] Moisture dependent TKW estimation

[0226] In an embodiment, the seed color and the seed size information can be used for computer-implemented estimation of a moisture dependent TKW value from color image(s) of a seed sample.

[0227] In an embodiment, image-based estimation of characteristic value for a seed sample comprises inputting color image data preprocessed as described above, including optional metadata, into an algorithm / a model trained for determination of TKW value in relation to moisture content value based on feature extraction from color image data, said feature extraction considering at least seed color and seed size, optionally metadata for the seed sample. This algorithm / model is further referred to as “moisture dependent TKW model”. The embodiment aims to predict, from data (images and metadata) collected for the seed sample comprising immature / humid seeds (green and swallowed with water), the TKW the seed sample would have after the plants are swathed, dried and ready to harvest i.e. when the seeds are dried / mature (black and around 8% moisture).

[0228] In an embodiment, the moisture dependent TKW model can be one or more machine learning algorithms, in particular neural networks (NN). In an embodiment, one or more machine learning algorithms trained as regressor / trained for solving regression problem are used. Specifically, examples of CNNs are mobile net, vanilla network, inception v3, and ResNet50. Among the transformer based NNs, vision transformer method can be used.

[0229] In an embodiment, supervised correlation between image features like object area and color, moisture content value and TKW value can be achieved using trained CNN, e. g. ResNet50, as moisture dependent TKW model.

[0230] In an embodiment, training of a moisture dependent TKW prediction model is achieved by self-learning CNN with a training algorithm for feature discovery in a training data set.

[0231] Additionally or alternatively, one or more mechanistic model(s) can be generated that correlate the seed color with specific moisture content value and specific time point with a TKW value of dry seeds after harvest.

[0232] Trained / generated models can be stored for later use in a model repository of the computer system, or in a model repository connected to the computer system.

[0233] In an embodiment, said moisture dependent TKW model is obtained by training / generating a model with a training data set comprising image data for seed sample and at least measured ground truth TKW value and ground truth moisture content value as metadata and a training algorithm for feature correlation. Training data comprises color images of a plurality of samples comprising humid or mature seeds, wherein ground truth moisture content value and ground truth TKW value were measured for each sample. For generating a training data set, seed samples comprising an adequate number of seeds are collected at different ripeness / maturity period (time point t); ground truth values are obtained for imaged seed samples by respective ground truth measurement methods described above.

[0234] In an embodiment, the moisture dependent TKW model is trained with one or more additional influencing predictors. In an embodiment, influencing predictors are one or more of the metadata selected from seed genotype, time point of image acquisition in crop season, location of sample acquisition (field, field plot, sampling coordinates), weather data, historical field management information (field treatment, irrigation, soil information or the like), landscape information (orientation, topography, shadowing, ponding information, etc.).

[0235] In an embodiment, model information for storage in the model repository comprises one or more additional influencing predictors as parameters to be considered / parameter values to be acquired for use of the moisture dependent TKW model on new images.

[0236] In an embodiment, influencing predictor is genotype of the seed sample considered in the training data set. The training data set comprises images and ground truth values for seeds samples for different genotypes. In an embodiment, 20 to 30 inbreeds are considered.

[0237] In an embodiment, influencing predictors is location of sample acquisition (related to landscape, soil, climate information for the location). The training data set comprises images and ground truth values for seeds samples at different locations.

[0238] In an embodiment, the training data set comprises image data and related metadata over several crop seasons for better training.

[0239] The skilled person will appreciate that the seed growing process can be influenced by many factors; it may be advisable to retrain the models for value determination when valuable new data, in particular for a new season for a location or an area, become available, for better consideration of model influencing factors.

[0240] Running the solution on a new sample

[0241] In an embodiment, the solution described above is used for new color images of a seed sample.

[0242] In an embodiment, user may select one or more measurement(s) and / or data processing of interest from a list of available determination based on acquired images. The one or more processors are configured to run sub-processes based on selection and cause presentation of corresponding results.

[0243] In an embodiment, the one or more processors are configured to prompt the user for metadata entry and / or to cause (automatic) acquisition of values for influencing predictors from external sources (e.g. from data repository, external services, sensors, etc...) in consideration of selected measurement or data processing and / or definitions relating to sub-area(s) or area(s) for example limits thereof. Routines for such data acquisition are well known in the art.

[0244] In an embodiment, user can select determination of seed information for one or more of sample location(s), sub-area(s) or area(s) comprising the sample location(s).

[0245] In an embodiment, user can instruct storage of images and metadata in the data repository, e. g. for a crop season. Metadata for storage can comprises input metadata described above and / or measurements determined by the solution.

[0246] Results and further data processing

[0247] In the present application, “results” comprises one or more estimated values obtained from the computer-implemented analysis of image data.

[0248] As it is preferred that more than one images are acquired for each seed sample, eventually a set of images is acquired for a sample. In an embodiment, image analysis is conducted on each image of the set; results of image analysis can be averaged for the seed sample.

[0249] The one or more processors can be configured for causing display and / or transmission of the results. Said results can be provided to the user via peripherals. Additionally or alternatively, result can be further used for computer-implemented determination of further field relevant values and / or to generate scripts for controlling one or more field management devices (together referred to as further data processing).

[0250] In an embodiment, that the one or more processors can be configured to send an alarm signal if the automated analysis of the image data shows that a specific target value has been identified, for example estimated values being below and / or above predefined thresholds. Specific targets may be one or more of target moisture content value for pre-harvest, target moisture content value for harvest, target TKW value for (pre-)harvest, but not limited thereto.

[0251] In an embodiment, user can select one or more mode of presentation for the results. Examples of options and related sub-processes for data processing are further described below. However, the solution is not limited thereto.

[0252] In an embodiment, the result of image analysis is estimated value for each seed image. As already mentioned, it is preferred that several images are acquired for a seed sample from different sides (obtained by wobbling the image device between images). Alternatively or additionally, batch may be used, that is several samples may be acquired and imaged for a location. In other words, color images may show color variations.

[0253] In an embodiment, the solution comprises further data processing for the estimated values. Further data processing can comprise one or more of: Averaging estimated moisture content values for a seed sample and / or for a batch; Determining a moisture content value range for a seed sample and / or for a batch; and / or Determining a variability of the determined seed moisture content value for a seed sample.

[0254] When running the solution for an area or subarea, and data are acquired at several location further data processing may comprise determining averaged value, value range or value distribution for the area or sub-area.

[0255] In an embodiment, the solution is configured to generate one or more maps of the area comprising markings of sampling location(s) and / or distribution of the parameter values determined by image analysis (moisture, TKW, moisture-based TKW, seed color, seed size) for the area.

[0256] In an embodiment, the system may be configured to identify sub-areas or locations, wherein sampling and / or batch sampling may be advantageous or advisable.

[0257] In an embodiment, user can instruct determination of seed information for one or more areas or sub-areas.

[0258] In an embodiment, the system may be configured for: Acquiring, from a data repository, determined values for a seed sample parameter at a plurality of locations within an area or sub-area; Estimating a variability for the one or more determined values for seed samples within the area or sub-area; and / or Creating a map for the one or more determined values for seed samples within the area or sub-area.

[0259] In an embodiment, the system may be configured to issue a warning when said variability is beyond a predefined threshold, for example but not limited to providing advice / recommendation for field management measure.

[0260] In an embodiment, recommendation for a field management measure for one or more sub-areas can be displayed on a map.

[0261] In an embodiment, color images and metadata are acquired over time of a crop season.

[0262] In an embodiment, results of value determination as described above can be stored in a data repository. Best is storing determined value(s) with corresponding color images.

[0263] In an embodiment, the one or more processors can be configured to build one or more curves for determined values over time by way of: Acquiring, by the computer system from a data repository, determined values for a seed sample parameter at a plurality of time points t-x to t for a crop season at a location of interest; Generating a temporal measurement curve of the determined values for the season from the time point t-x to time point t for the crop season.

[0264] In an embodiment, the one or more processors are configured to cause display of the one or more of the determined single values and / or curves using a peripheral.

[0265] In an embodiment, the solution can be implemented to provide a time proposal for field management measure based on a forecasting for one or more of the parameter values determined by image analysis described above (moisture content, TKW, moisture dependent TKW, seed color, seed size).

[0266] Forecasting change for seed information / supporting timing for field management measures

[0267] In an embodiment, the solution is configured for determining evolution of one or more of determined seed parameter values - moisture content value, TKW value, moisture content dependent TKW value, seed size, and / or seed color - over time based on estimated seed information described above in consideration of forecast for influencing factors (weather forecast, irrigation plan, or the like).

[0268] In an embodiment, evolution is determined using previous determined values of the crop season to create a temporal measurement curve for the season.

[0269] In an embodiment, the solution is implemented to predict evolution of one or more of determined seed parameter values within a forecast period t+y (also referred to as prediction period t+y) after time point t of last measurement of said seed parameter value (image acquisition).

[0270] In an embodiment, forecasts are created for a future prediction period t+y, wherein y is 1 to 20 days into the future, more preferably for y of 1 to 10 days into the future (also referred to as forecasting period).

[0271] In an embodiment, for the purpose of forecasting, the solution further comprises: From a data repository, acquiring, by the computer system, determined values for a seed sample parameter at a plurality of time points t-x to t for a crop season at a location of interest; Generating a temporal measurement curve of the determined values for the season from the time point t-x to time point t for the crop season; Acquiring, by the computer system, for the location, one or more of climate data, weather data and field data for a future period t+y; Extrapolating the temporal measurement curve for the future period t+y in consideration of the one or more of climate data / , weather and / or field data for the location using an algorithm.

[0272] Climate data as used therein refers to as averaged temperature, humidity, atmospheric pressure, wind, and precipitation over a long period for a location.

[0273] Weather data as used therein refers to current weather, weather in the previous days, weather forecast for the next few days - that can be acquired from weather data services.

[0274] Field data refers to data acquired from field sensors; field data can comprise field weather data acquired from sensors for weather parameter (rain gauge, thermometer, soil sensors, and the like).

[0275] In an embodiment, the one or more processors are configured to compare weather data from weather services and field weather data acquired from field sensors. In an embodiment, the one or more processors are configured to cause warning the user that weather or climate data may false forecasting, e. g. in case weather data, climate data and field data are inconsistent or in case required prediction period is beyond reliable forecasting period.

[0276] In an embodiment, forecasting can be obtained using a model configured to determine (positive or adverse) effects weather could have on the seed moisture content value, the TKW value, the moisture dependent TKW value, etc....

[0277] An example of a solution for a machine learning-based seed harvest moisture predictions to improve computer-assisted agricultural farm operation is described in WO2020 / 086750 for com seeds. One or more of the seed values, in particular moisture content value, determined by the solution described herein can be used in the solution described in WO2020 / 086750.

[0278] In an embodiment, the one or more processors can be configured to identify a measure for field management which can be recommended to the user. For example, but not limited to, the one or more processors may be configured to determine a proposal for watering, preharvesting and / or harvesting time point or favorable period for the deployment of other field management measure. If growing / maturing effects of the coming weather on the seed are determined to be predominant, the solution may advise postponing harvesting or pre-harvesting.

[0279] In an embodiment, the user may select the field management measure for implementation by an agricultural device.

[0280] In an embodiment, the solution can be configured to generate scripts for controlling an agricultural device.

[0281] In an embodiment, image analysis is conducted by the smartphone using a computer program comprising necessary processors and instructions.

[0282] The computer system can comprise or be connected to peripherals for displaying test images, acquired images and / or results of image analysis. The system can comprise peripherals for displaying test images, acquired images and / or results of image analysis. The system can be configured to receive the results of image analysis from a remote computer system for display to the user on said peripheral.

[0283] In an embodiment, the device and / or the computer system can comprise a data repository for storage of the seed sample images and / or results of analysis thereof.

[0284] Objects of the application are therefore:

[0285] A system comprising: One or more peripherals; One or more data repository; One or more processors configured for: Causing receiving image data from a scene comprising a seed sample of a plurality of seeds, said image data comprising quantized values representing color of image pixels; Assigning each image pixel of the image data to seeds or background using a segmentation algorithm; Analyzing the part of the image pixels assigned to seeds, by way of one or more of: Determining, using a moisture content model trained to solve a regression problem of correlating image features with seed moisture content value, a seed moisture content value for the seed sample based on feature analysis of the image pixels assigned to seeds; and / or Determining, using a TKW model trained to solve a regression problem of correlating image features with TKW value, a TKW value for the seed sample based on feature analysis of the image pixels assigned to seeds; Causing output of one or more of the determined value(s).

[0286] In an embodiment, the system further comprises peripherals.

[0287] In an embodiment, the one or more processors are configured for acquiring metadata for the image data, said metadata being one or more of identification for seed sample, seed genotype, sample location, time point of image acquisition, weather data, historical field management information, and / or landscape information.

[0288] In an embodiment, the one or more processors are configured for pretreatment of the image data, said pretreatment comprising one or more of: Checking if a seed count value for the plurality of seeds is within a defined seed count range; Determining a blurriness coefficient for the image and filtering out images which blurriness coefficient is below a defined threshold; Cropping image data as defined by cropping parameters; Resizing image data as defined by resizing parameters.

[0289] In an embodiment, the moisture content model and / or the TKW model are machine learning algorithms.

[0290] In an embodiment, the one or more processors are configured for generating the trained moisture content model and / or the trained TKW model by way of Acquiring a training data set comprising color image data of a scene comprising a seed sample of a plurality of seeds and ground truth value for moisture content or TKW respectively measured for the seed sample using a ground truth moisture measurement method; Introducing the training data set into a model and training the model for feature extraction using a training algorithm to obtain a trained model; Saving the trained model in the model repository.

[0291] In an embodiment, the one or more processors are configured for determining a moisture-dependent TKW value based on feature analysis of the image pixels assigned to seeds using a moisture-dependent TKW model trained to solve a regression problem of correlating image features with TKW value in relation with seed moisture content value.

[0292] In an embodiment, the one or more processors are configured for generating the trained moisture-dependent TKW model by way of Acquiring a training data set comprising color image data of a scene comprising a seed sample of a plurality of seeds and at least both ground truth moisture content value and ground truth TKW value measured for the seed sample using a ground truth moisture measurement method and a ground truth TKW measurement method respectively; Introducing the training data set into a model and training the model for feature extraction using a training algorithm to obtain a trained moisture content model; Saving the trained moisture-dependent TKW model in a model repository.

[0293] In an embodiment, the one or more processors are configured for estimating one or more seed size values, by way of Determining a total number of pixels assigned to seeds in the image; Determining centroids for the pixels assigned to seeds by the segmentation algorithm; Determining a number of identified centroids; Counting for each centroid the number of pixels assigned to the centroid; Determining a range for seed size values based on the counted number of pixels per centroid; and / or Determining an average seed size by dividing the total number of pixels assigned to seeds by the number of centroids; Causing output of the one or more seed size values.

[0294] In an embodiment, for image data of a seed sample collected at time point t, the one or more processors are configured to predict evolution of one or more of determined seed parameter values within a prediction period t+y by way of: From a data repository, acquiring one or more of the determined values for a seed sample parameter at a plurality of time points t-x to t for a crop season at a location of interest; Generating a temporal measurement curve of the determined values for the season from the time point t-x to time point t for the crop season; Acquiring, for the location, one or more of climate data, weather data and field data for a future period t+y; Extrapolating the temporal measurement curve for the future period t+y in consideration of the one or more of climate data / , weather and / or field data for the location using an algorithm.

[0295] In an embodiment, the system further comprises one or more imaging device, said imaging device comprising: A seed collecting region configured to receive a seed sample; An opaque chamber delimited by a chamber floor comprising the seed collecting region and a chamber top connected by chamber side walls between the chamber floor and the chamber top; An imaging unit comprising a camera positioned to acquire one or more images of a scene comprising at least part of a seed sample positioned in the seed collecting region.

[0296] In an embodiment, the system further comprises a control unit connected to a peripheral, said control unit comprising one or more processors configured for one or more of: Prompting user on the peripheral for introducing or confirming identificator for seed sample; Prompting user for wobbling the imaging device and triggering the imaging unit of the imaging device to generate images of the seed collecting region after wobling; Checking positioning of the seeds on the seed imaging surface by way of activating a light source and / or a function of test imaging; and / or Prompting user on the peripheral that image acquisition is completed; Prompting user on the peripheral that (new) sample can be introduced.

[0297] In an embodiment, the system can comprise a transmitter unit, and the control unit can be configured for triggering the transmitter unit to send the images and / or information regarding the images via a network to one or more device or computer system for image analysis, display of the information, estimated seed information and / or generating instructions for a field management device.

[0298] Further object of the application is a computer-implemented method comprising: Causing receiving image data from a scene comprising a seed sample of a plurality of seeds, said image data comprising quantized values representing color of image pixels; Assigning each image pixel of the image data to seeds or background using a segmentation algorithm; Analyzing the part of the image pixels assigned to seeds, by way of one or more of: Determining, using a moisture content model trained to solve a regression problem of correlating image features with a seed moisture content value, a seed moisture content value for the seed sample based on feature analysis of the image pixels assigned to seeds; and / or Determining, using a TKW model trained to solve a regression problem of correlating image features with a TKW value, a TKW value for the seed sample based on feature analysis of the image pixels assigned to seeds; Causing output of one or more of the determined value(s).

[0299] In an embodiment, the computer-implemented method further comprising one or more of: Checking if a seed count value for the plurality of seeds is within a defined seed count range; Determining a blurriness coefficient for the image and filtering out images which blurriness coefficient is below a defined threshold; Cropping image data as defined by cropping parameters; Resizing image data as defined by resizing parameters.

[0300] In an embodiment, the moisture content model and / or the TKW model are machine learning algorithms.

[0301] In an embodiment, the computer-implemented method further comprises generating the trained moisture content model and / or the trained TKW model by way of: Acquiring a training data set comprising color image data of a scene comprising a seed sample of a plurality of seeds and ground truth value for moisture content or TKW respectively, wherein said ground truth value is measured for the seed sample using a ground truth moisture measurement method; Introducing the training data set into a model and training the model for feature extraction using a training algorithm to obtain a trained model; Saving the trained model in the model repository.

[0302] In an embodiment, the computer-implemented method further comprising determining a moisture-dependent TKW value based on feature analysis the image pixels assigned to seeds using a moisture-dependent TKW model trained to solve a regression problem of correlating image features with TKW value in relation with seed moisture content value.

[0303] In an embodiment, the computer-implemented method according further comprises generating the trained moisture-dependent TKW model by way of: Acquiring a training data set comprising color image data of a scene comprising a seed sample of a plurality of seeds and at least both ground truth moisture content value and ground truth TKW value measured for the seed sample using a ground truth moisture measurement method and a ground truth TKW measurement method respectively; Introducing the training data set into a model and training the model for feature extraction using a training algorithm to obtain a trained moisture content model; Saving the trained moisture-dependent TKW model in a model repository.

[0304] In an embodiment, the computer-implemented method further comprises estimating one or more seed size values, by way of: Determining a total number of pixels assigned to seeds in the image; Determining centroids for the pixels assigned to seeds by the segmentation algorithm; Determining a number of identified centroids; Counting for each centroid the number of pixels assigned to the centroid; Determining a range for seed size values based on the counted number of pixels per centroid; and / or Determining an average seed size by dividing the total number of pixels assigned to seeds by the number of centroids; Causing output of the one or more seed size values.

[0305] In an embodiment, for image data of a seed sample collected at time point t, the method further comprises predicting evolution of one or more of determined seed parameter values within a prediction period t+y by way of: From a data repository, acquiring one or more of the determined values for a seed sample parameter at a plurality of time points t-x to t for a crop season at a location of interest; Generating a temporal measurement curve of the determined values for the season from the time point t-x to time point t for the crop season; Acquiring, for the location, one or more of climate data, weather data and field data for a future period t+y; Extrapolating the temporal measurement curve for the future period t+y in consideration of the one or more of climate data / , weather and / or field data for the location using an algorithm.

[0306] In an embodiment, the method further comprises: Placing a seed sample on a seed collecting region of an imaging device, said imaging device comprising the seed collecting region configured to receive the seed sample, an opaque chamber delimited by a chamber floor comprising the seed collecting region and a chamber top connected by chamber side walls between the chamber floor and the chamber top, an imaging unit comprising a camera positioned to acquire an image of a scene comprising at least part of a seed sample positioned in the seed collecting region; Triggering the imaging unit to generate one or more images of a scene comprising whole or part of the seeds collecting region, preferably the whole seeds collecting region; and Acquiring the one or more images and transferring the one or more images to the one or more processors.

[0307] In an embodiment, the method further comprises one or more of: Prompting user on the peripheral for introducing or confirming identifier for a seed sample; Prompting user on the peripheral that a (new) sample can be introduced; Triggering the imaging unit of the imaging device to generate one or more image of the seed collecting region; Checking positioning of the seeds on the seed imaging surface by way of activating a light source and / or a function of test imaging; and / or Prompting user on the peripheral that image acquisition is completed.

[0308] Further subject matter of the application is a computer program product comprising a computer program which can be loaded into a memory of a computer, where it prompts the computer to execute the method described above.

[0309] Further subject matter of the application is a non-transitory computer-readable media storing one or more sequences of instructions, which when executed cause one or more processors to execute the method described above.

[0310] Figures and example:

[0311] The invention is elucidated in detail hereinafter by figures and examples, without any intention to restrict the invention to the features and combinations of features specified in the figures and examples.

[0312] Short description of the figures:

[0313] Figure 1 shows a 3-dimensional schematic view of image acquisition device 1 (also referred to as imaging device) from above front.

[0314] Figure 2 shows a 3-dimensional view from above of an empty drawer and insert.

[0315] Figure 3 shows a detailed schematic view from above of an insert contained a seed sample.

[0316] Fig. 4 shows the architecture of system of the invention and its components.

[0317] Fig. 5 shows a flow chart summarizing the computer-implemented method for analysis of an acquired image of a seed sample.

[0318] Fig. 6 shows a flow chart of the computer-implemented method for training of a moisture content model with a training data set.

[0319] Fig. 7 shows the validity of the results obtained with the image-based method for moisture content measurement compared to ground truth measurement values.

[0320] Fig. 8 shows intermediary steps of the solution for image-based TKW determination implemented for OSR seeds and obtained results compared to respective measured ground truth.

[0321] Fig. 9 shows the correlation of computed seed count and ground truth seed count obtained with a seed counting machine.

[0322] Fig. 10 shows the validity of the results obtained with the image-based method for TKW value measurement compared to ground truth measurement values.

[0323] Fig. 11 shows correlation between predicted and actual TKW values for the moisturedependent model.

[0324] Fig. 12 shows a comparison of predicted and actual TKW in relation to moisture

[0325] References: Image acquisition device (1) Seed / Seed sample (2) Chamber wall (5) delimiting a chamber room (6) Drawer (7) Drawer handle (8) Insert (10) Seed collecting region (11) Structured surface of seed collecting region (Ila) Tilted walls (12) Rills of tilted walls (13) Imaging unit (20) Holder for imaging unit (21) Handle (30)

[0326] Detailed description of the figures

[0327] The solution was implemented for computer-implemented analysis of RGB images of OSR seed sample comprising 300 to 1000 seeds per sample.

[0328] Figure 1 shows a 3-dimensional schematic view of Image acquisition device 1 comprising a drawer 7 which can be opened / closed by way of drawer handle 8. The chamber room 6 delimited by chamber walls 5 has a truncated pyramid form with square base. When closed, the drawer 7 covers the floor of chamber room 6. At the top of chamber room 6 (truncated top of the pyramid), an imaging unit 20 (not shown) is positioned in a detachable way using holder 21. Imaging unit 20 for use in the arrangement of Fig. 1 is a smartphone (not shown). For convenient handling of the device, the holder 21 is connected to an ergonomic handle 30. An eyelet 31 (not shown) can be added to handle 30 to insert a strap for easy carrying of the device 1 as a portable device 1.

[0329] Figure 2 shows details of drawer 7 wherein insert 10 comprising a planar seed collecting region 11 framed by tilted walls 12 is positioned. Drawer 7 can be opened / closed by way of drawer handle 8. The drawer 7 and insert 10 of Fig. 2 are square shaped.

[0330] Figure 3 shows the tilted walls 12 and lowered comer of insert 10 for better controlled wobbling and pouring of seed sample 2 in and from seed collecting region 11.

[0331] Fig. 4 show the architecture of system of the invention and its sub-processes, wherein: Image acquisition device (1) Imaging unit (20) GPS unit (50) User interface (60) Control unit (70) Weather data server (80) Field sensors (90) Data repository (405) Model repository (406) One or more processors 410 for: 411 Receiving image 412 Image pretreatment 413 Image segmentation 420 Model-based determination of moisture content value based on color image data 421 Training moisture content model 430 Model-based determination of TKW value 431 Training TKW model 440 Model-based determination of moisture dependent TKW value 441 Training moisture dependent TKW model 450 Determination of seed size 460 Further data processing

[0332] The skilled person will appreciate that the hardware hosting the imaging control unit, the one or more processors for image pretreatment and / or the one or more processors for image analysis may be separate devices connected with each other for transmission of data in one or both directions. System architecture is typically chosen depending on computing power needed / available for the respective computer-implemented subprocesses.

[0333] Fig. 5 shows an exemplary flow chart of the computer-implemented method for analysis of an acquired color image of a seed sample, wherein: SOI Causing receiving color image data from a scene comprising seeds in the collecting region of the imaging device, said image data comprising quantized values representing color of image pixels; Image pretreatment S02-S04 S02 Determining a blurriness coefficient for the image and filtering out images which blurriness coefficient is below a defined threshold; (Filtering blur image out is not shown) S03 Cropping image data as defined by cropping parameters; S04 Resizing image data as defined by resizing parameters S05 Assigning each image pixel of the color image data to seeds or background using a trained U-Net segmentation algorithm, selecting image pixels assigned to seeds; Analyzing the pixel assigned to seeds, by way of one or more of: S06 Determining a seed moisture content value for the seed sample based on analysis of the image pixels assigned to seeds using a moisture content ResNet50 model; S07 Determining a TKW value for the seed sample based on analysis of image pixels assigned to seeds using a TKW ResNet50 model; S08 Determining a moisture-dependent TKW value for the seed sample based on analysis of image pixels assigned to seeds using a moisture-dependent TKW ResNet50 model; S09 Determining seed size distribution for the sample; S10 Storing determined results of image analysis for further processing; Reiterating for new image for the seed sample until the set of images for a sample is analyzed; SI 1 Averaging results for the set of images for a seed sample; S12 Causing output of one or more of the determined result(s) for a seed sample and / or storage of said results in a data repository together with corresponding image data.

[0334] Fig. 6 shows a flow chart of the computer-implemented method for training of a ResNet50 model as a moisture content model described above, wherein. S101 Acquiring training data set from data repository, said data set comprising a color image data and measured ground truth content value obtained using ground truth measurement method; SI02 Determining a blurriness coefficient for the image and filtering out images which blurriness coefficient is below a defined threshold; (Filtering blur image out is not shown) SI03 Cropping image with adequate blurriness coefficient as defined by cropping parameters SI04 Resizing image as defined by resizing parameters; SI05 Assigning each image pixel of the color image data to seeds or background using a segmentation algorithm; reiterating for the training data set SI06 Inputting training data set into ResNet50 model for finetuning; SI07 Training the model using a training algorithm optimizing features extraction from image data and metadata; SI08 Storing trained model and information related thereto in a model repository. Information related to the model comprises one or more of: imaging device information (camera information, focus), threshold for blurriness coefficient, cropping parameters, resizing parameters and the like.

[0335] The flow chart applies similarly for training of the TKW model and / or the moisturedependent TKW model.

[0336] Example - solution implemented for oilseed rape (OSR) seeds

[0337] The solution described above was implemented for oilseed rape seeds (OSR). Details of implementation are provided below. The skilled person will appreciate that the solution is not limited thereto.

[0338] A / Image acquisition

[0339] OSR seed samples were collected in a field at different maturity stage. Seeds were separated manually and deposited in a blue insert 10 engaged in drawer 7 of imaging device 1 described above comprising a smartphone Type iPhone SE 2020 alternatively SE 2022 (imaging unit) fixed at the top of the chamber room 6 for image acquisition. Drawer 7 was engaged in the drawer cabinet and closed, so chamber room 6 is tight to external illumination. A set of RGB images of seed samples positioned in the insert was acquired with the imaging unit. Said smartphone comprised a GPS module, an imaging control application comprising options and capabilities for introducing metadata relating to the seed sample, supporting image acquisition, conducting image acquisition, transmitting acquired data for later use, but not limited thereto. For each OSR seed sample collected in the field, 5 RGB images were acquired, wherein the imaging device 1 was wobbled between each image. Fig. 8A shows an example of a RGB image of a OSR seed sample in the collecting region of the imaging device.

[0340] Images were pretreated as described above. Fig. 8B shows an example of cropped and resized image with defined parameters.

[0341] B / Generating a training set of segmented images of seeds samples for training of U-Net segmentation model

[0342] A training set of 25 segmented images of OSR seeds on background (blue collecting region) in RGB images was obtained by way of manually segmenting acquired images of seed samples using software solution FlJI-ImageJ for manual segmentation of training images. Then, these seed sample images with the corresponding segmented images were used to train a U-Net model from scratch, wherein training the CNN from scratch comprises varying one or more of input image size, number and order of CNN layers, training options, etc.... [Training A Model From Scratch - MATLAB & Simulink (mathworks.com); https: / / de.mathworks.com / solutions / deep-learning / examples / training-a-model-from- scratch.html]. Trained U-Net model and image processing parameters (cropping, resizing) were stored for later use in pretreatment of new images.

[0343] The obtained trained U-Net model was tested to select the seed region in new pretreated images and create a spatial heatmap of the seed region therein. Fig. 8C shows spatial heatmap obtained after segmentation of the preprocessed image using trained U-Net model; heatmap can be generated within 1 second, which is acceptable response time for user acceptance.

[0344] Fig. 8D shows spatial heatmap with color RGB filters showing color distribution of the segmented seeds.

[0345] B / Generating a seed moisture content model

[0346] A ResNet50 model was selected for moisture content model for computer-implemented analysis of RGB images of OSR seed samples.

[0347] Training of ResNet50 model was achieved using training data set described below.

[0348] The training data set for generating the moisture content model was chosen to comprise RGB images of seed samples and respective metadata of 20-30 OSR inbreds, at different maturity status from humid to dry (moisture approximately from 65% to 30%, with increments of 5-10%).

[0349] The humid seeds were obtained from plants grown in an open field composed of independent plots of 4 rows of 9m long, with an interrow spacing of 0.6m. Each plot was planted with a different inbred. For each moisture value, 3 batches of 20 pods of each inbred were harvested and shelled to obtain a “seed sample” of approximately 500 clean seeds. For each seed sample 5 pictures were acquired with the imaging device using the method described above. The RGB images were preprocessed according to the parameters for cropping and resizing.

[0350] Ground truth moisture content was measured as following: for each imaged sample, ground-truth TKW was obtained with an OptoAgri machine from OptoMachines (https: / / optomachines.fr / home / optomachines-equipment-for-seeds-analysis / opto-agri-tsw-seed-biometry / ). The ground-truth moisture was determined by measuring the weight of the sample before and after the incubation at 130°C for 2h (the method is destructive) using an oven from             Chopin             technologies             (https: / / assets-global. website files. com / 63cf34956bc59159af577c42 / 63cf34956bc591ba0e57854b_EM10_leaflet_EN_2016 1208.pdf).

[0351] Acquired RGB images, respective ground-truth moisture for each batch were stored and used for training of an “aggregation moisture content model”. Trained model was stored in a model repository for later use.

[0352] Obtained trained moisture content model was validated. Fig. 8 shows the correlation of computed (predicted) moisture content of OSR seed samples with ground truth for a crop season for a training data set with a finetuned moisture content ResNet50 model and validation data obtained using a test set of color images. Figures clearly show that, within range of interest (30 to 80% moisture content), the predicted / computed moisture value for the seeds are quite close to the ground-truth values obtained in laboratory measurements.

[0353] C) Generating TKW model

[0354] A ResNet50 model was selected as TKW model for computer-implemented analysis of RGB images of OSR seed samples.

[0355] Training of ResNet50 model was achieved using training data set described below.

[0356] Samples of clean seeds with known ground truth TKW were provided for imaging. Known ground truth TKW had been obtained with an OptoAgri machine from OptoMachines (https: / / optomachines.fr / home / optomachines-equipment-for-seeds-analysis / opto-agri-tsw-seed-biometry / ).

[0357] Four different inbreds with known ground truth TKW from the smallest available to the biggest available were used (3.0-7.5 g). The moisture of these seeds was generally around 8%; this samples are also referred to as “dry seed samples”. Ground truth TKW was stored as metadata for each sample / image.

[0358] Samples of four inbred batches containing from 50 to 1000 seeds were prepared, with increments of 50 seeds using a seed counting machine (such as https: / / www.pfeuffer.com / de / produkt / contador). The number of seeds was stored as metadata for each sample as ground-truth number of seeds.

[0359] Five pictures were acquired for each sample with the imaging device according to the method described above.

[0360] In total, approximately 1200 images with respective ground truth TKW were used for training and approximately 900 images were used for testing / validating the TKW model; respective ground truth TKW were used for validation.

[0361] Experiments showed that the trained model runs more accurately with images of a seed sample of minimum 300 seeds.

[0362] Figure 8 shows the results of applying ResNet50 network as TKW model for OSR seeds. An example of image data transformation, from an RGB image as acquired to estimated values via intermediate images by applying the described algorithms is presented.

[0363] Accuracy of prediction results is satisfactory for both seed counting and TKW estimations compared to the methods of the art.

[0364] Results of image-based determined values for TKW and seed count compared ground truths for TKW and seed count show reliability of the method.

[0365] Fig. 9 shows data validating the solution for seed counting within range of interest (0 to 1000 seeds). The correlation of computed seed count obtained RGB images of seed samples and ground truth seed count obtained with a seed counting machine is shown for the training data set (upper curve) and a test set for validation (lower curve).

[0366] Fig. 10 shows validity of the results obtained with the image-based method for TKW measurement compared to ground truth measurement values within range of interest (30 to 65% TKW). Correlation curve between computed value and ground truth value is presented for the training data set (upper curve, root-mean-square error (RMSE) for training set: 0.05) and for a test set for validation (lower curve; RMSE for test set: 0.11).

[0367] Figures clearly show that the predicted TKW value for the seeds are quite close to the ground-truth values obtained in laboratory measurement for TKW as described above.

[0368] D) Generating a moisture dependent TKW model

[0369] The TKW of dry seeds depends on when the plants have been swathed.

[0370] In an example, the method is implemented to predict, from data (images and metadata) collected for the seeds that are still immature / humid (green and swallowed with water), the TKW they would have after the plants is swathed, dried and ready to harvest i.e. when the seeds would be mature and dry (black and around 8% moisture).

[0371] A ResNet50 model is selected as moisture dependent TKW model for computer-implemented analysis of RGB images of OSR seed samples.

[0372] Training of ResNet50 model is achieved using training data set described below.

[0373] The training set is designed to comprise 2 samples for each condition, one condition being a combination of a given moisture for a given inbred. For each condition, 2 batches of 20 pods of each inbred are harvested at the same time in the same plot.

[0374] The first batch, also referred to as “immature seeds sample”, is treated similarly as described above for the moisture content model: The humid seeds are obtained from plants grown in an open field. For each seed sample 5 pictures are acquired with the imaging device using the method described above. The RGB images are preprocessed according to parameters for cropping and resizing. Ground truth moisture content is measured as following: for each imaged sample, ground-truth TKW was obtained with an OptoAgri machine from OptoMachines            (https: / / optomachines.fr / home / optomachines-equipment-for-seeds- analysis / opto-agri-tsw-seed-biometry / ). The ground-truth moisture is determined by measuring the weight of the sample before and after the incubation at 130°C for 2h (the method is destructive) using an oven from Chopin technologies (https: / / assets-global.website files.com / 63cfi4956bc59159af577c42 / 63cfi4956bc591ba0e57854b_EM10_leaflet_EN_2016 1208.pdf).

[0375] The second batch, so called “dry seeds sample”, is obtained by letting the plant / pods dry in open air (like regular field conditions) until that the seeds reach 8% moisture (regular moisture content value for harvest time). Then the seeds were harvested / cleaned and the ground-truth TKW was obtained using an OtipAgri machine as described above.

[0376] The training data set for training a moisture dependent TKW model comprised the ground truth for response variable (dry TKW) as well as the following predictors: moisture level, weather variables sums (GDU as defined below, precipitation, solar radiation) calculated across various temporal windows (date of F10 to date of EF, date of EF to date of the Picture, date of the picture acquisition to date of harvest, date of F10 to harvest, date of EF to date of harvest.

[0377] For a better model performance, data was collected over several crop seasons.

[0378] The training dataset was first preprocessed (filtering to years 2022 to 2024 and standardized by year and by location of experiment to remove environmental effects, manual removal of obvious outliers) and an exploratory multiple factor analysis (MFA) was conducted to understand variable relationships. The dataset was then split into training and testing subsets and a preprocessing pipeline was applied, including normalization of numerical predictors.

[0379] To optimize model performance, we systematically tested multiple predictor combinations through cross-validation, fitting an XGBoost regression model to each subset and evaluating performance using mean absolute error (MAE). Hyperparameter tuning was conducted using a grid search approach to refine model parameters, ensuring optimal predictive capability.

[0380] The best model was set up in the following way: TKW ~ mst + GDUFlOtoEF + GDUEFtoPicture + GDUFlOtoPicture + prec FlO toPicture + rad FlOtoEF wherein: F10 refers to date of 10% flowering (when 10% of the plant have at least one flower; EF refers to date of End of Flowering: when 95% of plants still have at least one flower; Picture refers to the date of picture acquisition. GDU refers to Growing Degree Unit (calculation: difference of daily temperature maximal and minimum / 2);. mst refers to seed moisture in % weight; prec refers to sum of precipitation; 5 rad refers to sum of radiation; As a matter of example, GDUFlOtoEF refers to the Growing Degree Unit between the date of 10% flowering and the end of flowering. Similarly, prec FlOtoPicture refers to the sum of precipitation between the date of 10% flowering and the date of picture acquisition. TKW stands for dry TKW, that is a value of TKW when the seeds are at 8% moisture, ready 10 for harvesting.

[0381] The skilled person will recognize that optimizing the selection of weather variables and temporal windows may be suitable for improving the model.

[0382] Fig. 11 shows acceptable correlation between predicted and actual TKW values for the moisture-dependent model. 15

[0383] Fig. 12 shows a comparison of predicted and actual TKW in relation to moisture.

Claims

1. A system comprising:One or more peripherals;One or more data repository;One or more processors configured for:Causing receiving image data from a scene comprising a seed sample of a plurality of seeds, said image data comprising quantized values representing color of image pixels;Assigning each image pixel of the image data to seeds or background using a segmentation algorithm;Analyzing the part of the image pixels assigned to seeds, by way of one or more of: Determining, using a moisture content model trained to solve a regression problem of correlating image features with seed moisture content value, a seed moisture content value for the seed sample based on feature analysis of the image pixels assigned to seeds; and / orDetermining, using a TKW model trained to solve a regression problem of correlating image features with TKW value, a TKW value for the seed sample based on feature analysis of the image pixels assigned to seeds;Causing output of one or more of the determined value(s).

2. The system of claim 1, wherein the one or more processors are configured for acquiring metadata for the image data, said metadata being one or more of identification for seed sample, seed genotype, sample location, time point of image acquisition, weather data, historical field management information, and / or landscape information.

3. The system of claim 1, wherein the one or more processors are configured for pretreatment of the image data, said pretreatment comprising one or more of:Checking if a seed count value for the plurality of seeds is within a defined seed count range;Determining a blurriness coefficient for the image and filtering out images which blurriness coefficient is below a defined threshold;Cropping image data as defined by cropping parameters;Resizing image data as defined by resizing parameters.

4. The system of claim 1, wherein the moisture content model and / or the TKW model are machine learning algorithms.

5. The system of claim 1, wherein the one or more processors are configured for generating the trained moisture content model and / or the trained TKW model by way of Acquiring a training data set comprising color image data of a scene comprising a seed sample of a plurality of seeds and ground truth value for moisture content or TKW respectively measured for the seed sample using a ground truth moisture measurement method;Introducing the training data set into a model and training the model for feature extraction using a training algorithm to obtain a trained model;Saving the trained model in the model repository.

6. The system according to claim 1, wherein the one or more processors are configured for determining a moisture-dependent TKW value based on feature analysis the image pixels assigned to seeds using a moisture-dependent TKW model trained to solve a regression problem of correlating image features with TKW value in relation with seed moisture content value.

7. The system of claim 6, wherein the one or more processors are configured for generating the trained moisture-dependent TKW model by way ofAcquiring a training data set comprising color image data of a scene comprising a seed sample of a plurality of seeds and at least both ground truth moisture content value and ground truth TKW value measured for the seed sample using a ground truth moisture measurement method and a ground truth TKW measurement method respectively;Introducing the training data set into a model and training the model for feature extraction using a training algorithm to obtain a trained moisture content model;Saving the trained moisture-dependent TKW model in a model repository.

8. The system according to claim 1, wherein the one or more processors are configured for estimating one or more seed size values, by way ofDetermining a total number of pixels assigned to seeds in the image;Determining centroids for the pixels assigned to seeds by the segmentation algorithm;Determining a number of identified centroids;Counting for each centroid the number of pixels assigned to the centroid;Determining a range for seed size values based on the counted number of pixels per centroid; and / orDetermining an average seed size by dividing the total number of pixels assigned to seeds by the number of centroids;Causing output of the one or more seed size values.

9. The system according to claim 1, wherein for image data of a seed sample collected at time point t, the one or more processors are configured to predict evolution of one or more of determined seed parameter values within a prediction period t+y by way ofFrom a data repository, acquiring one or more of the determined values for a seed sample parameter at a plurality of time points t-x to t for a crop season at a location of interest;Generating a temporal measurement curve of the determined values for the season from the time point t-x to time point t for the crop season;Acquiring, for the location, one or more of climate data, weather data and field data for a future period t+y;Extrapolating the temporal measurement curve for the future period t+y in consideration of the one or more of climate data / , weather and / or field data for the location using an algorithm.

10. A computer-implemented method comprising:Causing receiving image data from a scene comprising a seed sample of a plurality of seeds, said image data comprising quantized values representing color of image pixels;Assigning each image pixel of the image data to seeds or background using a segmentation algorithm;Analyzing the part of the image pixels assigned to seeds, by way of one or more of:Determining, using a moisture content model trained to solve a regression problem of correlating image features with a seed moisture content value, a seed moisture content value for the seed sample based on feature analysis of the image pixels assigned to seeds; and / orDetermining, using a TKW model trained to solve a regression problem of correlating image features with a TKW value, a TKW value for the seed sample based on feature analysis of the image pixels assigned to seeds;Causing output of one or more of the determined value(s).

11. The computer-implemented method according to any one of the preceding claims, further comprising one or more of:Checking if a seed count value for the plurality of seeds is within a defined seed count range;Determining a blurriness coefficient for the image and filtering out images which blurriness coefficient is below a defined threshold;Cropping image data as defined by cropping parameters;Resizing image data as defined by resizing parameters.

12. The computer-implemented method according to any one of the preceding claims, wherein the moisture content model and / or the TKW model are machine learning algorithms.

13. The computer-implemented method according to any one of the preceding claims, further comprising generating the trained moisture content model and / or the trained TKW model by way of:Acquiring a training data set comprising color image data of a scene comprising a seed sample of a plurality of seeds and ground truth value for moisture content or TKW respectively, wherein said ground truth value is measured for the seed sample using a ground truth moisture measurement method;Introducing the training data set into a model and training the model for feature extraction using a training algorithm to obtain a trained model;Saving the trained model in the model repository.

14. The computer-implemented method according to any one of the preceding claims, further comprising determining a moisture-dependent TKW value based on feature analysis the image pixels assigned to seeds using a moisture-dependent TKW model trained to solve a regression problem of correlating image features with TKW value in relation with seed moisture content value.

15. The computer-implemented method according to any one of the preceding claims, further comprising generating the trained moisture-dependent TKW model by way of: Acquiring a training data set comprising color image data of a scenecomprising a seed sample of a plurality of seeds and at least both ground truth moisture content value and ground truth TKW value measured for the seed sample using a ground truth moisture measurement method and a ground truth TKW measurement method respectively;Introducing the training data set into a model and training the model for feature extraction using a training algorithm to obtain a trained moisture content model;Saving the trained moisture-dependent TKW model in a model repository.

16. The computer-implemented method according to any one of the preceding claims, further comprising estimating one or more seed size values, by way of:Determining a total number of pixels assigned to seeds in the image;Determining centroids for the pixels assigned to seeds by the segmentation algorithm;Determining a number of identified centroids;Counting for each centroid the number of pixels assigned to the centroid;Determining a range for seed size values based on the counted number of pixels per centroid; and / orDetermining an average seed size by dividing the total number of pixels assigned to seeds by the number of centroids;Causing output of the one or more seed size values.

17. The computer-implemented method according to any one of the preceding claims, wherein for image data of a seed sample collected at time point t, the method further comprises predicting evolution of one or more of determined seed parameter values within a prediction period t+y by way of:From a data repository, acquiring one or more of the determined values for a seed sample parameter at a plurality of time points t-x to t for a crop season at a location of interest;Generating a temporal measurement curve of the determined values for the season from the time point t-x to time point t for the crop season;Acquiring, for the location, one or more of climate data, weather data and field data for a future period t+y;Extrapolating the temporal measurement curve for the future period t+y in consideration of the one or more of climate data / , weather and / or field data for the location using an algorithm.

18. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method according to any of the claims 10 to 17.

Images