Plant root growth condition analysis system and method

Abstract

The invention relates to the field of artificial intelligence, in particular to a plant root growth condition analysis system and method, and the method comprises the steps: inputting image blocks into an encoder to obtain the coding output of each image block, and carrying out the clustering of the coding outputs of the image blocks, and obtaining a plurality of classifications; inputting the class embedding and the coding output of each image block into a decoder to obtain the decoder output corresponding to each image block and the decoder output of each class embedding; according to the corresponding relation between the clustering result and the class embedding, feature fusion is carried out on the output of the decoder and the corresponding class embedding, and a root system segmentation result is output after up-sampling; and comparing the development conditions of the same root system in two adjacent images in the sequence to obtain the growth condition of the plant root system. According to the method, the root system identification accuracy is improved by introducing multiple class embedding, and the root system analysis precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and particularly to a system and method for analyzing the growth status of plant roots. Background Art

[0002] By studying the distribution, depth, and density of roots, it is possible to gain in-depth understanding of how plants absorb water and nutrients under different soil conditions. This information is crucial for formulating more precise irrigation and fertilization plans, as different crops and soil types require different management strategies. For example, deep-rooted crops may require less but deeper irrigation to promote root growth downward, while shallow-rooted crops may require more frequent but shallower irrigation. Similarly, based on the distribution of root density, fertilization can be carried out more targeted, ensuring that nutrients are directly supplied to areas with dense roots, thereby reducing fertilizer waste and improving the overall health and yield of crops.

[0003] The main methods for analyzing plant roots include direct observation method, root trap and rhizotron method, underground imaging method, etc. Among them, the underground imaging method does not require damaging the roots and can continuously observe the growth of roots, and is applied more and more widely. However, the analysis of roots mainly relies on manual work, especially the annotation of roots in images. The growth of roots is analyzed by annotating images at different time periods. Summary of the Invention

[0004] To solve the above problems, the present invention provides a method for analyzing the growth status of plant roots based on multiple images. The method includes the following steps:

[0005] Obtain a sequence of plant root images at the same location in different periods. Segment each image in the root image sequence to obtain image patches of each image. Input the image patches into an encoder to obtain the encoded output of each image patch. Cluster the encoded output of the image patches to obtain multiple classifications;

[0006] Input the class embeddings and the encoded output of each image patch into a decoder to obtain the decoder output corresponding to each image patch and the decoder output of each class embedding; According to the correspondence between the clustering results and the class embeddings, fuse the features of the decoder output and the corresponding class embeddings, and output the segmentation result of the roots after upsampling; where the number of clusters is the same as the number of class embeddings, and the number of class embeddings is a hyperparameter;

[0007] Compare the development of the same root in two adjacent images in the sequence to obtain the growth status of the plant roots.

[0008] Preferably, clustering the encoded output of the image patches specifically includes:

[0009] Cluster the image patches according to the similarity of the image patches, and the number of clusters is the number of class embeddings; for each clustering result, retain the image patch closest to the cluster center;

[0010] Output the encoding corresponding to the retained image patch as the initial cluster center, and cluster the encoding outputs of the image patches according to the similarity of the image patches and the similarity of the encoding outputs corresponding to the image patches.

[0011] Preferably, the clustering of the encoding outputs of the image patches according to the similarity of the image patches and the similarity of the encoding outputs corresponding to the image patches is specifically as follows:

[0012] Set a first weight and a second weight, calculate the distance from the remaining image patches to each cluster center by using the first weight, the second weight, the similarity of the image patches, and the similarity of the encoding outputs corresponding to the image patches, and assign the encoding output of the image patch to the cluster closest to the cluster center; wherein, the sum of the first weight and the second weight is 1.

[0013] Preferably, the feature fusion of the output of the decoder and the corresponding class embedding according to the correspondence between the clustering result and the class embedding is specifically as follows:

[0014] Calculate the average value of the encoding outputs of all image patches in each clustering result, and establish the correspondence between the clustering result and the class embedding according to the average value;

[0015] Find the class embedding corresponding to each output of the decoder based on the correspondence; perform feature fusion on the class embedding and the corresponding decoder output.

[0016] Preferably, the growth condition of the plant root system is obtained by comparing the development conditions of the same root system in two adjacent images in the sequence, specifically as follows:

[0017] Retain the segmentation result of each image, and obtain the change of the root system between two adjacent images according to the differential calculation result of two adjacent images;

[0018] Taking the segmentation result of the first image in the sequence as the benchmark, dynamically display the growth condition of the root system on the benchmark.

[0019] In addition, the present invention also provides a plant root system growth condition analysis system based on multiple images, and the system includes the following modules:

[0020] An encoding clustering module, configured to obtain a sequence of plant root system images taken at the same position in different periods, segment each image in the root system image sequence to obtain image patches of each image, input the image patches into an encoder to obtain the encoding output of each image patch, and cluster the encoding outputs of the image patches to obtain multiple classifications;

[0021] An image segmentation module, which is used to input the class embedding and the encoded output of each image patch into a decoder to obtain the decoder output corresponding to each image patch and the decoder output of each class embedding; according to the correspondence between the clustering result and the class embedding, fuse the features of the decoder output and the corresponding class embedding, and output the segmentation result of the root system after upsampling; wherein, the number of clusters is the same as the number of class embeddings, and the number of class embeddings is a hyperparameter;

[0022] A root system growth status analysis module, which is used to compare the development of the same root system in two adjacent images in the sequence to obtain the growth status of the plant root system.

[0023] Preferably, the clustering of the encoded output of the image patches is specifically as follows:

[0024] Cluster the image patches according to the similarity of the image patches, and the number of clusters is the number of class embeddings; for each clustering result, retain the image patch closest to the cluster center;

[0025] Use the encoded output corresponding to the retained image patch as the initial cluster center, and cluster the encoded output of the image patches according to the similarity of the image patches and the similarity of the encoded output corresponding to the image patches.

[0026] Preferably, the clustering of the encoded output of the image patches according to the similarity of the image patches and the similarity of the encoded output corresponding to the image patches is specifically as follows:

[0027] Set a first weight and a second weight, and use the first weight, the second weight, the similarity of the image patches, and the similarity of the encoded output corresponding to the image patches to calculate the distance of the remaining image patches to each cluster center, and assign the encoded output of the image patches to the cluster closest to the cluster center; wherein, the sum of the first weight and the second weight is 1.

[0028] Preferably, the feature fusion of the decoder output and the corresponding class embedding according to the correspondence between the clustering result and the class embedding is specifically as follows:

[0029] Calculate the average value of the encoded output of all image patches in each clustering result, and establish the correspondence between the clustering result and the class embedding according to the average value;

[0030] Find the class embedding corresponding to each output of the decoder based on the correspondence; fuse the class embedding and the corresponding decoder output.

[0031] Preferably, the comparison of the development of the same root system in two adjacent images in the sequence to obtain the growth status of the plant root system is specifically as follows:

[0032] Retain the segmentation results of each image, and obtain the root system changes between two adjacent images based on the difference calculation results of the two adjacent images;

[0033] Taking the segmentation result of the first image in the sequence as a reference, dynamically display the growth status of the root system on the reference.

[0034] In addition, the present invention also provides a computer program product, which implements the method as described above when executed by a processor.

[0035] In order to improve the accuracy of plant root system segmentation, the present invention inputs image patches into an encoder to obtain the encoded output of each image patch, clusters the encoded output of the image patches to obtain multiple classifications; inputs the class embeddings and the encoded output of each image patch into a decoder to obtain the decoder output corresponding to each image patch and the decoder output of each class embedding; according to the correspondence between the clustering results and the class embeddings, perform feature fusion on the output of the decoder and the corresponding class embeddings, and output the segmentation result of the root system after upsampling.

[0036] The present invention has the following advantages: First, according to the output of the decoder, different class embeddings are used for different outputs, avoiding the use of the same class embedding in the Segmenter model; second, multiple class embeddings are set, and different decoder outputs use different class embeddings, avoiding the problem of less content learned in the class embeddings, improving the segmentation accuracy, and greatly improving the growth analysis of plant root systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Is the flowchart of Embodiment 1;

[0038] Figure 2 Is the root system diagram of the same location at different times taken by underground imaging;

[0039] Figure 3 Is the structural diagram of the encoder;

[0040] Figure 4 Is the structural diagram of the decoder;

[0041] Figure 5 Is the schematic diagram of the extracted root structure;

[0042] Figure 6 Is the structural diagram of Embodiment 2. DETAILED DESCRIPTION OF THE INVENTION

[0043] In this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element. In the present invention, if personal privacy data is collected, such as face, mobile phone usage information, etc., personal permission will be obtained in advance, including but not limited to oral reminders, posting posters, mobile phone reminders, etc.; if it involves conflicts with laws and regulations, it will be produced or used within the scope permitted by laws and regulations.

[0044] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0045] Embodiment 1, the present invention provides a method for analyzing the growth status of plant roots based on multiple images, as Figure 1 shown, the method includes the following steps:

[0046] S1, obtain a sequence of plant root images at the same position in different periods, segment each image in the root image sequence to obtain image blocks of each image, input the image blocks into an encoder to obtain the encoded output of each image block, and cluster the encoded output of the image blocks to obtain multiple classifications;

[0047] Use a plant root growth monitoring system to capture an image sequence of plant roots. For example, using the CI-600 system, the growth conditions of plant roots at the same position at different time points can be obtained, that is, images at different times. Among them, the plant root image sequence is composed of images at different time points arranged in time.

[0048] After obtaining the image, preprocess the image. The preprocessing includes but is not limited to denoising, image enhancement, etc. In a specific embodiment, the preprocessing further includes removing capillary roots in the image. Segment the preprocessed image into multiple image patches (patches), and flatten each image patch to obtain the vector corresponding to the image patch. Input the vector corresponding to the image patch into the encoder. Among them, the encoder preferably uses a Transformer Encoder, and the structure of the encoder is as Figure 3 shown. When using the Segmenter model, it involves class embeddings (cls). The class embedding (cls) is a learnable vector, and each class corresponds to a class embedding (cls). However, the class embedding (cls) is a vector of a fixed size, and the content it can learn is limited. Based on this, the present invention sets multiple class embeddings (cls) for the same class.

[0049] After inputting the image patch into the encoder, the encoder will output multiple outputs, and each output corresponds to the input image patch. After obtaining the encoded output of the image patch, cluster the encoded output of the image patch to obtain multiple classifications. There are various clustering methods, such as clustering according to the vectors output by the encoder. In a more specific embodiment, the clustering of the encoded output of the image patch is specifically:

[0050] Cluster the image patches according to the similarity of the image patches, and the number of clusters is the number of class embeddings; for each clustering result, retain the image patch closest to the cluster center;

[0051] The number of class embeddings is the same as the number of clusters, that is, each cluster corresponds to a class embedding (cls). When clustering, first cluster the image patches. The information contained in similar image patches is generally the same or similar. For example, they all contain root systems. When clustering, first cluster the image patches, and then retain the image patch closest to the cluster center. For example, after clustering according to the similarity of the image patches, they are clustered into 5 categories in total. Among them, select an image patch closest to the cluster center from the first cluster.

[0052] In an alternative embodiment, cluster the image patches according to the similarity of the encoder output corresponding to the image patches, or cluster the image patches according to the similarity of the image patches and the similarity of the encoder output corresponding to the image patches.

[0053] Use the encoded output corresponding to the retained image patch as the initial cluster center, and cluster the encoded output of the image patch according to the similarity of the image patches and the similarity of the encoder output corresponding to the image patches.

[0054] After obtaining the initial cluster center, cluster the encoded output of the image patch into multiple categories according to the similarity of the image patches and the similarity of the encoder output corresponding to the image patches.

[0055] In a more specific embodiment, clustering the encoded outputs of image patches according to the similarity of the image patches and the similarity of the encoded outputs corresponding to the image patches is specifically as follows:

[0056] Set a first weight and a second weight, and calculate the distances from the remaining image patches to each cluster center using the first weight, the second weight, the similarity of the image patches, and the similarity of the encoded outputs corresponding to the image patches, and assign the encoded output of the image patch to the cluster closest to the cluster center; where the sum of the first weight and the second weight is 1.

[0057] In this embodiment, the similarity of the image patches and the similarity of the vectors of the encoder outputs are considered simultaneously, and different weights are set: the first weight and the second weight, and the sum of the products of the weights and the similarities is used as the distance. For example, the similarity between image patch 1 and image patch 2 is 0.7, the similarity between the vectors corresponding to the encoder outputs of image patch 1 and image patch 2 is 0.6, the first weight is 0.6, and the second weight is 0.4, then the distance between these two image patches is: 0.66. When clustering, whenever a new image patch is added to a cluster, the center of the corresponding cluster is updated.

[0058] S2. Input the class embeddings and the encoded outputs of each image patch into the decoder to obtain the decoder outputs corresponding to each image patch and the decoder outputs of each class embedding; according to the correspondence between the clustering results and the class embeddings, perform feature fusion on the decoder outputs and the corresponding class embeddings, and output the segmentation result of the root system after upsampling; where the number of clusters is the same as the number of class embeddings, and the number of class embeddings is a hyperparameter;

[0059] After obtaining the results of the encoder, input the encoder results into the decoder. Preferably, the decoder uses MaskTransformer, and Mask Transformer also has multiple class embeddings cls inputs, and multiple class embeddings cls will be output after passing through Mask Transformer.

[0060] In a more specific embodiment, performing feature fusion on the decoder outputs and the corresponding class embeddings according to the correspondence between the clustering results and the class embeddings is specifically as follows:

[0061] Calculate the average value of the encoded outputs of all image patches in each clustering result, and establish the correspondence between the clustering result and the class embedding according to the average value;

[0062] When using the model, for the average value of the encoded outputs corresponding to all image patches in each cluster, establish the correspondence between the clustering result and the class embedding according to the average value. For example, clustering result A corresponds to class embedding cls1, and clustering result B corresponds to class embedding cls3.

[0063] Among them, the average value calculation method is to sum the vectors of the encoded outputs corresponding to all image patches in each cluster bit by bit, and then calculate the average value of each bit of the vector. For example, if the encoded output vectors corresponding to the first image patch are [6, 2, 126] and [2, 8, 20], then the average value is [4, 5, 73]. Since the relationship between the class embeddings and the average values is recorded during training, when using the model for segmentation, the class embeddings corresponding to each cluster can be found according to the recorded relationship.

[0064] Find the class embedding corresponding to each output of the decoder based on the corresponding relationship; perform feature fusion on the class embedding and the corresponding decoder output.

[0065] After obtaining the corresponding relationship, perform feature fusion on the class embedding and the output of the decoder. As Figure 4 shown, if the image patch corresponding to an output vector of the decoder belongs to cluster B, then the corresponding class embedding is cls3. When performing feature fusion, the decoder output vector corresponding to this image patch will be fused with cls3. Since the decoder outputs of the image patches in different clusters correspond to different class embeddings, and the class embeddings are obtained through learning, setting multiple class embeddings can learn more knowledge and perform more accurate segmentation during segmentation. It should be noted that in the present invention, there is only one category: root system, but there are multiple class embeddings for this category. This is different from setting different class embeddings according to different segmentation types in the prior art. For example, in the prior art, if it is necessary to segment trees, sidewalks, and people in an image, three class embeddings, Tree, Sidewalk, and Person, will be set. The present invention only needs to segment one category: root system, and the class embeddings set are also of one type, but there are multiple. Figure 5 Shows the result of segmenting the Figure 2 right part through the present invention.

[0066] S3. Compare the development of the same root system in two adjacent images in the sequence to obtain the growth status of the plant root system.

[0067] In the same sequence, the growth of the root system will be shown in the images. By segmenting the images of the root system through the above segmentation method and then comparing the segmentation results of two adjacent images, the growth of the root system can be obtained. In a more specific embodiment, the step of comparing the development of the same root system in two adjacent images in the sequence to obtain the growth status of the plant root system is specifically:

[0068] Retain the segmentation results of each image, and obtain the change of the root system between two adjacent images according to the difference calculation result of two adjacent images;

[0069] Based on the segmentation result of the first image in the sequence, dynamically display the growth status of the root system on the basis.

[0070] Before using the model, the model needs to be trained. During the training process, the clustering relationship corresponding to each class embedding will be stored, such as the average value of the encoded outputs corresponding to all image patches in the cluster. During training, this average value is continuously updated. During the use of the model, find the average value that is closest to the average value corresponding to the class embedding, or establish the corresponding relationship between the class embedding and the cluster in other ways. The present invention does not make specific limitations on this.

[0071] Embodiment 2, the present invention also provides a system for analyzing the growth status of plant roots based on multiple images, as Figure 6 shown, the system includes the following modules:

[0072] An encoding and clustering module, configured to obtain a sequence of plant root images taken at the same position at different times, segment each image in the root image sequence to obtain image patches of each image, input the image patches into an encoder to obtain the encoded outputs of each image patch, and cluster the encoded outputs of the image patches to obtain multiple classifications;

[0073] An image segmentation module, configured to input the class embedding and the encoded output of each image patch into a decoder to obtain the decoder output corresponding to each image patch and the decoder output of each class embedding; according to the corresponding relationship between the clustering result and the class embedding, perform feature fusion on the output of the decoder and the corresponding class embedding, and output the segmentation result of the root system after upsampling; wherein, the number of clusters is the same as the number of class embeddings, and the number of class embeddings is a hyperparameter;

[0074] A root growth status analysis module, configured to compare the development of the same root system in two adjacent images in the sequence to obtain the growth status of the plant root system.

[0075] Preferably, the clustering of the encoded outputs of the image patches is specifically:

[0076] Cluster the image patches according to the similarity of the image patches, and the number of clusters is the number of class embeddings; for each clustering result, retain the image patch closest to the cluster center;

[0077] Use the encoded output corresponding to the retained image patch as the initial cluster center, and cluster the encoded outputs of the image patches according to the similarity of the image patches and the similarity of the encoded outputs corresponding to the image patches.

[0078] Preferably, the clustering of the encoded outputs of the image patches according to the similarity of the image patches and the similarity of the encoded outputs corresponding to the image patches is specifically:

[0079] Set a first weight and a second weight, and calculate the distances from the remaining image patches to each cluster center by using the first weight, the second weight, the similarity of the image patches, and the similarity of the encoded outputs corresponding to the image patches. Then, assign the encoded output of the image patch to the cluster center with the closest distance; wherein, the sum of the first weight and the second weight is 1.

[0080] Preferably, according to the correspondence between the clustering result and the class embedding, the output of the decoder and the corresponding class embedding are subjected to feature fusion, specifically:

[0081] Calculate the average value of the encoded outputs of all image patches in each clustering result, and establish the correspondence between the clustering result and the class embedding according to the average value;

[0082] Based on the correspondence, find the class embedding corresponding to each output of the decoder; perform feature fusion on the class embedding and the corresponding decoder output.

[0083] Preferably, comparing the development of the same root system in two adjacent images in the sequence to obtain the growth condition of the plant root system, specifically:

[0084] Retain the segmentation result of each image, and obtain the root system change between two adjacent images according to the differential calculation result of the two adjacent images;

[0085] Taking the segmentation result of the first image in the sequence as a reference, dynamically display the growth condition of the root system on the reference.

[0086] Embodiment 3, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and the computer program realizes the method as described above when being executed by a processor.

[0087] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of adding a necessary general hardware platform, and of course, can also be implemented by a combination of hardware and software. Based on such an understanding, the above technical solutions essentially or the part that contributes to the prior art can be embodied in the form of a computer product, and the present invention can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Other embodiments can also be adopted. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A plant root growth status analysis system, characterized in that: The system includes the following modules: The coding clustering module is used to obtain a sequence of plant root images taken at the same location at different times, segment each image in the sequence of root images to obtain image blocks of each image, input the image blocks into an encoder to obtain a coded output of each image block, and cluster the coded outputs of the image blocks to obtain multiple classifications; An image segmentation module is used to input the class embedding and the encoding output of each image block into a decoder to obtain the decoder output corresponding to each image block and the decoder output of each class embedding; According to the corresponding relationship between clustering results and class embeddings, the decoder output and the corresponding class embedding are feature fused, and the root segmentation result is output after upsampling. The number of clusters is the same as the number of class embeddings, and the number of class embeddings is a hyperparameter. The root growth status analysis module is used to compare the development status of the same root system in two adjacent images in the sequence to obtain the growth status of the plant root system.

2. The system according to claim 1, characterized in that The encoding output of the image block is clustered, specifically: Cluster the image blocks according to their similarity, and the number of clusters is the number of class embeddings; for each clustering result, retain the image block closest to the cluster center; The coding output corresponding to the reserved image block is used as an initial clustering center, and the coding output of the image block is clustered according to the similarity of the image block and the similarity of the coding output corresponding to the image block.

3. The system according to claim 2, characterized in that The clustering of the coded outputs of the image blocks according to the similarity of the image blocks and the similarity of the coded outputs corresponding to the image blocks is specifically as follows: A first weight and a second weight are set, and the distances from the remaining image blocks to each cluster center are calculated using the first weight, the second weight, the similarity of the image blocks, and the similarity of the encoded outputs corresponding to the image blocks, and the encoded outputs of the image blocks are assigned to the clusters closest to the cluster centers; wherein the sum of the first weight and the second weight is 1.

4. The system according to claim 1, characterized in that According to the correspondence between the clustering results and the class embeddings, the output of the decoder and the corresponding class embeddings are feature fused, specifically: Calculate the average value of all image block encoding outputs in each clustering result, and establish a corresponding relationship between the clustering result and the class embedding according to the average value; Based on the corresponding relationship, the class embedding corresponding to each decoder output is found; and the class embedding and the corresponding decoder output are feature fused.

5. A method for analyzing plant root growth conditions, characterized in that: The method comprises the following steps: Acquire a sequence of plant root images taken at the same location at different times, segment each image in the sequence of root images to obtain image blocks of each image, input the image blocks into an encoder to obtain an encoding output of each image block, and cluster the encoding outputs of the image blocks to obtain multiple classifications; The class embedding and the encoding output of each image block are input into the decoder to obtain the decoder output corresponding to each image block and the decoder output of each class embedding; according to the corresponding relationship between the clustering result and the class embedding, the decoder output and the corresponding class embedding are feature fused, and the root segmentation result is output after upsampling; the number of clusters is the same as the number of class embeddings, and the number of class embeddings is a hyperparameter; The growth status of the plant root system is obtained by comparing the development status of the same root system in two adjacent images in the sequence.

6. The method according to claim 5, characterized in that The encoding output of the image block is clustered, specifically: Cluster the image blocks according to their similarity, and the number of clusters is the number of class embeddings; for each clustering result, retain the image block closest to the cluster center; The coding output corresponding to the reserved image block is used as an initial clustering center, and the coding output of the image block is clustered according to the similarity of the image block and the similarity of the coding output corresponding to the image block.

7. The method according to claim 6, characterized in that The clustering of the coded outputs of the image blocks according to the similarity of the image blocks and the similarity of the coded outputs corresponding to the image blocks is specifically as follows: A first weight and a second weight are set, and the distances from the remaining image blocks to each cluster center are calculated using the first weight, the second weight, the similarity of the image blocks, and the similarity of the encoded outputs corresponding to the image blocks, and the encoded outputs of the image blocks are assigned to the clusters closest to the cluster centers; wherein the sum of the first weight and the second weight is 1.

8. The method according to claim 5, characterized in that According to the correspondence between the clustering results and the class embeddings, the output of the decoder and the corresponding class embeddings are feature fused, specifically: Calculate the average value of all image block encoding outputs in each clustering result, and establish a corresponding relationship between the clustering result and the class embedding according to the average value; Based on the corresponding relationship, the class embedding corresponding to each decoder output is found; and the class embedding and the corresponding decoder output are feature fused.

9. The method according to claim 5, characterized in that The step of comparing the development of the same root system in two adjacent images in the sequence to obtain the growth status of the plant root system is specifically as follows: The segmentation results of each image are retained, and the root system changes between the two adjacent images are obtained based on the difference calculation results of the two adjacent images; The segmentation result of the first image in the sequence is used as a benchmark, and the growth status of the root system is dynamically displayed on the benchmark.

10. A computer storage device, wherein a computer program is stored on the storage device, characterized in that: When the computer program is executed by a processor, the computer program implements the method according to any one of claims 1 to 7.

Images