A method for efficiently hardening yellow flower tissue culture seedlings

Through the two-particle root growth model and transformer network optimization, the problem of low efficiency of traditional yellow flower tissue culture seedling refining methods is solved, and the automatic optimization of efficient seedling refining conditions is achieved, which significantly improves the survival rate and seedling quality of daylily.

CN118038447BActive Publication Date: 2024-11-15达州市农业科学研究院
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410158385.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-04
Publication Date
2024-11-15
Estimated Expiration
2044-02-04

AI Technical Summary

Technical Problem

The traditional yellow flower tissue culture seedling cultivation method is inefficient, it is difficult to systematically grasp the influence laws of multiple conditions, and it is difficult to efficiently optimize based on experience.

Method used

The two-particle root growth model is used to extract the overall morphology and single root detail characteristics through the root image data, and the learning model is used to quantitatively predict the value range of seedling refining conditions, and combined with the transformer network to optimize the model parameters to achieve automatic optimization of seedling refining conditions.

Benefits of technology

The screening efficiency of seedling refining conditions has been improved, the survival rate and seedling quality of daylily has been significantly improved, and the intelligent optimization method driven by traditional trial and error methods has been transformed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118038447B_ABST
    Figure CN118038447B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of computer-aided agricultural condition optimization, and discloses a method for efficient hardening of daylily tissue culture seedlings, by constructing a root system image training library, designing a root system dual-granularity feature extraction model, and learning the overall root system morphology and single root detail feature representation respectively. Through this feature learning network, an attention mechanism is used to optimize the end-to-end training of the morphological loss function, and a deep representation describing the root system characteristics is extracted. At the same time, multi-target regression is achieved through embedding and feature fusion, and an optimization scheme is automatically given, indicating that the method can provide quantitative hardening guidance for daylily based on morphological learning, and provide reliable technical support for improving cultivation efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of computer-aided farming condition optimization, and in particular to a method for efficiently hardening yellow flower tissue culture seedlings. Background Art

[0002] Hardening is an important part of seedling production, which is the process of transferring tissue culture seedlings or cuttings to a new substrate or environment for adaptation and growth. Hardening can help seedlings sprout more roots and healthy upper tissues to adapt to the growth environment after subsequent transplanting. A correct and efficient hardening process is crucial to ensuring seedling quality and production efficiency. Root growth and survival rate directly affect the subsequent growth and development of seedlings. Traditionally, hardening conditions for yellow flowers are mostly determined by trial and error, but the efficiency is low and difficult to control.

[0003] Traditionally, the exploration of conditions for hardening of yellow flower tissue culture seedlings is mainly based on a large number of experimental experiments. By changing one or several condition parameters such as temperature, humidity, substrate type, etc., the growth conditions of yellow flower seedlings under different conditions, such as rooting conditions, leaf conditions, etc., are observed, and their growth effects, such as survival rate and other indicators, are recorded. After a long period of repeated attempts, the more ideal conditions for hardening are summarized from successful and failed cases. However, this method is inefficient and it is difficult to systematically grasp the influence of multiple conditions. In addition, the error is large, and it is difficult to accurately locate the degree of optimization of each condition. At the same time, since the condition setting relies on experience, it is difficult to optimize efficiently.

[0004] In recent years, advances in computational science have begun to have an impact on the field of agricultural science, especially the emergence of methods for optimizing agricultural conditions based on computing power and data-driven. This method can solve the shortcomings of traditional methods. By learning from a large number of samples, it can automatically match the laws of agricultural optimization conditions, quantitatively predict the value range of each condition based on the model, and provide automated optimization plan recommendations. This data-driven method mines implicit laws, is highly efficient, and provides refined and personalized solutions. It is no longer a simple trial and error, but an intelligent way to optimize agricultural conditions guided by data quantification. This data-driven approach has become an industry consensus. Based on this, we proposed an efficient method for hardening yellow flower tissue culture seedlings, and based on this, we obtained an optimization condition for hardening yellow flower tissue culture seedlings. Summary of the invention

[0005] The present invention proposes an efficient method for hardening daylily tissue culture seedlings. Specifically, we have developed a dual-granularity root growth model to extract the overall morphological characteristics of daylily roots and the detailed characteristics of single roots, and then based on these characteristics, a learning model is used to quantitatively predict the value range of each condition.

[0006] First, we obtain a series of root system image data of daylily seedlings;

[0007] Next, each item in the root system image data of the daylily is encoded and input into VIT to obtain the encoding features of the root system image of the daylily;

[0008] Next, the dimension reduction of the encoded features of the daylily root image is performed to obtain its dimension reduction sequence;

[0009] Next, the dimension-reduced sequence is directly input into the first transformer model to obtain a further feature sequence;

[0010] Next, we obtain the overall morphology of the daylily root system image and the detailed signs of a single root.

[0011] Below, we explain in detail how the model obtains the overall morphological characteristics of the daylily root system image and the detailed characteristics of a single root.

[0012] In order to make the model learn the overall morphological characteristics of the daylily root system image, theoretically, the daylilies that adopt the same seedling hardening conditions (such as open-cover placement conditions, rooting water treatment conditions, substrate treatment conditions, and greenhouse hardening conditions) are more similar overall than the daylilies that adopt different seedling hardening conditions. Therefore, we use the daylily root system images of the same seedling hardening conditions as positive samples, and the daylily root system images of the same seedling hardening conditions as negative samples.

[0013] Then the overall morphological sign loss function is as follows:

[0014]

[0015] r j Indicates the root morphology code corresponding to sample j;

[0016] r p Indicates the root morphology code of the same cultivation data condition p as sample j, as a positive sample;

[0017] r a Represents the root morphology encoding of other different cultivation data conditions a, as negative samples;

[0018] Similarity(r j , r p ) represents the similarity between the root encoding of sample j and positive sample p;

[0019] Similarity(r j, r a ) represents the similarity between the root encoding of sample j and negative sample a;

[0020] A(j) represents the index of other samples that are different from sample j except the positive sample set;

[0021] 1 / NS represents the normalization factor, NS is the number of samples;

[0022] M represents the set of all samples;

[0023] The purpose of loss1 is to make the root morphology of sample j more similar to that of the positive sample p, and to make the root morphology of sample j more different from that of the negative sample a, so as to optimize the accuracy of root encoding in extracting the overall morphological characteristics of the root system.

[0024] Specifically, the loss function of a single detail feature is as follows:

[0025]

[0026] The parameters of the loss2 loss function have the following meanings:

[0027] u represents the shape of a randomly sampled root;

[0028] u p represents the positive morphological samples from the same root as u;

[0029] u mj represents the negative morphological samples from other roots, j represents the index of the negative sample;

[0030] Similarity(u,u p ) represents u and positive sample u p The morphological similarity of

[0031] Similarity(u,u mj ) represents u and negative sample u mj Morphological similarity;

[0032] J represents the number of negative morphological samples of a single root sample;

[0033] According to loss1 and loss2, we get the comprehensive loss function loss=loss1+loss2;

[0034] The seedling hardening condition corresponding to the current sample is subjected to WordEmbdding to obtain the word embedding of the current seedling hardening condition; the word embedding is passed through the second transformer to obtain the encoding feature of the seedling hardening condition corresponding to the current sample;

[0035] The seedling hardening condition coding features and the single root detail signs corresponding to the current sample are spliced ​​into comprehensive features, which are input into the third transformer, and the entire model is trained according to the comprehensive loss function to obtain the model parameters of the dual-granularity root growth model;

[0036] Obtaining the dual-grained root growth model according to the model parameters;

[0037] Using the dual-size root growth model to predict multiple groups of seedling hardening conditions;

[0038] Artificial seedling hardening is carried out according to the multiple groups of seedling hardening conditions to obtain the best seedling hardening conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A model structure diagram of a dual-grained root growth model for the highly efficient hardening method for yellow lily tissue culture seedlings provided by the present invention;

[0040] Figure 2 Different yellow flower seedling images are input into the model;

[0041] Figure 3 These are tissue culture seedlings of yellow flower. DETAILED DESCRIPTION

[0042] like Figure 1 As shown, it shows the model structure diagram of the dual-grained root growth model. Figure 2 For Figure 1 Enter a zoomed-in image of the model example in .

[0043] First, we obtained a large number of daylily root growth images from multiple daylily cultivation experiments, such as Figure 2 The images in the figure are from the planting of daylily under different hardening conditions. Each image records the corresponding specific hardening condition parameter settings.

[0044] We collected thousands of daylily root images as sample data for model learning and training.

[0045] Next, in order to improve the consistency of model learning, we pre-processed each daylily root image, including segmenting the image to locate and extract the root part to remove the image background and impurities, and then standardized the root image to obtain the size and pixel points. Each regularized root image was input as a sample to the subsequent deep learning model.

[0046] For each root system image sample, we use the pre-trained ViT image encoder to encode it and extract the low-dimensional feature representation of the image. The ViT encoder uses a Transformer structure and learns the representation of global image features through an attention mechanism; at the same time, we also perform fine-grained encoding on a single root system to extract single root features.

[0047] Subsequently, we perform dimensionality reduction on the root image features and single root features extracted by ViT, and use a time series model (such as LSTM, not specifically limited here) to extract local sequential features; the reduced feature sequence is used as input and fed to the first transformer to further learn the root morphology representation.

[0048] Then, we query the root image feature sequence to obtain the features describing the overall root morphology, and query the single root feature sequence to extract the single root detail features.

[0049] We propose two loss functions to optimize the overall root system morphology learning and the individual root morphology learning respectively: the overall morphology loss function , where r j represents the root morphology code corresponding to sample j, r p represents the root morphology encoding of the same cultivation data condition p as sample j, as a positive sample, r a Represents the root morphology encoding of other different cultivation data conditions a, as negative samples, Similarity(r j , r p ) represents the similarity between the root encoding of sample j and positive sample p, Similarity(r j, r a ) represents the similarity between the root encoding of sample j and negative sample a, A(j) represents the index of other samples that are different from sample j except the positive sample set, 1 / NS represents the normalization factor, NS is the number of samples, and M represents all sample sets.

[0050] The purpose of loss1 is to make the root morphology of sample j more similar to that of the positive sample p, and to make the root morphology of sample j more different from that of the negative sample a, so as to optimize the accuracy of root encoding in extracting the overall morphological characteristics of the root system.

[0051] Features are learned by maximizing the similarity of root morphology under the same hardening condition, and the single root loss function , where u represents the shape of a randomly sampled root, u p represents the positive morphological samples from the same root as u, u mj represents the negative morphological samples from other roots, j represents the index of the negative sample, Similarity(u, u p ) represents u and positive sample u p The morphological similarity, Similarity(u,u mj ) represents u and negative sample u mj Morphological similarity, J represents the number of negative morphological samples of a single sample.

[0052] By optimizing these two loss terms simultaneously, the dual-granularity root image feature representation of daylily can be comprehensively learned.

[0053] Next, we embed the conditional word vector representation of the seedling training condition parameters corresponding to each image as a latent variable and input it into the second transformer network for deep learning to extract the feature vector.

[0054] Subsequently, we concatenate the root morphology feature vector, single root feature vector and conditional feature vector to obtain the feature glue representation, which is fed into the third transformer model as input for deep training. At the same time, we optimize the root morphology loss and conditional prediction loss to achieve a combination of feature learning and multi-objective regression.

[0055] After training a dual-granularity root image deep learning model that can extract root morphological features and predict parameters for seedling hardening conditions, we can input new samples. The model predicts multiple possible combinations of optimized seedling hardening conditions based on features, and selects the best combination for experimental verification. If the effect is not good, we can continue to optimize the model parameters to obtain more efficient seedling hardening guidance. The entire process realizes data-driven intelligent diagnosis and seedling hardening optimization of daylily root images.

[0056] According to the above scheme, we will set the following conditions:

[0057] Conditions for uncapping: 2 days, temperature 15 degrees, humidity 65%;

[0058] Rooting water treatment conditions: After washing the culture medium of the seedling roots, place them in rooting water for 8 minutes. The rooting water is 3.20% sodium nitrate and naphthylacetic acid aqueous solution (including 1.65% sodium nitrophenolate and 1.2% naphthylacetic acid) prepared into a 2000-fold solution.

[0059] Substrate treatment conditions: temperature was 18°C, leaves were sprayed with water once a day, and the substrate formula ratio was coconut bran: nutrient soil = 3:1;

[0060] Greenhouse seedling hardening conditions: After the yellow flower seedlings have grown indoors for 50-60 days, when the number of roots reaches more than 4 and the leaf length is more than 8 cm, the temperature is controlled at 15-20℃, the leaves are sprayed with water once every 3 days, and the soil pH is between 5.2-7.6.

[0061] Further optimization for:

[0062] Conditions for uncapping: 3 days, temperature 20 degrees, humidity 70%;

[0063] Rooting water treatment conditions: After washing the culture medium of the seedling roots, place them in rooting water for 10 minutes. The rooting water is a 2000-fold solution of 2.85% sodium nitrate and naphthylacetic acid (including 1.65% sodium nitrophenolate and 1.2% naphthylacetic acid).

[0064] Substrate treatment conditions: temperature was 20℃, leaves were sprayed with water once a day, and the substrate formula ratio was coconut bran: nutrient soil = 4:1;

[0065] Greenhouse seedling hardening conditions: After the yellow flower seedlings grow indoors for 40-50 days, when the number of roots reaches more than 5 and the leaf length is more than 10, the temperature is controlled at 15-20℃, the leaves are sprayed with water once every 3 days, and the soil pH is between 5.0-8.6.

[0066] According to the above optimized conditions, the survival rate and seedling quality of daylily have been significantly improved. Figure 3 As shown, it is the yellow flower tissue culture seedlings after optimizing the seedling hardening conditions.

[0067] The main contributions and beneficial effects of the present invention are as follows: the present invention adopts computer-aided means to optimize the conditions for hardening of day lily seedlings, and designs a dual-granularity feature extraction model for the root system, respectively learning the overall morphology of the root system and the detailed feature representation of a single root, and accelerating the optimization efficiency of the conditions for hardening of seedlings through data-driven condition optimization, and transforming the traditional method of finding a needle in a haystack verified by pure experiments into a method of finding a needle in a small pond, and the screening rate of the conditions for hardening of day lily seedlings is significantly improved.

Claims

1. A method for efficiently hardening yellow lily tissue culture seedlings, characterized in that: The method comprises the following steps: Establish a dual-size root growth model; Acquire the image data of daylily root system; Get the seedling hardening conditions corresponding to the current sample; Training the dual-granularity root growth model according to the daylily root system image data and the hardening conditions corresponding to the current sample to obtain a trained dual-granularity root system growth model; The loss function of the dual-grained root growth model is: loss = loss 1 + loss 2 , ,in, r j represents the root morphology code corresponding to sample j, r p represents the root morphology code of the same cultivation data condition p as sample j, as a positive sample, r a Represents the root morphology encoding of other different cultivation data conditions a, as negative samples, Similarity(r j , r p ) represents the similarity between the root encoding of sample j and positive sample p, Similarity(r j, r a ) Representation sample j and negative samples a The similarity of root encoding, A(j) Indicates that sample j is excluding the positive sample set and j Different other sample indexes, 1 / NS represents the normalization factor, NS is the sampling number, M represents the set of all samples, ,in, u represents the shape of a randomly sampled root, u p Representation and u Positive morphological samples from the same root, u mj represents negative morphological samples from other roots, j represents the index of negative samples, Similarity(u,u p ) express u With positive samples u p The morphological similarity of Similarity(u,u mj ) express u With negative samples u mj Morphological similarity, J It represents the number of negative morphological samples of a single sample; Optimizing the daylily seedling hardening conditions according to the trained dual-granularity root growth model to obtain multiple groups of optimized daylily seedling hardening conditions; According to the multiple groups of optimized day lily seedling hardening conditions, artificial experiments are carried out to select the final day lily seedling hardening conditions.

2. A method for efficiently hardening yellow lily tissue culture seedlings as claimed in claim 1, characterized in that: The dual granularity represents the overall morphological signs of the daylily and the detailed signs of a single root.

3. A method for efficiently hardening yellow lily tissue culture seedlings as claimed in claim 1, characterized in that: The conditions to be optimized include: open-lid placement conditions, rooting treatment conditions, substrate treatment conditions and greenhouse treatment conditions.