Wild rice salt tolerance intelligent evaluation method and system based on deep learning

Abstract

The embodiment of the invention provides an intelligent evaluation method and system for the salt tolerance of wild rice based on deep learning, and the method comprises the steps: S1, obtaining the wild rice, collecting the pictures of the wild rice, and making a data set; s2, using the data set to train an improved YOLOv8-seg model, and obtaining a wild rice salt tolerance intelligent evaluation model ST-YOLO, the improvement measures comprising: constructing a C2f-DBB module to replace a convolutional layer in the C2f module, introducing a convolution and attention fusion module CAFM, and using a deformable convolution and spatial information enhancement module to design a spatial pyramid pooling layer; s3, performing instance segmentation on an input image by using the ST-YOLO model, and outputting a green leaf pixel number, a yellow leaf pixel number and a total leaf pixel number of a single plant of wild rice leaf; and S4, obtaining the salt tolerance grade of the wild rice based on the pixel number. According to the method, by improving the YOLOv8-seg model, the feature extraction and fusion capability is improved, automatic salt tolerance grade evaluation is achieved, the precision is high, the real-time performance is high, and germplasm screening is assisted.

Description

Technical Field

[0001] The present invention relates to the field of crop breeding, and particularly to an intelligent evaluation method and system for the salt tolerance of wild rice based on deep learning. Background Art

[0002] As an important germplasm resource for rice genetic improvement, the phenotypic evaluation of the salt tolerance of wild rice seedlings is crucial for the breeding of salt-tolerant rice varieties and the utilization of saline-alkali land resources. Traditional methods evaluate the salt tolerance level by manually observing the dead leaf rate, which has problems such as low efficiency, strong subjectivity, and insufficient quantification accuracy, and it is difficult to meet the needs of large-scale germplasm resource screening. With the application of deep learning technology in agricultural phenotype analysis, image-based automated detection provides a new path for salt tolerance evaluation. However, existing models face multiple challenges: after wild rice leaves are stressed by salt, the feature changes are subtle (such as differences in leaf curling and yellowing degree), and there are complex morphologies such as leaf overlap and different sizes, resulting in difficult extraction of small target features; in addition, existing models have limitations in feature fusion and spatial representation capabilities, making it difficult to accurately capture subtle phenotypic differences, and the model lightweight is insufficient, and the computational efficiency cannot meet the requirements of real-time detection.

[0003] Existing technologies generally have defects such as insufficient feature extraction ability, low detection accuracy, or high computational complexity. There is an urgent need for a high-precision deep learning model to achieve efficient and accurate evaluation of the salt tolerance level of wild rice. Summary of the Invention

[0004] The present application provides an intelligent evaluation method and system for the salt tolerance of wild rice based on deep learning. Aiming at the problems of low detection accuracy and low efficiency of wild rice salt tolerance detection, the YOLOv8-seg model is improved through modules such as DBB and CAFM to enhance the feature extraction and fusion capabilities, realize automated salt tolerance level evaluation, with high accuracy and strong real-time performance, and assist in germplasm screening.

[0005] In the first aspect, an intelligent evaluation method for the salt tolerance of wild rice based on deep learning is provided. The method includes: S1: Obtain wild rice, collect pictures of the wild rice and make a data set; S2: Use the data set to train an improved YOLOv8-seg model to obtain an intelligent evaluation model ST-YOLO for the salt tolerance of wild rice. The improvement measures include: replacing the convolutional layer in the C2f module with a multi-branch structure DBB module to construct a C2f-DBB module, introducing a convolutional and attention fusion module CAFM, and designing a spatial pyramid pooling layer using deformable convolution and spatial information enhancement module; S3: Use the ST-YOLO model to perform instance segmentation on the input image, and output the number of green leaf pixels, the number of yellow leaf pixels, and the total number of leaf pixels of a single wild rice leaf; S4: Obtaining the salt tolerance level of the wild rice based on the number of pixels.

[0006] It should be understood that by constructing multi-branch feature extraction through the DBB module, fusing global and local features through the CAFM module, and optimizing spatial representation through deformable convolution, the model's ability to capture subtle salt stress characteristics of leaves can be improved, and the automated and accurate evaluation of the salt tolerance level of wild rice can be achieved. The detection has high accuracy and strong real-time performance, providing efficient technical support for the screening of germplasm resources.

[0007] In combination with the first aspect, in some implementations of the first aspect, step S1 includes: S101: Seed selection and dormancy breaking; S102: Disinfection and seed soaking and germination; S103: hydroponics and salt stress treatment; S104: Data collection to produce the data set.

[0008] It should be understood that "dormancy breaking" refers to the technical operation of breaking the dormancy of wild rice seeds through artificial intervention to enable them to germinate. Wild rice seeds are often in a physiological dormant state due to the dense structure of the seed coat and the presence of germination inhibitors. If dormancy is not broken, it will lead to low germination rate and inconsistent germination time, affecting the accuracy and repeatability of subsequent salt tolerance phenotypic experiments.

[0009] It should be understood that "seed soaking and germination" is a key processing step before wild rice seeds germinate. It refers to placing the disinfected seeds in a climate incubator for soaking, allowing the seeds to fully absorb water and swell, activate enzyme activity and soften the seed coat. They are then covered with a double layer of gauze and distilled water is added. Culture continues under the same temperature and shading conditions to encourage the radicle and sheath to break through the seed coat to form strong young shoots. This process ensures uniform seed germination and synchronous seedling development by controlling water, temperature and environmental conditions, providing consistent experimental materials for subsequent hydroponic growth and salt stress treatment, avoiding phenotypic data errors caused by differences in seed germination, and ensuring the reliability and accuracy of salt tolerance evaluation model training.

[0010] It should also be understood that "salt stress treatment" is an experimental operation that simulates the saline-alkali environment to impose salt stress on wild rice seedlings, causing the plants to suffer from salt ion toxicity and osmotic stress, inducing salt-tolerance related phenotypes such as yellowing, curling, and wilting of leaves, and prompting wild rice seedlings to show different degrees of salt damage characteristics (such as differences in dead leaf rates). It provides key phenotypic data for subsequent collection of leaf images, quantification of salt stress damage, and training of salt-tolerance phenotypic evaluation models, and is an important link in screening salt-tolerant germplasm resources and analyzing stress resistance mechanisms.

[0011] In combination with the first aspect, in some implementations of the first aspect, in the backbone network of the YOLOv8-seg model, a multi-branch structure DBB module is used to replace the bottleneck convolutional layer in the C2f module to construct a C2f-DBB module.

[0012] In combination with the first aspect, in some implementations of the first aspect, the DBB module includes a 1×1 convolutional module, a K×K convolutional module, and an average pooling module. During the inference stage, the DBB module is converted into a single K×K convolutional layer through the reparameterization technique to enhance the multi-dimensional feature extraction ability without increasing the computational cost.

[0013] It should be understood that the DBB module extracts multi-scale and multi-complexity features through a multi-branch structure (including parallel branches such as 1×1 convolution, K×K convolution, and average pooling) during the training stage to enrich the feature representation. During the inference stage, the multi-branch is merged into a single K×K convolutional layer through the reparameterization technique to avoid increasing the computational cost. By replacing the bottleneck convolutional layer of C2f with the DBB module in the backbone network to construct the C2f-DBB module, the above technical solution can significantly enhance the multi-dimensional extraction ability of the subtle salt stress features of wild rice leaves without increasing the model parameters and inference time, improve the feature diversity and expression ability, thereby improving the detection accuracy of the model for leaves with complex shapes, and providing a basis for the accurate evaluation of salt tolerance phenotypes.

[0014] In combination with the first aspect, in some implementations of the first aspect, the convolution and attention fusion module CAFM is introduced between the backbone network and the detection head of the YOLOv8-seg model. The convolution and attention fusion module CAFM fuses global and local features through a multi-scale feed-forward network and an attention mechanism to improve the detection performance of small target leaf features.

[0015] It should be understood that the CAFM module realizes feature enhancement through a multi-scale feed-forward network and an attention mechanism: the input feature is branched into three paths of query (Q), key (K), and value (V), and local features (such as leaf edge details) are extracted through deformable convolution respectively, combined with the global features (such as the distribution of the whole plant leaves) processed by 1×1 convolution, and multi-scale information is dynamically fused through attention weights. The technical solution of this application can effectively solve the problem of small targets in the detection of wild rice leaves by introducing CAFM between the backbone network and the detection head: on the one hand, the deformable convolution adaptively adjusts the sampling point positions to accurately capture the features of small leaves that are occluded or overlapped; on the other hand, the attention mechanism suppresses background noise and strengthens the response to subtle leaf changes (such as local yellowing) caused by salt stress, significantly improving the detection recall rate and localization accuracy of small target leaves while maintaining the overall efficiency of the model, and providing a more reliable feature basis for the automated evaluation of wild rice salt tolerance phenotypes.

[0016] In combination with the first aspect, in some implementations of the first aspect, the design of the spatial pyramid pooling layer includes: using deformable convolution and a spatial information enhancement module to adaptively adjust the receptive field to fit the complex morphology of the leaves and strengthen the spatial representation ability of the target features.

[0017] It should be understood that the spatial pyramid pooling layer dynamically adjusts the sampling points of the convolution kernel through deformable convolution to adapt to the complex morphology of wild rice leaves with different sizes and curling and overlapping (such as curved leaf edges and local occlusion areas). Combining with the spatial information enhancement module to integrate multi-scale spatial features (such as the global contour and local texture of the leaves), it realizes adaptive feature extraction for leaves of different morphologies. This design effectively solves the limitations of traditional fixed convolution kernels in the extraction of irregular leaf features, significantly strengthens the model's representation ability for the spatial position, edge details, and local damage features (such as yellowing patches) of the leaves, thereby improving the instance segmentation accuracy and providing a more reliable spatial feature basis for accurately calculating the dead leaf rate and salt tolerance level classification. It has important practical value for the phenotypic detection of wild rice in complex field environments.

[0018] In combination with the first aspect, in some implementations of the first aspect, the step S4 includes: S401: Calculate the ratio of yellow leaf pixels to the total leaf pixels according to the number of pixels obtained in step S3. The calculation formula is as follows: , where represents the area of the segmented dead leaves, represents the total area of the segmented leaves; S402: Evaluate the salt tolerance level and the severity of the impact of the wild rice leaves due to salt stress according to the ratio of the dead leaf area to the total leaf area.

[0019] It should be understood that step S4 quantifies the salt stress damage by calculating the ratio of the dead leaf area (number of yellow leaf pixels / total number of leaf pixels), divides the salt tolerance level according to the ratio, converts the leaf phenotypic characteristics into quantifiable salt tolerance evaluation indicators, realizes objective and efficient automatic level determination, solves the problems of strong subjectivity and low efficiency in traditional manual evaluation, and provides a standardized quantitative basis for wild rice germplasm screening.

[0020] In the second aspect, a wild rice salt tolerance intelligent evaluation system based on deep learning is provided. The system includes: An image input module, and the image acquisition module is used to input the wild rice image; A detection and evaluation module, and the detection and evaluation module is used to detect the input wild rice image and obtain the salt tolerance level of the wild rice. The detection and evaluation module can execute the implementation methods described in any item of the first aspect.

[0021] In combination with the second aspect, in some implementations of the second aspect, the system adopts a web - based interaction design, supporting remote operations and data management for multiple users and multiple roles. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flowchart for implementing a method for intelligent evaluation of salt tolerance of wild rice based on deep learning provided for the application embodiments.

[0023] Figure 2 It is a flowchart for implementing a method for obtaining wild rice and making a wild rice data set provided for the application embodiments. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] The terms used in the following embodiments are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification and appended claims of the present application, the singular forms "a", "an", "the", "above", "said", "this" are also intended to include the forms such as "one or more", unless there is a clear contrary indication in the context. It should also be understood that in the following embodiments of the present application, "at least one", "one or more" refer to one, two or more than two. The term "and / or" is used to describe the association relationship of associated objects and means that three relationships can exist; for example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone, where A and B can be singular or plural. The character " / " generally means that the associated objects before and after are in an "or" relationship.

[0025] The reference to "one embodiment" or "some embodiments" etc. described in this specification means that a specific feature, structure or characteristic described in combination with the embodiment is included in one or more embodiments of the present application. Thus, the statements "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments" etc. that appear in different places in this specification do not necessarily all refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "comprise", "include", "have" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0026] As the core germplasm for rice genetic improvement, the evaluation of the salt tolerance phenotype at the seedling stage of wild rice is crucial for salt tolerance breeding and the development of saline-alkali land. The traditional method of manually observing the dead leaf rate is inefficient, subjective, and lacks sufficient quantification accuracy, making it difficult to meet the needs of large-scale germplasm screening. Deep learning technology provides a new path for automated phenotype detection. However, there are problems such as subtle characteristics (such as yellowing and curling) and complex morphology (such as overlapping and different sizes) in the salt-stressed leaves of wild rice, resulting in difficulties in extracting small target features for existing models, and limitations in feature fusion, spatial representation, and lightweight efficiency. Therefore, targeted improvements are urgently needed.

[0027] The embodiments of this application provide an intelligent evaluation method and system for the salt tolerance of wild rice based on deep learning. To address the problem of evaluating the salt tolerance level at the seedling stage of wild rice, the YOLOv8-seg model is improved (by introducing modules such as DBB multi-branch, CAFM attention, and deformable convolution), enhancing the ability to extract complex leaf features and detect small targets. Combining pixel quantization enables automated evaluation of the salt tolerance level, solving the problems of low efficiency and strong subjectivity of traditional methods, and providing high-precision and real-time technical support for germplasm screening.

[0028] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings.

[0029] Figure 1 It is a flowchart for implementing an intelligent evaluation method for the salt tolerance of wild rice based on deep learning provided by the embodiments of this application.

[0030] In some examples, the method includes: S1: Obtain wild rice, collect pictures of the wild rice and make a dataset; S2: Use the dataset to train the improved YOLOv8-seg model to obtain the intelligent evaluation model ST-YOLO for the salt tolerance of wild rice. The improvement measures include: replacing the convolutional layer in the C2f module with a multi-branch structure DBB module to construct a C2f-DBB module, introducing a convolutional and attention fusion module CAFM, and using a deformable convolution and spatial information enhancement module to design a spatial pyramid pooling layer; S3: Use the ST-YOLO model to perform instance segmentation on the input image, and output the number of green leaf pixels, the number of yellow leaf pixels, and the total number of leaf pixels of a single wild rice leaf; S4: Obtain the salt tolerance level of the wild rice based on the number of pixels.

[0031] Figure 2 It is a flowchart for implementing a method of obtaining wild rice and making a wild rice dataset provided by the embodiments of this application.

[0032] In some examples, the step S1 includes: S101: Seed selection and dormancy breaking; S102: Disinfection and seed soaking for germination acceleration; S103: Hydroponics and salt stress treatment; S104: Data collection to produce the said data set.

[0033] In a possible implementation manner, the embodiments of the present application select the wild rice variety No. 1 - 254 provided by the Institute of Crop Science, Chinese Academy of Agricultural Sciences. 30 plump seeds are taken for each variety, and the dormancy is broken by high-temperature treatment at 45°C - 50°C for 72 hours. After disinfection with 1% NaClO solution for 25 minutes, the seeds are soaked for 48 hours and germinated for 48 hours (covered with double-layer gauze and distilled water added) under the condition of shading at 28°C. 20 germinated seeds are selected and placed in a 96-well plastic hydroponic box, and hydroponically cultured for 14 days to the two-leaf and one-heart stage at 30°C and 12000 Lx light (12 h / day) with Yoshida culture solution. Subsequently, it is transferred to Yoshida culture solution containing 10‰ NaCl for treatment for 7 days, and then restored with NaCl-free culture solution for 7 days. Each variety is replicated 8 plants, and images are collected: The vertical shooting method is adopted (the device is 42.5 cm away from the plant), and JPG format images are obtained with a 50-million-pixel lens (two-fold digital zoom) against a black background. The green leaf and yellow leaf areas are labeled through Labelme to generate a data set, which is divided into a training set, a validation set, and a test set according to a ratio for training the improved YOLOv8-seg model (ST-YOLO) to realize the automatic evaluation of the salt tolerance phenotype of wild rice seedlings.

[0034] In some examples, in the backbone network of the YOLOv8-seg model, the bottleneck convolutional layer in the C2f module is replaced with a multi-branch structure DBB module to construct a C2f-DBB module.

[0035] In some examples, the DBB module includes a 1×1 convolutional module, a K×K convolutional module, and an average pooling module. The DBB module is converted into a single K×K convolutional layer through the reparameterization technique during the inference stage to enhance the multi-dimensional feature extraction ability without increasing the computational amount.

[0036] In some examples, the convolutional and attention fusion module CAFM is introduced between the backbone network and the detection head of the YOLOv8-seg model. The convolutional and attention fusion module CAFM fuses global and local features through a multi-scale feedforward network and an attention mechanism to improve the detection performance of small target leaf features.

[0037] In some examples, the design of the spatial pyramid pooling layer includes: adopting deformable convolution and a spatial information enhancement module to adaptively adjust the receptive field to adapt to the complex morphology of the leaves and strengthen the spatial representation ability of target features.

[0038] In some examples, the step S4 includes: S401: Calculate the ratio of yellow leaf pixels to total leaf pixels according to the number of pixels obtained in step S3 above. The calculation formula is as follows: , wherein, represents the area of the segmented dead leaves, represents the total area of the segmented leaves; S402: Evaluate the salt tolerance level and the severity of the impact of the wild rice leaves due to salt stress according to the ratio of the dead leaf area to the total leaf area.

[0039] The embodiments of the present application also provide an intelligent evaluation system for the salt tolerance of wild rice based on deep learning. The system includes: An image input module, and the image acquisition module is used to input the wild rice image; A detection and evaluation module, and the detection and evaluation module is used to detect the input wild rice image and obtain the salt tolerance level of the wild rice. The detection and evaluation module can execute the method described in any example of the foregoing embodiments.

[0040] In some examples, the system adopts a web-based interactive design, supporting remote operations and data management for multiple users and multiple roles.

[0041] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modification or change made by those of ordinary skill in the art according to the content disclosed in the present invention shall be included in the protection scope recorded in the claims.

Claims

1. An intelligent evaluation method for the salt tolerance of wild rice based on deep learning, characterized in that, The method includes: S1: Obtain wild rice, collect pictures of the wild rice and make a dataset; S2: Use the dataset to train an improved YOLOv8-seg model to obtain a wild rice salt tolerance intelligent evaluation model ST-YOLO. The improvement measures include: replacing the convolutional layer in the C2f module with a multi-branch structure DBB module to construct a C2f-DBB module, introducing a convolutional and attention fusion module CAFM, and designing a spatial pyramid pooling layer using deformable convolution and spatial information enhancement module; S3: Use the ST-YOLO model to perform instance segmentation on the input image, and output the number of green leaf pixels, the number of yellow leaf pixels and the total number of leaf pixels of a single wild rice plant leaf; S4: Obtain the salt tolerance level of the wild rice based on the pixel numbers.

2. The method according to claim 1, wherein The step S1 includes: S101: Seed selection and dormancy breaking; S102: Disinfection and soaking for germination acceleration; S103: Hydroponics and salt stress treatment; S104: Data collection to make the dataset.

3. The method according to claim 1, wherein In the backbone network of the YOLOv8-seg model, replace the bottleneck convolutional layer in the C2f module with a multi-branch structure DBB module to construct a C2f-DBB module.

4. The method according to claim 3, characterized in that The DBB module includes a 1×1 convolutional module, a K×K convolutional module and an average pooling module. The DBB module is converted into a single K×K convolutional layer through reparameterization technology during the inference stage to enhance the multi-dimensional feature extraction ability without increasing the computational load.

5. The method according to claim 1, characterized in that, Introduce the convolutional and attention fusion module CAFM between the backbone network and the detection head of the YOLOv8-seg model. The convolutional and attention fusion module CAFM fuses global and local features through a multi-scale feedforward network and an attention mechanism to improve the detection performance of small target leaf features.

6. The method according to claim 1, characterized in that The design of the spatial pyramid pooling layer includes: using deformable convolution and spatial information enhancement module to adaptively adjust the receptive field to adapt to the complex morphology of the leaves and strengthen the spatial representation ability of target features.

7. The method according to claim 1, characterized in that, The step S4 includes: S401: Calculate the ratio of yellow leaf pixels to total leaf pixels according to the pixel numbers obtained in the step S3. The calculation formula is as follows: ， Among them, represents the area of the dead leaves that have been segmented, represents the total area of the segmented leaves; S402: Evaluate the salt tolerance level and the severity of the impact of the wild rice leaves due to salt stress according to the ratio of the dead leaf area to the total leaf area.

8. An intelligent evaluation system for the salt tolerance of wild rice based on deep learning, characterized in that, The system includes: An image input module, which is used to input the wild rice image; A detection and evaluation module, which is used to detect the input wild rice image and obtain the salt tolerance level of the wild rice. The detection and evaluation module can execute the method according to any one of claims 1 to 7.

9. The system according to claim 8, wherein The system adopts a web-based interactive design, supporting remote operation and data management for multiple users and multiple roles.

Images