General crop stomata phenotype extraction method based on in-situ imaging and deep learning

Abstract

The application discloses a kind of general crop stomata phenotyping extraction methods based on in situ imaging and deep learning, comprising the following steps: S1, with monitoring terminal acquisition stomata image and carry out pixel-level labeling and data enhancement, construct stomata image dataset;S2, in semantic segmentation network model introduction efficient multi-scale attention module and deformable convolution module;Through stomata image dataset training, output segmentation mask;S3, extract the full-dimension phenotyping index of stomata;S4, the stomata image to be measured is input into the trained semantic segmentation network model, and the corresponding visualization chart and standard form are output.The application constructs in situ non-destructive monitoring terminal, can carry out cross-species in situ microscopic monitoring under the premise of not destroying leaf physiological state;Also improve semantic segmentation network model, solve the identification problem brought by the different morphologies of multiple crop stomata and complex background, and construct phenotyping system, realize the accurate differentiation and quantification between different species and different varieties of the same species.

Description

Technical Field

[0001] This invention relates to the field of agricultural technology, specifically to a general method for extracting crop stomatal phenotypes based on in-situ imaging and deep learning. Background Technology

[0002] Stomata are tiny pores on plant epidermal cells surrounded by guard cells, serving as crucial gateways for the exchange of gases such as water and carbon dioxide between plants and their external environment. The phenotypic characteristics of stomata significantly determine a plant's water use efficiency and abiotic stress response mechanisms, thus substantially influencing crop growth, development, yield, and quality. Therefore, accurately exploring stomatal morphology and its spatiotemporal evolution is of great importance for analyzing the mechanisms of plant environmental adaptation.

[0003] However, traditional stomatal phenotypic acquisition mainly relies on destructive sampling methods such as imprinting, which are not only time-consuming and labor-intensive with extremely low throughput, but also lead to stomatal morphology distortion, making it impossible to achieve in-situ, continuous dynamic monitoring of living organisms. Although deep learning technology has been applied to stomatal recognition in recent years, existing networks often suffer from severe "species specificity," meaning that models trained on a single crop have difficulty generalizing across species. In particular, there are significant morphological differences between gramineous food crops and oilseed crops, resulting in poor generalization ability, blurred edge segmentation, and low accuracy of existing models when faced with complex backgrounds and multi-species differences. Furthermore, current stomatal evaluation systems are often limited to simple static basic indicators (such as stomatal count and area), lacking refined quantification of advanced morphological indicators and dynamic kinematic features. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to propose a general crop stomatal phenotypic extraction method based on in-situ imaging and deep learning. This method involves constructing an in-situ non-destructive monitoring terminal to perform cross-species in-situ microscopic monitoring without damaging leaf physiological states. Furthermore, it improves the semantic segmentation network model to solve the recognition challenges posed by diverse stomatal morphologies and complex backgrounds across multiple crops. A phenotypic measurement system is also constructed to achieve accurate quantification and differentiation between different species and different varieties of the same species. In addition, the trained model is combined with external software to achieve a high degree of automation in the testing process.

[0005] This was achieved through the following technical solutions: A general method for extracting crop stomatal phenotypes based on in-situ imaging and deep learning includes the following steps: S1. Use monitoring terminals to collect stomatal images from living leaves of various crops, and perform pixel-level annotation and data augmentation on the stomatal images to construct a stomatal image dataset. S2. Construct a semantic segmentation network model based on the SegFormer architecture, and introduce an efficient multi-scale attention module and a deformable convolution module into the encoder front end of the semantic segmentation network model; train the semantic segmentation network model with the pore image dataset in step S1, and output the segmentation mask of pores and gaps. S3. Based on the segmentation mask in step S2, extract the full-dimensional phenotypic indices of stomata; the full-dimensional phenotypic indices include the following parameter types: basic static parameters and advanced morphological parameters. S4. Input the image of the stomata to be tested into the trained semantic segmentation network model, and output the corresponding visualization charts and standard tables according to the parameter types in step S3.

[0006] Optionally, in step S1, the monitoring terminal includes: an image acquisition module, a leaf fixing module, and an illumination module; the image acquisition module is used to capture images of the stomata of any live leaf; the leaf fixing module uses a double-layer clamping mechanism to fix live leaves of various crops, the upper layer of the double-layer clamping mechanism is a grid-shaped aluminum alloy plate with a central observation window, and the lower layer of the double-layer clamping mechanism is an acrylic plate with transparent round holes; the illumination module uses a pure white cold light source located below the leaf fixing module to provide transmitted light to the live leaves.

[0007] Optionally, the multiple crops in step S1 include at least wheat, corn, cotton, soybean, rapeseed, and peanut; stomatal images include still images and video clips.

[0008] Optionally, in step S2, the efficient multi-scale attention module is used to group and reorganize the channel dimensions of the input features of the semantic segmentation network model, extract channel interaction features and local spatial features using parallel convolutional paths, and aggregate across spatial dimensions in the horizontal and vertical directions using two-dimensional global average pooling to generate a multi-scale attention weight map.

[0009] Optionally, in step S2, the deformable convolution module is connected in series with the efficient multi-scale attention module; the deformable convolution module adaptively learns the sampling offset and modulation scalar to fit the receptive field to the irregular boundary of the pore.

[0010] Optionally, the semantic segmentation network model in step S2 is further equipped with a decoder, which is used to receive the output of the encoder and extract multi-scale contextual information through the dilated spatial convolution pooling pyramid module, and finally outputs a pixel-level segmentation mask by the classification segmentation convolutional layer.

[0011] Optionally, in step S3, the basic static parameters include at least: the area, perimeter, length, width, aspect ratio, equivalent diameter and orientation of pores or voids at the individual scale, and the total number of pores, open ratio, density, average pore distance and coefficient of variation at the population scale.

[0012] Optionally, in step S3, the advanced morphological parameters include at least: the roundness, eccentricity, and aperture of stomata or pores at the individual scale, and the Thiessen polygon area, Thiessen polygon uniformity, coordination number, cell area, and homogeneity of stomata or pores at the population scale.

[0013] Optionally, before inputting the pore image to be tested into the trained semantic segmentation network model in step S4, a preprocessing step is also included: scanning the edges of the pore image to be tested, filtering out truncated pores and tiny noise points with an area smaller than a preset threshold.

[0014] Optionally, in step S4, the visualization charts include: nearest neighbor distance histogram, comprehensive phenotypic radar chart, rose diagram showing stomatal orientation, stomatal distribution topology network diagram, and Thiessen polygon spatial distribution diagram.

[0015] The beneficial effects of this invention compared to the prior art are: 1) Achieve in-situ non-destructive high-throughput monitoring of multiple species: The developed hardware system overcomes the shortcomings of traditional destructive sampling and can carry out cross-species in-situ microscopic monitoring without damaging the physiological state of leaves.

[0016] 2) Achieving extremely high segmentation accuracy and generalization: The semantic segmentation network model solves the recognition problem caused by the different stomatal morphologies of various crops and complex backgrounds by combining efficient multi-scale attention modules and deformable convolution modules. Experiments show that the model achieves mIoU accuracy of 92.31% and 87.18% for stomata and pores, respectively. It can also achieve high-precision stomatal extraction for other species that did not participate in training in the case of zero samples.

[0017] 3) A comprehensive and multi-dimensional phenotypic measurement system was constructed: This method not only extracts basic static parameters, but also innovatively incorporates a number of new advanced spatial indicators such as Voronoi cell area and column alignment score, which effectively realizes the accurate differentiation and quantification between different species and different varieties of the same species.

[0018] 4) Fully automated processing: The trained model, combined with the accompanying software, automates the process from image import and model inference to chart drawing and phenotypic index export, greatly reducing the technical threshold of phenomics analysis and improving the data conversion efficiency of stomatal research. Attached Figure Description

[0019] Figure 1 This is a flowchart of a general crop stomatal phenotype extraction method based on in-situ imaging and deep learning; Figure 2 This is a schematic diagram of the hardware components of a monitoring terminal; Figure 3This is a schematic diagram illustrating data extraction, phenotypic analysis, and software workflow. Figure 4 This is a schematic diagram showing the ablation experiment and comparison experiment results of a semantic segmentation network model; Figure 5 A comprehensive quantitative analysis diagram of stomatal morphology, spatial arrangement patterns, and dynamic behavior of six major grain and oil crops; Figure 6 A comparison chart showing the values ​​extracted by the semantic segmentation network model for the eight basic parameters and the values ​​measured manually; Figure 7 This is a schematic diagram illustrating the segmentation performance verification of the semantic segmentation network model on various zero-shot plants. Figure 8 This is a schematic diagram of a software interface. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] like Figure 1 The diagram shows a flowchart of a general crop stomatal phenotypic extraction method based on in-situ imaging and deep learning. The in-situ non-destructive monitoring terminal constructed by this method can perform cross-species in-situ microscopic monitoring without damaging the physiological state of the leaves. It also improves the semantic segmentation network model to solve the recognition problem caused by the different stomatal morphologies and complex backgrounds of multiple crops, and constructs a phenotypic measurement system to achieve accurate differentiation and quantification between different species and different varieties of the same species. In addition, the trained model is combined with external software to achieve a high degree of automation in the testing process.

[0022] The specific steps of this general method for extracting crop stomatal phenotypes are as follows: S1. The first stage, in-situ microscopic image acquisition and multi-crop dataset construction, mainly includes: using monitoring terminals to collect stomatal images from living leaves of various crops, and performing pixel-level annotation and data augmentation on the stomatal images to construct a stomatal image dataset.

[0023] To overcome the morphological distortion caused by in vitro sampling of various crops in traditional methods, this embodiment employs a specially developed monitoring terminal for multi-species stomatal phenotypes, which collects data through an image acquisition module, a leaf fixation module, and an illumination module. Combined with... Figure 2 The diagram shown illustrates the hardware composition of the monitoring terminal. Figure 2 'a' represents a physical prototype of the monitoring terminal equipment. Figure 2 b is the corresponding engineering design drawing. Figure 2 c represents the main hardware configuration of the image acquisition module. Figure 2 d represents the structural composition of the blade fixing module and the pure white cold light illumination module.

[0024] Researchers used a leaf-fixing module to clamp and secure live leaves of various crops. The module consists of a mesh-like aluminum alloy interlayer with a central rectangular observation window on top, and an acrylic base plate with transparent circular holes below. This double-layered, perforated mechanical clamping design not only prevents scintillation caused by wind or plant vibration but also ensures natural gas exchange and light transmission on the lower epidermis of the leaf. A pure white, cool light source is placed beneath the leaf as an illumination module, providing transmitted light to the live leaf while preventing heat buildup from causing stress-induced closure of the stomata.

[0025] In this embodiment, during the data acquisition process, the types of various crops covered at least six major grain and oil crops: wheat, corn, cotton, soybean, rapeseed, and peanut. For example, it covered 228 varieties of wheat, and each of the other crops covered 10 to 15 varieties. Meanwhile, to significantly improve the environmental adaptability and generalization performance of the trained model, the data acquisition personnel deliberately retained some non-ideal image samples with complex backgrounds such as out-of-focus blur, leaf surface impurities, highlights, and shadows when constructing the stomatal image dataset.

[0026] For each stomatal image, it was imported into the ISAT data annotation tool, where the polygon tool was used to perform extremely fine pixel-level edge annotation of the stomata and pore diameter. After annotation, the effective dataset size was expanded by simultaneously performing geometric enhancement transformations such as random rotation, stretching, and scaling on the original images and label files. Finally, it was strictly divided into training, testing, and validation sets in a 7:2:1 ratio to prepare for subsequent local model training.

[0027] S2. The second stage, the construction and training of the semantic segmentation network model, mainly includes: building a semantic segmentation network model based on the SegFormer architecture, introducing an efficient multi-scale attention module and a deformable convolution module at the front end of the encoder of the semantic segmentation network model; training the semantic segmentation network model through the pore image dataset in step S1, and outputting the segmentation mask of pores and gaps.

[0028] Since the high-precision segmentation of pores and their pores plays a decisive role in the subsequent calculation of area and orientation phenotype, this embodiment, after comprehensively considering the number of model parameters and adaptability to complex backgrounds, selects the SegFormer architecture as the baseline network to construct the semantic segmentation network model, and abandons the conventional structure in the feature encoder part, and performs two deep optimization designs.

[0029] For semantic segmentation network models, firstly, considering that images in stomatal image datasets are highly susceptible to noise interference such as dust particles, shadows from metal edges of leaf clips, and high-gloss reflections from water droplets, an efficient multi-scale attention module (EMA module) is integrated at the front end of the encoder of the semantic segmentation network model. The EMA module works by regularly grouping and reorganizing the channel dimensions of the input features, and independently extracting channel interaction features and local spatial features between different groups using parallel 1×1 and 3×3 convolutional paths. Subsequently, a two-dimensional global average pooling layer is used to aggregate the features across spatial dimensions in both horizontal and vertical directions, ultimately generating a recalibrated multi-scale attention weight map. This mechanism essentially teaches the encoder to "visually focus," adaptively ignoring irrelevant background interference and concentrating entirely on the stomata and the minute textures of their apertures.

[0030] Secondly, addressing the significant geometrical differences between different crops (such as the extremely long, dumbbell-shaped stomata of wheat and the nearly circular, kidney-shaped stomata of soybeans), and the resulting blurring of mask edges, this embodiment immediately connects a third-generation deformable convolutional (DCNv3) module after the EMA module. Traditional convolutional neural networks use a fixed square grid for receptive field sampling, which is difficult to fit the irregular curved edges of biological tissues. However, the deformable convolutional module introduced in this solution allows the network to adaptively learn the spatial offset and modulation scalar of each sampling grid point. This enables the model's receptive field to dynamically "deform" according to the actual curvature and opening / closing state of the stomata in the image, tightly wrapping the irregular boundaries of the stomata like a rubber band, thus completely solving the problems of target edge loss and the adhesion of different instances.

[0031] In this embodiment, the semantic segmentation network model is also equipped with a decoder (e.g., a lightweight ALL-MLP decoder). The decoder is used to receive the output of the encoder and extract multi-scale contextual information through the dilated spatial convolution pooling pyramid module. Finally, the classification and segmentation convolutional layer outputs a pixel-level segmentation mask.

[0032] S3. The third stage involves the construction of a multi-dimensional stomatal phenotypic evaluation index system, which mainly includes: extracting full-dimensional phenotypic indices of stomata based on the segmentation mask in step S2; the full-dimensional phenotypic indices include basic static parameters and advanced morphological parameters.

[0033] Traditional stomatal phenotypic analysis is often limited to counting the number of stomata in the field of view or measuring the length and width of a single stoma, which seriously lacks in-depth analysis of population dynamics and spatial distribution uniformity. Therefore, this embodiment establishes a comprehensive phenotypic evaluation system consisting of 59 parameters based on the high-precision masking results output by the semantic segmentation network model, and analyzes stomata and pores at both the individual and population scales.

[0034] The comprehensive phenotypic evaluation system is divided into two parts: First, there are 40 basic static parameters (i.e., basic parameters). Based on the mask coordinates, the area, perimeter, equivalent diameter, direction, major axis length, minor axis width, and aspect ratio of each pore and aperture in the image are automatically calculated. At the macroscopic group scale, the total number of pores, opening ratio, density, average pore distance, and coefficient of variation in the current field of view are calculated.

[0035] Secondly, to meet the needs of higher-dimensional analysis, this system also constructs 19 advanced morphological parameters (i.e., advanced morphological spatial parameters). For example, graph theory algorithms are used to connect adjacent stomatal centroids to establish a topological network model of stomatal spatial distribution, from which the "Row Alignment Score" is extracted. This index can objectively assess the degree of highly regular arraying of stomata along leaf veins in gramineous crops. In addition, Voronoi spatial distribution polygon maps can be generated based on coordinate points, dividing the epidermal region where each stoma is located into independent and non-overlapping control units. By calculating the "Voronoi polygon cell area" and its spatial homogeneity, the potential physical area controlled by a single stoma for obtaining survival resources (water, carbon exchange) and the intensity of competition among populations can be intuitively and quantitatively assessed. At the same time, by combining dynamic indicators of area change analysis with continuous video frame extraction, the entire phenotypic evaluation system is elevated from geometric measurement of a single static image to dynamic mechanism evaluation with spatiotemporal attributes.

[0036] For example, the specific contents included in the comprehensive phenotypic evaluation system are shown in Table I below: Table I:

[0037] like Figure 4 The diagram shown illustrates the ablation experiment and comparison experiment results of a semantic segmentation network model. Figure 4 A bar chart showing the performance comparison before and after integrating different modules, comparing the performance in four aspects: F1 score, recall, intersection-over-union ratio, and precision. Figure 4 b is a comparison of the training process curves, showing the relationship between the number of training rounds and the accuracy of stomata in different experimental processes; Figure 4 c is the confusion matrix between the baseline model and the semantic segmentation network model, showing the predicted and measured values ​​for background, pores, and voids. To verify the performance of the semantic segmentation network model, this approach compares it with other state-of-the-art (SOTA) networks under the same dataset and experimental conditions. Figure 4The comprehensive results prove that the model of this scheme has achieved the optimal solution and has the highest accuracy in pore and void segmentation.

[0038] S4, the fourth stage, is based on the trained semantic segmentation network model. The model is combined with software (EasyStoma) to achieve automated packaging and deployment of the application, which can quickly measure the pore images to be tested and finally output visualized charts and standard tables.

[0039] To bridge the technical gap between deep learning models and practical agricultural applications, and to facilitate rapid adoption by users, this embodiment also includes a standalone client software with a fully graphical user interface: EasyStoma. (See reference...) Figure 8 The image shown is a schematic diagram of a software interface. Figure 8 a is the main interface layout and functional areas, and users can perform corresponding operations according to the functional instructions on the interface; Figure 8 b is the nearest neighbor distance histogram of the current image; Figure 8 c is the comprehensive phenotypic radar chart; Figure 8 d is the stomatal distribution topology; Figure 8 e is a rose diagram showing the orientation of the stomata; Figure 8 f is the Voronoi diagram, which shows the spatial distribution of the Thiessen polygons.

[0040] Researchers using EasyStoma software daily require no programming background. They simply open the main interface and use the "1. Open Folder" module to batch select and load folders containing in-situ images of various formats. They can then customize the contrast colors for stomatal and pore size mask rendering using the color palette in the "2. Configure Color" module. Next, by selecting the edge and minimal area filtering functions in the "3. Preprocessing Settings" module, the system automatically scans the edges of the stomatal images before analysis and removes incomplete stomatal halves truncated due to the microscope's edge field of view. This prevents these incomplete geometric fragments from lowering or contaminating the overall area statistical mean.

[0041] After setting the parameters, simply click the "4. Batch Analysis" button on the main interface, and the software will immediately mount and call the pre-trained EasyStomaFormer model. This module adopts an efficient multi-threaded concurrent architecture, automatically reading images in the folder one by one. It not only completes high-precision instance segmentation mask output in the background, but also simultaneously extracts and calculates all 59 basic and advanced phenotypic indicators mentioned above. All calculations are completed silently and at high speed in the background. After the analysis task is completed, researchers can switch between different drop-down menus in the "5. Data Analysis" data visualization module on the right side of the interface to directly review the nearest neighbor distance histogram of the current image, the comprehensive radar chart comparing phenotypic differences between different crop varieties, and the rose diagram and Voronoi spatial boundary map that intuitively show the physical orientation of stomata, which are generated in real time by the software. Finally, the software will export the summary results, which include all detailed stomatal measurement data, population statistical mean, and advanced graph theory indicators, along with intuitive images plotted with segmented color blocks, into standardized Excel spreadsheet files in one click. This completes the fully automated closed-loop analysis process from in-situ image capture to multidimensional phenotypic data storage.

[0042] In this embodiment, experiments were also conducted on six major grain and oil crops: wheat, corn, cotton, soybean, rapeseed, and peanut. The experiments covered 228 varieties of wheat, and 10 to 15 varieties for each of the other crops. A total of 9,720 stomatal images were collected, imported into the ISAT data annotation tool for annotation and data augmentation, and then expanded to 38,800 images to form a stomatal image dataset. These datasets were then divided into training, testing, and validation sections in a 7:2:1 ratio.

[0043] All training experiments for the semantic segmentation model were conducted on a Windows 11 64-bit environment. The deep learning framework used was PyTorch, and the hardware configuration consisted of an AMD Ryzen 7 9700X CPU, 64 GB of RAM, and an NVIDIA RTX 4080 Super graphics card with 16 GB of VRAM. CUDA deep acceleration was enabled throughout the training process. The AdamW optimizer was used during training to ensure model convergence stability. The initial learning rate was finely set to 6e-5, and a decay factor of 0.8 was introduced to fine-tune the network weights in the later stages of training. The batch size was set to 8, and the model underwent a total of 200 epochs of training.

[0044] Figure 3 This is a schematic diagram illustrating data extraction, phenotypic analysis, and software workflow. Figure 3 It demonstrates the image segmentation and phenotypic data extraction for six major grain and oil crops, and showcases the usage process of the corresponding software. Figure 3 'a' involves inputting the in-situ stomatal image into the ESF model for segmentation. Figure 3 b is the generated pore segmentation mask. Figure 3 c is the calculation of the length and width of the pores and the pore gaps (taking an open pore as an example). Figure 3 d is an example of calculating the Thiessen polygon (Voronoi diagram) and pore spacing. Figure 3 "e" is the main interface of the EasyStoma software. Figure 3 f represents the complete workflow for stomatal phenotypic data extraction and analysis. Figure 5 A comprehensive quantitative analysis diagram of stomatal morphology, spatial arrangement patterns, and dynamic behavior of six major grain and oil crops; Figure 5 'a' represents a comparison between the stomatal length-to-width ratio (LW Ratio) and the stomatal angle variation coefficient (Angle CV). Figure 5 b represents the correlation between the average stomatal area and the overall stomatal density. Figure 5 c is a visualization of advanced stomatal spatial indices, including stomatal masks and their corresponding topological networks and the spatial distribution of Voronoi polygons. Figure 5 d represents a comparison of nine advanced spatial morphology indicators. The data are uniformly normalized to [0, 1], and the gray dots represent the average value. Figure 5 e represents the dynamic change in the percentage of average stomatal area during stomatal opening and closing; the shaded area represents the range of change. Based on Figure 5 Comprehensive quantitative analysis can compare the length, width, area, and number of pores and voids with actual manual measurements. The comparison results are as follows: Figure 6 As shown, in terms of length, width, area, and quantity, the recognition performance of the semantic segmentation model in this scheme is generally very close to the measured values, verifying the effectiveness and accuracy of this scheme.

[0045] Furthermore, it should be noted that the semantic segmentation network model trained based on the general crop stomatal phenotype extraction method of this scheme can also be used to segment and identify other crops that did not participate in the model training. (See reference...) Figure 7 The diagram illustrates the segmentation performance verification of the semantic segmentation network model on various zero-sample plants. It showcases the segmentation results of this semantic segmentation network model on eight other crops with zero samples, demonstrating high-precision stomatal extraction even in zero-sample scenarios. The eight crops are Chrysanthemum, Cucumber, Arabidopsis, Tobacco, Rice, Tomato, Liriodendron, and Sesbania.

[0046] In summary, this invention constructs an in-situ non-destructive monitoring terminal, overcoming the shortcomings of traditional destructive sampling. It enables cross-species in-situ microscopic monitoring without damaging the physiological state of leaves. Secondly, by combining an efficient multi-scale attention module and a deformable convolution module in the semantic segmentation network model, it solves the recognition challenges posed by the diverse stomatal morphologies and complex backgrounds of various crops, achieving extremely high segmentation accuracy and generalization. Furthermore, based on the extraction of basic static parameters, it innovatively incorporates several novel advanced spatial indicators such as Voronoi cell area and column alignment score, constructing a comprehensive and multi-dimensional phenotypic measurement system that effectively achieves accurate differentiation and quantification between different species and different varieties of the same species. In addition, the trained model, combined with supporting software, automates the process from image import and model inference to chart drawing and phenotypic indicator export, greatly reducing the technical threshold of phenomics analysis and improving the data conversion efficiency of stomatal research, demonstrating significant progress.

[0047] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A general method for extracting crop stomatal phenotypes based on in-situ imaging and deep learning, characterized in that, Includes the following steps: S1. Use monitoring terminals to collect stomatal images from living leaves of various crops, and perform pixel-level annotation and data augmentation on the stomatal images to construct a stomatal image dataset. S2. Construct a semantic segmentation network model based on the SegFormer architecture, and introduce an efficient multi-scale attention module and a deformable convolution module into the encoder front end of the semantic segmentation network model; train the semantic segmentation network model with the pore image dataset in step S1, and output the segmentation mask of pores and gaps. S3. Based on the segmentation mask in step S2, extract the full-dimensional phenotypic indices of stomata; the full-dimensional phenotypic indices include the following parameter types: basic static parameters and advanced morphological parameters. S4. Input the image of the stomata to be tested into the trained semantic segmentation network model, and output the corresponding visualization charts and standard tables according to the parameter types in step S3.

2. The method according to claim 1, characterized in that, In step S1, the monitoring terminal includes: an image acquisition module, a leaf fixing module, and an illumination module; the image acquisition module is used to capture images of the stomata of any living leaf; the leaf fixing module uses a double-layer clamping mechanism to fix living leaves of various crops, the upper layer of the double-layer clamping mechanism is a grid-shaped aluminum alloy plate with a central observation window, and the lower layer of the double-layer clamping mechanism is an acrylic plate with transparent round holes; the illumination module uses a pure white cold light source located below the leaf fixing module to provide transmitted light to the living leaves.

3. The method according to claim 1, characterized in that, The various crops in step S1 include at least wheat, corn, cotton, soybean, rapeseed, and peanut; stomatal images include still images and video clips.

4. The method according to claim 1, characterized in that, In step S2, the efficient multi-scale attention module is used to group and reorganize the channel dimensions of the input features of the semantic segmentation network model, extract channel interaction features and local spatial features using parallel convolutional paths, and aggregate across spatial dimensions in the horizontal and vertical directions using two-dimensional global average pooling to generate a multi-scale attention weight map.

5. The method according to claim 1, characterized in that, In step S2, the deformable convolution module is connected in series with the efficient multi-scale attention module; the deformable convolution module adaptively learns the sampling offset and modulation scalar to fit the receptive field to the irregular boundary of the pore.

6. The method according to claim 1, characterized in that, The semantic segmentation network model in step S2 also includes a decoder. The decoder receives the output of the encoder and extracts multi-scale contextual information through the dilated spatial convolutional pooling pyramid module. Finally, the classification and segmentation convolutional layer outputs a pixel-level segmentation mask.

7. The method according to claim 1, characterized in that, In step S3, the basic static parameters include at least the area, perimeter, length, width, aspect ratio, equivalent diameter and orientation of pores or voids at the individual scale, and the total number of pores, open ratio, density, average pore distance and coefficient of variation at the population scale.

8. The method according to claim 1, characterized in that, In step S3, the advanced morphological parameters include at least: the roundness, eccentricity, and aperture of stomata or pores at the individual scale, and the Thiessen polygon area, Thiessen polygon uniformity, coordination number, cell area, and homogeneity of stomata or pores at the population scale.

9. The method according to claim 1, characterized in that, Before inputting the image of the pore to be tested into the trained semantic segmentation network model in step S4, a preprocessing step is also included: scanning the edges of the image of the pore to be tested, filtering out truncated pores and tiny noise points with an area smaller than a preset threshold.

10. The method according to claim 1, characterized in that, In step S4, the visualization charts include: nearest neighbor distance histogram, comprehensive phenotypic radar chart, rose diagram showing stomatal orientation, stomatal distribution topology network diagram, and Thiessen polygon spatial distribution diagram.

Images