Greenhouse plant growth state recognition method and planting management and control system

Abstract

The invention relates to the technical field of machine learning, in particular to a greenhouse plant growth state recognition method and a planting management and control system.The greenhouse plant growth state recognition method comprises the steps that firstly, a greenhouse plant image is obtained, brightness and chromaticity separation is conducted on the greenhouse plant image, gradient enhancement is conducted on a brightness feature subspace, and color difference processing is conducted on a chromaticity feature subspace; the method comprises the following steps: firstly, processing a brightness feature subspace, secondly, performing weighted fusion on the processed brightness feature subspace, chromaticity feature subspace and an original greenhouse plant image to obtain an enhanced image of the greenhouse plant, finally, inputting the enhanced image of the greenhouse plant into a preset growth state detection model, and finally, outputting an identification result. The comprehensive tracking and recording of the plant growth state are realized, the plant planting scheme is intelligently decided, and the efficiency and scientificity of planting management are improved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method for identifying the growth status of greenhouse plants and a planting control system. Background Technology

[0002] As modern agriculture develops towards digitalization and intelligence, greenhouse cultivation has gradually become an important mode of efficient agricultural production. The greenhouse environment can provide a relatively stable environment for plant growth by artificially controlling conditions such as light, temperature, humidity, and carbon dioxide concentration, thereby improving crop yield and quality. However, in actual planting, the growth status of plants is affected by a combination of factors, such as fluctuations in environmental parameters, pest and disease infestations, nutritional imbalances, and light shading, which can still lead to uneven growth, reduced yield, or even plant death. In recent years, computer vision and deep learning technologies have been widely applied in the field of smart agriculture. By collecting image data of greenhouse plants and combining feature extraction, pattern recognition, and classification models, automatic identification of plant leaf health status, disease characteristics, light stress, and nutritional abnormalities can be achieved. However, most existing methods rely on traditional RGB images or direct convolutional feature extraction, which are easily affected by uneven lighting, complex backgrounds, and leaf shading, resulting in insufficient capture of subtle texture features, color differences, and edge structures. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a method for identifying the growth status of greenhouse plants and a planting control system, which enables comprehensive tracking and recording of plant growth status, intelligent decision-making on plant planting plans, improves the efficiency and scientific nature of planting management, and promotes the improvement of product yield and quality.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A method for identifying the growth status of greenhouse plants, the method comprising:

[0006] Acquire image data of the greenhouse plants;

[0007] The image data is input into a preset growth state detection model, and the recognition result is output through the growth state detection model.

[0008] The main structure of the growth state detection model is the YOLO target detection model, which includes a high-resolution detection layer used to extract target feature information and spatial feature information.

[0009] The high-resolution detection layer also includes a prediction head, which is obtained through a Swin-Transformer structure. The prediction head is used to identify a portion of the image data through a self-attention mechanism, extracting leaf features, flower features, fruit features, and spatial feature information. The portion of the image data is a region of interest, which is obtained by the high-resolution detection layer through response filtering of the feature map of the image data.

[0010] The Neck portion of the YOLO object detection model also includes a GS-ELAN structure constructed using GSConv, which is used to reduce noise in the output of the high-resolution detection layer.

[0011] Methods for identifying the growth status of greenhouse plants also include:

[0012] The image data of the greenhouse plants are separated into luminance and chrominance to obtain a luminance feature subspace and a chrominance feature subspace. Gradient enhancement is performed on the luminance feature subspace, and color difference processing is performed on the chrominance feature subspace.

[0013] The processed luminance feature subspace, chrominance feature subspace and the original greenhouse plant image data are weighted and fused to obtain the greenhouse plant image enhancement data;

[0014] The image enhancement data of the greenhouse plants is input into a preset growth status detection model, and the recognition result is output through the growth status detection model.

[0015] The image data of the greenhouse plants are subjected to luminance and chrominance separation to obtain luminance feature subspace and chrominance feature subspace, including:

[0016] The image data of the greenhouse plants are subjected to noise suppression processing;

[0017] The processed greenhouse plant image data is decomposed into a luminance feature subspace and a chrominance feature subspace using a reversible mapping strategy. The luminance feature subspace is used to preserve the spatial structure information of the greenhouse plant image data, including texture and edge features, while the chrominance feature subspace is used to characterize the color distribution information of the greenhouse plant image data, including red-green hue components and yellow-blue hue components.

[0018] Gradient enhancement of the brightness feature subspace includes:

[0019] Calculate the gradient magnitude matrix and gradient direction matrix based on the brightness feature subspace;

[0020] Based on the gradient magnitude matrix and gradient direction matrix, the neighborhood of the gradient magnitude is determined, and the average, maximum, and minimum values ​​of all gradient magnitudes within the neighborhood are calculated. A semi-dilation matrix and a semi-erosion matrix are then constructed. When constructing the semi-dilation matrix, if the gradient magnitude is greater than or equal to the average value, it is replaced with the maximum value within the neighborhood; if the gradient magnitude is less than the average value, it is replaced with the average value within the neighborhood. Similarly, when constructing the semi-erosion matrix, if the gradient magnitude is less than or equal to the average value, it is replaced with the minimum value within the neighborhood; if the gradient magnitude is greater than the average value, it is replaced with the average value within the neighborhood. The morphological gradient matrix is ​​obtained by comparing the semi-dilation matrix and the semi-erosion matrix.

[0021] The morphological gradient matrix is ​​normalized and used as a weight matrix to be fused with the gradient magnitude matrix to obtain the gradient enhancement matrix.

[0022] Numerical mapping is performed on the gradient enhancement matrix to map the gradient enhancement matrix values ​​to the range of image pixel values, resulting in the processed brightness feature subspace.

[0023] The neighborhood of each gradient magnitude is designed as an elliptical structure, where the minor axis of the ellipse is along the gradient direction corresponding to the gradient magnitude. The range of the elliptical neighborhood of each gradient magnitude is adjusted according to the gradient direction matrix, and the values ​​in the gradient magnitude matrix are replaced according to the elliptical neighborhood.

[0024] Color difference processing is performed on the chromaticity feature subspace, including:

[0025] The color vector of each pixel in the chromaticity feature subspace is mapped to a high-dimensional feature space through a kernel function. The color vector is represented by red-green chromaticity components and yellow-blue chromaticity components.

[0026] Calculate the similarity of pixels in the high-dimensional feature space, construct a global similarity matrix, and cluster the pixels in the high-dimensional feature space according to the global similarity matrix to form color clusters;

[0027] For intra-cluster weights, the enhancement weights are calculated and set based on the distance between the pixel and the cluster center of the color cluster. For inter-cluster weights, the enhancement weights are calculated and set based on the distance between the cluster centers of the color clusters. The weight matrix is ​​obtained based on the intra-cluster weights and inter-cluster weights.

[0028] The high-dimensional feature space is weighted as a whole according to the weight matrix to achieve color difference processing and obtain a color-enhanced high-dimensional feature space.

[0029] The high-dimensional feature space of color enhancement is reduced to obtain an enhanced two-dimensional feature space, which is then mapped to the range of image pixel values ​​to obtain the processed chromaticity feature subspace.

[0030] A greenhouse plant cultivation and management system, the system comprising a plant digital monitoring module, a knowledge graph construction module, a plant growth status identification module, and an intelligent management decision-making module;

[0031] The plant digital monitoring module collects plant growth environment data in real time through multiple types of sensors and monitors plant growth status in real time by installing network camera equipment in the demonstration park.

[0032] The knowledge graph construction module integrates historical planting data and morphological and environmental data throughout the entire growth cycle of plants, and constructs a plant growth knowledge graph through modular construction, dynamic reasoning, and multi-source data fusion.

[0033] The plant growth status recognition module preprocesses the image data of greenhouse plants, inputs the processed plant images into a preset growth status detection model, and outputs the current growth status of the plants.

[0034] The intelligent management decision-making module formulates decision rules based on the plant growth knowledge graph and the real-time growth status of plants, and outputs plant planting decision-making plans based on the current growth status of plants and the decision rules.

[0035] The knowledge graph construction module includes a historical data processing unit, an entity-attribute-relationship mapping unit, a knowledge storage unit, and a knowledge iteration and update unit;

[0036] The historical data processing unit extracts and integrates plant surface feature data and environmental sensor data to form the basic dataset for knowledge graph construction.

[0037] The entity-attribute-relationship mapping unit extracts core elements of plant knowledge from the basic dataset through a deep learning model, including entities, entity attributes, and relationships between entities.

[0038] The knowledge storage unit uses the Neo4j graph database to modularly store the extracted entities, attributes and relationships, and establishes hierarchical relationships between entities through Cypher statements to realize entity query and semantic reasoning.

[0039] The knowledge iteration and update unit receives newly collected plant data and updates entities, attributes, or relationships.

[0040] The intelligent management decision-making module includes a rule reasoning and execution unit, an association rule mining unit, and a decision scheme generation unit;

[0041] The rule-based reasoning execution unit transforms the preset plant planting decision rules into computer-executable logic, and uses a rule-based reasoning algorithm to match and reason about the input plant growth status to generate preliminary decision results.

[0042] The association rule mining unit uses the Apriori association rule mining algorithm to mine hidden growth relationships from historical planting data and incorporates them as supplementary rules into the decision-making model to improve decision accuracy.

[0043] The decision-making scheme generation unit combines the results of rule-based reasoning with supplementary association rules to output decision-making schemes that include suggestions for agricultural operations, anomaly warnings, plant protection, water and fertilizer management, and harvesting.

[0044] Compared with existing technologies, the proposed solution first enhances the original images of greenhouse plants. By introducing improvements in image gradient enhancement and color difference optimization, it not only effectively improves the overall clarity and detail of the images, but also fully preserves the subtle texture features and color differences in the images. This allows the growth status detection model to obtain more comprehensive and accurate input information during the analysis process, thereby demonstrating higher accuracy and stability in identifying plant morphological features, judging growth status, and monitoring abnormal changes. Attached Figure Description

[0045] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0046] Figure 1 This is a flowchart illustrating the greenhouse plant growth status identification method of Embodiment 1 of the present invention;

[0047] Figure 2 This is a diagram showing the generation of the morphological gradient matrix in Embodiment 1 of the present invention;

[0048] Figure 3 This is a schematic diagram of the elliptical neighborhood in Embodiment 1 of the present invention;

[0049] Figure 4 This is a system module diagram of the greenhouse plant planting and management system of Embodiment 2 of the present invention. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0051] Example 1

[0052] Please see Figure 1 The present invention provides an embodiment of a method for identifying the growth status of greenhouse plants, the specific steps of which are as follows:

[0053] S1: Acquire image data of the greenhouse plants;

[0054] In this embodiment, high-definition network cameras are deployed in the greenhouses and key areas of the demonstration park to achieve real-time monitoring of strawberry growth and facility operation. Each greenhouse is equipped with a high-definition PTZ camera to monitor the leaves and flowers of the plants, and a high-definition bullet camera with an intelligent sliding rail to monitor the shape, quantity, and number of deformed fruits. Through this video monitoring deployment, not only can the crops in the greenhouse be monitored in real time around the clock, but the growth status of each plant can also be continuously tracked and recorded.

[0055] Specifically, this application uses monitoring equipment deployed in the greenhouse environment to collect images in real time, and extracts key frame images from the continuous video stream according to the set time interval or specific event triggering conditions, so as to ensure that the acquired images can fully reflect the state changes of plants at different growth stages.

[0056] S2: Separate the brightness and chromaticity of the image data of the greenhouse plants to obtain a brightness feature subspace and a chromaticity feature subspace. Perform gradient enhancement processing on the brightness feature subspace and color difference processing on the chromaticity feature subspace.

[0057] In this embodiment, noise suppression processing is performed on the image data of the greenhouse plants to reduce the impact of noise and environment on texture and edge information. Then, the image is decomposed into a brightness feature subspace and a chromaticity feature subspace through a reversible mapping strategy, and enhancement processing is performed on each. For the brightness feature subspace, the gradient magnitude matrix and gradient direction matrix are calculated, an elliptical neighborhood is designed based on the gradient direction, a semi-dilation matrix and a semi-erosion matrix are constructed to obtain the morphological gradient matrix, which is then normalized and finally fused to obtain the gradient enhancement matrix. The processed brightness feature subspace is then obtained through numerical mapping. For the chromaticity feature subspace, the color vector is mapped to a high-dimensional feature space through a kernel function, the similarity between pixels is calculated and a global similarity matrix is ​​constructed, color clusters are obtained through clustering, and the high-dimensional feature space is weighted as a whole by combining intra-cluster and inter-cluster weights to achieve color difference processing. Finally, the dimensionality is reduced and mapped back to the two-dimensional chromaticity feature subspace and mapped to the pixel value range to obtain the processed chromaticity feature subspace.

[0058] Furthermore, gradient enhancement in the brightness feature subspace can effectively highlight leaf texture, vein structure, and edge contours, making fine structural features more obvious. Color difference enhancement in the chromaticity feature subspace can highlight color changes, lesions, or nutritional abnormalities in plant leaves, improving color distinguishability. Overall, enhancing brightness and chromaticity separately not only preserves the structural and color information of the image but also strengthens key features, significantly improving the enhanced greenhouse plant image in terms of detail expression, edge clarity, and color distinguishability.

[0059] S3: The processed luminance feature subspace, chrominance feature subspace and the original greenhouse plant image data are weighted and fused to obtain the greenhouse plant image enhancement data;

[0060] In this embodiment, the luminance feature subspace can highlight the texture structure and detail differences in greenhouse plant images, while the chroma feature subspace can better reflect the color distribution and color changes of plants. By processing these two types of features, not only can the image's expressiveness at the detail level be enhanced, but the integrity of color information can also be effectively preserved. Then, a reversible mapping strategy is used to inversely map these two feature subspaces back to RGB representation, thereby realizing the recombination and restoration of structural and color information. The inversely mapped image is then weighted and fused with the original greenhouse plant image data. This fusion process comprehensively considers texture details, color differences, and overall visual effects, so that the generated enhanced image maintains realism while significantly improving information expression capabilities.

[0061] S4: Input the image enhancement data of the greenhouse plants into a preset growth state detection model, and output the recognition result through the growth state detection model. The main structure of the growth state detection model is a YOLO target detection model. The YOLO target detection model includes a high-resolution detection layer, which is used to extract target feature information and spatial feature information. The high-resolution detection layer also includes a prediction head, which is obtained through a Swin-Transformer structure. The prediction head is used to identify a part of the image data through a self-attention mechanism, and extract detailed features and spatial feature information. The part of the image data is the region of interest, which is obtained by the high-resolution detection layer through response filtering of the feature map of the image data. The Neck part of the YOLO target detection model also includes a GS-ELAN structure constructed through GSConv. The GS-ELAN structure is used to reduce the noise of the output of the high-resolution detection layer.

[0062] In this embodiment, the LabelImg tool is used to manually label the leaves, flowers, fruits, and water exudation of strawberries under abnormal conditions, constructing a standardized dataset, which is then divided into a training set, a validation set, and a test set, providing reliable data support for the training and optimization of the algorithm model.

[0063] Specifically, in order to accurately identify the abnormal growth state of strawberries, and to address issues such as small target size, dense overlapping positions, and similar target and background colors, the recognition algorithm is based on the YOLO target detection model. A high-resolution detection layer is built in the shallow layer. Starting from the feature integrity and spatial information richness of small targets, features are extracted and predicted from the high-resolution feature map, enhancing the expressive power of the feature map and significantly improving the model's detection performance for small targets.

[0064] Furthermore, a Swin-Transformer prediction head is introduced into the high-resolution detection layer to perform self-attention recognition on the region of interest, effectively extracting rich details and spatial features, reducing the interference of dense target overlap and color similarity occlusion on the recognition results, thereby improving the robustness of the model in complex scenes.

[0065] Furthermore, GSConv is introduced in the Neck section to construct a GS-ELAN structure, which is used to suppress shallow noise caused by the high-resolution detection layer. At the same time, it alleviates the problem of increased parameter quantity introduced by the high-resolution detection layer and the Swin-Transformer prediction head. While ensuring improved detection accuracy, it effectively reduces the number of model parameters and computational complexity. The combination of the above improvements with the YOLO model significantly improves the overall performance and reliability of the model in detecting abnormal growth status in strawberries.

[0066] The specific steps for S2 are as follows:

[0067] S2.1: Perform luminance and chrominance separation on the image data of the greenhouse plants to obtain luminance feature subspace and chrominance feature subspace. The specific steps are as follows:

[0068] S2.1.1: Perform noise suppression processing on the image data of the greenhouse plants;

[0069] Specifically, in the implementation process, various image denoising methods can be adopted, such as filtering-based temporal smoothing methods and spatial convolution processing methods, or frequency-domain wavelet transform methods and adaptive threshold suppression strategies. For continuous image frames extracted from video, the influence of random noise can be further reduced by combining temporal correlation and using multi-frame information fusion methods, thereby effectively suppressing background noise and environmental interference while preserving image details.

[0070] Furthermore, image acquisition in greenhouse environments is often affected by factors such as uneven lighting, environmental dust, humidity fluctuations, and sensor imaging defects, which can easily lead to random noise and structural artifacts in the images. This noise not only masks the subtle texture features of plants but also weakens the expression of color differences, thereby interfering with subsequent image enhancement and feature extraction processes. If left untreated, it will directly affect the accuracy and stability of the plant growth status detection model. Through effective noise suppression processing, the signal-to-noise ratio can be significantly improved while ensuring image clarity, allowing the leaf contours, texture structures, and color information of plants to be preserved more realistically and completely.

[0071] S2.1.2: Through a reversible mapping strategy, the processed greenhouse plant image data is decomposed into a luminance feature subspace and a chrominance feature subspace. The luminance feature subspace is used to retain the spatial structure information of the greenhouse plant image data, including texture and edge features. The chrominance feature subspace is used to characterize the color distribution information of the greenhouse plant image data, including red-green hue components and yellow-blue hue components.

[0072] Specifically, reversible mapping refers to a mapping method that can both convert the original image data into a feature subspace and completely recover the original image from the feature subspace. Unlike traditional unidirectional mapping, reversible mapping maintains information without loss during the transformation process, ensuring consistency between decomposition and reconstruction.

[0073] Furthermore, using a reversible mapping strategy, the noise-suppressed greenhouse plant image data is decomposed into a luminance feature subspace and a chromaticity feature subspace. The luminance component is typically calculated based on a weighted combination of the original RGB three channels. The human eye is most sensitive to green, followed by red, and then blue. The luminance feature subspace mainly preserves the overall light intensity distribution and spatial structure features of the image, including the texture details, vein structure, and edge contour information of plant leaves. The chromaticity component is obtained by converting redundant information in the RGB space into color contrast information, typically including two components: red-green difference and yellow-blue difference. The chromaticity feature subspace is used to characterize the distribution and differences in plant colors, such as yellowing of leaves, red spots, or other color anomalies. Through this decomposition method, the structural features and color features of the image are effectively distinguished, facilitating independent processing and flexible fusion in subsequent stages.

[0074] Furthermore, by adopting a reversible mapping strategy, on the one hand, it can avoid mutual interference between different features, so that texture, edges and colors can be fully preserved during the enhancement process. On the other hand, reversibility ensures the traceability of the processing process and the stability of the results, avoiding the quality degradation caused by information loss. The resulting image enhancement data is clearer and more natural in visual effect, and at the same time provides high-quality input for plant growth status detection, which helps to improve the accuracy and robustness of the detection model.

[0075] S2.2: Perform gradient enhancement processing on the brightness feature subspace, the specific steps of which are as follows:

[0076] S2.2.1: Calculate the gradient magnitude matrix and gradient direction matrix based on the brightness feature subspace;

[0077] Specifically, the calculation methods for gradient magnitude and gradient direction are based on pixel-level operations in the brightness feature subspace. Operators such as Sobel, Prewitt, or Scharr are typically used to calculate the gradient components in the horizontal and vertical directions of the image, respectively. Then, the gradient magnitude matrix and gradient direction matrix are obtained from the horizontal and vertical gradient components. The specific calculation formulas are as follows:

[0078] ;

[0079] ;

[0080] in, It is the gradient component of the image pixel (x,y) in the horizontal direction. It is the gradient component of the image pixel (x,y) in the vertical direction. It is the gradient magnitude of the image pixel (x, y), representing the strength of the brightness change. It represents the gradient direction of an image pixel (x,y), indicating the direction of the fastest change in brightness.

[0081] Furthermore, gradient magnitude and gradient direction are important features for describing local gray-level changes in an image. Gradient magnitude reflects the intensity of pixel brightness changes and is used to highlight the texture and edge structure of the image, making subtle features in the image more obvious. Gradient direction represents the direction of the fastest gray-level change, i.e. the normal direction of the edge, providing directional information for subsequent semi-dilation, semi-erosion matrix construction and adaptive neighborhood design. By combining magnitude and direction, the local intensity and structural orientation of the image can be grasped simultaneously.

[0082] S2.2.2: Based on the gradient magnitude matrix and gradient direction matrix, determine the neighborhood of the gradient magnitude and calculate the average, maximum, and minimum values ​​of all gradient magnitudes within the neighborhood. Construct a semi-dilation matrix and a semi-erosion matrix. When constructing the semi-dilation matrix, if the gradient magnitude is greater than or equal to the average value, replace the gradient magnitude with the maximum value within the neighborhood; if the gradient magnitude is less than the average value, replace the gradient magnitude with the average value within the neighborhood. This yields the semi-dilation matrix. When constructing the semi-erosion matrix, if the gradient magnitude is less than or equal to the average value, replace the gradient magnitude with the minimum value within the neighborhood; if the gradient magnitude is greater than the average value, replace the gradient magnitude with the average value within the neighborhood. This yields the semi-erosion matrix. The morphological gradient matrix is ​​obtained by the difference between the semi-dilation matrix and the semi-erosion matrix.

[0083] Specifically, the core function of the morphological gradient matrix is ​​to highlight the edges and details in the image. Since the dilation operation can enhance the area with high gradient value, while the erosion operation can weaken the area with low gradient value, the subtraction of the two can significantly enhance the gray-scale changes at the boundary, thereby making the edge contour clearer. In greenhouse plant image processing, the morphological gradient matrix can highlight leaf veins, edge contours and local structural features.

[0084] Furthermore, for greenhouse plants such as strawberries, the fruit surface has a large number of tiny seeds, the leaf veins are fine and the edges are serrated, and the fruit itself is small and easily obscured. If the traditional dilation operation is used directly, all gradient values ​​in the neighborhood will be replaced with the maximum value, resulting in excessive expansion of local edges. This will cause the seeds of the fruit to be magnified, overlapped or even merged in the image, the surface features of the fruit will be lost, and the details of the leaf veins will be blurred due to the dilation, making the originally clear serrated edges smooth and indistinct. The traditional erosion operation, on the other hand, replaces the local gradient values ​​with the minimum value, which weakens or even eliminates the soft transition areas of the petals or the tiny textures on the surface of the fruit.

[0085] This method employs a semi-dilation matrix and a semi-erosion matrix to enhance significant edges while avoiding excessive edge expansion and loss of local features. The semi-dilation and semi-erosion matrices selectively enhance local gradients by referencing the neighborhood average, achieving a balance between edge expansion and detail preservation. This selective enhancement maintains the independence of fruit seed points, highlights leaf vein details, preserves soft petal edges, and avoids edges merging or blurring, thus better adapting to the image characteristics of strawberry fruits that are small, numerous, and easily occluded.

[0086] Furthermore, the generation of the morphological gradient matrix first calculates the average value of all gradient magnitudes in the neighborhood to achieve a balance between edge enhancement and weak texture suppression. The average value, as a representative of the local overall level, can be used to determine whether the current pixel belongs to a prominent edge region or a smooth region. When the gradient magnitude is greater than or equal to the average value, it is replaced with the maximum value to enhance prominent edges. When the gradient magnitude is less than the average value, it is replaced with the average value to smooth the background and weak textures. This mechanism effectively reduces the situation where noise points are misjudged as edges. Similarly, for the construction of the semi-erosion matrix, the minimum value in the neighborhood is used as a reference, combined with the average value for conditional replacement, thereby avoiding the loss of details caused by excessive erosion and improving the robustness of edge detection. This ensures that the morphological gradient matrix has both high sensitivity and good stability.

[0087] Please refer to Figure 2 , Figure 2 Generate a graph for the morphological gradient matrix. Figure 2An elliptical neighborhood is defined with all pixel gradient directions pointing to the right. Each pixel coordinate in the matrix is ​​numbered, with the first row labeled (1,1), (1,2), (1,3), (1,4), and so on. Taking pixel (1,1) as an example, its elliptical neighborhood includes the coordinates (1,1), (1,2), (2,1), (2,2), and (3,1), corresponding to gradient magnitudes of 3, 4, 1, 2, and 3 respectively. The average gradient magnitude of each pixel within the elliptical neighborhood is calculated to be 2.6. Comparing the gradient magnitude of pixel (1,1) with the average, since its gradient magnitude is greater than the average, the gradient magnitude is replaced with the maximum value within the elliptical neighborhood. The maximum gradient magnitude within the elliptical neighborhood is 4. The value 4 is used as the semi-dilation value of pixel (1,1) in the semi-dilation matrix. For the construction of the semi-erosion matrix, since the gradient magnitude of pixel (1,1) is greater than the average value, the gradient magnitude is replaced with the average magnitude of the elliptical neighborhood. That is, the average magnitude of 2.6 is used as the semi-erosion value of pixel (1,1) in the semi-erosion matrix. All gradient magnitudes in the gradient magnitude matrix are calculated in the above manner to obtain the semi-dilation matrix and the semi-erosion matrix. Then, the morphological gradient matrix is ​​calculated by the difference between the semi-dilation matrix and the semi-erosion matrix. Taking pixel (1,1) as an example, the value of the semi-dilation matrix corresponding to pixel (1,1) is 4, and the value of the semi-erosion matrix is ​​2.6. The difference is used to obtain the value of the morphological gradient matrix corresponding to pixel (1,1).

[0088] S2.2.2.1: Design the neighborhood of each gradient magnitude as an elliptical structure, where the minor axis of the ellipse is along the gradient direction corresponding to the gradient magnitude. Adjust the elliptical neighborhood range of each gradient magnitude according to the gradient direction matrix, and perform a replacement operation on the values ​​in the gradient magnitude matrix according to the elliptical neighborhood.

[0089] Specifically, taking strawberries as an example of greenhouse plants, strawberry leaves have complex vein patterns, mostly radial or reticulate, with obvious directional differences between the main vein and lateral veins. The leaf edges are serrated, and the petal edges are mostly rounded. The overall volume of strawberry fruit is relatively small, exhibiting a typical obovate or nearly heart-shaped shape, with a pointed upper end and gradually becoming fuller at the lower end. These morphological characteristics give the strawberry's edge contour a clear directional visual appearance, and also make the top and bottom of the fruit have obvious differences in curvature and structure. The elliptical structure is chosen in the neighborhood design, rather than the traditional rectangular or circular structure, because the elliptical structure can more flexibly adapt to the directional characteristics of the gradient distribution in the image. If a circular neighborhood is used, it samples consistently in all directions, which can easily introduce redundant information, causing the strawberry leaf veins to become blurred or the edge details to be masked. The sharp features at the top of the fruit are smoothed, resulting in the loss of sharp features. The elliptical neighborhood can set different scales in the main direction and vertical direction, enabling differentiated processing between the main gradient direction and the vertical direction, making the edge detection results more consistent with the true distribution of the strawberry image.

[0090] Please refer to Figure 3 , Figure 3 This diagram illustrates an elliptical neighborhood. Taking the orange pixel as an example, first, based on its position in the gradient magnitude matrix, the corresponding gradient direction is found in the gradient direction matrix. The minor axis of the elliptical neighborhood is set to 3 pixels, and the major axis to 5 pixels. Then, the minor axis is arranged along the corresponding gradient direction, i.e., the direction of the red line in the diagram. Since image edges are usually perpendicular to the gradient direction, and the major axis is perpendicular to the minor axis, the major axis represents the direction of the image edge, i.e., the direction of the green line in the diagram. When a pixel is located within the elliptical coverage area and its coverage area exceeds half, it is determined that the pixel belongs to the elliptical neighborhood. The gray area in the diagram represents the orange pixel. The elliptical neighborhood range of a point is calculated using only the orange pixel as a simple example. The neighborhood ranges of all other pixels are calculated in the same way. Arranging the minor axis of the ellipse along the gradient direction can preserve the details in the vertical direction of the edge to the maximum extent. This is because the gradient direction essentially reflects the steepest direction of brightness change. If the neighborhood value is too large in this direction, it can easily lead to blurred edges or crossing edge regions. By arranging the minor axis along the gradient direction, the neighborhood range can be strictly limited in the direction most sensitive to the edge, avoiding interference across the edge, while retaining a larger sampling range in the edge direction to enhance the coherence of the structure.

[0091] Furthermore, this gradient-direction-based elliptical neighborhood design has significant adaptive advantages. The neighborhood shape of each pixel is no longer fixed in direction, but is dynamically rotated and adjusted according to the gradient direction of that pixel, so that the minor axis of the ellipse always remains consistent with the gradient direction. In the edge region neighborhood, it can conform to the direction distribution of the local structure and avoid information confusion caused by crossing the edge. In the smooth region neighborhood, it also maintains coordination with the local gradient direction, so that the processing result is more consistent with the intrinsic characteristics of the image. Finally, after the gradient magnitude replacement operation, the gradient magnitude matrix can not only highlight edge and texture features, but also improve the robustness of edge detection and enhancement as a whole.

[0092] S2.2.3: Normalize the morphological gradient matrix and use it as a weight matrix to fuse with the gradient magnitude matrix to obtain the gradient enhancement matrix;

[0093] Specifically, normalizing the morphological gradient matrix is ​​to map values ​​from different ranges to a unified interval. Linear normalization is generally used, where the minimum value is subtracted from each element of the matrix, and then the result is divided by the difference between the maximum and minimum values, ensuring the result falls within the range [0,1]. The normalized matrix not only facilitates numerical calculations with other feature matrices but also avoids computational biases caused by excessive differences in numerical scales. Furthermore, normalization allows the morphological gradient matrix to serve as a standardized weighting factor, more reasonably reflecting the edge saliency of each pixel in the image. The normalized morphological gradient matrix is ​​then used as a weighting matrix and fused with the original gradient magnitude matrix. The gradient magnitude is multiplied by the corresponding weight value to obtain a new gradient enhancement matrix.

[0094] Furthermore, a normalized morphological gradient matrix is ​​used as a standardized weighting factor. In areas with significant edges, the normalized morphological gradient matrix has a higher weight, further amplifying the original gradient magnitude and thus strengthening edge features. In smooth areas, the normalized morphological gradient matrix has a lower weight, which can reduce the influence of noise and weak texture. Through this fusion strategy, the generated gradient enhancement matrix retains the original gradient information while highlighting key structural features.

[0095] S2.2.4: Perform numerical mapping on the gradient enhancement matrix, mapping the gradient enhancement matrix values ​​to the range of image pixel values ​​to obtain the processed brightness feature subspace.

[0096] Specifically, numerical mapping of the gradient enhancement matrix is ​​necessary. After processing and fusion, the value range of the gradient magnitude matrix is ​​often no longer within the range of normal image pixel values. If it is used directly as luminance information, it will cause abnormal image display or failure to be correctly fused with chrominance information. Therefore, it is necessary to use numerical mapping to uniformly scale the value of the gradient enhancement matrix to the range allowed by pixel values. This ensures that the processed data can be correctly rendered and stored.

[0097] Furthermore, using a gradient enhancement matrix as the luminance feature subspace can bring significant advantages. Ordinary luminance components can only reflect grayscale levels, while after gradient enhancement, the luminance feature subspace has higher contrast and sensitivity in edge, texture and detail areas. This enhancement can not only strengthen the texture and edge contours of plant leaves, but also effectively distinguish subtle structural differences.

[0098] S2.3: Perform color difference processing on the chromaticity feature subspace, the specific steps of which are as follows:

[0099] S2.3.1: Map the color vector of each pixel in the chromaticity feature subspace to a high-dimensional feature space using a kernel function. The color vector is represented using red-green chromaticity components and yellow-blue chromaticity components.

[0100] Specifically, each pixel color vector in the chromaticity feature subspace consists of red-green chromaticity components and yellow-blue chromaticity components, which can be regarded as a two-dimensional feature vector. By introducing a kernel function, the low-dimensional color vector is mapped to the high-dimensional feature space. This mapping process is equivalent to a non-linear extension of the color features, so that subtle chromaticity differences can be expressed and measured more clearly.

[0101] Furthermore, in the original chromaticity feature subspace, some color differences may be difficult to distinguish effectively due to similar distribution or component values. However, in the high-dimensional space, these differences are amplified through kernel function mapping, making color features that were originally difficult to distinguish in the low-dimensional space more linearly separable in the high-dimensional space. This enhances the model's sensitivity to detailed color differences. For greenhouse plant images, this high-dimensional representation can better capture color differences caused by leaf color changes, pest and disease spots, or malnutrition.

[0102] S2.3.2: Calculate the similarity of pixels in the high-dimensional feature space, construct a global similarity matrix, and cluster the pixels in the high-dimensional feature space according to the global similarity matrix to form color clusters;

[0103] Specifically, for high-dimensional feature vectors obtained by mapping the chroma feature subspace through a kernel function, the similarity between pixels can be calculated to measure their proximity in the high-dimensional space. Common similarity measurement methods include inner product, cosine similarity, and Gaussian kernel similarity. Among them, the inner product can quickly reflect the overall similarity between vectors, cosine similarity focuses more on the consistency of pixel color distribution direction, while Gaussian kernel similarity can better capture non-linear relationships. Through these methods, the color similarity between any two pixels can be quantified. The larger the value, the closer the color distribution, and thus the more likely they belong to the same color cluster.

[0104] Furthermore, after obtaining the similarity between pixels, a global similarity matrix can be constructed, with a dimension of N. N, where N is the total number of pixels in the image, can be used to intuitively determine which pixels in the image are more likely to cluster together to form regions of the same color, and which pixels show obvious differences.

[0105] Furthermore, a suitable clustering algorithm, such as spectral clustering, graph cut, or K-means clustering, is selected based on the global similarity matrix. Through these clustering methods, pixels with high similarity are grouped into the same color cluster, thereby achieving reasonable grouping of color features.

[0106] S2.3.3: For intra-cluster weights, the enhancement weights are calculated and set based on the distance between the pixel and the cluster center of the color cluster. For inter-cluster weights, the enhancement weights are calculated and set based on the distance between the cluster centers of the color clusters. The weight matrix is ​​obtained based on the intra-cluster weights and inter-cluster weights.

[0107] Specifically, for each color cluster, its cluster center is usually the average of the high-dimensional feature vectors of all pixels within the cluster. The cluster center represents the typical color features of the cluster and serves as the benchmark for subsequent calculation of intra-cluster weights. For each pixel within the cluster, the enhancement weight can be determined based on the distance between its high-dimensional feature vector and the cluster center. Pixels with smaller distances receive higher weights, which can highlight typical features within the cluster while suppressing outliers or edge noise, thereby achieving consistent enhancement within the cluster. For the calculation of inter-cluster weights, the center position of each cluster needs to be obtained first, and then the distance between clusters is calculated as the basis for measuring the similarity between different color clusters. Based on the distance between cluster centers, inter-cluster weights are assigned to each pixel, thereby introducing global constraints in the enhancement process. Inter-cluster weights can ensure that the enhancement process not only focuses on local features within the cluster but also considers the overall differences between color clusters, making the enhancement effect more balanced and globally consistent.

[0108] Furthermore, by fusing intra-cluster weights and inter-cluster weights through weighting or multiplication, the resulting weight matrix can be used to enhance the chroma of the image. This ensures that the enhancement level of each pixel reflects its typicality within the cluster while also taking into account the color differences between clusters, thereby achieving overall color balance and enhancement while preserving local details.

[0109] S2.3.4: The high-dimensional feature space is weighted as a whole according to the weight matrix to achieve color difference processing and obtain a color-enhanced high-dimensional feature space;

[0110] Specifically, for each pixel vector in the high-dimensional feature space, it is multiplied or weighted with the corresponding weight in the weight matrix to obtain the enhanced high-dimensional feature space. The weight matrix, as a control factor, can apply differentiated adjustments to the high-dimensional feature vectors of different pixels, thereby strengthening the typical color features and improving the distinguishability of color features.

[0111] S2.3.5: The high-dimensional feature space of color enhancement is reduced to obtain an enhanced two-dimensional feature space, which is then mapped to the range of image pixel values ​​to obtain the processed chromaticity feature subspace.

[0112] Specifically, after obtaining the high-dimensional feature space for color enhancement, it is dimensionality reduced to map the high-dimensional features back to a two-dimensional space, resulting in an enhanced two-dimensional feature space. Dimensionality reduction can be achieved through methods such as principal component analysis, multidimensional scaling, or kernel dimensionality reduction, which preserves the main color difference information while reducing computational complexity. After dimensionality reduction, the obtained two-dimensional feature space needs to be mapped to the pixel value range of the image to ensure that it corresponds to the image data. This step ensures that the enhanced chromaticity information can be integrated into the color channels of the image, thereby achieving chromaticity enhancement and visualization of the image.

[0113] Furthermore, enhancing the chromaticity feature subspace can better highlight the color differences between different regions in plant images, making subtle color variations no longer overlooked. For example, in greenhouse plant monitoring, slight yellowing of leaves, lesions, or differences in growth status can be made more apparent through enhancement of the chromaticity subspace. At the same time, the enhanced chromaticity features also improve the image's color distinguishability, providing richer and more reliable input information for the automatic detection and intelligent analysis of plant growth status.

[0114] Example 2:

[0115] Please see Figure 4 A greenhouse plant cultivation and management system, the system comprising a plant digital monitoring module, a knowledge graph construction module, a plant growth status identification module, and an intelligent management decision-making module;

[0116] The plant digital monitoring module collects plant growth environment data in real time through multiple types of sensors, and monitors plant growth status in real time by installing high-definition network cameras in key areas of the demonstration park.

[0117] The knowledge graph construction module integrates historical planting data and morphological and environmental data throughout the entire growth cycle of plants, and constructs a plant growth knowledge graph through modular construction, dynamic reasoning, and multi-source data fusion.

[0118] The plant growth status recognition module preprocesses the image data of greenhouse plants, inputs the processed plant images into a preset growth status detection model, and outputs the current growth status of the plants.

[0119] The intelligent management decision-making module formulates decision rules based on the plant growth knowledge graph and the real-time growth status of plants, and outputs plant planting decision-making plans based on the current growth status of plants and the decision rules.

[0120] The knowledge graph construction module includes a historical data processing unit, an entity-attribute-relationship mapping unit, a knowledge storage unit, and a knowledge iteration and update unit.

[0121] The historical data processing unit extracts and integrates plant surface feature data, environmental sensor data, expert experience database and open literature to form the basic dataset for knowledge graph construction.

[0122] The entity-attribute-relationship mapping unit extracts core elements of plant knowledge from the basic dataset through a deep learning model, including entities, entity attributes, and relationships between entities.

[0123] The knowledge storage unit uses the Neo4j graph database to modularly store the extracted entities, attributes and relationships, and establishes hierarchical relationships between entities through Cypher statements to realize entity query and semantic reasoning.

[0124] The knowledge iteration and update unit receives newly collected plant data, automatically matches it with existing knowledge, and updates entities, attributes, or relationships.

[0125] In this embodiment, the construction of the strawberry knowledge graph is based on multi-source data fusion technology and graph database technology: First, entities, attributes and relationships are extracted by acquiring strawberry knowledge data and base ledger data over many years to construct a knowledge pattern hierarchy. Finally, entities, attributes and relationships are sequentially imported into the Neo4j graph database to achieve structured knowledge storage and management.

[0126] The specific steps for constructing the knowledge graph are as follows: First, integrate historical data such as strawberry planting water and fertilizer records, pest and disease archives, etc., combined with the experience database of strawberry planting experts, public literature, and morphological data such as temperature, humidity, and light collected by environmental sensors and plant height, stem diameter, and number of leaves collected by machine vision, to construct a knowledge structure table covering the entire life cycle of strawberries. Through remote supervision and deep learning models, entities, attributes, and relationships are extracted from the integrated data to construct a schema layer. The schema layer is designed with strawberry as the first-level entity, and varieties, growth period, common diseases, and common pests as second-level entities. Each second-level entity is further expanded into third-level entities, such as "Hongyan" and "Xiangye" as attribute values ​​for varieties, and powdery mildew and gray mold as attribute values ​​for common diseases. Fourth-level entities, such as overview and characteristics, are used to describe "Hongyan", and symptoms, prevention and control, and occurrence environment are used to describe powdery mildew, etc. Entities and attribute values ​​are associated with "attribute is" relationships to achieve hierarchical knowledge modeling. Secondly, a Neo4j graph database is used to construct the knowledge storage architecture. Through a modular import strategy, node modules such as variety, growth period, disease, and environmental control are created in stages to achieve efficient knowledge organization. The CREATE statement in the Cypher language is used to establish hierarchical relationships between entities, supporting semantic reasoning and dynamic updates. The construction process combines top-down and bottom-up methods: the top layer defines core ontology, such as plant parameters, physiological processes, and environmental factors; the bottom layer fills in knowledge through multi-source data, achieving a structured and scalable knowledge system. In the verification phase, the Cypher query engine combined with an LSTM prediction model is used to evaluate the accuracy and reasoning ability of the knowledge graph. Finally, knowledge services are integrated through a cloud platform to form a dynamically linked knowledge network of "environment-morphology-physiology," providing a complete decision-making loop for precision strawberry cultivation. This method, through modular construction, dynamic reasoning, and multimodal data fusion, solves the problem of fragmented knowledge in traditional agriculture and provides a reusable technological paradigm for the whole life cycle management of strawberries.

[0127] The intelligent management decision-making module includes a rule reasoning and execution unit, an association rule mining unit, and a decision scheme generation unit;

[0128] The rule-based reasoning execution unit transforms the preset plant planting decision rules into computer-executable logic, and uses a rule-based reasoning algorithm to match and reason about the input plant growth status to generate preliminary decision results.

[0129] The association rule mining unit uses the Apriori association rule mining algorithm to mine hidden growth relationships from a large amount of historical planting data and integrates them as supplementary rules into the decision-making model to improve decision accuracy.

[0130] The decision-making scheme generation unit combines the results of rule-based reasoning with supplementary association rules to output specific decision-making schemes that include suggestions for agricultural operations, anomaly warnings, plant protection, water and fertilizer management, and harvesting.

[0131] In this embodiment, the strawberry smart management decision model uses data analysis and artificial intelligence algorithms to achieve intelligent management of the entire strawberry planting process. The model combines rule-based reasoning algorithms and machine learning methods to transform the formulated decision rules into computer-executable logic. At the same time, it uses the Apriori association rule algorithm to mine hidden patterns from historical planting data and incorporates them as supplementary rules into the model, thereby improving the scientific nature and accuracy of decision-making.

[0132] The model's input data comes from three sources: first, real-time environmental data collected by sensors, such as temperature, humidity, light intensity, and carbon dioxide concentration; second, planting knowledge stored in the strawberry knowledge graph, including variety characteristics, growth stage requirements, pest and disease characteristics, and control methods; and third, abnormal status data monitored by the strawberry growth status recognition algorithm, such as growth period, fruit, flower, and water exudation. By comprehensively analyzing this multi-source information, the model can formulate reasonable decision-making rules based on strawberry growth patterns and real-time status.

[0133] Based on input data and decision rules, the model can output specific planting decision plans, including agricultural operation suggestions, abnormal growth warnings, plant protection measures, water and fertilizer management, and harvesting arrangements, which are presented to growers intuitively through the system interface. The application of this model realizes intelligent decision-making in strawberry planting, which not only improves management efficiency and scientificity, but also effectively reduces planting risks, promotes yield and quality improvement, and provides reliable technical support for precision agriculture.

Claims

1. A method for identifying the growth status of greenhouse plants, characterized in that, The method includes: Acquire image data of the greenhouse plants; The image data is input into a preset growth state detection model, and the recognition result is output through the growth state detection model. The main structure of the growth state detection model is the YOLO target detection model, which includes a high-resolution detection layer used to extract target feature information and spatial feature information.

2. The method for identifying the growth status of greenhouse plants according to claim 1, characterized in that, The high-resolution detection layer also includes a prediction head, which is obtained through a Swin-Transformer structure. The prediction head is used to identify a portion of the image data through a self-attention mechanism, extracting leaf features, flower features, fruit features, and spatial feature information. The portion of the image data is a region of interest, which is obtained by the high-resolution detection layer through response filtering of the feature map of the image data.

3. The method for identifying the growth status of greenhouse plants according to claim 1, characterized in that, The Neck portion of the YOLO object detection model also includes a GS-ELAN structure constructed using GSConv, which is used to reduce noise in the output of the high-resolution detection layer.

4. The method for identifying the growth status of greenhouse plants according to claim 1, characterized in that, The method further includes: The image data of the greenhouse plants are separated into luminance and chrominance to obtain a luminance feature subspace and a chrominance feature subspace. Gradient enhancement processing is performed on the luminance feature subspace, and color difference processing is performed on the chrominance feature subspace. The processed luminance feature subspace, chrominance feature subspace and the original greenhouse plant image data are weighted and fused to obtain the greenhouse plant image enhancement data; The image enhancement data of the greenhouse plants is input into a preset growth status detection model, and the recognition result is output through the growth status detection model.

5. The method for identifying the growth status of greenhouse plants according to claim 4, characterized in that, The image data of the greenhouse plants are subjected to luminance and chrominance separation to obtain luminance feature subspace and chrominance feature subspace, including: The image data of the greenhouse plants are subjected to noise suppression processing; The processed greenhouse plant image data is decomposed into a luminance feature subspace and a chrominance feature subspace using a reversible mapping strategy. The luminance feature subspace is used to preserve the spatial structure information of the greenhouse plant image data, including texture and edge features, while the chrominance feature subspace is used to characterize the color distribution information of the greenhouse plant image data, including red-green hue components and yellow-blue hue components.

6. The method for identifying the growth status of greenhouse plants according to claim 4, characterized in that, Gradient enhancement processing is performed on the brightness feature subspace, including: Calculate the gradient magnitude matrix and gradient direction matrix based on the brightness feature subspace; Based on the gradient magnitude matrix and gradient direction matrix, the neighborhood of the gradient magnitude is determined, and the average, maximum, and minimum values ​​of all gradient magnitudes within the neighborhood are calculated. A semi-dilation matrix and a semi-erosion matrix are then constructed. When constructing the semi-dilation matrix, if the gradient magnitude is greater than or equal to the average value, it is replaced with the maximum value within the neighborhood; if the gradient magnitude is less than the average value, it is replaced with the average value within the neighborhood. Similarly, when constructing the semi-erosion matrix, if the gradient magnitude is less than or equal to the average value, it is replaced with the minimum value within the neighborhood; if the gradient magnitude is greater than the average value, it is replaced with the average value within the neighborhood. The morphological gradient matrix is ​​obtained by comparing the semi-dilation matrix and the semi-erosion matrix. The morphological gradient matrix is ​​normalized and used as a weight matrix to be fused with the gradient magnitude matrix to obtain the gradient enhancement matrix. Numerical mapping is performed on the gradient enhancement matrix to map the gradient enhancement matrix values ​​to the range of image pixel values, resulting in the processed brightness feature subspace.

7. The method for identifying the growth status of greenhouse plants according to claim 6, characterized in that, The neighborhood of each gradient magnitude is designed as an elliptical structure, where the minor axis of the ellipse is along the gradient direction corresponding to the gradient magnitude. The range of the elliptical neighborhood of each gradient magnitude is adjusted according to the gradient direction matrix, and the values ​​in the gradient magnitude matrix are replaced according to the elliptical neighborhood.

8. The method for identifying the growth status of greenhouse plants according to claim 4, characterized in that, Color difference processing is performed on the chromaticity feature subspace, including: The color vector of each pixel in the chromaticity feature subspace is mapped to a high-dimensional feature space through a kernel function. The color vector is represented by red-green chromaticity components and yellow-blue chromaticity components. Calculate the similarity of pixels in the high-dimensional feature space, construct a global similarity matrix, and cluster the pixels in the high-dimensional feature space according to the global similarity matrix to form color clusters; For intra-cluster weights, the enhancement weights are calculated and set based on the distance between the pixel and the cluster center of the color cluster. For inter-cluster weights, the enhancement weights are calculated and set based on the distance between the cluster centers of the color clusters. The weight matrix is ​​obtained based on the intra-cluster weights and inter-cluster weights. The high-dimensional feature space is weighted as a whole according to the weight matrix to achieve color difference processing and obtain a color-enhanced high-dimensional feature space. The high-dimensional feature space of color enhancement is reduced to obtain an enhanced two-dimensional feature space, which is then mapped to the range of image pixel values ​​to obtain the processed chromaticity feature subspace.

9. A greenhouse plant cultivation management system, characterized in that, The system includes: a plant digital monitoring module, a knowledge graph construction module, a plant growth status identification module, and a smart management decision-making module; The plant digital monitoring module collects plant growth environment data in real time through multiple types of sensors and monitors plant growth status in real time by installing network camera equipment in the demonstration park. The knowledge graph construction module integrates historical planting data and morphological and environmental data throughout the entire growth cycle of plants, and constructs a plant growth knowledge graph through modular construction, dynamic reasoning, and multi-source data fusion. The plant growth status recognition module preprocesses the image data of greenhouse plants, inputs the processed plant images into a preset growth status detection model, and outputs the current growth status of the plants. The intelligent management decision-making module formulates decision rules based on the plant growth knowledge graph and the real-time growth status of plants, and outputs plant planting decision-making plans based on the current growth status of plants and the decision rules.

10. The greenhouse plant cultivation and control system according to claim 9, characterized in that, The knowledge graph construction module includes a historical data processing unit, an entity-attribute-relationship mapping unit, a knowledge storage unit, and a knowledge iteration and update unit; The historical data processing unit extracts and integrates plant surface feature data and environmental sensor data to form the basic dataset for knowledge graph construction. The entity-attribute-relationship mapping unit extracts core elements of plant knowledge from the basic dataset through a deep learning model, including entities, entity attributes, and relationships between entities. The knowledge storage unit uses the Neo4j graph database to modularly store the extracted entities, attributes and relationships, and establishes hierarchical relationships between entities through Cypher statements to realize entity query and semantic reasoning. The knowledge iteration and update unit receives newly collected plant data and updates entities, attributes, or relationships.

11. The greenhouse plant planting control system according to claim 9, characterized in that, The intelligent management decision-making module includes a rule reasoning and execution unit, an association rule mining unit, and a decision scheme generation unit; The rule-based reasoning execution unit transforms the preset plant planting decision rules into computer-executable logic, and uses a rule-based reasoning algorithm to match and reason about the input plant growth status to generate preliminary decision results. The association rule mining unit uses the Apriori association rule mining algorithm to mine hidden growth relationships from historical planting data and incorporates them as supplementary rules into the decision-making model to improve decision accuracy. The decision-making scheme generation unit combines the results of rule-based reasoning with supplementary association rules to output decision-making schemes that include suggestions for agricultural operations, anomaly warnings, plant protection, water and fertilizer management, and harvesting.

Images