Corn seed environment adaptability cultivation method based on edge calculation

Abstract

The invention discloses a corn seed environment adaptability cultivation method based on edge calculation, which comprises the following steps: step 1, acquiring environment factor data of a corn seed sowing area, and preprocessing the environment factor data; 2, extracting the edge contour of the corn seedling by adopting Canny edge detection and a Sobel operator based on the RGB image; 3, constructing a multi-source heterogeneous data sample set according to time and space coordinate alignment; 4, constructing a causal structure diagram by adopting an improved NOTEARS algorithm; 5, carrying out comprehensive sorting to obtain a key factor path; step 6, inputting the multi-dimensional environment factor vector of the target area to the starting node of the key factor path to obtain a predicted phenotypic feature vector; and 7, comparing the actual phenotypic feature vector with the predicted phenotypic feature vector, and counting a prediction error. According to the method, causal modeling and image analysis are fused, and the breeding efficiency and the accuracy of regional suitability judgment are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural breeding technology, and in particular to a method for environmentally adaptable breeding of maize seeds based on edge computing. Background Technology

[0002] In traditional maize seed breeding, the assessment of environmental adaptability typically relies on long-term, multi-location, and multi-season field trials. This involves observing the growth performance of different varieties in different regions to infer their adaptability to specific ecological environments. While this approach has some practical value, it suffers from drawbacks such as long cycles, high costs, limited experimental site locations, and a single dimension of information acquisition. Especially when faced with complex interactions of environmental factors, traditional methods struggle to accurately analyze the specific impact pathways of each environmental factor on maize phenotypic traits and to predict adaptability in untested areas, severely limiting the efficiency and intelligence of maize breeding.

[0003] Existing intelligent breeding technologies have gradually incorporated image recognition and data analysis methods. For example, images of maize seedlings are used to assist in assessing emergence rate, growth status, or disease conditions, while environmental sensors collect field soil and meteorological data for breeding decisions. However, most of these methods focus only on a single type of data source, such as relying solely on image information or a single environmental indicator, lacking the ability to uniformly model multi-source heterogeneous data. Furthermore, existing technologies generally lack a causal analysis framework, with most employing correlation modeling or black-box prediction models, making it difficult to identify clear causal paths between specific environmental factors and phenotypic traits. Consequently, they struggle to provide interpretable decision support for breeding results.

[0004] Furthermore, in terms of adaptive prediction, existing methods often rely on well-trained machine learning models to generalize predictions for target areas. However, these models lack dynamic update capabilities and cannot self-correct their structure based on actual trial planting feedback data. This leads to the accumulation of biases in long-term adaptive evaluation results, affecting the accuracy and stability of recommended varieties.

[0005] Therefore, how to provide an environmentally adaptable breeding method for maize seeds based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for breeding maize seeds based on edge computing to improve their environmental adaptability. This invention integrates environmental perception, image recognition, and causal modeling technologies to construct an environment-phenotype causal path, enabling intelligent prediction of maize seed adaptability in target areas and identification of key factors. It also supports error feedback and path updates after trial planting, thereby improving the accuracy, interpretability, and efficiency of the breeding process.

[0007] The edge computing-based method for cultivating maize seeds to adapt to environmental conditions, according to an embodiment of the present invention, includes the following steps: Step 1: Deploy an agricultural IoT sensor network in the corn experimental field to acquire environmental factor data of the corn seed sowing area, and preprocess the data to generate a multi-dimensional environmental factor vector; Step 2: Use fixed-point camera equipment and drones to acquire RGB image sequences for several consecutive days after corn sowing. Based on the RGB images, use Canny edge detection and Sobel operator to extract the edge contours of corn seedlings and obtain phenotypic feature vectors. Step 3: Align the environmental factor vector and the phenotypic feature vector according to the time and spatial coordinates, construct a multi-source heterogeneous data sample set, and uniformly encode it into a spatial-environment-phenotypic three-dimensional data tensor; Step 4: The improved NOTEARS algorithm is used to model the space-environment-phenotype three-dimensional data tensor, and a causal structure graph containing environmental factor nodes and phenotypic feature nodes is constructed. Step 5: Based on the cause-effect structure graph, accumulate the weight values ​​corresponding to each directed edge in each directed path, and sort them comprehensively to obtain the key factor paths; Step 6: Obtain the multidimensional environmental factor vector of the target region and input it into the starting node of the key factor path. Perform a node-by-node forward propagation operation to obtain the predicted phenotypic feature vector. Step 7: After completing the trial planting in the target area, collect the actual phenotypic feature vectors of maize seeds, compare the actual phenotypic feature vectors with the predicted phenotypic feature vectors, and calculate the prediction error.

[0008] Optionally, the environmental factor data includes soil temperature, soil moisture, electrical conductivity, pH value, organic matter content, air temperature, air humidity, rainfall, wind speed, light intensity, and plot spatial location information; the preprocessing involves imputing missing values ​​and normalizing different types of data to generate multidimensional environmental factor vectors.

[0009] Optionally, step two specifically includes: Gaussian filtering is performed on the RGB images collected after corn sowing to remove noise generated by environmental factors or during image acquisition, and grayscale processing is performed to obtain grayscale images. The Canny edge detection algorithm calculates the rate of change of grayscale values ​​of pixels in the surrounding neighborhood for each pixel in a grayscale image. The surrounding neighborhood is a window of a preset size. The Sobel operator is used to perform convolution operations to obtain the horizontal and vertical gradients of each pixel, and then the gradient intensity and gradient direction of each pixel are obtained. The gradient is the rate of change of gray value. In the gradient direction, only the local maximum gradient value is retained to extract edge details and remove false edge points in the grayscale image; The details of the extracted edges specifically include: Two thresholds, a high threshold and a low threshold, are set to classify each pixel in the grayscale image; If the gradient value is greater than the high threshold, then this pixel is considered an edge; If the gradient value is less than the low threshold, then the pixel is considered not to belong to the edge. Pixels falling between the two thresholds are determined based on the connection between adjacent pixels; The specific steps for judging based on the connection of adjacent pixels are as follows: compare the gradient intensity of this pixel with the two neighboring pixels along the gradient direction. If the gradient intensity of this pixel is greater than or equal to its neighboring pixels, it is considered to be part of the edge and is retained as an edge; otherwise, it is not. The pixels identified as edges are connected to form the edge outline of the corn seedling; The DBSCAN clustering algorithm is used to cluster edge pixels. A minimum number of points and a distance threshold are set to form multiple independent pixel clusters. Each pixel cluster represents a mask region of a corn seedling. Multiple masked regions are stitched together to obtain a masked image of a single corn seedling, and phenotypic feature values ​​are calculated based on each masked region. The phenotypic feature values ​​include emergence rate, growth period change rate, ear length, and disease score. The germination rate is the ratio of the number of masked areas to the preset number of sowing points; The growth period change rate is linearly fitted to the daily mask area and its time series, and the slope value is calculated. The mask area is the sum of the pixel area of ​​the mask region of each corn seedling. The ear length is the maximum vertical edge pixel distance in the mask region; The disease score is the percentage of color-abnormal areas within the masked area, which is combined with a preset percentage standard to generate a disease grading score. The emergence rate, growth period change rate, panicle length and disease score were standardized using Z-score, and the vectors were concatenated in a fixed order to form phenotypic feature vectors.

[0010] Optionally, step three specifically includes: The spatial coordinate information of each corn planting area is obtained, and the spatial coordinate information is used as an identifier to bind the corresponding environmental factor vector and phenotypic feature vector respectively. The spatial coordinate information includes the plot number, the latitude and longitude coordinates of the center point, the plot elevation value and the slope angle. Time alignment processing is performed on the environmental factor vector. The time alignment processing involves using a linear interpolation algorithm to fill in missing time points and converting data from different sampling frequencies into data sequences with a uniform time step. The acquisition time of the RGB image sequence is timestamped, and the phenotypic feature vectors extracted from the RGB images are synchronously matched with the environmental factor vector time according to the corresponding shooting time. Under the condition that the time error does not exceed the set threshold, a one-to-one correspondence is established. Based on spatial coordinate consistency and timestamp matching, the environmental factor vector and phenotypic feature vector of the same plot at the same time are combined to form a set of sample data, and a multi-source heterogeneous data sample set consisting of multiple plots and multiple time slices is constructed. The multi-source heterogeneous data sample set is structured and organized according to a preset dimensional order. Each group of multi-source heterogeneous data samples is uniformly encoded into a ternary structure form, which includes spatial coordinate information, environmental factor vector, and phenotypic feature vector. All multi-source heterogeneous data samples are arranged in temporal and spatial order and uniformly encoded into a spatial-environment-phenotypic three-dimensional metadata tensor.

[0011] Optionally, the improved NOTEARS algorithm is specifically as follows: Extract the environmental factor value and phenotypic feature value corresponding to each sample from the space-environment-phenotype three-dimensional data tensor, and concatenate them into a sample vector set in a unified order; The improved NOTEARS algorithm performs linear regression modeling for each pair of variables in the sample vector set; The linear regression modeling operation involves specifying one variable as the target variable and another variable as the input variable, and then calculating the residual squared difference between the predicted value and the actual observed value of the target variable after linear fitting. Based on the magnitude of the residual squared difference, when the residual squared difference between the input variable and the target variable is less than the preset residual threshold, it is determined that the input variable has a significant linear effect on the target variable, and a directed connection path from the input variable to the target variable is established. After adding each directed connection path, check in real time whether the current path set forms a directed loop. When a circular path with the first and last connected variables is detected between any three or more variables, delete the connection path with the largest residual squared difference in the circular path. Repeatedly perform residual calculations between variable pairs, directed connection path establishment, and loop elimination operations until the dependency relationships of all variable pairs are determined, and finally construct a causal structure graph containing all environmental factor nodes and phenotypic feature nodes. The directed edges in the causal structure graph represent causal influence paths from environmental factors to phenotypic characteristics, and the weight values ​​of the directed edges are the corresponding residual squared differences.

[0012] Optionally, step five specifically includes: Perform a traversal analysis on the causal structure graph, extract all directed paths that start from environmental factor nodes and eventually point to any phenotypic feature node, and record the node sequence and the connection direction of the directed edges for each directed path. Count the number of paths that each environmental factor node can reach to all phenotypic feature nodes, and record the path length, which is the number of directed edges contained in the path. The weight values ​​corresponding to each directed edge in each directed path are summed to obtain the total weight value of each directed path; All directed paths are sorted by a weighted average of their total weight and path length. Based on the comprehensive ranking results, the top-ranked directed paths of a predetermined number are selected as key factor paths. The key factor paths represent the dominant influence links of specific environmental factors on the target phenotypic characteristics.

[0013] Optionally, step six specifically includes: Multidimensional environmental factor data of the target area are collected, and environmental factor vectors of the target area are generated after standardizing the values ​​of each environmental factor. The target area is the area where maize seed breeding experiments are to be carried out. Based on the key factor paths already determined in the causal structure diagram, identify the starting node, ending node, and intermediate nodes of each key factor path, where the starting node is an environmental factor node and the ending node is a phenotypic feature node. The values ​​corresponding to the starting nodes in the environmental factor vector of the target region are used as the initial input values. The forward propagation is carried out node by node along each critical path according to the node connection direction of the critical factor path. The forward propagation is carried out according to the direction of each directed edge when constructing the causal structure graph. If multiple input values ​​in the forward propagation path point to the same node, a weighted summation operation is performed on all the input values ​​to synthesize the input value of this node and continue to propagate to the next node; When the propagation endpoint reaches the phenotypic feature node, the phenotypic feature vector corresponding to this phenotypic feature node is used as the predicted phenotypic feature vector of the target region.

[0014] Optionally, step seven specifically includes: After completing the corn seed trial planting in the target area, the actual phenotypic feature vectors of the target area were collected. The predicted phenotypic feature vector is compared one-to-one with the actual phenotypic feature vector, and the prediction error of each phenotypic feature value is calculated. The prediction error of each phenotypic feature value is compared with a preset error threshold. Phenotypic features with errors exceeding the threshold are selected and the corresponding key factor paths in the causal structure diagram are traced. The weight of each directed edge in the corresponding key factor path is reduced, and the reduction value is the absolute value of the difference between the prediction error and the preset error threshold. Save the updated cause-effect diagram as the current version.

[0015] The beneficial effects of this invention are: This invention achieves structured fusion of multi-source heterogeneous agricultural data by constructing a unified spatial-environmental-phenotypic three-dimensional data tensor. It also utilizes an improved NOTEARS causal structure learning algorithm to establish a directed acyclic causal graph between environmental factors and maize phenotypic characteristics, clearly revealing the influence paths of environmental factors on key phenotypes such as emergence rate, growth period variation, ear length, and disease occurrence. By screening key factor paths in the graph and performing forward propagation calculations on environmental factors in the target area, the phenotypic response of maize varieties can be accurately predicted without actual planting trials, significantly improving breeding efficiency and the accuracy of regional suitability assessment. Furthermore, this method supports feeding back actual observed phenotypic data into the causal structure graph after planting trials, enabling automatic correction of path weights and structural updates, and possessing continuous optimization capabilities. Compared to traditional breeding methods that rely on empirical judgment or static models, this invention offers advantages such as interpretable prediction processes, updatable models, strong environmental adaptability, and flexible deployment. It significantly reduces the cost of field trials while ensuring the scientific rigor of breeding, accelerating the screening and promotion of high-quality maize seeds. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is an overall flowchart of the edge computing-based method for cultivating maize seeds to adapt to environmental conditions, as proposed in this invention. Figure 2 This is a flowchart of the process for extracting the outline of maize seedlings and constructing phenotypic feature vectors from RGB images of the maize seed environmental adaptability cultivation method based on edge computing proposed in this invention. Figure 3 This is a schematic diagram of the causal structure graph constructed by the improved NOTEARS algorithm for the edge computing-based maize seed environmental adaptability breeding method proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-3 An edge computing-based method for cultivating maize seeds to adapt to environmental conditions includes the following steps: Step 1: Deploy an agricultural IoT sensor network in the corn experimental field to acquire environmental factor data of the corn seed sowing area, and preprocess the data to generate a multi-dimensional environmental factor vector; Step 2: Use fixed-point camera equipment and drones to acquire RGB image sequences for several consecutive days after corn sowing. Based on the RGB images, use Canny edge detection and Sobel operator to extract the edge contours of corn seedlings and obtain phenotypic feature vectors. Step 3: Align the environmental factor vector and the phenotypic feature vector according to the time and spatial coordinates, construct a multi-source heterogeneous data sample set, and uniformly encode it into a spatial-environment-phenotypic three-dimensional data tensor; Step 4: The improved NOTEARS algorithm is used to model the space-environment-phenotype three-dimensional data tensor, and a causal structure graph containing environmental factor nodes and phenotypic feature nodes is constructed. Step 5: Based on the cause-effect structure graph, accumulate the weight values ​​corresponding to each directed edge in each directed path, and sort them comprehensively to obtain the key factor paths; Step 6: Obtain the multidimensional environmental factor vector of the target region and input it into the starting node of the key factor path. Perform a node-by-node forward propagation operation to obtain the predicted phenotypic feature vector. Step 7: After completing the trial planting in the target area, collect the actual phenotypic feature vectors of maize seeds, compare the actual phenotypic feature vectors with the predicted phenotypic feature vectors, and calculate the prediction error.

[0020] This invention integrates edge environment perception and image recognition technologies to construct a unified environment-phenotype data tensor and introduces an improved NOTEARS algorithm to establish a causal structure graph, enabling phenotypic prediction and key factor identification of maize seeds under different regional environments. Compared with traditional experimental breeding methods, this method can assess variety adaptability in advance, improve prediction accuracy and breeding efficiency, and has advantages such as strong interpretability, clear data structure, and wide applicability.

[0021] In this embodiment, the environmental factor data includes soil temperature, soil moisture, electrical conductivity, pH value, organic matter content, air temperature, air humidity, rainfall, wind speed, light intensity, and plot spatial location information; the preprocessing involves imputing missing values ​​and normalizing different types of data to generate multidimensional environmental factor vectors.

[0022] This step involves comprehensive collection of soil and climate environmental factors, combined with missing value imputation and normalization preprocessing techniques, to construct a stable and reliable multidimensional environmental factor vector. This improves the spatiotemporal integrity and modeling consistency of environmental data, providing a high-quality and structurally unified input foundation for subsequent causal inference and regional adaptive prediction, and enhancing the model's adaptability and generalization ability.

[0023] In this embodiment, step two specifically includes: Gaussian filtering is performed on the RGB images collected after corn sowing to remove noise generated by environmental factors or during image acquisition, and grayscale processing is performed to obtain grayscale images. The Canny edge detection algorithm calculates the rate of change of grayscale values ​​of pixels in the surrounding neighborhood for each pixel in a grayscale image. The surrounding neighborhood is a window of a preset size. The Sobel operator is used to perform convolution operations to obtain the horizontal and vertical gradients of each pixel, and then the gradient intensity and gradient direction of each pixel are obtained. The gradient is the rate of change of gray value. In the gradient direction, only the local maximum gradient value is retained to extract edge details and remove false edge points in the grayscale image; The details of the extracted edges specifically include: Two thresholds, a high threshold and a low threshold, are set to classify each pixel in the grayscale image; If the gradient value is greater than the high threshold, then this pixel is considered an edge; If the gradient value is less than the low threshold, then the pixel is considered not to belong to the edge. Pixels falling between the two thresholds are determined based on the connection between adjacent pixels; The specific steps for judging based on the connection of adjacent pixels are as follows: compare the gradient intensity of this pixel with the two neighboring pixels along the gradient direction. If the gradient intensity of this pixel is greater than or equal to its neighboring pixels, it is considered to be part of the edge and is retained as an edge; otherwise, it is not. The pixels identified as edges are connected to form the edge outline of the corn seedling; The DBSCAN clustering algorithm is used to cluster edge pixels. A minimum number of points and a distance threshold are set to form multiple independent pixel clusters. Each pixel cluster represents a mask region of a corn seedling. Multiple masked regions are stitched together to obtain a masked image of a single corn seedling, and phenotypic feature values ​​are calculated based on each masked region. The phenotypic feature values ​​include emergence rate, growth period change rate, ear length, and disease score. The germination rate is the ratio of the number of masked areas to the preset number of sowing points; The growth period change rate is linearly fitted to the daily mask area and its time series, and the slope value is calculated. The mask area is the sum of the pixel area of ​​the mask region of each corn seedling. The ear length is the maximum vertical edge pixel distance in the mask region; The disease score is the percentage of color-abnormal areas within the masked area, which is combined with a preset percentage standard to generate a disease grading score. The emergence rate, growth period change rate, panicle length and disease score were standardized using Z-score, and the vectors were concatenated in a fixed order to form phenotypic feature vectors.

[0024] This step significantly improves the robustness and accuracy of image edge extraction by introducing an image processing method combining Gaussian filtering, Canny edge detection, and the Sobel operator. DBSCAN clustering intelligently divides edge pixels into independent plant units, and further extracts phenotypic indicators such as seedling emergence rate, growth period change rate, ear length, and disease score to construct a high-dimensional phenotypic feature vector. This method achieves high-precision, automated maize seedling image recognition and feature quantization without relying on complex network models, improving the efficiency of image analysis and the reliability of environmental adaptability evaluation in edge computing scenarios.

[0025] In this embodiment, step three specifically includes: The spatial coordinate information of each corn planting area is obtained, and the spatial coordinate information is used as an identifier to bind the corresponding environmental factor vector and phenotypic feature vector respectively. The spatial coordinate information includes the plot number, the latitude and longitude coordinates of the center point, the plot elevation value and the slope angle. Time alignment processing is performed on the environmental factor vector. The time alignment processing involves using a linear interpolation algorithm to fill in missing time points and converting data from different sampling frequencies into data sequences with a uniform time step. The acquisition time of the RGB image sequence is timestamped, and the phenotypic feature vectors extracted from the RGB images are synchronously matched with the environmental factor vector time according to the corresponding shooting time. Under the condition that the time error does not exceed the set threshold, a one-to-one correspondence is established. Based on spatial coordinate consistency and timestamp matching, the environmental factor vector and phenotypic feature vector of the same plot at the same time are combined to form a set of sample data, and a multi-source heterogeneous data sample set consisting of multiple plots and multiple time slices is constructed. The multi-source heterogeneous data sample set is structured and organized according to a preset dimensional order. Each group of multi-source heterogeneous data samples is uniformly encoded into a ternary structure form, which includes spatial coordinate information, environmental factor vector, and phenotypic feature vector. All multi-source heterogeneous data samples are arranged in temporal and spatial order and uniformly encoded into a spatial-environment-phenotypic three-dimensional metadata tensor.

[0026] This step achieves unified matching and spatiotemporal alignment of multi-source heterogeneous data by highly binding spatial coordinate information of maize planting areas with environmental and phenotypic data, overcoming the temporal asynchrony and spatial misalignment problems between environmental data and phenotypic responses in traditional methods. By introducing linear interpolation for temporal alignment, using spatial coordinates as data fusion identifiers, and finally constructing a unified spatial-environment-phenotypic three-dimensional data tensor, a high-quality data foundation is provided for subsequent causal relationship modeling and path reasoning, significantly improving the accuracy and stability of breeding decision analysis.

[0027] In this embodiment, the improved NOTEARS algorithm is specifically as follows: Extract the environmental factor value and phenotypic feature value corresponding to each sample from the spatial-environment-phenotype three-dimensional data tensor, and concatenate them into a sample vector set in a uniform order; The improved NOTEARS algorithm performs linear regression modeling for each pair of variables in the sample vector set; The linear regression modeling operation involves specifying one variable as the target variable and another variable as the input variable, and then calculating the residual squared difference between the predicted value and the actual observed value of the target variable after linear fitting. Based on the magnitude of the residual squared difference, when the residual squared difference between the input variable and the target variable is less than the preset residual threshold, it is determined that the input variable has a significant linear effect on the target variable, and a directed connection path from the input variable to the target variable is established. After adding each directed connection path, check in real time whether the current path set forms a directed loop. When a circular path with the first and last connected variables is detected between any three or more variables, delete the connection path with the largest residual squared difference in the circular path. Repeatedly perform residual calculations between variable pairs, directed connection path establishment, and loop elimination operations until the dependency relationships of all variable pairs are determined, and finally construct a causal structure graph containing all environmental factor nodes and phenotypic feature nodes. The directed edges in the causal structure graph represent causal influence paths from environmental factors to phenotypic characteristics, and the weight values ​​of the directed edges are the corresponding residual squared differences.

[0028] This step involves modeling the spatial-environment-phenotypic three-dimensional data tensor using an improved NOTEARS algorithm. Residual squared differences are used as an evaluation index for causal influence between variables, enabling efficient construction of acyclic causal structure graphs. Compared to traditional methods that infer causal relationships based on correlation or empirical rules, this algorithm can dynamically eliminate potential closed-loop paths while maintaining graph accuracy, ensuring the effectiveness and interpretability of the causal graph structure. The resulting causal structure graph not only clarifies the direct driving relationship between key environmental factors and maize phenotypic responses but also provides a scientific basis for subsequent path analysis and predictive inference, significantly improving the accuracy and reliability of breeding adaptability modeling.

[0029] In this embodiment, step five specifically includes: Perform a traversal analysis on the causal structure graph, extract all directed paths that start from environmental factor nodes and eventually point to any phenotypic feature node, and record the node sequence and the connection direction of the directed edges for each directed path. Count the number of paths that each environmental factor node can reach to all phenotypic feature nodes, and record the path length, which is the number of directed edges contained in the path. The weight values ​​corresponding to each directed edge in each directed path are summed to obtain the total weight value of each directed path; All directed paths are sorted by a weighted average of their total weight and path length. Based on the comprehensive ranking results, the top-ranked directed paths of a predetermined number are selected as key factor paths. The key factor paths represent the dominant influence links of specific environmental factors on the target phenotypic characteristics.

[0030] After constructing the causal structure diagram, this step systematically identifies high-influence paths from environmental factors to phenotypic characteristics through path traversal and weighted analysis strategies, accurately selecting dominant factor links. Compared to traditional methods relying on single-factor correlation analysis, this invention comprehensively considers the path length and weight strength in the causal chain, ensuring that the selected paths are both global and significant. The extracted key factor paths not only possess clear physical interpretability but can also serve as a basis for subsequent prediction simulations and precise policy implementation, thereby significantly improving the scientific rigor and controllability of maize seed environmental adaptability assessment and optimized breeding.

[0031] In this embodiment, step six specifically includes: Multidimensional environmental factor data of the target area are collected, and environmental factor vectors of the target area are generated after standardizing the values ​​of each environmental factor. The target area is the area where maize seed breeding experiments are to be carried out. Based on the key factor paths already determined in the causal structure diagram, identify the starting node, ending node, and intermediate nodes of each key factor path, where the starting node is an environmental factor node and the ending node is a phenotypic feature node. The values ​​corresponding to the starting nodes in the environmental factor vector of the target region are used as the initial input values. The forward propagation is carried out node by node along each critical path according to the node connection direction of the critical factor path. The forward propagation is carried out according to the direction of each directed edge when constructing the causal structure graph. If multiple input values ​​in the forward propagation path point to the same node, a weighted summation operation is performed on all the input values ​​to synthesize the input value of this node and continue to propagate to the next node; When the propagation endpoint reaches the phenotypic feature node, the phenotypic feature vector corresponding to this phenotypic feature node is used as the predicted phenotypic feature vector of the target region.

[0032] This step involves collecting environmental factor data from the target area and combining it with key factor paths in a pre-constructed causal structure graph to achieve forward prediction of maize seed phenotypic characteristics in the target area. This method fully utilizes the causal transmission relationships inherent in the key factor paths, achieving a node-by-node inferential propagation process based on edge computing. It can accurately simulate key phenotypic indicators such as emergence rate, growth period, ear length, and diseases before trial planting, significantly improving the efficiency of early-stage decision-making and the rationality of resource allocation in breeding work, and effectively reducing the planting failure rate caused by environmental mismatch.

[0033] In this embodiment, step seven specifically includes: After completing the corn seed trial planting in the target area, the actual phenotypic feature vectors of the target area were collected. The predicted phenotypic feature vector is compared one-to-one with the actual phenotypic feature vector, and the prediction error of each phenotypic feature value is calculated. The prediction error of each phenotypic feature value is compared with a preset error threshold. Phenotypic features with errors exceeding the threshold are selected and the corresponding key factor paths in the causal structure diagram are traced. The weight of each directed edge in the corresponding key factor path is reduced, and the reduction value is the absolute value of the difference between the prediction error and the preset error threshold. Save the updated cause-effect diagram as the current version.

[0034] This step involves collecting the actual phenotypic feature vectors of the target region and comparing them with the prediction results item by item to dynamically adjust and self-correct the path weights of key factors in the causal structure graph. This strategy effectively enhances the model's feedback learning ability and environmental adaptability, proactively reducing the causal influence of paths with large prediction errors. This improves the prediction accuracy and generalization ability of the causal structure graph in subsequent applications, providing more timely and robust causal reasoning support for iterative breeding of maize seeds.

[0035] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-site maize breeding experiment in a typical hilly terrain in western province. This region has varied topography, with significant differences in soil type, light conditions, wind speed, and wind direction between plots. Traditional methods struggle to accurately assess the environmental adaptability of maize varieties, often leading to a mismatch between planting layout and breeding strategy, and low seed selection efficiency. Therefore, multiple sets of edge sensing nodes were deployed in the experimental area. Each node integrated environmental sensors for soil temperature, humidity, electrical conductivity, pH value, organic matter, air temperature and humidity, light intensity, wind speed, and rainfall, collecting real-time microenvironmental data for each plot and recording the spatial coordinates, elevation, and slope aspect of the plots via a BeiDou module.

[0036] Simultaneously, high-resolution RGB images of maize during the seedling stage are periodically acquired using lightweight drones. Gaussian filtering, Canny edge detection, and Sobel convolution are performed on edge devices to further extract edge contours. Individual plant mask regions are then automatically clustered using the DBSCAN algorithm to calculate phenotypic features such as seedling emergence rate, growth period variation rate, ear length, and disease score for each plant. All data undergoes local preprocessing and temporal alignment at the edge nodes, and is then uniformly encoded into a "space-environment-phenotype" ternary tensor. This tensor is then input into the improved NOTEARS causal structure learning algorithm to generate a directed graph from environmental factors to phenotypic indicators. Paths are sorted based on the weight values ​​of the squared differences of the directed edge residuals, selecting the causal paths with the most significant impact on phenotype, and identifying key influence chains such as "light intensity → seedling emergence rate," "pH value → ear length," and "wind speed → disease score."

[0037] Before trial planting of new maize varieties in the target area, the system collects environmental parameters of the target plots, constructs a standardized environmental factor vector, and performs causal forward propagation based on the aforementioned critical paths to generate a predicted phenotypic vector for the region under the current environment. After the trial planting, actual data on emergence rate, growth progress, and ear length are collected and compared with the predicted values. The system automatically identifies causal paths with large prediction errors and lowers their connection weights, achieving adaptive structural updates and continuously optimizing prediction accuracy.

[0038] To visually demonstrate the effectiveness of the method of this invention, three typical plots (A, B, and C) from the experiment were selected for comparative analysis, and systematically compared with three conventional breeding evaluation methods—traditional empirical estimation method, single-factor analysis method, and multiple linear regression method—resulting in the following table: Table 1. Comparison of the effects of the method of the present invention and the traditional method in maize breeding in multiple plots.

[0039] As shown in Table 1, the phenotypic prediction error of the method of this invention is significantly lower than that of the three traditional methods on all plots, especially in the prediction of disease scores, where the error is controlled within 1 point, while the traditional methods are generally between 2 and 3 points. The prediction stability score also consistently remains above 9 points, far superior to other methods. In addition, this method can continuously optimize its structure through model adjustment mechanisms, demonstrating high adaptability. These results fully demonstrate the practicality and innovation of edge intelligent environment data acquisition, multi-source heterogeneous data modeling, and causal path dynamic learning mechanisms in agricultural seed breeding scenarios.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for cultivating maize seeds to adapt to environmental conditions based on edge computing, characterized in that, Includes the following steps: Step 1: Deploy an agricultural IoT sensor network in the corn experimental field to acquire environmental factor data of the corn seed sowing area, and preprocess the data to generate a multi-dimensional environmental factor vector; Step 2: Use fixed-point camera equipment and drones to acquire RGB image sequences for several consecutive days after corn sowing. Based on the RGB images, use Canny edge detection and Sobel operator to extract the edge contours of corn seedlings and obtain phenotypic feature vectors. Step 3: Align the environmental factor vector and the phenotypic feature vector according to the time and spatial coordinates, construct a multi-source heterogeneous data sample set, and uniformly encode it into a spatial-environment-phenotypic three-dimensional data tensor; Step 4: The improved NOTEARS algorithm is used to model the space-environment-phenotype three-dimensional data tensor, and a causal structure graph containing environmental factor nodes and phenotypic feature nodes is constructed. Step 5: Based on the cause-effect structure graph, accumulate the weight values ​​corresponding to each directed edge in each directed path, and sort them comprehensively to obtain the key factor paths; Step 6: Obtain the multidimensional environmental factor vector of the target region and input it into the starting node of the key factor path. Perform a node-by-node forward propagation operation to obtain the predicted phenotypic feature vector. Step 7: After completing the trial planting in the target area, collect the actual phenotypic feature vectors of maize seeds, compare the actual phenotypic feature vectors with the predicted phenotypic feature vectors, and calculate the prediction error.

2. The method for cultivating maize seeds based on edge computing for environmental adaptability according to claim 1, characterized in that, The environmental factor data includes soil temperature, soil moisture, electrical conductivity, pH value, organic matter content, air temperature, air humidity, rainfall, wind speed, light intensity, and spatial location information of the plot; the preprocessing involves imputing missing values ​​and normalizing different types of data to generate multidimensional environmental factor vectors.

3. The method for cultivating maize seeds based on edge computing for environmental adaptability according to claim 1, characterized in that, Step two specifically involves: Gaussian filtering is performed on the RGB images collected after corn sowing to remove noise generated by environmental factors or during image acquisition, and grayscale processing is performed to obtain grayscale images. The Canny edge detection algorithm calculates the rate of change of grayscale values ​​of pixels in the surrounding neighborhood for each pixel in a grayscale image. The surrounding neighborhood is a window of a preset size. The Sobel operator is used to perform convolution operations to obtain the horizontal and vertical gradients of each pixel, and then the gradient intensity and gradient direction of each pixel are obtained. The gradient is the rate of change of gray value. In the gradient direction, only the local maximum gradient value is retained to extract edge details and remove false edge points in the grayscale image; The details of the extracted edges specifically include: Two thresholds, a high threshold and a low threshold, are set to classify each pixel in the grayscale image; If the gradient value is greater than the high threshold, then this pixel is considered an edge; If the gradient value is less than the low threshold, then the pixel is considered not to belong to the edge. Pixels falling between the two thresholds are determined based on the connection between adjacent pixels; The specific steps for judging based on the connection of adjacent pixels are as follows: compare the gradient intensity of this pixel with the two neighboring pixels along the gradient direction. If the gradient intensity of this pixel is greater than or equal to its neighboring pixels, it is considered to be part of the edge and is retained as an edge; otherwise, it is not. The pixels identified as edges are connected to form the edge outline of the corn seedling; The DBSCAN clustering algorithm is used to cluster edge pixels. A minimum number of points and a distance threshold are set to form multiple independent pixel clusters. Each pixel cluster represents a mask region of a corn seedling. Multiple masked regions are stitched together to obtain a masked image of a single corn seedling, and phenotypic feature values ​​are calculated based on each masked region. The phenotypic feature values ​​include emergence rate, growth period change rate, ear length, and disease score. The germination rate is the ratio of the number of masked areas to the preset number of sowing points; The growth period change rate is linearly fitted to the daily mask area and its time series, and the slope value is calculated. The mask area is the sum of the pixel area of ​​the mask region of each corn seedling. The ear length is the maximum vertical edge pixel distance in the mask region; The disease score is the percentage of color-abnormal areas within the masked area, which is combined with a preset percentage standard to generate a disease grading score. The emergence rate, growth period change rate, panicle length and disease score were standardized using Z-score, and the vectors were concatenated in a fixed order to form phenotypic feature vectors.

4. The method for cultivating maize seeds based on edge computing for environmental adaptability according to claim 1, characterized in that, Step three specifically involves: The spatial coordinate information of each corn planting area is obtained, and the spatial coordinate information is used as an identifier to bind the corresponding environmental factor vector and phenotypic feature vector respectively. The spatial coordinate information includes the plot number, the latitude and longitude coordinates of the center point, the plot elevation value and the slope angle. Time alignment processing is performed on the environmental factor vector. The time alignment processing involves using a linear interpolation algorithm to fill in missing time points and converting data from different sampling frequencies into data sequences with a uniform time step. The acquisition time of the RGB image sequence is timestamped, and the phenotypic feature vectors extracted from the RGB images are synchronously matched with the environmental factor vector time according to the corresponding shooting time. Under the condition that the time error does not exceed the set threshold, a one-to-one correspondence is established. Based on spatial coordinate consistency and timestamp matching, the environmental factor vector and phenotypic feature vector of the same plot at the same time are combined to form a set of sample data, and a multi-source heterogeneous data sample set consisting of multiple plots and multiple time slices is constructed. The multi-source heterogeneous data sample set is structured and organized according to a preset dimensional order. Each group of multi-source heterogeneous data samples is uniformly encoded into a ternary structure form, which includes spatial coordinate information, environmental factor vector, and phenotypic feature vector. All multi-source heterogeneous data samples are arranged in temporal and spatial order and uniformly encoded into a spatial-environment-phenotypic three-dimensional metadata tensor.

5. The method for cultivating maize seeds based on edge computing for environmental adaptability according to claim 1, characterized in that, The improved NOTEARS algorithm is specifically as follows: Extract the environmental factor value and phenotypic feature value corresponding to each sample from the spatial-environment-phenotype three-dimensional data tensor, and concatenate them into a sample vector set in a uniform order; The improved NOTEARS algorithm performs linear regression modeling for each pair of variables in the sample vector set; The linear regression modeling operation involves specifying one variable as the target variable and another variable as the input variable, and then calculating the residual squared difference between the predicted value and the actual observed value of the target variable after linear fitting. Based on the magnitude of the residual squared difference, when the residual squared difference between the input variable and the target variable is less than the preset residual threshold, it is determined that the input variable has a significant linear effect on the target variable, and a directed connection path from the input variable to the target variable is established. After adding each directed connection path, check in real time whether the current path set forms a directed loop. When a circular path with the first and last connected variables is detected between any three or more variables, delete the connection path with the largest residual squared difference in the circular path. Repeatedly perform residual calculations between variable pairs, directed connection path establishment, and loop elimination operations until the dependency relationships of all variable pairs are determined, and finally construct a causal structure graph containing all environmental factor nodes and phenotypic feature nodes. The directed edges in the causal structure graph represent causal influence paths from environmental factors to phenotypic characteristics, and the weight values ​​of the directed edges are the corresponding residual squared differences.

6. The method for cultivating maize seeds based on edge computing for environmental adaptability according to claim 1, characterized in that, Step five specifically involves: Perform a traversal analysis on the causal structure graph, extract all directed paths that start from environmental factor nodes and eventually point to any phenotypic feature node, and record the node sequence and the connection direction of the directed edges for each directed path. Count the number of paths that each environmental factor node can reach to all phenotypic feature nodes, and record the path length, which is the number of directed edges contained in the path. The weight values ​​corresponding to each directed edge in each directed path are summed to obtain the total weight value of each directed path; All directed paths are sorted by a weighted average of their total weight and path length. Based on the comprehensive ranking results, the top-ranked directed paths of a predetermined number are selected as key factor paths. The key factor paths represent the dominant influence links of specific environmental factors on the target phenotypic characteristics.

7. The method for cultivating maize seeds based on edge computing for environmental adaptability according to claim 1, characterized in that, Step six specifically includes: Multidimensional environmental factor data of the target area are collected, and the environmental factor vector of the target area is generated after standardizing the values ​​of each environmental factor. The target area is the area where the maize seed breeding experiment is to be carried out. Based on the key factor paths already determined in the causal structure diagram, identify the starting node, ending node, and intermediate nodes of each key factor path, where the starting node is an environmental factor node and the ending node is a phenotypic feature node. The values ​​corresponding to the starting nodes in the environmental factor vector of the target region are used as the initial input values. The forward propagation is carried out node by node along each critical path according to the node connection direction of the critical factor path. The forward propagation is carried out according to the direction of each directed edge when constructing the causal structure graph. If multiple input values ​​in the forward propagation path point to the same node, a weighted summation operation is performed on all the input values ​​to synthesize the input value of this node and continue to propagate to the next node; When the propagation endpoint reaches the phenotypic feature node, the phenotypic feature vector corresponding to this phenotypic feature node is used as the predicted phenotypic feature vector of the target region.

8. The method for cultivating maize seeds based on edge computing for environmental adaptability according to claim 1, characterized in that, Step seven specifically includes: After completing the corn seed trial planting in the target area, the actual phenotypic feature vectors of the target area were collected. The predicted phenotypic feature vector is compared one-to-one with the actual phenotypic feature vector, and the prediction error of each phenotypic feature value is calculated. The prediction error of each phenotypic feature value is compared with a preset error threshold. Phenotypic features with errors exceeding the threshold are selected and the corresponding key factor paths in the causal structure diagram are traced. The weight of each directed edge in the corresponding key factor path is reduced, and the reduction value is the absolute value of the difference between the prediction error and the preset error threshold. Save the updated cause-effect diagram as the current version.

Images