Method and system for automatically identifying plants based on neural network model

Abstract

The invention provides an automatic plant identification method and system based on a neural network model, and relates to the technical field of intelligent agriculture, and the method comprises the steps: inputting a composite feature map into a compressed and optimized lightweight convolutional neural network model, and outputting a plant type identification result and a confidence value in real time; dynamically binding a plant type identification result with a geographic information system (GIS) coordinate, generating an endangered plant space distribution thermodynamic diagram and triggering a real-time alarm; images are automatically collected and marked based on newly-added plant sample positions marked in the spatial distribution thermodynamic diagram, and weight parameters of the lightweight convolutional neural network model are periodically updated. The plant identification and protection capability is comprehensively improved through the links of data acquisition diversification and standardization, image processing high efficiency and scientificity, feature extraction and identification accuracy, endangered plant protection and dynamic monitoring, model continuous optimization and adaptability and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent agriculture, and in particular to a method and system for automatically identifying plants based on a neural network model. Background Art

[0002] With the growing demand for ecological protection, agricultural monitoring, and biodiversity research, accurate identification of plant species has become a key technical issue. However, some traditional plant identification methods rely on manual observation of plant morphological characteristics, such as leaf shape, flower structure, and plant height, combined with knowledge of plant taxonomy for identification. This approach is not only inefficient and highly dependent on professional personnel, but is also prone to identification errors when faced with large-scale plant samples or morphologically similar species. This makes it difficult to meet the needs of modern ecological protection and agricultural production for rapid and accurate plant identification. In recent years, plant identification technology based on computer vision and machine learning has gradually emerged, realizing automatic identification of plant species through image acquisition and analysis algorithms; however, existing technologies may have limitations: First, in the image acquisition process, a single device (such as relying solely on drones or mobile phone terminals) may find it difficult to simultaneously obtain the macro-community distribution and micro-texture details of plants, resulting in missing feature information and affecting recognition accuracy; second, some images in natural environments are easily affected by factors such as uneven lighting and occlusion (such as overlapping branches and leaves, shadow coverage), and traditional image preprocessing methods may not be able to dynamically adjust processing parameters according to the scene, making it difficult to effectively eliminate image quality problems; in the field of plant distribution monitoring, some existing systems simply associate plant identification results with geographic location, which may lack consideration of data timeliness and cannot promptly reflect dynamic changes in plant distribution, making it difficult to provide effective decision-making support for the protection of endangered plants. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and system for automatic plant identification based on a neural network model, which comprehensively improves the plant identification and protection capabilities through the diversification and standardization of data collection, the efficiency and scientificity of image processing, the accuracy of feature extraction and identification, the protection and dynamic monitoring of endangered plants, and the continuous optimization and adaptability of the model.

[0004] In order to solve the above technical problems, the technical solutions of the present invention are as follows: In a first aspect, a method for automatically identifying plants based on a neural network model comprises: Step S1: Generate an image set of the overall plant morphology and community distribution through high-altitude aerial photography by a drone, and simultaneously generate an image set of the plant's leaf texture and flower structure details through a portable intelligent terminal. In the process of generating the image set, four specific fixed marking points are set within the image acquisition area, and a closed figure is formed by the four marking points. The closed figure is then segmented, and an adjustment value is generated based on the grid characteristics formed after the segmentation. Step S2: Using the adjustment value, perform targeted illumination compensation and occlusion restoration processing parameter adjustments on the image set, perform illumination compensation and occlusion restoration processing, and output a standardized image set that eliminates overexposure, blur, and occlusion interference; Step S3: performing spatial feature alignment and multi-scale fusion on the standardized image set to generate a composite feature map containing macro-morphology and micro-texture; Step S4: inputting the composite feature map into the compressed and optimized lightweight convolutional neural network model, and outputting the plant species recognition result and confidence value in real time; Step S5: Dynamically bind the plant species identification results with the geographic information system (GIS) coordinates to generate a heat map of the spatial distribution of endangered plants and trigger a real-time alarm; Step S6: Based on the newly added plant sample locations marked in the spatial distribution heat map, images are automatically collected and annotated, and the weight parameters of the lightweight convolutional neural network model are periodically updated.

[0005] Furthermore, step S1: generating an image set of the overall morphology and community distribution of plants through high-altitude aerial photography by a drone, and generating an image set of the leaf texture and flower structure details of the plants through a portable intelligent terminal; in the process of generating the image set, four specific fixed marking points are set in the image acquisition area, a closed figure is formed by the four marking points, and the closed figure is segmented, and an adjustment value is generated according to the grid characteristics formed after the segmentation, including: Four physical markers with fixed geographic coordinates were evenly distributed throughout the target vegetation area, covering the core distribution area of ​​the plant community and samples at different growth stages. Simultaneously, the following were performed: drone aerial photography was used to obtain a set of images of the overall vegetation morphology and community distribution including the markers; and a portable smart terminal was used to capture a set of images of leaf texture and flower structure details of plants around the markers. Based on the geographic coordinates of the four marker points, they are connected in spatial order to form a closed graph; Adaptively segmenting the closed figure into grids, dynamically adjusting the size of the segmented grids according to the density of plants within the closed figure, and obtaining segmented grid units; For the grid cells formed by the segmentation, based on the distribution of plant features within each grid cell, the plant coverage, the consistency of plant morphological features and the complexity of leaf texture are calculated respectively to generate a feature vector for each grid cell; The feature vector of the grid unit is input into a pre-trained weight calculation model, which extracts spatial distribution pattern features through a convolutional layer and calculates the overall attribute offset of the region through regression calculation of a fully connected layer, and outputs an adjustment value for the current image acquisition area.

[0006] Furthermore, step S2: using the adjustment value, adjusting the parameters of illumination compensation and occlusion restoration processing for the image set, performing illumination compensation and occlusion restoration processing, and outputting a standardized image set that eliminates overexposure, blur, and occlusion interference, including: Based on a preset mapping relationship table, a correspondence between the adjustment value and the illumination compensation parameter and the occlusion restoration parameter is established. The mapping relationship table includes the illumination intensity adjustment coefficient, contrast enhancement parameter, and occlusion area prediction threshold corresponding to different adjustment value intervals; According to the adjustment value output in step S1, the corresponding parameter combination is matched in the mapping relationship table to determine the illumination compensation parameters and occlusion restoration parameters that match the current image set; Using the matched illumination compensation parameters, the original image set is processed with histogram equalization and adaptive gamma correction to eliminate overexposed areas and enhance brightness uniformity, and the illumination-compensated image set is output. Based on the matched occlusion restoration parameters, the occlusion region prediction is performed on the illumination-compensated image set, i.e., superpixel segmentation regions of the image are extracted. The texture similarity difference between each region and the surrounding area is calculated. If the difference is greater than T, the region is marked as an occlusion candidate. The occlusion candidate region is then restored by edge-guided diffusion restoration, with pixels with the highest texture continuity with adjacent healthy areas being prioritized. For large occluded areas, a context encoder is used to generate structurally coherent restoration content, and the restored image set is output. Based on the restored image set, a bilateral filtering algorithm is used for noise reduction to remove the noise introduced during the processing and output a standardized image set that eliminates overexposure, blur and occlusion interference.

[0007] Furthermore, step S3: performing spatial feature alignment and multi-scale fusion on the standardized image set to generate a composite feature map containing macro-morphology and micro-texture, including: Based on a standardized image set, local key points are extracted from drone aerial images and detail images captured by portable terminals, and directional gradient histogram descriptors of the key points are calculated. The calculated directional gradient histogram descriptors are then filtered out using a bidirectional nearest neighbor matching algorithm to select descriptor pairs with a similarity greater than 0.85. An affine transformation matrix is ​​fitted based on the matching point pairs corresponding to the filtered descriptor pairs, minimizing the reprojection error. The fitted affine transformation matrix is ​​applied to the detail images, mapping them to the reference coordinate system of the aerial images, and outputting a spatially aligned image set. Taking the spatially aligned image set as input, a multi-scale feature layer set including a low-frequency feature layer and a high-frequency feature layer is generated through a multi-scale spatial decomposition operation, wherein the low-frequency feature layer is obtained by performing Gaussian pyramid downsampling on the aerial image, extracting the morphological skeleton of the vegetation coverage area, and calculating the skeleton branch angle distribution histogram as the macro-morphological feature; the high-frequency feature layer is obtained by performing Laplacian pyramid decomposition on the detail image, retaining the high-frequency components of the 2nd to 4th layers, extracting the local binary pattern histogram of the high-frequency components, and quantifying the leaf texture and petal edge sharpness; The multi-scale feature layer set is integrated through an adaptive weight fusion algorithm, and fusion weights are dynamically allocated according to the morphological discriminability of the low-frequency feature layer and the texture discriminability of the high-frequency feature layer. A composite feature map that retains both macroscopic morphological representation capabilities and microscopic texture details is generated through feature layer reconstruction.

[0008] Furthermore, step S4: inputting the composite feature map into the compressed and optimized lightweight convolutional neural network model, outputting the plant species recognition results and confidence values ​​in real time, including: Inputting the composite feature map into a compressed and optimized lightweight convolutional neural network model, extracting deep features step by step through multi-layer cascade feature transformation operations, wherein each level of feature transformation includes feature channel reorganization and spatial dimensionality reduction processing, and outputting a set of feature maps with hierarchical abstract relationships; A double enhancement process is performed on the feature map set of the hierarchical abstract relationship: Perform global average pooling on the highest-level feature map to generate a channel description vector; learn channel weights through two fully connected layers, and use the Sigmoid function to output the enhancement coefficient of each channel; multiply the enhancement coefficient by the highest-level feature map channel by channel to obtain the enhanced discriminative feature expression feature map; Spatial pyramid pooling is applied to the feature map to enhance the expression of discriminative features. Average pooling of three grid scales (4×4, 2×2, and 1×1) is performed in parallel to obtain pooling results of different scales. These pooling results of different scales are then concatenated to obtain a multi-scale contextual feature vector, which is then output as an abstract feature vector that integrates channel and spatial information. The abstract feature vector is input into the fully connected layer and mapped to the plant species dimensional space. The probability distribution of each category is generated through probability normalization processing based on the exponential function. The plant species corresponding to the maximum probability value is selected as the recognition result, and the probability value is output as the confidence.

[0009] Furthermore, step S5: dynamically binding the plant species identification results with the geographic information system (GIS) coordinates to generate a heat map of the spatial distribution of endangered plants and trigger a real-time alarm, including: Based on the plant species identification result and confidence value output in step S4, combined with the geographic information system (GIS) coordinates obtained during the drone aerial image acquisition, the identification result is dynamically bound to the geographic location data through timestamp marking and spatial coding conversion to generate a plant identification record set containing species identification, confidence value, timestamp, and geocoding fields; Based on the plant identification record set, identification records within the same geographic grid are aggregated according to species identification and geocoding fields; for each identification record within each grid, the number of records is adjusted by applying a preset time decay weight function based on the interval between the record timestamp and the current time; the number of all time decay adjusted records within the grid is accumulated to obtain a time decay record density; the time decay record density is weighted and corrected based on the confidence value of each record to generate a grid distribution intensity value; the distribution intensity values ​​of all grids are mapped to a geographic spatial coordinate system to generate a dynamically updated spatial distribution heat map of endangered plants; Based on the endangered plant spatial distribution heat map, when it is detected that the distribution intensity of endangered species in the preset protection area exceeds the preset threshold, a real-time alarm is automatically triggered, and alarm information including species location, distribution intensity and confidence is pushed to the supervision terminal.

[0010] Furthermore, step S6: based on the newly added plant sample locations marked in the spatial distribution heat map, automatically collect and annotate images, and periodically update the weight parameters of the lightweight convolutional neural network model, including: Based on the coordinates of the newly added sample locations marked in the endangered plant spatial distribution heat map, a scheduling instruction for an image acquisition device is automatically generated to control the device to capture original images containing multi-view features of the target plant at the target geographic location, and the geographic location information and acquisition timestamp are integrated to output a set of newly added plant sample images; Extracting the morphological contour boundaries of the target plant in a single image of the newly added plant sample image set; associating the plant species identification result corresponding to the image as a species label; generating a spatiotemporal identifier by combining the geographic coordinates and timestamp recorded when the image was acquired; integrating the morphological contour boundaries, species labels, and spatiotemporal identifiers to construct a structured annotation data unit; aggregating the structured annotation data units of all images to output an annotated newly added training sample set; Merge the newly labeled training sample set with the historical training set, freeze the feature extraction layer weights of the lightweight convolutional neural network, fine-tune the parameters of the fully connected classification layer, optimize the model's ability to recognize new samples through the backpropagation algorithm, and output the lightweight convolutional neural network model with updated weight parameters; After a preset time period, the lightweight convolutional neural network model with updated weight parameters is deployed to the recognition system to start a new recognition cycle.

[0011] In a second aspect, a system for automatically identifying plants based on a neural network model comprises: The acquisition module is used to generate an image set of the overall morphology and community distribution of plants through high-altitude aerial photography by a drone, and simultaneously generate an image set of plant leaf texture and flower structure details through a portable smart terminal. In the process of generating the image set, four specific fixed marking points are set in the image acquisition area, and a closed figure is formed by the four marking points. The closed figure is then segmented, and an adjustment value is generated based on the grid characteristics formed after segmentation. A normalization module is used to adjust the parameters of the illumination compensation and occlusion restoration processing of the image set using the adjustment value, and perform illumination compensation and occlusion restoration processing to output a normalized image set that eliminates overexposure, blur, and occlusion interference; The fusion module is used to perform spatial feature alignment and multi-scale fusion on the standardized image set to generate a composite feature map containing macro-morphology and micro-texture; The confidence module is used to input the composite feature map into the compressed and optimized lightweight convolutional neural network model and output the plant species recognition results and confidence values ​​in real time; Dynamic binding module, used to dynamically bind plant species identification results with geographic information system (GIS) coordinates, generate a heat map of the spatial distribution of endangered plants, and trigger real-time alerts; The periodic module is used to automatically collect and annotate images based on the locations of newly added plant samples marked in the spatial distribution heat map, and periodically update the weight parameters of the lightweight convolutional neural network model.

[0012] According to a third aspect, a computing device includes: one or more processors; The storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0013] In a fourth aspect, a computer-readable storage medium stores a program, which implements the method when executed by a processor.

[0014] The above solution of the present invention includes at least the following beneficial effects: Drone aerial photography captures the macroscopic morphology of plant communities, while portable terminals capture microscopic details. The combination of the two provides complete multi-scale feature information, avoids the loss of features from a single perspective, and improves the ability to distinguish similar plants and the accuracy of recognition. Marker points are set to divide the grid to generate adjustment values, dynamically adjust illumination compensation and occlusion repair parameters, eliminate image overexposure, blur, and occlusion problems, and achieve image quality standardization. Spatial feature alignment eliminates perspective differences, and multi-scale fusion integrates macro and micro features to generate a composite feature map. A lightweight convolutional neural network with compression optimization is used to reduce the amount of calculation and model parameters while ensuring recognition accuracy, achieving real-time recognition of plant species and meeting the rapid processing requirements of scenarios such as drone inspections. Recognition results are combined with GIS coordinates to generate a heat map, and the time decay mechanism reflects the latest distribution. Automatic alarms are triggered when the intensity of endangered species exceeds the threshold, and protected areas are accurately located. New samples are automatically collected and labeled based on the heat map, the model is periodically updated, and the classification layer is fine-tuned to enable the model to quickly adapt to new features and continuously optimize recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of a method for automatically identifying plants based on a neural network model provided by an embodiment of the present invention.

[0016] Figure 2 1 is a schematic diagram of a system for automatically identifying plants based on a neural network model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0018] like Figure 1 As shown, an embodiment of the present invention provides a method for automatically identifying plants based on a neural network model, the method comprising the following steps: Step S1: Generate an image set of the overall plant morphology and community distribution through high-altitude aerial photography by a drone, and simultaneously generate an image set of the plant's leaf texture and flower structure details through a portable intelligent terminal. In the process of generating the image set, four specific fixed marking points are set within the image acquisition area, and a closed figure is formed by the four marking points. The closed figure is then segmented, and an adjustment value is generated based on the grid characteristics formed after the segmentation. Step S2: Using the adjustment value, perform targeted illumination compensation and occlusion restoration processing parameter adjustments on the image set, perform illumination compensation and occlusion restoration processing, and output a standardized image set that eliminates overexposure, blur, and occlusion interference; Step S3: performing spatial feature alignment and multi-scale fusion on the standardized image set to generate a composite feature map containing macro-morphology and micro-texture; Step S4: inputting the composite feature map into the compressed and optimized lightweight convolutional neural network model, and outputting the plant species recognition result and confidence value in real time; Step S5: Dynamically bind the plant species identification results with the geographic information system (GIS) coordinates to generate a heat map of the spatial distribution of endangered plants and trigger a real-time alarm; Step S6: Based on the newly added plant sample locations marked in the spatial distribution heat map, images are automatically collected and annotated, and the weight parameters of the lightweight convolutional neural network model are periodically updated.

[0019] In an embodiment of the present invention, macroscopic and microscopic data are acquired through drone aerial photography and portable smart terminals, and rasterized calibration data of regional marker points is combined; adjustment values ​​are used to perform targeted illumination compensation and occlusion repair to generate a standardized image set, thereby improving image clarity and enhancing data applicability; multi-scale features are integrated to accurately capture the unique characteristics of plant species, improve recognition accuracy, and use a lightweight model to achieve real-time and efficient recognition, reducing computing resource consumption; the recognition results are bound to GIS to generate a heat map and set alarms, visually displaying the distribution status and change trends of endangered plants; samples are automatically collected based on the heat map to update model parameters, ensuring continuous optimization of the system and improving plant recognition and protection capabilities.

[0020] In a preferred embodiment of the present invention, step S1: generating an image set of the overall morphology and community distribution of plants through high-altitude aerial photography by a drone, and simultaneously generating an image set of plant leaf texture and flower structure details through a portable intelligent terminal; in the process of generating the image set, four specific fixed marking points are set in the image acquisition area, a closed figure is formed by the four marking points, and the closed figure is segmented. Based on the grid characteristics formed after the segmentation, an adjustment value is generated, which may include: Step S11: evenly distribute four physical markers with fixed geographic coordinates in the target vegetation area, covering the core distribution area of ​​the plant community and samples at different growth stages; simultaneously execute: using drone aerial photography to obtain a set of images of the overall vegetation morphology and community distribution including the markers; and using a portable smart terminal to capture a set of image details of the leaf texture and flower structure of plants around the markers; Step S12, based on the geographic coordinates of the four marking points, connect them in spatial order to form a closed figure; Step S13, performing adaptive grid segmentation on the closed figure, dynamically adjusting the grid size after segmentation according to the plant density within the closed figure, and obtaining grid units formed by segmentation; Step S14, for the grid cells formed by the segmentation, based on the distribution of plant features in each grid cell, respectively calculating plant coverage, plant morphological feature consistency and leaf texture complexity, and generating a feature vector for each grid cell; In step S15, the feature vector of the grid unit is input into a pre-trained weight calculation model. The weight calculation model extracts the spatial distribution pattern features through the convolution layer and calculates the overall attribute offset of the region through regression calculation of the fully connected layer, and outputs the adjustment value for the current image acquisition area.

[0021] In an embodiment of the present invention, the selection of marker points takes into account the diversity of plant communities, ensures sample representativeness to reduce recognition errors, and lays a data foundation for image calibration and feature analysis; constructs clear image analysis area boundaries to facilitate targeted processing of closed area image data, improves data processing efficiency and accuracy, and defines the scope for grid segmentation and feature extraction; flexibly adjusts the grid size according to the actual distribution of plants, uses small grids to accurately capture features in densely planted areas, and uses large grids to improve computing efficiency in sparse areas, thereby balancing processing accuracy and efficiency; quantifies plant features within the grid from multiple dimensions to form a comprehensive feature description, providing data support for adjustment value calculation; and automatically generates actual adjustment values ​​with the help of a weight calculation model.

[0022] In the embodiments of the present invention, when applied specifically, it can be achieved through the following technical solutions, for example: In the above step S11, the geographic coordinate system obtained from the high-altitude aerial images of the UAV is used to clarify the specific scope of the image acquisition area, and the area is accurately marked on the map through professional geographic information system software. At the same time, the boundaries are verified and fine-tuned in combination with the actual topography.

[0023] Carry out on-site surveys of the flora, bring professional plant illustrations, measuring tools and recording equipment, and conduct systematic inspections within the designated area; distinguish target plants at different growth stages, such as seedlings, mature plants, and flowering plants, based on indicators such as plant morphological characteristics, growth height, leaf color and shape, and flower status, and comprehensively record their distribution locations, growth conditions, and quantity information.

[0024] After completing the survey, the grid division method is used to evenly divide the area into several small grids within the typical area of ​​plant community distribution. Four representative grids are selected through a comprehensive evaluation of factors such as the richness of the plant community, species composition, and the proportion of the growth stage of the target plants in each grid. At the center of the selected grid, positioning equipment (such as a high-precision GPS locator) is used to obtain the geographic coordinates of the point and keep detailed records.

[0025] Within a radius of 1-2 meters around each marking point, check the sample status of the plant growth stage again. If there is a problem of missing or unrepresentative samples, adjust the position of the marking point appropriately until the point covers samples of all major growth stages of the target plant; set up obvious and durable markers at the marking points, such as metal stakes with numbers and positioning information, landmark flags of special colors, etc., to facilitate subsequent image acquisition and data processing to quickly and accurately find the corresponding marking points. At the same time, the detailed information of each marking point (including geographic coordinates, surrounding plant growth conditions, marker characteristics, etc.) should be organized into documents and properly preserved.

[0026] In the above step S12, the specific coordinate positions of the four marking points in the geographic coordinate system are obtained, and the four marking points are connected in sequence by means of coordinate lines to form a closed figure. The shape of the closed figure (such as a quadrilateral, trapezoid, etc.) depends entirely on the spatial position relationship of the four marking points.

[0027] In the above step S13, the distribution of plants within the closed figure is analyzed, and the number of plants in different areas is counted to determine the plant density. For areas with high plant density, the grid size is set to a smaller value. For example, in areas with dense plant groups, the grid is divided into squares with smaller side lengths. In sparse areas with low plant density, the grid size is increased, such as being set to squares with larger side lengths. In this way, the closed figure is adaptively segmented.

[0028] In the above step S14, for each segmented grid, the ratio of the area occupied by plants in the grid to the total area of ​​the grid is calculated to obtain the plant coverage; the similarity of plant morphology in the grid is analyzed, such as the consistency of features such as plant height and leaf shape, to evaluate the consistency of morphological features; the texture complexity is obtained by calculating the complexity of plant texture in the grid, such as the complexity of leaf veins and the fineness of flower patterns; and these three indicators are integrated into a feature vector.

[0029] In the above step S15, a multi-dimensional plant feature vector is determined as input, including key indicators such as plant coverage (such as the proportion of leaf area), morphological feature consistency (such as the standard deviation of plant height), and texture complexity (such as the grayscale change value of leaf veins), while integrating environmental parameters such as geographic coordinates and light intensity to form structured input data; the adjustment value is set as the model output target, which needs to reflect the comprehensive characteristics of the image acquisition area (such as brightness correction coefficient, resolution adjustment parameter, etc.) for parameter calibration of subsequent image processing.

[0030] Based on the nonlinear correlation characteristics of plant features, a multi-layer neural network architecture (such as a fully connected neural network or a convolutional neural network) is adopted to abstractly map the feature vectors through hidden layers. For example, the convolutional layer is used to extract image texture features, and the fully connected layer integrates the weight relationship between coverage and morphological features, enabling the model to process multi-source data.

[0031] Integrate plant image data from different regions, seasons, and growth stages, covering typical scenes such as seedlings, mature plants, and flowering periods, and simultaneously collect corresponding geographic coordinates, environmental parameters, and manually annotated adjustment values ​​(such as optimization parameters manually set by experts based on image quality); expand the image dataset through rotation, scaling, brightness adjustment, etc., simulate different shooting angles and lighting conditions, and improve the model's adaptability to complex scenes; for example, add Gaussian noise to the same plant image to generate multiple groups of samples with different texture complexity.

[0032] The dataset is divided into training set, validation set, and test set in a ratio of 7:2:1 to ensure that each subset contains sample distributions at different growth stages and coverage. When initializing the model weights, random normal distribution is used to avoid training falling into local optimality due to initial parameter deviation.

[0033] The loss function is designed with the mean square error between the predicted adjustment value and the manually labeled value as the optimization target, combined with regularization terms (such as L2 regularization) to suppress model overfitting and ensure the generalization ability of weight parameters for new samples; gradient descent optimization uses a batch gradient descent algorithm, inputs training data in batches, and updates the weights of each layer through the back propagation algorithm. For example, when the model predicts a large deviation in the adjustment value of the high coverage area, it automatically increases the update amplitude of the weight parameter corresponding to the coverage feature.

[0034] During the training process, the validation set is used to periodically evaluate the model performance and adjust hyperparameters such as the learning rate and the number of hidden layer nodes. For example, if the weight of the texture complexity feature is always low, the model's sensitivity to texture information can be enhanced by increasing the diversity of this feature in the training samples.

[0035] The trained model is finally evaluated using the test set, with a focus on verifying the accuracy of its predictions of adjustment values ​​in non-training areas (such as new plant communities and extreme climate scenarios). If the model has large prediction errors in sparsely vegetated areas, such samples are supplemented and retrained.

[0036] In a preferred embodiment of the present invention, step S2: using the adjustment value to adjust the parameters of the illumination compensation and occlusion restoration processing for the image set, performing illumination compensation and occlusion restoration processing, and outputting a standardized image set with overexposure, blur, and occlusion interference eliminated, may include: Step S21: establishing a correspondence between the adjustment value and the illumination compensation parameter and the occlusion restoration parameter based on a preset mapping relationship table, wherein the mapping relationship table includes illumination intensity adjustment coefficients, contrast enhancement parameters, and occlusion area prediction thresholds corresponding to different adjustment value intervals; Step S22, matching corresponding parameter combinations in the mapping relationship table according to the adjustment values ​​outputted in step S1, and determining illumination compensation parameters and occlusion restoration parameters matched to the current image set; Step S23, using the matched illumination compensation parameters, performing histogram equalization and adaptive gamma correction on the original image set to eliminate overexposed areas and enhance brightness uniformity, and outputting an illumination-compensated image set; Step S24: Based on the matched occlusion restoration parameters, occlusion region prediction is performed on the illumination-compensated image set, i.e., superpixel segmentation regions of the image are extracted. The texture similarity difference between each region and the surrounding region is calculated. If the difference is greater than T, the region is marked as an occlusion candidate. The occlusion candidate region is then restored by performing edge-guided diffusion restoration on the candidate region, prioritizing pixels with the highest texture continuity with adjacent healthy regions. For large occluded regions, a context encoder is used to generate structurally coherent restoration content, and the restored image set is output. In step S25 , a bilateral filtering algorithm is used to perform noise reduction processing based on the restored image set to remove the noise introduced during the processing and output a standardized image set that eliminates overexposure, blur and occlusion interference.

[0037] In an embodiment of the present invention, a preset mapping relationship table is used to achieve standardization and automation of parameter adjustment, avoid the randomness of manual debugging, and ensure that different adjustment values ​​correspond to optimal processing parameters; dynamically match parameters based on the adjustment values ​​to make illumination compensation and occlusion repair more in line with the actual characteristics of the image; enhance global brightness uniformity through histogram equalization, and combine adaptive gamma correction to repair local overexposure or underexposure, so that the image brightness is closer to the visual effect of the human eye; identify and repair occluded areas to avoid the loss of plant features due to leaf overlap or foreign object occlusion, and ensure the integrity of plant morphology in the image; bilateral filtering retains the edges and texture details of plant leaves while reducing noise, avoiding image blurring caused by traditional Gaussian filtering, so that the standardized image set is clean without losing feature details.

[0038] In the embodiments of the present invention, when applied specifically, it can be achieved through the following technical solutions, for example: In the above step S21, the adjustment value is divided into several intervals (such as 0-2, 2-4, 4-6, etc.) according to the numerical range, and each interval corresponds to image characteristics under different scenes (such as low light, medium occlusion, high overexposure, etc.).

[0039] For each interval, the optimal parameters are determined through historical data statistics and manual debugging. For example, when the adjustment value ∈ [0-2], the light intensity adjustment coefficient is set to +15% (brightening), the contrast enhancement parameter is set to 1.2 (enhancing details), and the occlusion prediction threshold is set to 30 (a lower threshold, prioritizing the detection of small-area occlusions); when the adjustment value ∈ [4-6], the light intensity adjustment coefficient is -10% (darkening overexposed areas), the contrast parameter is 0.8, and the occlusion threshold is increased to 50 (focusing on detecting large-area occlusions); the adjustment value interval and parameter combination are stored in correspondence to form a mapping relationship table that can be directly queried.

[0040] In step S22 , the adjustment value of the current image set (e.g., 3.5) is obtained, the interval range of the mapping table is traversed, and the interval to which it belongs (e.g., 2-4) is determined; a preset parameter combination (e.g., illumination coefficient +5%, contrast 1.0, occlusion threshold 40) is taken from the corresponding interval to form the processing parameter set for the current image.

[0041] In step S23, the pixel distribution of each brightness value of the image is counted to generate a brightness histogram; the original brightness values ​​are mapped to a new range to make the pixel distribution of bright and dark areas more uniform (for example, the brightness values ​​concentrated in the dark area are stretched to the full range), thereby eliminating the phenomenon of local over-darkness or over-brightness.

[0042] The image is divided into multiple sub-regions (such as 8×8 pixel blocks). The average brightness of each sub-region is calculated. If the brightness is lower than a threshold (such as 50), a smaller gamma value (such as 0.8) is applied to brighten it. If it is higher than a threshold (such as 200), a larger gamma value (such as 1.2) is applied to darken it, dynamically adjusting the brightness curve.

[0043] In step S24 above, a superpixel segmentation algorithm (such as SLIC or TurboPixels) is used to divide the illumination-compensated image set into multiple compact small regions with similar characteristics (color, texture, and brightness). Each region is called a superpixel. Spatial proximity and feature consistency must be maintained during segmentation to ensure the accuracy of subsequent texture analysis.

[0044] For each superpixel region, its "surrounding area" is defined as the adjacent superpixels (usually four or eight neighborhoods); the texture features of each superpixel region and its surrounding areas are extracted using local texture description methods (such as gray-level co-occurrence matrix, local binary pattern LBP or Gabor filter); the texture feature similarity between the current region and each surrounding region is calculated (such as Euclidean distance, cosine similarity), and the average value is taken as the overall similarity between the region and the surrounding area; the texture similarity difference between the current region and the surrounding area is calculated (such as the absolute difference between the current region similarity and the surrounding average similarity).

[0045] A threshold T is set. If the texture similarity difference between a superpixel area and its surroundings is greater than T, the area is determined to be a possible occlusion area and marked as an "occlusion candidate area".

[0046] For marked occlusion candidate areas, edge pixels are prioritized: Pixels at the edge of the candidate area are extracted and analyzed for edge continuity (e.g., edge direction and gradient consistency) with adjacent healthy (non-occluded) areas. Filling is performed pixel by pixel from the edge of the candidate area toward the interior, selecting pixel values ​​that have the highest continuity with the texture features (color, texture pattern) of the adjacent healthy area. For example, by searching for pixels in the healthy area that are most similar to edge pixels, their values ​​are diffused into the interior of the candidate area to maintain texture and structural coherence. The filling process follows an "outside-to-in" order to ensure a natural transition between the repaired area and its surroundings, avoiding obvious boundaries.

[0047] For occlusion candidates exceeding a set threshold (e.g., occupying >5% of the image), a context encoder model is used. Image blocks of the occluded area and its surrounding healthy areas are fed into the model. The model learns the global structure and local texture features of the image to generate restoration content that is structurally coherent with the surrounding content. For example, the model predicts plausible content in the occluded area (e.g., missing wall texture, object edges) based on the contours and texture patterns of surrounding objects. The resulting restoration content is then fused with the surrounding area, adjusting the pixel value transition boundaries to ensure color and brightness consistency between the restoration area and the original image, avoiding visual abruptness.

[0048] After completing the restoration process for all occluded candidate areas, all areas of the image are integrated to generate the final set of restored images, ensuring that the occluded areas of each image are naturally filled and the overall visual effect is coherent.

[0049] In the above step S25, for each pixel, the Euclidean distance between it and its neighboring pixels is calculated. The closer the distance, the higher the weight (e.g., in a 3×3 neighborhood, the weight of the center pixel is 1, the weight of the adjacent pixels is 0.8, and the weight of the diagonal pixels is 0.5).

[0050] Compare the grayscale value difference between the current pixel and the neighboring pixels. The smaller the difference, the higher the weight (for example, the weight is 1 when the grayscale difference is ≤ 10, and the weight is 0.2 when the difference is ≥ 30). Multiply the spatial distance weight by the grayscale difference weight to obtain a comprehensive weight. Update the current pixel value after taking the weighted average of the neighboring pixels to preserve the edge while filtering out noise (for example, removing the salt and pepper noise introduced by lighting processing).

[0051] In a preferred embodiment of the present invention, the above step S3: performing spatial feature alignment and multi-scale fusion on the standardized image set to generate a composite feature map containing macro-morphology and micro-texture may include: Step S31: Based on the standardized image set, local key points of the drone aerial image and the detail image taken by the portable terminal are extracted, and directional gradient histogram descriptors of the key points are calculated; the calculated directional gradient histogram descriptors are filtered out using a bidirectional nearest neighbor matching algorithm to select descriptor pairs with a similarity greater than 0.85; an affine transformation matrix is ​​fitted based on the matching point pairs corresponding to the filtered descriptor pairs, and the reprojection error is minimized; the fitted affine transformation matrix is ​​applied to the detail image, the detail image is mapped to the reference coordinate system of the aerial image, and the spatially aligned image set is output; Step S32: Using the spatially aligned image set as input, a multi-scale feature layer set including a low-frequency feature layer and a high-frequency feature layer is generated through a multi-scale spatial decomposition operation, wherein the low-frequency feature layer is obtained by performing Gaussian pyramid downsampling on the aerial image, extracting the morphological skeleton of the vegetation coverage area, and calculating the skeleton branch angle distribution histogram as the macro-morphological feature; and the high-frequency feature layer is obtained by performing Laplacian pyramid decomposition on the detail image, retaining the high-frequency components of the 2nd to 4th layers, extracting the local binary pattern histogram of the high-frequency components, and quantifying the leaf texture and petal edge sharpness. In step S33, the multi-scale feature layer set is integrated through an adaptive weight fusion algorithm, fusion weights are dynamically allocated according to the morphological discriminability of the low-frequency feature layer and the texture discriminability of the high-frequency feature layer, and a composite feature map is generated by feature layer reconstruction that retains both macroscopic morphological representation capabilities and microscopic texture details.

[0052] In an embodiment of the present invention, spatial alignment is used to eliminate perspective deviations and position differences between different images, so that the same parts of the same plant are in a relatively fixed position in the image, laying a spatial consistency foundation for subsequent multi-image feature fusion and avoiding feature extraction errors caused by spatial dislocation; multi-scale fusion simultaneously retains the macroscopic morphology and microscopic texture of the plant, avoiding feature loss caused by single-scale analysis; the adaptive weight mechanism enhances the expression of key features, suppresses background noise, and enables the composite feature map to more comprehensively reflect the multi-dimensional characteristics of the plant, thereby improving the accuracy of subsequent classification and recognition tasks.

[0053] In the embodiments of the present invention, when applied specifically, it can be achieved through the following technical solutions, for example: In the above step S31, a key point detection algorithm (such as SIFT, SURF or ORB) is applied to the drone aerial image and the detail image taken by the portable terminal in the standardized image set to extract local key points.

[0054] For each key point, its Histogram of Oriented Gradients (HOG) descriptor is calculated, and the local image gradient direction distribution around the key point is quantized into a multi-dimensional vector for subsequent matching.

[0055] A bidirectional nearest neighbor matching algorithm is used to match the descriptors of the aerial image and the detail image: for each descriptor in the aerial image, the two closest and second closest descriptors are found in the detail image, and their distance ratio is calculated; if the ratio is less than 0.85, the match is considered reliable and retained as a candidate matching point pair.

[0056] Based on the screened matching point pairs, the random sample consensus (RANSAC) algorithm is used to fit the affine transformation matrix, which contains rotation, translation, and scaling parameters and is used to map the detail image to the coordinate system of the aerial image. During the iterative optimization process, the reprojection error of the matching point pairs (i.e., the Euclidean distance between the transformed detail image points and the corresponding points in the aerial image) is minimized to eliminate erroneous matching points and improve the transformation accuracy.

[0057] Using the fitted affine transformation matrix, each pixel in the detail image is mapped to the reference coordinate system of the aerial image to generate a spatially aligned image set, ensuring that the detail image and the aerial image are accurately spatially corresponding.

[0058] In step S32, the spatially aligned aerial image is downsampled using a Gaussian pyramid: image layers of different resolutions are generated by applying Gaussian filtering and downsampling multiple times; the morphological skeleton of the vegetation-covered area is extracted from the low-resolution layer of the pyramid: the vegetation area is simplified to a centerline representation through morphological erosion and thinning operations, preserving its topological structure; and the angular distribution histogram of the skeleton branches is calculated: the directional angles of each branch in the skeleton are counted and quantified into histogram features, which serve as macroscopic morphological features of the vegetation (such as crown shape and branch direction).

[0059] The spatially aligned detail image is subjected to Laplacian pyramid decomposition: the high-frequency components (detail information) of each layer are obtained by subtracting adjacent layers of the Gaussian pyramid; the high-frequency components of layers 2-4 are retained, which contain medium-scale detail features such as leaf texture and petal edges; for each retained high-frequency component, a local binary pattern (LBP) histogram is calculated: each pixel is compared with its neighborhood to generate a binary pattern, the pattern distribution is statistically analyzed, and the roughness of leaf texture and the sharpness of petal edges are quantified.

[0060] In the above step S33, the morphological discriminability of the low-frequency feature layer is calculated: the effectiveness of the morphological skeleton and angular distribution histogram in distinguishing different vegetation types (such as trees and shrubs) is analyzed, and quantified by indicators such as inter-class variance or information gain; the texture discriminability of the high-frequency feature layer is calculated: the ability of the LBP histogram to distinguish subtle differences between similar vegetation (such as healthy leaves and diseased leaves) is evaluated, and the classification performance indicator is also used for quantification.

[0061] Based on the evaluation results of morphological discriminability and texture discrimination, fusion weights are dynamically assigned to the low-frequency and high-frequency feature layers: if the morphological features are more important to the current classification task, the weight of the low-frequency layer is increased; if the texture details are more critical, the weight of the high-frequency layer is increased; weight allocation can be achieved through learning algorithms (such as support vector machines, neural networks) or heuristic methods (such as based on task prior knowledge).

[0062] The multi-scale feature layer set is linearly combined or nonlinearly fused according to the assigned weights (such as weighted summation, cascading and passing through a fully connected layer); the fused feature layer is restored to the image space through inverse transformation (such as Laplacian pyramid reconstruction) to generate a composite feature map that retains both macro-morphology and micro-texture for subsequent analysis (such as vegetation classification and health assessment).

[0063] In a preferred embodiment of the present invention, the above step S4: inputting the composite feature map into the compressed and optimized lightweight convolutional neural network model to output the plant species recognition result and confidence value in real time may include: Step S41: Input the composite feature map into a compressed and optimized lightweight convolutional neural network model, extract deep features step by step through multi-layer cascade feature transformation operations, wherein each level of feature transformation includes feature channel reorganization and spatial dimensionality reduction processing, and output a set of feature maps with hierarchical abstract relationships; Step S42: performing a double enhancement process on the feature map set of the hierarchical abstract relationship: Perform global average pooling on the highest-level feature map to generate a channel description vector; learn channel weights through two fully connected layers, and use the Sigmoid function to output the enhancement coefficient of each channel; multiply the enhancement coefficient by the highest-level feature map channel by channel to obtain the enhanced discriminative feature expression feature map; Spatial pyramid pooling is applied to the feature map to enhance the expression of discriminative features. Average pooling of three grid scales (4×4, 2×2, and 1×1) is performed in parallel to obtain pooling results of different scales. These pooling results of different scales are then concatenated to obtain a multi-scale contextual feature vector, which is then output as an abstract feature vector that integrates channel and spatial information. In step S43, the abstract feature vector is input into the fully connected layer and mapped to the plant species dimensional space. The probability distribution of each category is generated through probability normalization processing based on an exponential function. The plant species corresponding to the maximum probability value is selected as the recognition result, and the probability value is output as the confidence value.

[0064] In an embodiment of the present invention, a lightweight convolutional neural network model is used to compress the model parameters and improve the calculation speed while maintaining the feature extraction capability, thereby meeting the real-time processing requirements in drone aerial photography scenarios; multi-level feature extraction enables the model to capture both micro-textures and understand macro-semantics, providing multi-dimensional feature support for subsequent recognition; the channel attention mechanism automatically selects the most effective feature channels for recognition, thereby improving the pertinence of feature expression; spatial pyramid pooling solves the traditional CNN's dependence on fixed input size, and through multi-scale context integration, enables the model to adapt to plant images at different shooting distances. The Softmax function is used to convert the feature vector into an interpretable probability distribution, and the confidence value provides a reliability assessment for the recognition result, facilitating the system to select high-confidence results or perform secondary processing on low-confidence results; the nonlinear mapping capability of the fully connected layer enables the model to learn the implicit association between plant features and categories, thereby improving classification accuracy.

[0065] In the embodiments of the present invention, when applied specifically, it can be achieved through the following technical solutions, for example: In step S41 above, the core goal of the lightweight convolutional neural network is to significantly reduce the number of parameters and computational complexity through compression techniques (channel pruning, weight quantization) and efficient modules (depthwise separable convolution, channel shuffling) while maintaining classification accuracy; The standard convolution is decomposed into depthwise convolution (DepthwiseConv) and pointwise convolution (PointwiseConv); depthwise convolution applies a convolution kernel to each input channel separately to extract only spatial features; pointwise convolution uses 1×1 convolution to integrate information between channels; compared with standard convolution, the number of parameters and computational complexity is reduced to the original 1 / N+1 / M (N is the number of input channels, M is the number of output channels).

[0066] The channels of the feature map are rearranged after group convolution to break the independence between channels and enhance the flow of information across groups. For example, after dividing the channels into g groups, the channels are shuffled through reshaping and transposition operations to enable channels from different groups to interact.

[0067] Hierarchical feature extraction architecture First level transformation (H1): Input: composite feature map (size W×H×C in ); Operation: Apply depthwise separable convolution, output feature map size W×H×C1; Divide C1 channels into g groups, each with C1 / g channels, and perform channel shuffling; use convolution or maximum pooling with a stride of 2 to reduce the spatial size to W / 2×H / 2×C1 to generate H1.

[0068] Second level transformation (H2): Input: H1 (size W / 2 × H / 2 × C1); Operation: Repeat depth-wise separable convolution + channel shuffling, output feature map size W / 2×H / 2×C2 (C2>C1); use stride 2 again to reduce the spatial size to W / 4×H / 4×C2, generating H2.

[0069] The third level transformation (H3): Input: H2 (size W / 4×H / 4×C2); Operation: Apply residual connections (adding H2 to the transformed H2) or attention mechanisms (such as SE modules) to enhance feature representation.

[0070] Finally, spatial dimensionality reduction (such as global average pooling) is performed to generate H3 (size 1×1×C3) to capture global semantic information.

[0071] During training, the importance of each channel is evaluated (e.g., through gradients or activation values), and channels with smaller contributions are removed. For example, C2 is pruned from 512 channels to 256 channels to reduce the amount of subsequent computation.

[0072] Convert 32-bit floating-point weights to 8-bit integers or binary values ​​(±1) to reduce storage requirements and computational complexity. For example, using INT8 quantization reduces memory access by a factor of four and increases multiplication speed.

[0073] Input layer → H1: The composite feature map extracts preliminary features through depthwise separable convolution (with BatchNorm and ReLU activation); the channel shuffling operation reorganizes the channels to enhance feature expression; spatial dimensionality reduction (with a step size of 2) reduces the size of the feature map and reduces the amount of computation.

[0074] H1→H2: Repeat depth-wise separable convolution + channel shuffling to further extract features; then perform spatial dimensionality reduction again to make the feature map more abstract and capture information from a larger area.

[0075] H2→H3: Residual connection or attention mechanism enhances the semantic expression ability of features and retains important information; global pooling compresses the spatial dimension to 1×1 and generates a feature vector H3 containing the global context.

[0076] By gradually reducing the spatial resolution (W / 2→W / 4→1), the computational complexity of subsequent layers is reduced; depthwise separable convolution and channel shuffling significantly reduce the number of parameters through parameter sharing and grouped calculation; operations of different scales (such as depthwise convolution and point-by-point convolution) can be executed in parallel to improve hardware utilization.

[0077] In the above step S42, global average pooling is performed on each channel of the H3 feature map, the spatial dimension is compressed to 1×1, a channel description vector is generated, and the global statistical information of each channel is retained; the channel description vector is input into two fully connected layers: the first layer uses the ReLU activation function to reduce the dimension, and the second layer restores the original number of channels.

[0078] The output value is mapped to the range of 0-1 through the Sigmoid function to generate an enhancement coefficient for each channel to characterize the importance of the channel; the enhancement coefficient is multiplied by the H3 feature map channel by channel to amplify the response of important channels, suppress redundant channels, and improve the contribution of classification-related features.

[0079] Three scales of average pooling are performed in parallel on the enhanced H3 feature map: 4×4 grid, 2×2 grid, and 1×1 grid, capturing local, medium, and global contextual information, respectively. The results of pooling at different scales are flattened into vectors and then concatenated to form a multi-scale contextual feature vector that contains rich spatial dependencies.

[0080] The concatenated feature vectors are further fused through a fully connected layer or a 1×1 convolution to generate the final abstract feature vector, which encodes channel importance and spatial context information.

[0081] In the above step S43, the abstract feature vector is input into the fully connected layer, and the feature dimension is mapped to the dimensional space of the number of plant species, where each dimension corresponds to a plant category; a probability normalization function based on an exponential function (such as Softmax) is applied to the output of the fully connected layer to convert the original score into a probability value between 0 and 1, and the sum of all category probabilities is 1.

[0082] The category with the largest probability value is selected as the final recognition result. The corresponding probability value is the confidence value, which indicates the degree of certainty of the model for the classification result.

[0083] The output format is: {plant species name, confidence value}, for example: {"Ginkgo",0.92}.

[0084] If the confidence level is lower than the set threshold (e.g. 0.7), a secondary identification or manual verification process can be triggered to improve identification reliability.

[0085] In a preferred embodiment of the present invention, the above step S5: dynamically binding the plant species identification results with the geographic information system (GIS) coordinates to generate a heat map of the spatial distribution of endangered plants and trigger a real-time alarm, may include: Step S51: Based on the plant species identification result and confidence value outputted in step S4, combined with the Geographic Information System (GIS) coordinates acquired during drone aerial image acquisition, the identification result is dynamically bound to the geographic location data through timestamp marking and spatial coding conversion to generate a plant identification record set containing species identification, confidence value, timestamp, and geocoding fields. Step S52: Based on the plant identification record set, identification records within the same geographic grid are aggregated according to species identification and geocoding fields; for each identification record within each grid, the number of records is adjusted by applying a preset time decay weight function based on the interval between the record timestamp and the current time; the number of all time decay adjusted records within the grid is accumulated to obtain a time decay record density; the time decay record density is weighted and modified based on the confidence value of each record to generate a grid distribution intensity value; the distribution intensity values ​​of all grids are mapped to a geographic spatial coordinate system to generate a dynamically updated endangered plant spatial distribution heat map; Step S53, based on the endangered plant spatial distribution heat map, when it is detected that the distribution intensity of endangered species in the preset protection area exceeds the preset threshold, a real-time alarm is automatically triggered, and alarm information including species location, distribution intensity and confidence is pushed to the monitoring terminal.

[0086] In this embodiment of the present invention, standardized grid coding and time indexing are used to convert plant identification results into traceable and comparable spatiotemporal data units, providing a foundation for subsequent spatial analysis while supporting historical data backtracking and trend analysis. Heat maps visually display the spatial distribution hotspots of endangered plants, helping managers quickly locate priority areas for protection. A time decay mechanism ensures that heat maps reflect the latest distribution dynamics, preventing historical data from interfering with real-time decision-making. An automated alarm mechanism enables real-time response to endangered plant monitoring, while a hierarchical strategy ensures that management resources are prioritized in high-risk areas. The combination of spatial overlay analysis and dynamic thresholds improves the accuracy and pertinence of alarms and reduces the cost of manual inspections.

[0087] In the embodiments of the present invention, when applied specifically, it can be achieved through the following technical solutions, for example: In the above step S51, a mapping relationship is established between the plant identification result (species name, confidence level) and the synchronously collected GIS coordinates (latitude and longitude), ensuring that each identification record corresponds to a unique geographic location.

[0088] Due to the time difference between image acquisition, transmission and recognition processing, for example, after the GPS positioning data is collected, there may be a delay of several seconds between the image transmission to the recognition system and the completion of processing. The system needs to integrate the plant recognition results, GIS coordinates and their corresponding timestamps; by comparing the timestamps, the recognition results with the closest time are associated with the GIS coordinates; for example, if the GPS positioning timestamp is "2025-06-16T10:30:25.000Z" and the plant recognition result timestamp is "2025-06-16T10:30:25.123Z", and the time difference between the two is within an acceptable range (such as no more than 1 second), then this set of data is considered a valid match, and a unique correspondence is established to ensure that each plant recognition result can be accurately associated with the geographic location at the time of collection.

[0089] A global discrete grid system (such as H3 and S2) is used to process the matched longitude and latitude coordinates. Taking the H3 grid system as an example, it divides the earth's surface into hexagonal grids of uniform size according to the set grid resolution (such as 1km×1km). The system uses a specific spatial calculation method to determine which hexagonal grid each geographic location falls into and assigns a unique identifier to the grid (such as `8928308280fffff`). This process is like covering the earth's surface with a hexagonal "coordinate grid". Each longitude and latitude coordinate can find its corresponding "grid room number", realizing standardized coding of geographic locations.

[0090] Combine grid codes with timestamps to generate a composite index. For example, combine the grid identifier `8928308280fffff` with the timestamp "2025-06-16T10:30:25Z" to create `8928308280fffff#2025-06-16T10:30:25Z`. This allows the system to quickly locate and filter relevant records, regardless of whether data is queried by geographic location (grid code) or by time range, enabling efficient management of spatiotemporal data.

[0091] The plant identification results (species name, confidence), grid code, timestamp, and original latitude and longitude coordinates are encapsulated into structured records to generate geotagged records.

[0092] In the above step S52, the current time is used as the reference point (denoted as `t0`), and the time difference between the two is calculated (`Δt=t0-tᵢ`) for the timestamp of each plant identification record (denoted as `tᵢ`). Based on this time difference, the system adopts two rules to determine the time decay weight: one is the exponential decay rule, that is, as the time difference increases, the weight decreases exponentially. For example, the attenuation coefficient `λ` is set to 0.05 / day. The longer the time, the lower the weight `w`; the other is the piecewise linear decay rule, which divides the time difference into different intervals and assigns different weights. For example, the weight of records within 7 days is 1, and the weight of records between 7 and 20 days decreases linearly with the number of days (decreasing by 0.05 per day). The weight of records over 20 days is uniformly set to 0.1.

[0093] Records are grouped by plant species and geographic grid code. For each record within each group, a weighted count is calculated, taking into account both the time decay weight and the identification confidence weight. The identification confidence weight is set based on the confidence level; for example, records with a confidence level of 0.9 or higher are weighted 1, while records with a confidence level between 0.7 and 0.9 are weighted 0.8. The weighted counts of all records within each group are summed to obtain the record density for that grid with time decay. To smooth the calculations and reduce density fluctuations caused by uneven sampling, the system also employs spatial neighborhood smoothing, for example, by selecting a 3×3 window around each grid cell and smoothing the density within that window.

[0094] The distribution intensity values ​​calculated for each grid are converted into color information according to pre-set mapping rules. For example, intensity values ​​between 0 and 2 are mapped to blue, between 2 and 5 to yellow, and between 5 and 10 to red. These colors are then applied to the corresponding grids using the GIS engine to form a visual heat map. To ensure the accuracy and contrast of the heat map's color display, the system automatically adjusts the color scale threshold based on the distribution range of recent data, allowing the heat map to more clearly display the distribution hotspots and changing trends of endangered plants.

[0095] In step S53, the generated heat map is spatially overlaid with the preset protected area boundary (stored as GeoJSON polygons) to identify the grids within the protected area; and the distribution intensity value of each endangered species in the protected area is extracted (for example, the intensity of "Davidia involucrata" in a certain grid is 8.5).

[0096] Independent thresholds are set for different endangered species (e.g. the threshold for national first-level protected plants is 7.0, and for second-level it is 5.0), and dynamic adjustment of thresholds based on seasonal and ecological environmental changes is supported; sliding window statistics are used (e.g. exceeding the threshold for three consecutive days) to filter out occasional monitoring anomalies and reduce false alarms.

[0097] Automatically classify the intensity exceeding the threshold and the duration (such as mild, moderate, severe) to generate alerts of different priorities; push the alert information to the associated monitoring terminal through a message queue (such as Kafka), including: species name, geographic location (latitude and longitude / grid code); distribution intensity value (such as 8.5), time decay factor (such as 0.85); highest confidence record value (such as 0.95) and corresponding image ID.

[0098] In a preferred embodiment of the present invention, the above step S6: automatically collecting and annotating images based on the newly added plant sample locations marked in the spatial distribution heat map, and periodically updating the weight parameters of the lightweight convolutional neural network model may include: Step S61, based on the coordinates of the newly added sample locations marked in the endangered plant spatial distribution heat map, automatically generates a scheduling instruction for an image acquisition device, controls the device to capture original images containing multi-view features of the target plant at the target geographic location, integrates the geographic location information and the acquisition timestamp, and outputs a set of newly added plant sample images; Step S62: extracting the morphological contour boundary of the target plant in a single image of the newly added plant sample image set; associating the plant species identification result corresponding to the image as a species label; generating a spatiotemporal identifier by combining the geographic coordinates and timestamp recorded when the image was acquired; integrating the morphological contour boundary, species label, and spatiotemporal identifier to construct a structured annotation data unit; aggregating the structured annotation data units of all images to output the annotated newly added training sample set; Step S63: Merge the newly labeled training sample set with the historical training set, freeze the feature extraction layer weights of the lightweight convolutional neural network, fine-tune the parameters of the fully connected classification layer, optimize the model's recognition ability for new samples through the backpropagation algorithm, and output the lightweight convolutional neural network model with updated weight parameters; Step S64: After a preset time period, the lightweight convolutional neural network model with updated weight parameters is deployed to the recognition system to start a new recognition cycle.

[0099] In the embodiment of the present invention, multi-angle acquisition ensures that the samples cover the full range of morphological characteristics of plants, and combined with the height adjustment of terrain adaptation, the image contains both macroscopic structure and microscopic details; structured annotation data integrates visual features, spatiotemporal information and environmental parameters, so that the model can learn the relationship between plants and geographical environment during training; the introduction of contour features enhances the model's geometric cognition of plant morphology and improves recognition stability under complex backgrounds. The frozen feature extraction layer retains the model's ability to learn general visual features, and the fine-tuning of the classification layer enables the model to quickly adapt to the recognition needs of new species; the incremental training strategy ensures that the model does not reduce the recognition accuracy of existing species when learning new categories; and periodic updates enable the model to continuously learn new plant species emerging in the wild, maintaining the timeliness of its recognition capabilities.

[0100] In the embodiments of the present invention, when applied specifically, it can be achieved through the following technical solutions, for example: In step S61, the geographic coordinates of the newly added samples (e.g., latitude and longitude 116.4074°, 39.9042°) are parsed from the heat map and mapped to the UAV flight control coordinate system through a coordinate conversion algorithm to generate the three-dimensional spatial coordinates (X, Y, Z) of the target point. The flight altitude is adjusted in combination with terrain data (e.g., DEM digital elevation model) to ensure that the UAV maintains a stable hover 5-10 meters above the target plant.

[0101] With the target plant as the center, a 360° circular shooting path is set, and a group of images is collected at every 45° interval (8 directions in total). Each group contains three photos from different perspectives: front view, 45° depression angle, and 90° depression angle. For tall plants (such as trees), an overhead shot at a height of 20-30 meters is added to capture the overall shape of the crown. For short plants (such as herbs), the height is lowered to 1-2 meters to collect detailed textures.

[0102] Real-time monitoring of the captured image clarity (assessed by gradient amplitude), illumination uniformity (histogram distribution), and plant ratio (target pixel area ratio). When any indicator does not meet the standard (such as clarity < threshold), the reshoot mechanism is automatically triggered.

[0103] In step S62, a semantic segmentation algorithm (such as a lightweight version of U-Net) is used to perform pixel-level segmentation on the newly added image, identify the boundary between the plant body and the background, and generate a binary contour mask. The segmentation result is then morphologically processed (dilation and erosion) to eliminate small noise holes (such as gaps between leaves) to ensure contour integrity.

[0104] The segmented plant outline is associated with the identification result (such as species "Davidia involucrata" and confidence level 0.95), and the geometric features of the outline (such as perimeter, area, and eccentricity) are extracted as auxiliary annotations. The geographic coordinates, timestamp (such as 2025-06-16 10:30:25) and environmental parameters (such as light intensity and temperature) at the time of collection are integrated to generate structured annotation data.

[0105] The automatic labeling results were manually checked (10%), and segmentation errors (such as misidentification of shadows as plants) were corrected using interactive labeling tools (such as LabelMe) to ensure a labeling accuracy rate of >95%.

[0106] In step S63, the newly labeled samples are merged with the historical training set (e.g., the original dataset containing 100,000 images) at a ratio of 1:9 to avoid training bias caused by too few new samples. To address the problem of a small number of samples of endangered species, data augmentation techniques (rotation, scaling, brightness adjustment) are used to generate 3-5 times expanded samples to ensure a balanced number of samples in each category (e.g., ≥500 images per category).

[0107] The weight parameters of the first five convolutional layers of the lightweight CNN model (such as the depthwise separable convolutional layers of MobileNet) are fixed. These layers have already learned common visual features such as edges and textures and do not require retraining. Only the last two fully connected classification layers (for example, 1024-dimensional → N-dimensional) are open for parameter updates, reducing the number of training parameters (by approximately 80%) and accelerating convergence.

[0108] Use mini-batch gradient descent (BatchSize=32) for 10-20 rounds of iterative training, with the learning rate gradually decaying from 0.001 (reduced to 1 / 10 of the original rate every 5 rounds). During training, monitor the accuracy of the validation set in real time (for example, use 20% of new samples as the validation set). Stop training automatically when the accuracy improvement is less than 0.5% to avoid overfitting.

[0109] In step S64 above, the model performance is evaluated on a test set (containing 10% newly added samples and 90% historical samples), with a focus on the recognition accuracy of new species (required to be ≥85%) and the retention rate of existing species (required to be ≥95%). Field flight tests are conducted, deploying the model in 3-5 different habitats (such as forests, grasslands, and wetlands) to verify its real-time recognition capability in complex environments (required inference speed ≥15 frames / second).

[0110] First, deploy the updated model to 10% of drone nodes for grayscale testing, collect actual operation data (such as recognition accuracy and system resource usage), and release it in full after 72 hours without any abnormalities; retain historical model versions (such as v1.0, v2.0), and implement one-click rollback through version control tools to ensure system stability.

[0111] The update cycle is set to 1 month (configurable). When the number of new samples exceeds 20% of the historical training set, an emergency update process is automatically triggered. A version report is generated after each update, recording indicators such as the number of new species, accuracy improvement, and training resource consumption for reference in subsequent optimization.

[0112] like Figure 2 As shown, an embodiment of the present invention further provides a system for automatically identifying plants based on a neural network model, comprising: The acquisition module is used to generate an image set of the overall morphology and community distribution of plants through high-altitude aerial photography by a drone, and simultaneously generate an image set of plant leaf texture and flower structure details through a portable smart terminal. In the process of generating the image set, four specific fixed marking points are set in the image acquisition area, and a closed figure is formed by the four marking points. The closed figure is then segmented, and an adjustment value is generated based on the grid characteristics formed after segmentation. A normalization module is used to adjust the parameters of the illumination compensation and occlusion restoration processing of the image set using the adjustment value, and perform illumination compensation and occlusion restoration processing to output a normalized image set that eliminates overexposure, blur, and occlusion interference; The fusion module is used to perform spatial feature alignment and multi-scale fusion on the standardized image set to generate a composite feature map containing macro-morphology and micro-texture; The confidence module is used to input the composite feature map into the compressed and optimized lightweight convolutional neural network model and output the plant species recognition results and confidence values ​​in real time; Dynamic binding module, used to dynamically bind plant species identification results with geographic information system (GIS) coordinates, generate a heat map of the spatial distribution of endangered plants, and trigger real-time alerts; The periodic module is used to automatically collect and annotate images based on the locations of newly added plant samples marked in the spatial distribution heat map, and periodically update the weight parameters of the lightweight convolutional neural network model.

[0113] It should be noted that this system is a system corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0114] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for automatically identifying plants based on a neural network model, characterized in that: The method comprises: Step S1: Generate an image set of the overall plant morphology and community distribution through high-altitude aerial photography by a drone, and simultaneously generate an image set of the plant's leaf texture and flower structure details through a portable intelligent terminal. In the process of generating the image set, four specific fixed marking points are set within the image acquisition area, and a closed figure is formed by the four marking points. The closed figure is then segmented, and an adjustment value is generated based on the grid characteristics formed after the segmentation. Step S2: Using the adjustment value, perform targeted illumination compensation and occlusion restoration processing parameter adjustments on the image set, perform illumination compensation and occlusion restoration processing, and output a standardized image set that eliminates overexposure, blur, and occlusion interference; Step S3: performing spatial feature alignment and multi-scale fusion on the standardized image set to generate a composite feature map containing macro-morphology and micro-texture; Step S4: inputting the composite feature map into the compressed and optimized lightweight convolutional neural network model, and outputting the plant species recognition result and confidence value in real time; Step S5: Dynamically bind the plant species identification results with the geographic information system (GIS) coordinates to generate a heat map of the spatial distribution of endangered plants and trigger a real-time alarm; Step S6: Based on the newly added plant sample locations marked in the spatial distribution heat map, images are automatically collected and annotated, and the weight parameters of the lightweight convolutional neural network model are periodically updated.

2. The method for automatic plant identification based on a neural network model according to claim 1, characterized in that: Step S1: Generate an image set of the overall morphology and community distribution of plants through high-altitude aerial photography by a drone, and simultaneously generate an image set of the leaf texture and flower structure details of the plants through a portable intelligent terminal. In the process of generating the image set, set four specific fixed marking points in the image acquisition area, form a closed figure through the four marking points, and segment the closed figure. According to the grid characteristics formed after segmentation, generate an adjustment value, including: Four physical markers with fixed geographic coordinates were evenly distributed throughout the target vegetation area, covering the core distribution area of ​​the plant community and samples at different growth stages. Simultaneously, the following were performed: drone aerial photography was used to obtain a set of images of the overall vegetation morphology and community distribution including the markers; and a portable smart terminal was used to capture a set of images of leaf texture and flower structure details of plants around the markers. Based on the geographic coordinates of the four marker points, they are connected in spatial order to form a closed graph; Adaptively segmenting the closed figure into grids, dynamically adjusting the size of the segmented grids according to the density of plants within the closed figure, and obtaining segmented grid units; For the grid cells formed by the segmentation, based on the distribution of plant features within each grid cell, the plant coverage, the consistency of plant morphological features and the complexity of leaf texture are calculated respectively to generate a feature vector for each grid cell; The feature vector of the grid unit is input into a pre-trained weight calculation model, which extracts spatial distribution pattern features through a convolutional layer and calculates the overall attribute offset of the region through regression calculation of a fully connected layer, and outputs an adjustment value for the current image acquisition area.

3. The method for automatic plant identification based on a neural network model according to claim 2, characterized in that: Step S2: Using the adjustment values, perform targeted illumination compensation and occlusion restoration processing parameter adjustments on the image set, perform illumination compensation and occlusion restoration processing, and output a standardized image set that eliminates overexposure, blur, and occlusion interference, including: Based on a preset mapping relationship table, a correspondence between the adjustment value and the illumination compensation parameter and the occlusion restoration parameter is established. The mapping relationship table includes the illumination intensity adjustment coefficient, contrast enhancement parameter, and occlusion area prediction threshold corresponding to different adjustment value intervals; According to the adjustment value output in step S1, the corresponding parameter combination is matched in the mapping relationship table to determine the illumination compensation parameters and occlusion restoration parameters that match the current image set; Using the matched illumination compensation parameters, the original image set is processed with histogram equalization and adaptive gamma correction to eliminate overexposed areas and enhance brightness uniformity, and the illumination-compensated image set is output. Based on the matched occlusion restoration parameters, the occlusion region prediction is performed on the illumination-compensated image set. This involves extracting superpixel segmented regions from the image and calculating the texture similarity difference between each region and the surrounding area. If the difference is greater than T, the region is marked as an occlusion candidate. The occlusion candidate region is then inpainted by performing edge-guided diffusion inpainting on the candidate region, prioritizing pixels with the highest texture continuity with adjacent healthy areas. For large occluded areas, a context encoder is used to generate structurally coherent inpainted content, and the inpainted image set is output. Based on the restored image set, a bilateral filtering algorithm is used for noise reduction to remove the noise introduced during the processing and output a standardized image set that eliminates overexposure, blur and occlusion interference.

4. The method for automatic plant identification based on a neural network model according to claim 3, characterized in that: Step S3: Perform spatial feature alignment and multi-scale fusion on the standardized image set to generate a composite feature map containing macro-morphology and micro-texture, including: Based on a standardized image set, local key points are extracted from drone aerial images and detail images captured by portable terminals, and directional gradient histogram descriptors of the key points are calculated. The calculated directional gradient histogram descriptors are then filtered out using a bidirectional nearest neighbor matching algorithm to select descriptor pairs with a similarity greater than 0.

85. An affine transformation matrix is ​​fitted based on the matching point pairs corresponding to the filtered descriptor pairs, minimizing the reprojection error. The fitted affine transformation matrix is ​​applied to the detail images, mapping them to the reference coordinate system of the aerial images, and outputting a spatially aligned image set. Taking the spatially aligned image set as input, a multi-scale feature layer set including a low-frequency feature layer and a high-frequency feature layer is generated through a multi-scale spatial decomposition operation, wherein the low-frequency feature layer is obtained by performing Gaussian pyramid downsampling on the aerial image, extracting the morphological skeleton of the vegetation coverage area, and calculating the skeleton branch angle distribution histogram as the macro-morphological feature; the high-frequency feature layer is obtained by performing Laplacian pyramid decomposition on the detail image, retaining the high-frequency components of the 2nd to 4th layers, extracting the local binary pattern histogram of the high-frequency components, and quantifying the leaf texture and petal edge sharpness; The multi-scale feature layer set is integrated through an adaptive weight fusion algorithm, and fusion weights are dynamically allocated according to the morphological discriminability of the low-frequency feature layer and the texture discriminability of the high-frequency feature layer. A composite feature map that retains both macroscopic morphological representation capabilities and microscopic texture details is generated through feature layer reconstruction.

5. The method for automatic plant identification based on a neural network model according to claim 4, characterized in that: Step S4: Input the composite feature map into the compressed and optimized lightweight convolutional neural network model, and output the plant species recognition results and confidence values ​​in real time, including: Inputting the composite feature map into a compressed and optimized lightweight convolutional neural network model, extracting deep features step by step through multi-layer cascade feature transformation operations, wherein each level of feature transformation includes feature channel reorganization and spatial dimensionality reduction processing, and outputting a set of feature maps with hierarchical abstract relationships; A double enhancement process is performed on the feature map set of the hierarchical abstract relationship: Perform global average pooling on the highest-level feature map to generate a channel description vector; learn channel weights through two fully connected layers, and use the Sigmoid function to output the enhancement coefficient of each channel; multiply the enhancement coefficient by the highest-level feature map channel by channel to obtain the enhanced discriminative feature expression feature map; Spatial pyramid pooling is applied to the feature map to enhance the expression of discriminative features. Average pooling of three grid scales (4×4, 2×2, and 1×1) is performed in parallel to obtain pooling results of different scales. These pooling results of different scales are then concatenated to obtain a multi-scale contextual feature vector, which is then output as an abstract feature vector that integrates channel and spatial information. The abstract feature vector is input into the fully connected layer and mapped to the plant species dimensional space. The probability distribution of each category is generated through probability normalization processing based on the exponential function. The plant species corresponding to the maximum probability value is selected as the recognition result, and the probability value is output as the confidence.

6. The method for automatic plant identification based on a neural network model according to claim 5, characterized in that: Step S5: Dynamically bind the plant species identification results to the geographic information system (GIS) coordinates to generate a heat map of the spatial distribution of endangered plants and trigger a real-time alarm, including: Based on the plant species identification result and confidence value output in step S4, combined with the geographic information system (GIS) coordinates obtained during the drone aerial image acquisition, the identification result is dynamically bound to the geographic location data through timestamp marking and spatial coding conversion to generate a plant identification record set containing species identification, confidence value, timestamp, and geocoding fields; Based on the plant identification record set, identification records within the same geographic grid are aggregated according to species identification and geocoding fields; for each identification record within each grid, the number of records is adjusted by applying a preset time decay weight function based on the interval between the record timestamp and the current time; the number of all time decay adjusted records within the grid is accumulated to obtain a time decay record density; the time decay record density is weighted and corrected based on the confidence value of each record to generate a grid distribution intensity value; the distribution intensity values ​​of all grids are mapped to a geographic spatial coordinate system to generate a dynamically updated spatial distribution heat map of endangered plants; Based on the endangered plant spatial distribution heat map, when it is detected that the distribution intensity of endangered species in the preset protection area exceeds the preset threshold, a real-time alarm is automatically triggered, and alarm information including species location, distribution intensity and confidence is pushed to the supervision terminal.

7. The method for automatic plant identification based on a neural network model according to claim 6, characterized in that: Step S6: Based on the newly added plant sample locations marked in the spatial distribution heat map, automatically collect and annotate images, and periodically update the weight parameters of the lightweight convolutional neural network model, including: Based on the coordinates of the newly added sample locations marked in the endangered plant spatial distribution heat map, a scheduling instruction for an image acquisition device is automatically generated to control the device to capture original images containing multi-view features of the target plant at the target geographic location, and the geographic location information and acquisition timestamp are integrated to output a set of newly added plant sample images; Extracting the morphological contour boundaries of the target plant in a single image of the newly added plant sample image set; associating the plant species identification result corresponding to the image as a species label; generating a spatiotemporal identifier by combining the geographic coordinates and timestamp recorded when the image was acquired; integrating the morphological contour boundaries, species labels, and spatiotemporal identifiers to construct a structured annotation data unit; aggregating the structured annotation data units of all images to output an annotated newly added training sample set; Merge the newly labeled training sample set with the historical training set, freeze the feature extraction layer weights of the lightweight convolutional neural network, fine-tune the parameters of the fully connected classification layer, optimize the model's ability to recognize new samples through the backpropagation algorithm, and output the lightweight convolutional neural network model with updated weight parameters; After a preset time period, the lightweight convolutional neural network model with updated weight parameters is deployed to the recognition system to start a new recognition cycle.

8. A system for automatic plant identification based on a neural network model, the system implementing the method according to any one of claims 1 to 7, characterized in that: include: The acquisition module is used to generate an image set of the overall morphology and community distribution of plants through high-altitude aerial photography by a drone, and simultaneously generate an image set of plant leaf texture and flower structure details through a portable smart terminal. In the process of generating the image set, four specific fixed marking points are set in the image acquisition area, and a closed figure is formed by the four marking points. The closed figure is then segmented, and an adjustment value is generated based on the grid characteristics formed after segmentation. A normalization module is used to adjust the parameters of the illumination compensation and occlusion restoration processing of the image set using the adjustment value, and perform illumination compensation and occlusion restoration processing to output a normalized image set that eliminates overexposure, blur, and occlusion interference; The fusion module is used to perform spatial feature alignment and multi-scale fusion on the standardized image set to generate a composite feature map containing macro-morphology and micro-texture; The confidence module is used to input the composite feature map into the compressed and optimized lightweight convolutional neural network model and output the plant species recognition results and confidence values ​​in real time; Dynamic binding module, used to dynamically bind plant species identification results with geographic information system (GIS) coordinates, generate a heat map of the spatial distribution of endangered plants, and trigger real-time alerts; The periodic module is used to automatically collect and annotate images based on the locations of newly added plant samples marked in the spatial distribution heat map, and periodically update the weight parameters of the lightweight convolutional neural network model.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images