Plateau crop disease and pest open space collaborative identification method and system

Abstract

The application relates to the technical field of intelligent agriculture, and discloses a highland crop disease and pest air-ground collaborative identification method and system. The method comprises the following steps: collecting image data, calibrating geographical coordinates to obtain a corrected field block boundary coordinate set; obtaining boundary region overlapping sampling points, determining the accurate range of the boundary overlapping region after grouping, extracting corresponding sub-images from the air-ground image in the range, calculating the sub-image definition score to obtain a high-definition sub-image group; extracting a feature vector and calculating a similarity to generate a similarity matrix and screen out a unique representative identification result; obtaining associated space-time label data, comparing label consistency to obtain a non-redundant disease and pest identification data set; extracting crop disease and pest feature points, fusing feature point intensity and position deviation to obtain a high-confidence disease and pest identification result, updating field block boundary records and obtaining an optimized air-ground collaborative identification result; and the application improves the accuracy of disease and pest identification.

Description

Technical Field

[0001] This application relates to the field of smart agriculture technology, and in particular to a method and system for aerial and ground-based collaborative identification of crop diseases and pests in high-altitude areas. Background Technology

[0002] With the development of smart agriculture technology, crop pest and disease monitoring has gradually shifted from traditional manual inspections to automated and intelligent identification based on remote sensing technology and the Internet of Things. The technological evolution has progressed from single ground-based sensor monitoring to pure UAV remote sensing analysis, and finally to the currently explored air-ground collaborative sensing model. In the unique environment of high-altitude mountainous areas, air-ground collaboration, by integrating continuous microscopic observations from ground monitoring vehicles with macroscopic instantaneous perception from UAVs, can theoretically achieve complementary advantages. However, this model still faces a series of technical challenges that urgently need to be addressed in practical applications: First, due to the significant topographical undulations in high-altitude mountainous areas, differences in viewing angle, sensor parameters, and elevation between ground-based and aerial equipment during image acquisition lead to systematic deviations in geographic coordinates. This makes it difficult to accurately align multi-source data spatially, directly affecting the accurate calibration of field boundaries. Second, in areas with overlapping boundaries, data collected by multiple devices contains substantial redundancy and noise. Traditional clustering algorithms are ineffective at grouping points with uneven density, resulting in inaccurate boundary delineation and consequently affecting the targeted nature of subsequent image region extraction. Furthermore, significant differences exist between ground-based and drone images in terms of resolution, lighting conditions, and shooting angles, leading to a lack of unified standards for image quality assessment and selection. This easily results in inconsistent clarity and poor feature comparability, impacting the reliability of recognition input. Moreover, when images from multiple devices exist for the same area, their feature representations exhibit heterogeneity. Direct recognition can lead to conflicting or repetitive results, and the lack of an effective feature fusion and redundancy removal mechanism makes it difficult to generate uniquely representative recognition results. Finally, the spatiotemporal label data generated by periodic inspections exhibits temporal overlap and inconsistent descriptions, and the intensity of pest and disease feature points and spatial location information are not effectively integrated, resulting in low confidence of the final identification results and an inability to provide a stable and reliable basis for precise operations.

[0003] Therefore, existing technologies lack a systematic solution to address the above problems and achieve a full-chain air-ground collaborative identification scheme for plateau crop diseases and pests, encompassing data acquisition, spatial calibration, quality assessment, feature fusion, and spatiotemporal redundancy removal. Summary of the Invention

[0004] To address the aforementioned technical challenges, this application provides a method and system for the collaborative identification of crop diseases and pests in high-altitude areas using both air and ground methods. This system achieves a complete closed loop, from precise multi-source data acquisition and spatial calibration to high-quality image screening, multi-device feature fusion and redundancy removal, spatiotemporal data consistency purification, and high-confidence identification result generation, ultimately outputting optimized decision-making. This improves the accuracy and operational efficiency of identifying crop diseases and pests in high-altitude areas.

[0005] Firstly, this application provides a method for air-ground collaborative identification of crop diseases and pests in high-altitude areas, the method comprising: Step S1: Acquire image data collected by ground monitoring vehicle and drone, calibrate geographic coordinates using elevation compensation algorithm, and obtain the corrected set of field boundary coordinates; Step S2: Based on the corrected set of field boundary coordinates, obtain the overlapping sampling points in the boundary area, group the overlapping sampling points, determine the precise range of the boundary overlapping area, and extract the corresponding sub-images from the images collected by the ground monitoring vehicle and the UAV within the precise range. Calculate the sharpness score of each sub-image to obtain the sub-image group with higher sharpness scores. Step S3: If the sub-images in the sub-image group come from multiple devices, extract the image feature vectors, calculate the similarity between each feature vector, generate a similarity matrix, and filter out the unique representative recognition result through the similarity matrix; Step S4: Based on the unique representative identification results, obtain the associated spatiotemporal label data, and use time series analysis to compare label consistency and filter out the pest and disease identification dataset without redundancy. Step S5: Based on the non-redundant pest and disease identification dataset, extract crop pest and disease feature points, fuse the feature point intensity and position deviation to obtain pest and disease identification results with high confidence, update the field boundary records in the database, and output the pest and disease identification results to the collaborative inspection system through a dynamic selection mechanism to obtain optimized air-ground collaborative identification results.

[0006] Secondly, this application provides a high-altitude crop pest and disease aerial-ground collaborative identification system, the system comprising: The image acquisition unit is used to acquire image data collected by ground monitoring vehicles and drones, and uses an elevation compensation algorithm to calibrate geographic coordinates to obtain a set of corrected field boundary coordinates. The sub-image acquisition unit is used to acquire overlapping sampling points in the boundary area based on the corrected set of field boundary coordinates, group the overlapping sampling points, determine the precise range of the boundary overlapping area, and extract corresponding sub-images from the images collected by the ground monitoring vehicle and the UAV within the precise range, calculate the sharpness score of each sub-image, and obtain the sub-image group with higher sharpness score. An image recognition unit is used to extract image feature vectors, calculate the similarity between each feature vector, generate a similarity matrix, and filter out a unique representative recognition result through the similarity matrix if the sub-images in the sub-image group come from multiple devices. The dataset identification unit is used to obtain associated spatiotemporal label data based on the unique representative identification result, and to compare label consistency using time series analysis methods to filter out non-redundant pest and disease identification datasets. The identification result acquisition unit is used to extract crop pest and disease feature points based on the non-redundant pest and disease identification dataset, fuse the feature point intensity and position deviation to obtain pest and disease identification results with high confidence, update the field boundary records in the database, and output the pest and disease identification results to the collaborative inspection system through a dynamic selection mechanism to obtain optimized air-ground collaborative identification results.

[0007] Compared with the prior art, the beneficial effects of the present invention are at least as follows: 1. This application effectively overcomes the coordinate system deviation caused by the complex terrain of the plateau by using an elevation compensation algorithm and multi-source data registration technology, and establishes a unified and accurate spatial reference frame, providing a reliable geographical basis for subsequent analysis; it adopts a clustering algorithm and a region description optimization method to accurately define the boundary overlapping area, which significantly improves the targeting of subsequent image cropping and target analysis and reduces invalid data processing; it establishes a two-level image quality screening mechanism, and ensures the consistency of input data quality through edge intensity detection and deep learning quality evaluation, providing a reliable guarantee for feature extraction.

[0008] 2. By employing feature vector similarity analysis and clustering optimization strategies, feature-level fusion and redundancy removal of images from multiple devices were achieved, resolving the conflict problem in multi-source data recognition results. Through spatiotemporal label consistency analysis and time series modeling, a non-redundant pest and disease identification dataset was constructed, significantly improving the spatiotemporal consistency and reliability of the data.

[0009] 3. By employing a confidence calculation formula to integrate feature point intensity and spatial location information, quantitative evaluation of the recognition results is achieved, significantly improving the credibility of decision-making basis. Through the combination of dynamic selection mechanism and path planning algorithm, intelligent allocation of inspection resources and autonomous optimization of operation paths are realized, significantly improving the overall operation efficiency of the system. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a method for air-ground collaborative identification of diseases and pests of highland crops in an embodiment of this application; Figure 2 This is a schematic diagram of data acquisition by the ground monitoring vehicle in an embodiment of this application; Figure 3 This is a schematic diagram of UAV aerial data acquisition in an embodiment of this application; Figure 4 This is a schematic diagram of the pest and disease distribution data processing flow in the embodiments of this application; Figure 5 This is a structural diagram of a ground-to-air collaborative identification system for diseases and pests of highland crops, as described in an embodiment of this application. Detailed Implementation

[0012] This application provides a method and system for the coordinated aerial and ground-based identification of crop diseases and pests in high-altitude areas. The terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] Example 1: In existing technologies, monitoring crop diseases and pests in high-altitude areas is difficult and the accuracy of data collection and analysis is not high. To address this technical problem, this application provides a ground-air collaborative identification method for crop diseases and pests in high-altitude areas. By integrating data from ground monitoring vehicles and UAVs, and combining elevation compensation algorithms and precise image processing technology, the monitoring and identification process of crop diseases and pests is optimized, thereby improving the accuracy of agricultural decision-making and operational efficiency.

[0014] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 The aerial-ground collaborative identification method for crop diseases and pests in high-altitude areas, as described in this application, includes: Step S1: Acquire image data collected by ground monitoring vehicle and drone, calibrate geographic coordinates using elevation compensation algorithm, and obtain the corrected set of field boundary coordinates.

[0015] Step S1 further includes: acquiring first image data through a ground monitoring vehicle; if the resolution of the first image data is lower than a preset resolution threshold, extracting edge features through a convolutional neural network to obtain second image data; acquiring third image data through a drone, and overlaying the second and third image data using an image registration method to obtain fourth image data; calculating the offset and performing elevation compensation based on the pixel values ​​of the fourth image data to obtain fifth image data; if the offset of the fifth image data exceeds a preset offset threshold, classifying outliers using a support vector machine to obtain sixth image data; and using the contour lines of the sixth image data to detect straight line segments using Hough transform to obtain the set of field boundary coordinates.

[0016] Specifically, in high-altitude mountainous areas, the systematic deviation of geographic coordinates during image acquisition by ground monitoring vehicles and drones leads to inaccurate field boundary calibration. Therefore, in this embodiment, the geographic coordinates are calibrated using image data acquired by ground monitoring vehicles and drones, combined with an elevation compensation algorithm, to obtain a corrected set of field boundary coordinates. The specific process includes: Figure 2 The deployed ground monitoring vehicle, equipped with a high-resolution camera and sensors, travels slowly along a preset path at a specified speed, such as 5 km / h, to collect first image data of the field in real time. It also simultaneously records the geographical coordinates provided by the GPS module and the attitude data acquired by the IMU sensor. If the resolution of the first image data is lower than a preset threshold, such as 1920×1080 pixels, a pre-trained U-Net convolutional neural network is used to extract edge features. The U-Net convolutional neural network takes the low-resolution first image data collected by the ground monitoring vehicle as input and outputs second image data with enhanced edge features through its encoder-decoder structure and skip connections. The U-Net convolutional neural network is obtained by using an image dataset containing labeled edge information, minimizing the binary cross-entropy loss function between the predicted edge and the real edge, and iteratively training it using the Adam optimizer.

[0017] At the same time, such as Figure 3The UAV shown flies in a grid pattern at a height of 50 meters and a speed of 8 m / s, collecting third image data, i.e., orthophotos, and uses a satellite positioning module to obtain initial geographic coordinates with centimeter-level accuracy. Subsequently, key points are extracted from the second and third image data using a scale-invariant feature transform algorithm, and matched and fused based on a random sample consensus method. Outliers are removed to form spatially unified fourth image data. Based on the pixel values ​​and corresponding geographic coordinates of the fourth image data, an elevation compensation algorithm is applied to calculate the offset: using a digital elevation model as a benchmark, the systematic deviation between the ground point cloud and the elevation data is fitted using the least squares method. The fitting coefficient is obtained by minimizing the sum of squared residuals from 1000 sample points. This offset is then applied to the original coordinates to obtain the corrected fifth image data. If the offset of the geographic coordinates in the fifth image data exceeds a preset threshold, for example, 0.3 meters, then... Anomalies are classified using a Support Vector Machine (SVM) classifier. The SVM classifier uses a radial basis function kernel and is trained on historical data to distinguish between normal and abnormal coordinate points, thus outputting refined sixth image data. Finally, Hough transform is used to detect contour line segments in the sixth image data. After connecting these line segments, a convex hull algorithm is used to generate a closed set of field boundary coordinates, thereby completing high-precision calibration of geographic coordinates and providing a reliable spatial basis for subsequent pest and disease identification. The core of Hough transform lies in a "voting mechanism," which maps pixels in the image space (XY plane) to the parameter space, also known as Hough space, for detection. The purpose of using Hough transform to detect contour line segments in the image data is to systematically transform and identify discrete, potentially discontinuous edge pixels in the image into a set of line segments with clear mathematical expressions, thus laying the geometric foundation for constructing closed field boundaries.

[0018] This embodiment, through data processing and elevation compensation, combined with data collected from the ground and drones, accurately corrected the boundary coordinates of the fields, providing higher-precision geographic information. This provides a reliable data foundation for subsequent pest and disease identification and agricultural decision support systems, ensuring the accuracy and efficiency of agricultural operations.

[0019] Step S2: Based on the corrected set of field boundary coordinates, obtain the overlapping sampling points in the boundary area, and group the overlapping sampling points using a clustering algorithm to determine the precise range of the boundary overlapping area. Within the precise range, extract the corresponding sub-images from the images collected by the ground monitoring vehicle and the UAV, and use an image quality assessment model to calculate the sharpness score of each sub-image to obtain the sub-image group with higher sharpness scores.

[0020] Determining the precise extent of the overlapping boundary region includes: extracting overlapping sampling points from the corrected set of field boundary coordinates, calculating the distance matrix between the points, and grouping them using the K-means clustering algorithm to obtain a refined point dataset; obtaining the center coordinates of each group from the refined point dataset, expanding the neighborhood range from the center coordinates of each group, and obtaining the pixel values ​​of the expanded area to obtain a preliminary overlapping region description; if the coefficient of variation of the pixel values ​​in the preliminary overlapping region description is higher than a preset coefficient of variation threshold, then performing dimensionality reduction processing on the preliminary overlapping region description using principal component analysis to obtain an optimized overlapping region description, calculating the boundary offset vector based on the optimized overlapping region description, and determining the precise boundary of the overlapping region.

[0021] Specifically, to address the issue of inaccurate boundary delineation and inconsistent image analysis data quality in high-altitude mountainous areas due to spatial overlap and noise interference in data collected by ground monitoring vehicles and drones at field boundary areas, the implementation process involves using a modified set of field boundary coordinates as a basis. This is achieved by extracting all sampling points within a preset distance (e.g., 2 meters) on both sides of the boundary through buffer analysis, forming an initial overlapping sampling point dataset. Then, the Euclidean distance matrix between every two points in this dataset is calculated, and the K-means clustering algorithm is applied to group the points. The number of clusters, K, is determined using the elbow rule, and iteration stops when the change in the center point is less than 0.01 meters, resulting in a refined dataset where each cluster represents a potential overlapping sub-region. The elbow rule is a method that determines the optimal number of clusters by finding the inflection point of the curve where the sum of squared errors within a cluster changes with the number of clusters. In this scheme, it is applied to the K-means clustering algorithm to automatically determine the number of clusters, K, for grouping overlapping sampling points, thereby ensuring the accuracy and stability of the boundary overlapping region delineation.

[0022] Subsequently, starting from the center coordinates of each cluster in the refined point dataset, a circular neighborhood is expanded, for example, with a radius of 1.5 meters. The RGB values ​​of all pixels within this neighborhood are extracted from the registered ground and UAV fused image, and their mean and standard deviation are calculated to form a preliminary overlapping region description. The preliminary overlapping region description is essentially a multidimensional feature vector used to characterize the texture consistency within the region. If the pixel value variation coefficient of this preliminary description, i.e., the ratio of standard deviation to mean, is higher than a preset threshold, such as 0.25, it indicates that the texture inside the region is complex or there is noise. Then, principal component analysis is used to reduce the dimensionality of this multidimensional description. By calculating the eigenvalues ​​and eigenvectors of the covariance matrix, the top k principal components with a cumulative contribution rate exceeding 85% are retained. After filtering out noise, an optimized overlapping region description is obtained. Based on this optimized description, the offset trend of the boundary points is fitted by the least squares method, and the boundary offset vector is calculated. For example, the offset in the main direction of change is determined according to the principal component loading, thereby correcting and outputting the precise boundary of the overlapping region as a polygonal vector sequence.

[0023] Finally, within this precise limit, based on the image georegistration results, corresponding sub-images are cropped from the ground monitoring vehicle image and the UAV image using bilinear interpolation. A pre-trained BRISQUE no-reference image quality assessment model is then used to calculate the sharpness score of each sub-image. The BRISQUE no-reference image quality assessment model takes the sub-images extracted from the ground monitoring vehicle and UAV images as input, calculates the natural scene statistical features of their local normalized brightness coefficients (such as a 66-dimensional feature vector containing mean, variance, and entropy), and inputs these features into a pre-trained support vector regression model to output a sharpness score. The training process of this model is achieved by minimizing the mean square error between the predicted score and the actual score on an image dataset containing human subjective quality scores, and by using radial basis function kernel function for nonlinear regression optimization. For example, a score threshold of 80 points is set to select sub-image groups with sharpness scores higher than 80 points, providing high-quality input for the feature extraction of pests and diseases in the following text. Among them, the BRISQUE no-reference image quality assessment model is a pipeline system consisting of a "feature extraction front end" and a "regression prediction back end". The support vector regression model is the "brain" or "decision core" in this pipeline that is responsible for finally giving the quality score.

[0024] The process of obtaining a sub-image group with a high sharpness score includes: obtaining a first sub-image sequence from the ground image based on the precise boundaries of the overlapping regions and the registration results of the UAV image; cropping a second sub-image sequence from the first sub-image sequence; obtaining a third sub-image sequence from the UAV image based on the region correspondence information of the second sub-image sequence; cropping a fourth sub-image sequence from the third sub-image sequence; calculating the edge intensity map of the fourth sub-image sequence using the Sobel operator to obtain a fifth sub-image sequence; if the edge intensity of the fifth sub-image sequence is lower than a preset edge intensity threshold, the fifth sub-image sequence is discarded to obtain a sixth sub-image sequence; calculating the sharpness score using a convolutional neural network model based on the sixth sub-image sequence and generating a score label sequence; selecting sub-images with sharpness scores exceeding a preset sixth threshold from the score label sequence to obtain a sub-image group with a high sharpness score.

[0025] Specifically, to address the issue of inconsistent sub-image quality in overlapping boundary areas during high-altitude mountainous collaborative identification due to differences in imaging equipment and environmental interference, which in turn affects the accuracy of subsequent pest and disease feature extraction, after obtaining the precise boundary of the overlapping area, the system first uses the polygon vertex coordinates of this boundary and the registration results of the ground monitoring vehicle and UAV images—that is, the unified spatial transformation matrix calculated by the Scale Invariant Feature Transform (SIFT) algorithm and the Random Sample Consensus (RANSAC) algorithm—to crop the first sub-image sequence corresponding to the precise boundary from the continuous frame images collected by the ground monitoring vehicle through perspective transformation and bilinear interpolation. To standardize the image size for subsequent processing, each sub-image in the first sub-image sequence is further cropped to 512x3 pixels. 84 pixels were used to obtain the second sub-image sequence. Based on the geographic coordinate metadata of each sub-image in the second sub-image sequence, the corresponding area was located in the georegistered UAV orthophoto database. The same spatial reference and cropping size of 1024x768 pixels were used to obtain the third sub-image sequence. After normalization and cropping, the fourth sub-image sequence was obtained, which was spatially aligned with the second sub-image sequence. The scale-invariant feature transformation algorithm initially established the correspondence between feature points in different images by extracting scale- and rotation-invariant key points in the image and generating their local feature descriptors. The random sampling consensus algorithm robustly removed mismatched outliers from these initial matches through iterative random sampling and model verification, thereby calculating the optimal geometric transformation model to achieve accurate image registration.

[0026] To perform preliminary quality screening, the Sobel operator is applied to each image in the fourth sub-image sequence to calculate its edge intensity map. The Sobel operator performs convolution using 3x3 horizontal and vertical gradient kernels and calculates its gradient magnitude. If the average edge intensity of a sub-image is lower than a preset threshold, such as 50, based on the intensity range of 0-255 for 8-bit grayscale images, it is considered to have blurred texture and low contribution to recognition, and is therefore removed, thus obtaining the sixth sub-image sequence with richer texture information. The Sobel operator is a discrete differential operator used for edge detection in the field of image processing. In this scheme, the application of the Sobel operator aims to quickly and effectively filter out images with rich texture details from the fourth sub-image sequence.

[0027] Finally, the sixth sub-image sequence is input into a pre-trained convolutional neural network model for fine-grained sharpness evaluation. This convolutional neural network adopts the lightweight MobileNetV2 architecture, whose input layer receives RGB sub-images with a size of 224x224x3. Features are extracted through inverse residual structure and linear bottleneck layer, and finally a sharpness score in the range of 0 to 100 is output by fully connected layer. The score label sequence is generated by forward propagation, and sub-images with scores exceeding a preset threshold, such as 80, are selected to form a sub-image group with high sharpness scores, providing a high-quality data foundation for feature fusion and accurate recognition of multi-source images in the following text.

[0028] Step S3: If the sub-images in the sub-image group come from multiple devices, then the image feature vectors are extracted by a convolutional neural network, the similarity between each feature vector is calculated, a similarity matrix is ​​generated, and the unique representative recognition result is selected by filtering the similarity matrix.

[0029] Step S3 further includes: obtaining the device identifier of each sub-image from the sub-image group and generating a device identifier sequence; if there are multiple different device identifiers in the device identifier sequence, extracting feature vectors from each sub-image using a convolutional neural network to obtain a feature vector set; calculating the cosine similarity between each pair of feature vectors in the feature vector set to generate a similarity matrix; selecting sub-images with similarity exceeding a preset similarity threshold from the similarity matrix to obtain a candidate sub-image group; performing histogram equalization on the candidate sub-image group to generate an equalized sub-image group; grouping the feature vectors in the equalized sub-image group using a clustering algorithm, determining the central feature vector of each group, and selecting a representative sub-image as the unique representative recognition result.

[0030] Specifically, due to differences in imaging perspective, resolution, and lighting conditions between ground monitoring vehicles and drones, multiple sub-images of the same scene exhibit feature heterogeneity, leading to redundancy or conflict in pest and disease identification results. In the aforementioned implementation, when the system detects that the device identification sequence of a sub-image group contains both ground monitoring vehicle and drone identifiers, it first uniformly scales each sub-image to 224×224 pixels and inputs it into a pre-trained ResNet-50 convolutional neural network. By removing its terminal fully connected layers, a 2048-dimensional feature vector is extracted from the global average pooling layer to form a feature vector set. Then, the cosine similarity between every two feature vectors in this set is calculated. The formula is the ratio of the dot product of the two vectors to the product of their respective L2 norms, generating an M×M similarity matrix (M being the number of sub-images). Then, a similarity value exceeding a preset threshold is selected. Sub-images with a value of 0.85 constitute a candidate sub-image group. To eliminate the influence of illumination differences, the candidate sub-image group undergoes contrast-limited adaptive histogram equalization to generate a more balanced sub-image group with more uniform illumination distribution. Subsequently, feature vectors are re-extracted from this group of images, and K-means clustering algorithm is used to group them according to the Euclidean distance of the feature vectors. The number of clusters is set to 2 and verified by the elbow rule. After iterating until the change of cluster center is less than 1e-4, the actual sub-image with the highest cosine similarity to the feature vector of the cluster center is selected as the representative sub-image. Finally, these representative sub-images are input into a pre-trained pest and disease classification model, and its output recognition result is used as the unique representative recognition result of the region under the current spatiotemporal conditions. This eliminates redundancy in the results while retaining the complementary advantages of multi-device data, ensuring the uniqueness and accuracy of the decision data.

[0031] Step S4: Based on the unique representative identification results, obtain the associated spatiotemporal label data, and use time series analysis to compare label consistency and filter out the pest and disease identification dataset without redundancy.

[0032] Step S4 further includes: obtaining spatiotemporal label data from the unique representative identification results; using a time series segmentation method, dividing the spatiotemporal label data into multiple time segments according to time windows to obtain a set of time series segments; performing autoregressive moving average model analysis on the time segments to obtain a set of time series feature vectors for each time segment; calculating the Euclidean distance between each pair of time series feature vectors to generate a distance matrix; if the Euclidean distance of the first pair of time series feature vectors in the distance matrix is ​​lower than a preset threshold, merging the time segments corresponding to the first pair of time series feature vectors to obtain a merged set of time series segments, where the first pair of time series feature vectors represents any pair of time series feature vectors in the distance matrix; using principal component analysis to extract the main time series features from the merged set of time series segments to obtain a simplified set of time series features; clustering the simplified set of time series features and determining the central feature vector of each group to obtain a non-redundant pest and disease identification dataset.

[0033] Specifically, due to temporal overlap and descriptive conflicts in the spatiotemporal label data generated by periodic inspections, the pest and disease identification dataset becomes redundant and inconsistent. In the specific implementation of step S4 above, the associated spatiotemporal label data, including pest and disease type, occurrence timestamp, and geographic coordinates, is first extracted from the unique representative identification result. A fixed-length time window segmentation method is used to divide the label data of 30 consecutive days into multiple time segments with a 7-day cycle, forming a time series segment set. For each time segment, a daily series of pest and disease occurrences is constructed, and its temporal characteristics are analyzed by an autoregressive moving average model. The parameters of the autoregressive moving average model are determined by grid search combined with the AIC criterion, and its autocorrelation coefficient and partial autocorrelation coefficient are extracted to form a 10-dimensional time series feature vector set. The Euclidean distance between each pair of time series feature vectors in this set is calculated to generate a distance matrix. When the Euclidean distance between any pair of feature vectors in the matrix is ​​lower than a preset threshold of 3.0, it is determined that the pest and disease occurrence patterns of the corresponding time segments are highly similar, and they are merged into the same event segment to obtain the merged time series segment set.

[0034] Subsequently, principal component analysis was used to reduce the 10-dimensional features of the set, retaining the top three principal components with a cumulative contribution rate exceeding 85%, resulting in a simplified temporal feature set. Finally, the simplified set was grouped using the K-means clustering algorithm, and the silhouette coefficient was set to determine the optimal number of clusters. After iterative calculation, the original data records corresponding to the feature vectors of each cluster center were selected to form a non-redundant pest and disease identification dataset. This dataset ensures that each independent pest and disease event has a unique representation in the spatiotemporal dimension, providing a purified data foundation for the confidence calculation in the following text.

[0035] Step S5: Based on the non-redundant pest and disease identification dataset, extract crop pest and disease feature points, and use the confidence calculation formula to fuse feature point intensity and positional deviation to obtain pest and disease identification results with high confidence. Update the field boundary records in the database, and output the pest and disease identification results to the collaborative inspection system through a dynamic selection mechanism to obtain optimized air-ground collaborative identification results.

[0036] The process of obtaining high-confidence pest and disease identification results includes: acquiring a set of feature points for crop pests and diseases from a non-redundant pest and disease identification dataset, generating an initial feature point matrix using a feature extraction method, calculating the intensity value of each feature point in the initial feature point matrix, and generating a feature point intensity set; removing feature points with intensity values ​​lower than a preset intensity threshold from the feature point intensity set to obtain a simplified feature point set; obtaining the positional deviation value of each feature point from the simplified feature point set, and fusing the feature point intensity and positional deviation value using a weighted average method to obtain a fused feature point set; classifying the feature points in the fused feature point set using a support vector machine algorithm to obtain a categorized feature point set; if the confidence of the first feature point in the categorized feature point set is lower than a preset confidence threshold, then removing the first feature point to obtain a high-confidence feature point set, where the first feature point represents any feature point in the categorized feature point set; and extracting the final pest and disease identification result from the high-confidence feature point set to generate a high-confidence pest and disease identification result.

[0037] Specifically, the low confidence level and insufficient decision reliability of the identification results are caused by the differences in intensity and spatial location deviation of the pest and disease feature points. In the specific implementation of step S5, samples containing RGB images and corresponding annotations are first obtained from the non-redundant pest and disease identification dataset. An improved VGG16 convolutional neural network is used as the feature extraction method. Its fully connected layer is replaced with a 1x1 convolutional layer for dense feature prediction. After outputting the feature map of spatial dimension, an initial feature point matrix is ​​generated through non-maximum suppression operation. Each feature point contains an intensity value, a corresponding feature map activation value, ranging from 0 to 1 and a geographic coordinate position, with an accuracy of 0.01 degrees. The intensity value of each feature point in the matrix is ​​calculated. After generating the feature point intensity set, weak response feature points are removed by a preset intensity threshold of 0.6 to obtain a simplified feature point set.

[0038] Next, based on this set, the Euclidean distance between each feature point and the average center of its respective field is calculated as the location deviation value. A weighted average method is used to fuse the intensity value and the location deviation value with weights of 0.7:0.3. The location deviation value is pre-processed using min-max normalization to form a fused feature point set that combines significance and spatial consistency. Subsequently, this set is classified using a support vector machine algorithm. This algorithm uses a radial basis function kernel and is trained on a training set containing 5000 labeled samples. For each feature point, the probability of it belonging to a specific pest or disease category is output as the confidence score. When the confidence score of any feature point in the classification feature point set is lower than a preset confidence threshold, such as 0.8, that feature point is removed, resulting in a high-confidence feature point set. The category labels of all feature points in this set are extracted, and a weighted voting mechanism, with the weights being the confidence scores, is used to determine the final pest or disease type, generating a high-confidence pest or disease identification result. Simultaneously, this result and its spatial distribution are associated with the corresponding field boundary records, achieving accurate database updates and visual support for field management decisions.

[0039] The process of obtaining optimized air-ground collaborative identification results includes: acquiring pest and disease distribution data from high-confidence pest and disease identification results, and generating a field distribution matrix using a grid partitioning method to obtain preliminary pest and disease distribution data; calculating field boundary coordinates using a boundary partitioning algorithm to update field boundary records in the database to obtain an updated field boundary dataset; extracting field boundary coordinates from the updated field boundary dataset and integrating the pest and disease distribution data with the field boundary coordinates using a weighted fusion method to obtain a fused distribution dataset; for the fused distribution dataset, if the matching degree of the distribution data is lower than a preset matching threshold, removing mismatched data points to obtain a simplified distribution dataset; and obtaining... Based on the distribution characteristics of pests and diseases, a dynamic selection mechanism is used to allocate inspection tasks to the collaborative inspection system, resulting in inspection task allocation results. According to these allocation results, the task priorities are sorted using a random forest algorithm to obtain an optimized inspection task sequence. Air-ground collaborative inspection paths are extracted from the optimized sequence, and a path planning algorithm is used to generate air-ground collaborative identification results, ultimately yielding optimized air-ground collaborative identification results. The random forest algorithm is used as a priority predictor, learning from historical data such as task urgency, equipment value, geographical location, and failure risk to predict or evaluate the comprehensive priority score of each inspection task. This enables intelligent sorting and optimization of the task sequence, providing a high-quality input sequence for path planning.

[0040] Specifically, the spatial heterogeneity of pest and disease distribution and the limited availability of operational resources lead to difficulties in dynamic decision-making and inefficient allocation of inspection resources. After obtaining pest and disease identification results with high confidence, spatial distribution data is first extracted from the included geographic coordinates and pest and disease type information, such as... Figure 4 The flowchart shown illustrates the data processing steps for pest and disease distribution. First, the target field is divided into uniform 0.5m x 0.5m grid cells using a grid partitioning method. The density and type of pest and disease points within each cell are statistically analyzed to generate a field distribution matrix, yielding preliminary pest and disease distribution data with spatial resolution. Next, the Delaunay triangulation algorithm is used to construct the topological relationships of the field boundary points. The α-shape algorithm is then used to extract the vertex coordinates of the boundary polygons, updating the field boundary records in the database and forming a field boundary dataset containing the latest spatial structure. The Delaunay triangulation algorithm is a computational geometry method that connects discretely distributed field boundary points into an optimal triangular mesh. The method, in this scheme, generates non-overlapping triangles using all boundary points as vertices, ensuring that the circumcircle of any triangle does not contain other vertices, thereby establishing adjacency relationships between points and shared edge relationships between triangles. Finally, it constructs a topological structure describing the outline of the field by filtering all "boundary edges" that belong to only one triangle. The α-shape algorithm is a general algorithm for extracting intuitive shape boundaries from a set of discrete points. In this scheme, it filters the edges of the triangles formed after Delaunay triangulation by setting a specific radius parameter α, retaining only edges with a length less than α and connecting their endpoints, thereby outlining the vertex coordinate sequence of the boundary polygons describing the irregular outline of the field.

[0041] Next, the updated boundary coordinates and the pest and disease distribution matrix are integrated using a weighted fusion method, with the boundary coordinates having a weight of 0.4 and the pest and disease density having a weight of 0.6, generating a fused distribution dataset that simultaneously includes geographical boundaries and pest and disease intensity. Spatial consistency is then checked on this dataset. If the matching degree between the pest and disease data and the boundary topology of a certain grid cell is lower than a preset threshold of 0.7, it is identified as a spatial outlier and removed, resulting in a simplified distribution dataset. The matching degree between the pest and disease data and the boundary topology is calculated by measuring the similarity between the two sets. That is, by quantifying the spatial overlap between "pest and disease locations" and "reasonable crop areas," obviously illogical outliers are automatically identified and removed, thus ensuring that the dataset on which the following analysis depends is clean and reliable.

[0042] Based on spatial features such as pest aggregation degree and distribution radius extracted from a simplified dataset, a dynamic selection mechanism is adopted to allocate inspection tasks: Based on the real-time location of the monitoring vehicle on the ground and the drone's remaining flight status, and using pest severity (weight 0.6) and equipment accessibility (weight 0.4) as decision factors, an inspection task allocation result containing task area and equipment type is generated. Subsequently, the result is prioritized using a random forest algorithm. This algorithm uses 100 decision trees, trained with task urgency, processing time, resource consumption, and task ranking results as features, and finally outputs an optimized task sequence arranged in descending order of comprehensive score. Finally, A... The path planning algorithm, using the current location of the equipment as the starting point and the center of gravity of each area in the task sequence as waypoints, calculates the shortest inspection path and generates an air-ground collaborative identification result that integrates spatial decision-making and path optimization, completing closed-loop control from data analysis to operation execution; among which, A The algorithm is a heuristic search algorithm that efficiently finds the shortest path between two points by comprehensively evaluating the actual cost of the current path and the estimated cost to the target point. In this embodiment, A The algorithm starts from the current location of the device and uses the centroids of each region in the task sequence optimized by random forest as navigation waypoints. By calculating the lowest cost path between two points, it finally generates a global inspection path with the shortest total distance or total time, thus integrating task priority decision-making with spatial path optimization to obtain the final collaborative recognition scheme.

[0043] By combining the above steps, this application improves the accuracy of identifying diseases and pests in highland crops.

[0044] Example 2: The above describes a method for air-ground collaborative identification of diseases and pests in high-altitude crops according to embodiments of this application. The following describes a system for air-ground collaborative identification of diseases and pests in high-altitude crops according to embodiments of this application. Please refer to [link / reference]. Figure 5 The aerial-ground collaborative identification system for crop diseases and pests in high-altitude areas, as described in this application embodiment, includes: The image acquisition unit is used to acquire image data collected by ground monitoring vehicles and drones, and uses an elevation compensation algorithm to calibrate geographic coordinates to obtain a set of corrected field boundary coordinates. The sub-image acquisition unit is used to acquire overlapping sampling points in the boundary area based on the corrected set of field boundary coordinates, group the overlapping sampling points, determine the precise range of the boundary overlapping area, and extract corresponding sub-images from the images collected by the ground monitoring vehicle and the UAV within the precise range, calculate the sharpness score of each sub-image, and obtain the sub-image group with higher sharpness score. The image recognition unit is used to extract image feature vectors if the sub-images in a sub-image group come from multiple devices, calculate the similarity between each feature vector, generate a similarity matrix, and select the unique representative recognition result through the similarity matrix. The dataset identification unit is used to obtain associated spatiotemporal label data based on unique and representative identification results, and to compare label consistency using time series analysis methods to filter out non-redundant pest and disease identification datasets. The identification result acquisition unit is used to extract crop pest and disease feature points based on a non-redundant pest and disease identification dataset, fuse feature point intensity and positional deviation to obtain pest and disease identification results with high confidence, update field boundary records in the database, and output the pest and disease identification results to the collaborative inspection system through a dynamic selection mechanism to obtain optimized air-ground collaborative identification results.

[0045] Through the synergistic cooperation of the above-mentioned components, the accuracy of identifying diseases and pests in highland crops has been further improved.

[0046] In summary, the aerial-ground collaborative identification method for crop diseases and pests in high-altitude areas provided in this application systematically solves key challenges in agricultural monitoring in high-altitude regions by constructing a complete technical closed loop. The method first achieves precise calibration of ground and aerial geographic coordinates through an elevation compensation algorithm, establishing a unified spatial benchmark. Then, it utilizes cluster analysis to accurately define overlapping boundary areas and combines a multi-level quality screening mechanism to obtain high-quality image data. For multi-source heterogeneous data, it innovatively employs feature fusion and similarity analysis techniques to generate uniquely representative identification results, effectively eliminating data redundancy. Furthermore, it constructs a purified disease and pest dataset through spatiotemporal label consistency analysis and time-series modeling. Finally, it integrates feature intensity and spatial location information for confidence calculation and generates optimized inspection plans based on a dynamic decision-making mechanism. This method overcomes the limitations of traditional single monitoring methods, achieving effective collaboration between aerial and ground resources and in-depth data value mining. It significantly improves the accuracy, timeliness, and reliability of disease and pest identification, providing complete technical support and a feasible implementation path for precision agriculture in high-altitude mountainous areas.

[0047] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0048] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0049] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for aerial-ground collaborative identification of crop diseases and pests in high-altitude areas, characterized in that, The method includes: Step S1: Acquire image data collected by ground monitoring vehicle and drone, calibrate geographic coordinates using elevation compensation algorithm, and obtain the corrected set of field boundary coordinates; Step S2: Based on the corrected set of field boundary coordinates, obtain the overlapping sampling points in the boundary area, group the overlapping sampling points, determine the precise range of the boundary overlapping area, and extract the corresponding sub-images from the images collected by the ground monitoring vehicle and the UAV within the precise range. Calculate the sharpness score of each sub-image to obtain the sub-image group with higher sharpness scores. Step S3: If the sub-images in the sub-image group come from multiple devices, extract the image feature vectors, calculate the similarity between each feature vector, generate a similarity matrix, and filter out the unique representative recognition result through the similarity matrix; Step S4: Based on the unique representative identification results, obtain the associated spatiotemporal label data, and use time series analysis to compare label consistency and filter out the pest and disease identification dataset without redundancy. Step S5: Based on the non-redundant pest and disease identification dataset, extract crop pest and disease feature points, fuse the feature point intensity and position deviation to obtain pest and disease identification results with high confidence, update the field boundary records in the database, and output the pest and disease identification results to the collaborative inspection system through a dynamic selection mechanism to obtain optimized air-ground collaborative identification results.

2. The method according to claim 1, characterized in that, Step S1 further includes: First image data is acquired by a ground monitoring vehicle. If the resolution of the first image data is lower than a preset resolution threshold, edge features are extracted using a convolutional neural network to obtain second image data. Third image data is acquired by a drone. The second and third image data are overlaid using an image registration method to obtain fourth image data. Based on the pixel values ​​of the fourth image data, the offset is calculated and elevation compensation is performed to obtain fifth image data. If the offset of the fifth image data exceeds a preset offset threshold, anomalies are classified using a support vector machine to obtain sixth image data. Through the contour lines of the sixth image data, Hough transform is used to detect straight line segments to obtain the set of field boundary coordinates.

3. The method according to claim 1, characterized in that, In step S2, determining the precise range of the boundary overlap region includes: Based on the corrected set of field boundary coordinates, overlapping sampling points are extracted, the distance matrix between the points is calculated, and the points are grouped using the K-means clustering algorithm to obtain a refined point dataset. The center coordinates of each group are obtained from the refined point dataset, the neighborhood range of each group is expanded from the center coordinates, and the pixel values ​​of the expanded region are obtained to obtain a preliminary overlapping region description. If the coefficient of variation of the pixel values ​​in the preliminary overlapping region description is higher than a preset coefficient of variation threshold, the preliminary overlapping region description is dimensionality reduced by principal component analysis to obtain an optimized overlapping region description. The boundary offset vector is calculated based on the optimized overlapping region description to determine the precise boundary of the overlapping region.

4. The method according to claim 3, characterized in that, In step S2, a sub-image group with a high sharpness score is obtained, including: Based on the precise boundaries of the overlapping regions and the image registration results, a first sub-image sequence is obtained from the ground image, and a second sub-image sequence is cropped from the first sub-image sequence. Based on the region correspondence information of the second sub-image sequence, a third sub-image sequence is obtained from the UAV image, and a fourth sub-image sequence is cropped from the third sub-image sequence. For the fourth sub-image sequence, the Sobel operator is used to calculate the edge intensity map to obtain a fifth sub-image sequence. If the edge intensity of the fifth sub-image sequence is lower than a preset edge intensity threshold, the fifth sub-image sequence is discarded to obtain a sixth sub-image sequence. Based on the sixth sub-image sequence, a convolutional neural network model is used to calculate the sharpness score and generate a score label sequence. Sub-images with sharpness scores exceeding a preset sixth threshold are selected from the score label sequence to obtain a sub-image group with higher sharpness scores.

5. The method according to claim 1, characterized in that, Step S3 further includes: The device identifiers of each sub-image are obtained from the sub-image group, and a device identifier sequence is generated. If there are multiple different device identifiers in the device identifier sequence, feature vectors are extracted from each sub-image using a convolutional neural network to obtain a feature vector set. The cosine similarity between each pair of feature vectors in the feature vector set is calculated to generate a similarity matrix. Sub-images with similarity exceeding a preset similarity threshold are selected from the similarity matrix to obtain candidate sub-image groups. Histogram equalization is performed on the candidate sub-image groups to generate equalized sub-image groups. The feature vectors in the equalized sub-image groups are grouped using a clustering algorithm to determine the central feature vector of each group, and a representative sub-image is selected as the unique representative recognition result.

6. The method according to claim 1, characterized in that, Step S4 further includes: Spatiotemporal label data is obtained from the unique representative identification results. The spatiotemporal label data is divided into multiple time segments according to time windows using a time series segmentation method to obtain a set of time series segments. The time segments are analyzed by an autoregressive moving average model to obtain a set of time series feature vectors for each time segment. The Euclidean distance between each pair of time series feature vectors is calculated to generate a distance matrix. If the Euclidean distance between the first pair of time-series feature vectors in the distance matrix is ​​lower than a preset threshold, the time segments corresponding to the first pair of time-series feature vectors are merged to obtain a merged time series segment set. The first pair of time-series feature vectors represents any pair of time-series feature vectors in the distance matrix. Principal component analysis is used to extract the main time-series features from the merged time series segment set to obtain a simplified time-series feature set. The simplified time-series feature set is clustered, and the central feature vector of each group is determined to obtain a non-redundant pest and disease identification dataset.

7. The method according to claim 1, characterized in that, In step S5, the pest and disease identification results with high confidence are obtained, including: A set of feature points for crop diseases and pests is obtained from the non-redundant disease and pest identification dataset. An initial feature point matrix is ​​generated using a feature extraction method. The intensity value of each feature point in the initial feature point matrix is ​​calculated, and a feature point intensity set is generated. Feature points with intensity values ​​lower than a preset intensity threshold are removed from the feature point intensity set to obtain a simplified feature point set. The positional deviation value of each feature point is obtained from the simplified feature point set, and the feature point intensity and positional deviation value are fused using a weighted average method to obtain a fused feature point set. The feature points in the fused feature point set are classified using the support vector machine algorithm to obtain a classification feature point set. If the confidence of the first feature point in the classification feature point set is lower than a preset confidence threshold, the first feature point is removed to obtain a high-confidence feature point set. The first feature point represents any feature point in the classification feature point set. The final pest and disease identification result is extracted from the high-confidence feature point set to generate a pest and disease identification result with high confidence.

8. The method according to claim 7, characterized in that, In step S5, the optimized air-ground cooperative identification result is obtained, including: Pest and disease distribution data are obtained from the high-confidence pest and disease identification results, and a field distribution matrix is ​​generated using a grid partitioning method to obtain preliminary pest and disease distribution data. For the preliminary pest and disease distribution data, the field boundary coordinates are calculated using a boundary partitioning algorithm, and the field boundary records in the database are updated to obtain an updated field boundary dataset. The field boundary coordinates are extracted from the updated field boundary dataset, and the pest and disease distribution data are integrated with the field boundary coordinates using a weighted fusion method to obtain a fused distribution dataset. For the fused distribution dataset, if the matching degree of the distribution data is lower than a preset matching threshold, mismatched data points are removed to obtain a simplified distribution dataset. The distribution characteristics of pests and diseases are obtained from the simplified distribution dataset. A dynamic selection mechanism is used to allocate inspection tasks to the collaborative inspection system to obtain the inspection task allocation results. Based on the inspection task allocation results, the task priorities are sorted using the random forest algorithm to obtain the optimized inspection task sequence. The air-ground collaborative inspection path is extracted from the optimized inspection task sequence, and the air-ground collaborative identification result is generated using the path planning algorithm to obtain the optimized air-ground collaborative identification result.

9. The method according to claim 5, characterized in that, A representative sub-image is selected as the unique representative recognition result, including: Based on the central feature vector of each group, the actual sub-image with the highest cosine similarity to the central feature vector of the group is selected as the representative sub-image.

10. A high-altitude crop pest and disease aerial-ground collaborative identification system, used to implement the high-altitude crop pest and disease aerial-ground collaborative identification method as described in any one of claims 1-9, characterized in that, The system includes: The image acquisition unit is used to acquire image data collected by ground monitoring vehicles and drones, and uses an elevation compensation algorithm to calibrate geographic coordinates to obtain a set of corrected field boundary coordinates. The sub-image acquisition unit is used to acquire overlapping sampling points in the boundary area based on the corrected set of field boundary coordinates, group the overlapping sampling points, determine the precise range of the boundary overlapping area, and extract corresponding sub-images from the images collected by the ground monitoring vehicle and the UAV within the precise range, calculate the sharpness score of each sub-image, and obtain the sub-image group with higher sharpness score. An image recognition unit is used to extract image feature vectors, calculate the similarity between each feature vector, generate a similarity matrix, and filter out a unique representative recognition result through the similarity matrix if the sub-images in the sub-image group come from multiple devices. The dataset identification unit is used to obtain associated spatiotemporal label data based on the unique representative identification result, and to compare label consistency using time series analysis methods to filter out non-redundant pest and disease identification datasets. The identification result acquisition unit is used to extract crop pest and disease feature points based on the non-redundant pest and disease identification dataset, fuse the feature point intensity and position deviation to obtain pest and disease identification results with high confidence, update the field boundary records in the database, and output the pest and disease identification results to the collaborative inspection system through a dynamic selection mechanism to obtain optimized air-ground collaborative identification results.

Images