Method and system for monitoring and identifying diseases and insect pests of jujube trees based on remote sensing technology

Abstract

The present application relates to the field of pest and disease monitoring and identification technology, and specifically to a method and system for monitoring and identifying pests and diseases in jujube trees based on remote sensing technology. The method comprises: collecting a jujube tree vegetation image, downsampling it through a neural network to obtain a first feature map; obtaining multi-scale feature maps of the first feature map, recorded as each second feature map; adding an adaptive fine-grained attention module to the neural network to perform weighted fusion on the second feature maps, specifically comprising: performing superpixel segmentation on each second feature map to obtain the pest probability of each superpixel block; obtaining the diseased spot area in each second feature map; determining the size of the local window divided by each second feature map, using a multi-head self-attention mechanism to obtain the weight of each second feature map; performing weighted fusion on all second feature maps, and obtaining the jujube tree pest and disease monitoring results through multi-stage feature extraction using a ResNet50 network structure. This improves the accuracy of jujube tree pest and disease monitoring and identification.

Description

Technical Field

[0001] The present application relates to the technical field of pest and disease monitoring and identification, and specifically to a method and system for monitoring and identifying pests and diseases of jujube trees based on remote sensing technology. Background Art

[0002] As the jujube cultivation industry develops, pests and diseases can cause a decline in the quality and quantity of jujubes. Monitoring jujube pests and diseases is a key measure to ensure healthy growth and improve yield and quality. Monitoring pests and diseases can detect jujube pest problems early and prevent them, making it a crucial component of the jujube cultivation industry. Remote sensing technology can cover a wide geographic area and provide real-time monitoring via platforms such as satellites and drones. This allows for rapid and comprehensive detection of pest and disease occurrences across large jujube cultivation areas, unrestricted by human intervention. Consequently, remote sensing technology is widely used in jujube pest and disease monitoring.

[0003] After obtaining jujube vegetation images using remote sensing technology, neural networks are usually used to identify disease and insect pest areas in the jujube vegetation images. The existing technology has low accuracy in identifying fine-grained diseased areas in jujube vegetation images. When the neural network performs multi-scale feature map fusion, the feature map is often divided into local windows of fixed size, and the self-attention mechanism is used to divide the weights of feature maps of different scales. However, the fixed-size local window may not be able to adapt to the scale or complexity of different features, resulting in information loss or failure to capture sufficient contextual information. Especially when processing high-resolution features, the suitability of the feature map weight setting is reduced, thereby affecting the accuracy of neural network feature extraction and the sensitivity of early jujube disease and insect pest identification, resulting in low accuracy in jujube disease and insect pest monitoring and identification. Summary of the Invention

[0004] In order to solve the above technical problems, the purpose of this application is to provide a method and system for monitoring and identifying jujube tree pests and diseases based on remote sensing technology. The technical solutions adopted are as follows:

[0005] In a first aspect, the present invention provides a method for monitoring and identifying pests and diseases of jujube trees based on remote sensing technology, the method comprising the following steps:

[0006] Collecting an image of jujube tree vegetation, inputting the image into a neural network, and obtaining a first feature map after downsampling; using multiple convolutional layers with different convolution kernel sizes to obtain multi-scale feature maps of the first feature map, recorded as respective second feature maps; wherein the neural network is a ResNet50 network structure;

[0007] Add an adaptive fine-grained attention module to the neural network to perform weighted fusion on the second feature map, including:

[0008] Perform superpixel segmentation on each second feature map, analyze the grayscale difference between each superpixel block and its adjacent superpixel blocks in each second feature map, and obtain the possibility of insect infestation of each superpixel block;

[0009] Based on the pest possibility, a target superpixel block in each second feature map is selected; based on the grayscale difference between the target superpixel block and its neighboring superpixel blocks, neighboring superpixel blocks of the target superpixel block are selected and merged with the target superpixel block, and a diseased spot area in each second feature map is obtained according to the morphological characteristics of the merged area.

[0010] The size of the lesion area in each second feature map is used to determine the size of the local window divided by each second feature map. After dividing each second feature map into local windows, a multi-head self-attention mechanism is used to obtain the weight of each second feature map;

[0011] All second feature maps are weightedly fused, and multi-stage feature extraction is performed using the ResNet50 network structure to obtain the output of the neural network and the pest and disease monitoring results of jujube trees.

[0012] In one embodiment, determining the insect infestation probability of each superpixel block includes:

[0013] For each second feature map, the grayscale mean of all pixels in each superpixel block is calculated, and the sum of the differences between the grayscale means of each superpixel block and all its adjacent superpixel blocks is calculated. The possibility of insect infestation of each superpixel block is determined by the proportion of the sum value of each superpixel block in the cumulative sum of the sum values ​​of all superpixel blocks in its second feature map.

[0014] In one embodiment, screening the target superpixel blocks in each second feature map based on the pest possibility includes:

[0015] A segmentation threshold of the pest possibility of all superpixel blocks in each second feature map is obtained, and a superpixel block whose pest possibility is greater than the segmentation threshold is used as a target superpixel block.

[0016] In one embodiment, the step of selecting a neighboring superpixel block of the target superpixel block and merging the target superpixel block with the target superpixel block comprises:

[0017] For any target super-pixel block, if the grayscale mean of its neighboring super-pixel block is greater than or equal to the grayscale mean of any target super-pixel block, its neighboring super-pixel block is merged with the any target super-pixel block.

[0018] In one embodiment, obtaining the diseased spot area in each second feature map includes:

[0019] Obtain a fitting circle of the merged area in each second feature map, calculate the overlap area ratio between the merged area and the fitting circle, and determine the lesion area in each second feature map based on the overlap area ratio.

[0020] In one embodiment, the diseased spot area in each second feature map is a merged area in which the proportion of the overlapping area is greater than a preset first threshold.

[0021] In one embodiment, determining the size of each second feature map divided into local windows includes:

[0022] For each second feature map, the minimum circumscribed matrix of each lesion area is obtained, and the length and width of the minimum circumscribed matrix with the largest area are used as the size of dividing the local window.

[0023] In one embodiment, determining the weight of each second feature map includes:

[0024] A multi-head self-attention mechanism is used to obtain the attention matrix of each second feature map, and the weight of each second feature map is the mean of all elements in the attention matrix.

[0025] In one embodiment, obtaining the pest and disease monitoring results of the jujube tree includes:

[0026] The output of the neural network includes the coordinates of the predicted boundary boxes of the disease and insect pest areas in the jujube tree vegetation image, and determines the proportion of the union area of ​​all the predicted boundary boxes of the disease and insect pest areas in the jujube tree vegetation image in the jujube tree vegetation image;

[0027] When the proportion is greater than a preset second threshold, it is determined that there is a pest and disease area in the corresponding area of ​​the jujube tree vegetation image, and the jujube trees need to be treated for pests and diseases. Otherwise, the jujube trees do not need to be treated for pests and diseases.

[0028] In the second aspect, an embodiment of the present application also provides a jujube tree disease and pest monitoring and identification system based on remote sensing technology, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the steps of any one of the above methods when executing the computer program.

[0029] This application has at least the following beneficial effects:

[0030] This application improves the adaptability to the size diversity of jujube tree pest and disease areas through multi-branch feature extraction with different convolution kernel sizes, and simultaneously captures the pathological features of different scales such as local yellowing, blackening, and lesions of jujube leaves, avoiding the problem of insufficient coverage of complex symptoms by a single receptive field; further, this application adds an adaptive fine-grained attention module to the neural network to perform weighted fusion of the second feature map, perform superpixel segmentation on each second feature map, analyze the grayscale difference between each superpixel block and its adjacent superpixel block in each second feature map, and obtain the possibility of pests in each superpixel block; enhances the refined recognition of the edge of the lesion The recognition ability of the method is improved to accurately locate the early lesions of the jujube tree leaves, suppress the interference of background noise, and improve the recognition ability of fine-grained lesions on the leaves; then, the target super-pixel blocks in each second feature map are screened based on the possibility of the pest; based on the grayscale difference between the target super-pixel block and its adjacent super-pixel blocks, the adjacent super-pixel blocks of the target super-pixel block are selected to be merged with the target super-pixel block, and the lesion area in each second feature map is obtained according to the morphological characteristics of the merged area; the accuracy and reliability of the identification of the lesion area are improved; by adding the shape characteristics of the early lesion area of ​​the leaf, it is helpful to improve the recognition and timeliness; further, the size of the lesion area in each second feature map is used to determine the size of the local window divided by each second feature map. After the local window division of each second feature map, the multi-head self-attention mechanism is used to obtain the weight of each second feature map; the local window size is dynamically adjusted based on the size of the lesion area, which improves the multi-head self-attention's ability to capture the long-range correlation of lesion features and optimizes the rationality of feature weight distribution, which helps to improve the recognition sensitivity of the neural network model in sparse lesion scenarios; finally, all the second feature maps are weighted fused and extracted through the multi-stage feature extraction of the ResNet50 network structure. , obtain the output of the neural network, and obtain the disease and pest monitoring results of jujube trees, which enhances the robustness of the neural network model to the complexity of the jujube tree growth environment. By capturing high-level feature information through the deep network, it effectively suppresses the negative impact of factors such as illumination changes and occlusion of branches and leaves on the detection results. In addition, the grayscale difference analysis mechanism of the adaptive fine-grained attention module improves the early warning ability of latent diseases and pests of jujube trees. By identifying the weak grayscale gradient features between superpixel blocks, it can issue an early warning before the symptoms visible to the naked eye appear, which wins critical time for grasping the window period for jujube tree prevention and control, and improves the accuracy of jujube tree disease and pest monitoring and identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0032] Figure 1 A flowchart of the steps of a method for monitoring and identifying pests and diseases of jujube trees based on remote sensing technology provided in one embodiment of the present application;

[0033] Figure 2 Functional flow chart of the adaptive fine-grained attention module;

[0034] Figure 3 It is the weighted fusion relationship graph of the second feature map. DETAILED DESCRIPTION

[0035] In order to further illustrate the technical means and effects adopted by this application to achieve the predetermined invention purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the method and system for monitoring and identifying pests and diseases of jujube trees based on remote sensing technology proposed in this application, its specific implementation methods, structures, features and effects. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics of one or more embodiments may be combined in any suitable form.

[0036] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0037] The specific scheme of the method and system for monitoring and identifying date tree pests and diseases based on remote sensing technology provided by this application is described in detail below with reference to the accompanying drawings.

[0038] See also Figure 1 , which shows a flowchart of a method for monitoring and identifying pests and diseases of jujube trees based on remote sensing technology provided by an embodiment of the present application, the method comprising the following steps:

[0039] S1, collect jujube tree vegetation images, input the jujube tree vegetation images into the neural network, and obtain a first feature map after downsampling; use multiple convolution layers with different convolution kernel sizes to obtain multi-scale feature maps of the first feature map, which are recorded as second feature maps.

[0040] In this example, a high-resolution RGB camera is mounted on a drone to capture images of jujube trees. These high-resolution images include leaves, fruits, and trunks, showing the characteristics of healthy plants and those with pests and diseases.

[0041] During the growth of jujube trees, black spot disease causes black spots to appear on the leaves. The death of some leaf tissues indicates the disappearance of chlorophyll, and yellowing of leaves appears around the spots. Therefore, pests and diseases can be detected and identified based on the difference in pixel values ​​between the diseased area and the healthy area in the jujube tree vegetation image.

[0042] Since the lesions appear in different sizes at different stages of jujube tree diseases, some of the images obtained from a large number of jujube tree vegetation images contain fine-grained lesion features, making it difficult to identify diseases and pests in the early stages.

[0043] In order to monitor and identify pests and diseases based on jujube tree vegetation images, this embodiment first preprocesses the jujube tree vegetation images. Specifically, the jujube tree vegetation images are enhanced using image enhancement technology to solve the problem of image clarity being affected by insufficient lighting conditions. This embodiment uses Retinex image enhancement technology, which is an existing technology. The specific process will not be repeated here. The implementer can choose other existing feasible image enhancement algorithms on his own, and this embodiment does not limit this.

[0044] The enhanced jujube tree vegetation image is normalized to obtain a preprocessed jujube tree vegetation image, normalizing the image pixel values ​​to between [0, 1] to facilitate better convergence during subsequent neural network training. In this embodiment, a maximum-minimum normalization method is used; implementers may choose other available normalization methods.

[0045] Furthermore, in this embodiment, the pre-processed jujube vegetation image is manually labeled with pest and disease areas to obtain label data, wherein the healthy vegetation area in the jujube vegetation image is labeled as 0, and the pest and disease vegetation area is labeled as 1.

[0046] This example inputs preprocessed jujube vegetation images and labeled data into a neural network model for training. A multi-scale, fine-grained neural network is then used to extract regional features of pests and diseases in the jujube vegetation images. Because the areas of diseased spots in jujube vegetation images vary within the same pixel grid size during neural network feature extraction, the neural network should assign different weights to the multi-scale feature maps based on image characteristics when capturing information at different scales. This allows the neural network to fuse the multi-scale information and capture early, fine-grained regional features of pests and diseases.

[0047] The neural network model in this embodiment consists of an encoder and a fully connected layer structure. The encoder specifically uses ResNet50 as the backbone network architecture, and the neural network model optimizer uses the Adam optimizer. The neural network model's loss function uses the GIoU loss and the binary cross-entropy loss. The output of the neural network model is a structured tensor, specifically including the confidence score, the coordinates of the predicted pest and disease area bounding box, and the classification probability.

[0048] From the ResNet50 network structure, we can see that it is a downsampling layer and a feature extraction of four-stage residual blocks. The four stages are respectively recorded as , where the network structure of ResNet50 is an existing well-known technology, and the specific content will not be repeated here.

[0049] The feature map obtained after the jujube vegetation image input into the neural network model passes through the downsampling layer is recorded as the first feature map and is expressed as The downsampling layer performs preliminary feature extraction on the jujube tree vegetation image to reduce the computational burden of subsequent layers.

[0050] Furthermore, the neural network model uses convolutional layers with different convolution kernel sizes to obtain the first feature map The multi-scale feature maps of the present embodiment use convolution kernel sizes of 1×1, 3×3, and 5×5, and the corresponding feature maps are represented as follows: , denoted as each second feature map. The multi-scale feature map contains feature information of different scales. When there are fine-grained disease spots in the jujube vegetation image, the convolution kernel with a smaller receptive field can focus on capturing the feature information of fine-grained disease spots.

[0051] S2, add an adaptive fine-grained attention module to the neural network to perform weighted fusion on the second feature map. The function flow chart of the adaptive fine-grained attention module is as follows: Figure 2 As shown in Figure 2, the adaptive fine-grained attention module specifically includes:

[0052] (1) Each second feature map is segmented into superpixels, and the grayscale difference between each superpixel block and its adjacent superpixel blocks in each second feature map is analyzed to obtain the possibility of insect infestation of each superpixel block.

[0053] The traditional weighted fusion method of multi-scale feature maps directly weights the feature maps, and the feature information of different scales is not distinguished, which will lead to the loss of fine-grained feature information, which is not conducive to the detection and identification of fine-grained diseased areas in jujube vegetation images. In order to better detect and identify fine-grained diseased areas in jujube vegetation images, this embodiment constructs an adaptive fine-grained attention module to The input is processed into the adaptive fine-grained attention module.

[0054] The adaptive fine-grained attention module assigns different weights to each second feature map for fusion based on the image feature information captured at different scales. It assigns higher weights to the fine-grained lesion area features captured within the receptive field of a small-scale convolution kernel, while the weights of the feature maps containing a large number of non-lesion areas captured within the receptive field of a large-scale convolution kernel are relatively lowered. This increases the focus on fine-grained features in the fusion of multi-scale feature maps.

[0055] The adaptive fine-grained attention module specifically includes the following:

[0056] First, yes Use convolution layers with a kernel size of 1×1 for position embedding, and then use the SLIC superpixel segmentation algorithm to segment each second feature map. Assume that in any second feature map containing small lesions, the lesions themselves appear black and the surrounding leaf tissues have lost their chlorophyll and appear yellow. Then, in the G superpixel blocks after segmentation, the first Superpixel blocks Contains the features of the lesion, then the superpixel block The grayscale value will be lower overall, and the super pixel block The grayscale value of the surrounding super pixel blocks will be higher than the super pixel block In this embodiment, G is set to 100, and the implementer can set it according to the actual situation. The SLIC superpixel segmentation algorithm is a well-known technology, and the specific process will not be described in detail.

[0057] For each second feature map, calculate the grayscale mean of all pixels in each super pixel block as the grayscale feature value of each super pixel block. The difference between the grayscale feature values ​​of a superpixel block and all its adjacent superpixel blocks is evaluated. The probability of black spots existing in a superpixel block is recorded as the probability of pest infestation. , the specific expression is:

[0058] Where, is the probability of pest infestation in the g-th superpixel block in each second feature map, is the grayscale eigenvalue of the g-th superpixel block in each second feature map, is the grayscale eigenvalue of the zth adjacent superpixel block in the gth superpixel block in each second feature map, G is the number of superpixel blocks in each second feature map, is the number of superpixel blocks adjacent to the g-th superpixel block in each second feature map.

[0059] It should be noted that, in this embodiment, the concept of adjacent means that there is a common boundary between two super-pixel blocks, that is, an adjacent super-pixel block represents a super-pixel block that has a common boundary with the g-th super-pixel block; the difference represents the degree of difference between two variables, which can be calculated specifically by using the difference, the absolute value of the difference, the square of the difference, the ratio, etc., and this embodiment does not impose any restrictions on this.

[0060] (2) Filter the target superpixel blocks in each second feature map based on the possibility of the pest; select the neighboring superpixel blocks of the target superpixel block and merge them with the target superpixel block based on the grayscale difference between the target superpixel block and its neighboring superpixel blocks, and obtain the diseased spot area in each second feature map according to the morphological characteristics of the merged area.

[0061] Next, for each second feature map, the pest infestation probability of all superpixels in it is calculated. The Otsu thresholding algorithm is then used to automatically determine the segmentation threshold. Superpixels with a pest infestation probability greater than the segmentation threshold are selected as target superpixels. The Otsu thresholding algorithm is a well-known technique, and the detailed process is omitted here.

[0062] For each target superpixel block, the superpixel blocks whose grayscale eigenvalues ​​are greater than or equal to the grayscale eigenvalues ​​of the target superpixel block in all adjacent superpixel blocks are merged with the target superpixel block. Since this application aims to improve the detection accuracy of fine-grained pest and disease areas, that is, early pest and disease areas are small and difficult to find, and early pest and disease areas generally appear in a circular shape, this embodiment further filters out superpixel blocks containing disease spots based on whether the merged result appears in an approximately circular shape.

[0063] For the merged superpixel block area, this embodiment uses the least squares circle fitting method to obtain the fitting circle of the merged area and calculates the overlap area ratio between the merged area and the fitting circle. ,like If the value of the combined area is greater than a preset first threshold, the combined area is determined to be a diseased area; otherwise, the combined area is determined to be a normal area. Least squares circle fitting is a well-known technique, and the specific process will not be described in detail. In this embodiment, the first threshold is set to 0.7, and the implementer can set it according to the specific situation.

[0064] (3) Using the size of the lesion area in each second feature map, determine the size of the local window for dividing each second feature map. After dividing each second feature map into local windows, use the multi-head self-attention mechanism to obtain the weight of each second feature map.

[0065] For all the lesion areas screened out in each second feature map, their minimum circumscribed matrix is ​​obtained, and the length and width of the matrix with the largest area among all the minimum circumscribed matrices in each second feature map are used as the size of the local window required to divide the corresponding second feature map.

[0066] After the second feature map is divided into local windows, the multi-head self-attention mechanism is used in each local window to calculate the self-attention matrix of the second feature map. The local block strategy of the second feature map can solve the problem that the traditional Softmax function lacks local bias in the global calculation of the entire second feature map, which leads to abnormal attention weights of some feature areas during the training of the neural network model. Among them, the multi-head self-attention mechanism is an existing well-known technology, and the specific process will not be described in detail. The calculation formula of the self-attention mechanism can be expressed as:

[0067] ;in, Represents the query vector, key vector, and value vector after the feature map is mapped by the linear layer. represents the matrix transpose, is the dimension of the key vector, is the normalization function.

[0068] Then, the attention matrix of each second feature map is used to calculate the weight of each second feature map. The calculation formula can be expressed as:

[0069] ;in, Represents the number of rows and columns of the attention matrix, is the first Rank Elements of a column.

[0070] After weight calculation, we get The weights are Therefore, the neural network model of this embodiment can dynamically and adaptively adjust the weights of each second feature map according to the feature information of the jujube tree vegetation image.

[0071] (4) Perform weighted fusion on all second feature maps.

[0072] pass Perform weighted fusion , get the feature map , specifically:

[0073]

[0074] The adaptive multi-scale feature map weight adjustment increases the weight of the local fine-grained feature map when there are a large number of small diseased areas in the jujube vegetation image, improves the neural network model's attention to fine-grained features, and reduces the influence of irrelevant background features. The second feature map weighted fusion relationship diagram is shown in Figure 3 shown.

[0075] S3, through the multi-stage feature extraction of the ResNet50 network structure, obtains the output of the neural network and obtains the pest and disease monitoring results of the jujube trees.

[0076] Feature map after weighted fusion Enter the ResNet50 network structure The new feature map obtained after feature extraction in the stage Repeat the above operation on the second feature map to obtain the fused feature map .

[0077] Correspondingly, in the ResNet50 network structure In the stage, the above feature maps are repeated in sequence The acquisition process gets the final feature map . Since the jujube vegetation images taken by drones contain a large number of healthy leaves, diseased and insect-spotted leaves, and the structure, texture, and grayscale features of the trunk, they constitute a scene with complex feature content. The neural network model of this embodiment uses a multi-stage feature extraction of the ResNet50 network structure to obtain high-level features in the jujube vegetation images, encode the features in the complex scene, and complete fine-grained detection and identification of disease and insect pest areas.

[0078] Feature Map A one-dimensional feature vector is obtained through a global average pooling layer and a flattening operation. Finally, a fully connected network is used to output the neural network model detection results. The neural network model detection results and the labeled data are compared using the Giounity Over Union (GIoU) ​​loss and binary cross entropy loss to calculate the loss. The neural network model parameters are then optimized to complete the training of the neural network model. Neural network training is a well-known technique, and the specific process will not be detailed here.

[0079] Finally, the output of the trained neural network can be used to determine whether there are pest and disease areas in the jujube vegetation image and outline the pest and disease areas. According to the coordinates of the positioning frame of the outlined pest and disease area, the union area of ​​all positioning frames in the jujube vegetation image is taken as the ratio to the overall image area of ​​the jujube vegetation image to calculate the proportion of pest and disease areas in the jujube vegetation image. Carry out disease and insect pest monitoring and identification of jujube trees. The larger it is, the more disease and insect pest areas there are in the jujube vegetation image, and the more serious the disturbance of disease and insect pests to the jujube trees is.

[0080] If the proportion If the value is greater than the preset second threshold, it is determined that there is a pest and disease area in the corresponding area of ​​the jujube tree vegetation image, and the jujube tree needs to be treated for pests and diseases. Otherwise, the jujube tree does not need to be treated for pests and diseases. In this embodiment, the second threshold is set to 0.3, and the implementer can set it according to the actual situation. This embodiment does not impose any restrictions on this.

[0081] Based on the same inventive concept as the above-mentioned method, an embodiment of the present application also provides a jujube tree disease and pest monitoring and identification system based on remote sensing technology, comprising a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, the steps of any one of the above-mentioned jujube tree disease and pest monitoring and identification methods based on remote sensing technology are implemented.

[0082] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0083] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0084] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for monitoring and identifying diseases and insect pests of jujube trees based on remote sensing technology, characterized in that: The method comprises the following steps: Collecting an image of jujube tree vegetation, inputting the image into a neural network, and obtaining a first feature map after downsampling; using multiple convolutional layers with different convolution kernel sizes to obtain multi-scale feature maps of the first feature map, recorded as respective second feature maps; wherein the neural network is a ResNet50 network structure; Add an adaptive fine-grained attention module to the neural network to perform weighted fusion on the second feature map, including: Perform superpixel segmentation on each second feature map, analyze the grayscale difference between each superpixel block and its adjacent superpixel blocks in each second feature map, and obtain the possibility of insect infestation of each superpixel block; Based on the pest possibility, a target superpixel block in each second feature map is screened; based on the grayscale difference between the target superpixel block and its neighboring superpixel blocks, a neighboring superpixel block of the target superpixel block is selected and merged with the target superpixel block; and a diseased spot area in each second feature map is obtained according to the morphological characteristics of the merged area, where the diseased spot area is a merged area having an overlapping area ratio greater than a preset first threshold, the overlapping area ratio being the overlapping area ratio between the merged area and a fitting circle, where the fitting circle is obtained by least squares circle fitting of the merged area. The size of the lesion area in each second feature map is used to determine the size of the local window divided by each second feature map. After dividing each second feature map into local windows, a multi-head self-attention mechanism is used to obtain the weight of each second feature map; All second feature maps are weightedly fused, and multi-stage feature extraction is performed using the ResNet50 network structure to obtain the output of the neural network and the pest and disease monitoring results of jujube trees.

2. The method for monitoring and identifying pests and diseases of jujube trees based on remote sensing technology as claimed in claim 1, wherein: The determination of the pest possibility of each superpixel block includes: For each second feature map, the grayscale mean of all pixels in each superpixel block is calculated, and the sum of the differences between the grayscale means of each superpixel block and all its adjacent superpixel blocks is calculated. The possibility of insect infestation of each superpixel block is determined by the proportion of the sum value of each superpixel block in the cumulative sum of the sum values ​​of all superpixel blocks in its second feature map.

3. The method for monitoring and identifying jujube tree diseases and insect pests based on remote sensing technology according to claim 1, wherein: The step of screening target superpixel blocks in each second feature map based on the pest possibility includes: A segmentation threshold of the pest possibility of all superpixel blocks in each second feature map is obtained, and a superpixel block whose pest possibility is greater than the segmentation threshold is used as a target superpixel block.

4. The method for monitoring and identifying jujube tree pests and diseases based on remote sensing technology as claimed in claim 2, wherein: The selecting a neighboring superpixel block of the target superpixel block and merging the target superpixel block with the target superpixel block comprises: For any target super-pixel block, if the grayscale mean of its neighboring super-pixel block is greater than or equal to the grayscale mean of any target super-pixel block, its neighboring super-pixel block is merged with the any target super-pixel block.

5. The method for monitoring and identifying pests and diseases of jujube trees based on remote sensing technology according to claim 1, wherein: Determining the size of each local window divided by the second feature map includes: For each second feature map, the minimum circumscribed matrix of each lesion area is obtained, and the length and width of the minimum circumscribed matrix with the largest area are used as the size of dividing the local window.

6. The method for monitoring and identifying pests and diseases of jujube trees based on remote sensing technology according to claim 1, wherein: Determining the weight of each second feature map includes: A multi-head self-attention mechanism is used to obtain the attention matrix of each second feature map, and the weight of each second feature map is the mean of all elements in the attention matrix.

7. The method for monitoring and identifying pests and diseases of jujube trees based on remote sensing technology according to claim 1, wherein: The pest and disease monitoring results of the jujube trees include: The output of the neural network includes the coordinates of the predicted boundary boxes of the disease and insect pest areas in the jujube tree vegetation image, and determines the proportion of the union area of ​​all the predicted boundary boxes of the disease and insect pest areas in the jujube tree vegetation image in the jujube tree vegetation image; When the proportion is greater than a preset second threshold, it is determined that there is a pest and disease area in the corresponding area of ​​the jujube tree vegetation image, and the jujube trees need to be treated for pests and diseases. Otherwise, the jujube trees do not need to be treated for pests and diseases.

8. A jujube tree pest and disease monitoring and identification system based on remote sensing technology, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

Images