Bursaphelenchus xylophilus epidemic wood monitoring method and system based on remote sensing satellite and storage medium

Abstract

The invention relates to the technical field of image data processing, and discloses a pine wood nematode disease epidemic wood monitoring method and system based on a remote sensing satellite and a storage medium. The model is trained by using the tile sample data containing the labels with different lesion degrees, so that the diseased wood with different lesion degrees can be identified, the efficiency is high, the selection of different feature combinations and the design of different monitoring methods for different lesion wood scenes are avoided, the obtained pine wood nematode disease change degree is more accurate, and the accuracy of the pine wood nematode disease change degree is improved. And clearer guidance is provided for subsequent epidemic wood removal and treatment work. And fusion of a remote sensing information extraction and monitoring technology means and a deep learning technology means is realized, image enhancement processing means such as radiation enhancement, resolution enhancement and the like are utilized, the saliency of the range graphic information of the pine wood nematode disease is expanded, and meanwhile, an all-around pseudo-pattern spot removal method is utilized, so that the method is more efficient and efficient. And the monitoring accuracy of the diseased wood is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to a remote sensing satellite-based method, system and storage medium for monitoring trees infected with pine wood nematodes. Background Art

[0002] Pine wilt disease, also known as pine wilt, is a common forest pest and a destructive pest that spreads quickly and is difficult to control. It has occurred in many areas and has become a disaster. Early detection and felling of plants infected with pine wilt disease can prevent the disease from spreading and greatly reduce the damage to the local ecological environment and the economic losses caused by the disease.

[0003] Traditional monitoring methods for pine wilt disease are mostly manual monitoring, but due to the complex topography and scattered distribution of pine wilt disease outbreak sites, manual monitoring methods cannot fundamentally solve the problem. In order to alleviate the pressure of manual monitoring, researchers at home and abroad have developed a variety of related technical methods for automatic monitoring, which can be divided into two categories according to the monitoring methods: generating Lab color space images and feature images based on satellite orthophoto fusion images, and comprehensively judging the scope of pests and diseases through segmentation algorithms and vegetation feature thresholds, but there is a problem of low efficiency, which is difficult to support the explosive growth of tasks and needs. In addition, with the continuous development of satellite sensors, methods such as intelligent extraction of pest and disease ranges based on satellite images, aerial photographs, and drone data have emerged, but intelligent extraction of satellite data can only obtain the approximate area of ​​pest-infested trees, and cannot accurately judge the severity of tree pests and diseases; intelligent extraction of aerial photographs and drone data has high resolution, but there are problems such as high cost and small range of single images. Summary of the invention

[0004] The object of the present invention is to provide a method, system and storage medium for monitoring pine wood nematode diseased trees based on remote sensing satellites which are highly efficient and capable of distinguishing the degree of pests and diseases.

[0005] In order to achieve the above object, the present invention provides a method for monitoring pine wood nematode diseased trees based on remote sensing satellites, comprising the following steps: S1: Acquire remote sensing image data collected by remote sensing satellites and pre-process the remote sensing image data; S2: Label the pine wilt diseased trees in the remote sensing image data, and give different labels according to the degree of disease and insect damage of the pine wilt diseased trees. The degree of disease and insect damage of the pine wilt diseased trees is divided into newly infected, diseased and dead, and obtain the pine wilt diseased tree label data containing the information of pine wilt diseased trees; S3: cutting the remote sensing image data processed in step S1 and the pine wood nematode diseased tree label data obtained in step S2 to obtain tile sample data, where the tile sample data includes image tiles and label image tiles; S4: dividing the tile sample data into a training tile sample set and a verification tile sample set, wherein the training tile sample set is an accurate and representative sample set, and the verification tile sample set is a sample set containing typical sample types; S5: Construct a pine wood nematode diseased wood extraction model, which is a deep convolutional neural network model; S6: Iteratively train the pine wood nematode diseased wood extraction model using the training tile sample set, and evaluate the pine wood nematode diseased wood extraction model using the validation tile sample set to obtain the trained pine wood nematode diseased wood extraction model; S7: acquiring remote sensing image data to be detected, extracting diseased and insect-infested trees from the remote sensing image data to be detected using the trained pine wood nematode disease-infested tree extraction model, and obtaining a grid probability map of the disease and insect pest range; S8: Combine the image characteristics of remote sensing image data and the grid probability map of pest and disease range, remove pseudo spots from the pest and disease area of ​​remote sensing image data, and obtain a grid result; S9: Based on the raster result obtained in step S8, the result is vectorized, and primary and secondary analysis, small face removal and vector thinning and smoothing processing are performed to obtain the final map containing pests and diseases in different disease stages.

[0006] As a preferred solution, in step S1, the remote sensing image data uses image data collected by Beijing-3N satellite.

[0007] As a preferred solution, in step S1, the seed point of the pine wilt diseased tree is outlined by gesture traction, and the Graph cut algorithm is used to achieve multi-scale segmentation around the seed point, and a weighted graph of the similarity between the surrounding pixel points and the seed point is established, and the feature difference threshold is calculated to obtain the range label of the pine wilt diseased tree.

[0008] As a preferred solution, in step S2, the remote sensing image data is homogenized using a spatial reference template of a remote sensing satellite, so that the remote sensing image data has uniform tone, moderate contrast, is close to natural true color, and has rich spectral information; Among them, the uniform light and color include: performing multi-scale feature statistics on the remote sensing image data based on the spatial reference template of the remote sensing satellite, and constructing the original eigenvalue space and the target eigenvalue space respectively; then performing Wallis transformation processing based on the original eigenvalue space and the target eigenvalue space to obtain the target eigenvalue corresponding to each pixel of the image; finally, the processed remote sensing image sample data is obtained through color mode conversion.

[0009] As a preferred solution, in step S5, the pine wood nematode diseased wood extraction model is based on an encoder-decoder structure and designs an MPT network, wherein the encoder part designs a CNN+Transformer network backbone MPT, and the Transformer part adopts a local parallel method to give the model the advantages of convolution while maintaining the advantages of Self-Attention. While enhancing the feature extraction capability, the model efficiency is also optimized; the decoder part uses UPerHead, which is composed of a detection head of a pyramid pooling module PPM and a top-down feature pyramid network FPN decoder.

[0010] As a preferred solution, in step S6, the training tile sample set obtained in step S4 is subjected to radiation enhancement processing and resolution enhancement processing, wherein the radiation enhancement processing includes brightness, hue, saturation, contrast, noise disturbance, and image blur, and the resolution enhancement processing includes random scaling, random horizontal flipping, and random vertical flipping; the training tile sample set obtained after the radiation enhancement processing and the resolution enhancement processing is used to train the pine wood nematode diseased wood extraction model; During training, the pine wilt diseased wood extraction model is input into the pine wilt diseased wood extraction model, and the binary cross entropy loss function is adopted. The training parameters are set for iterative training. When the training reaches the iterative stop condition, the trained pine wilt diseased wood extraction model is obtained and encapsulated. The set training parameters include learning rate, learning rate decay, and batch size. The initial learning rate is set to 0.001, the learning rate decay is set to 0.000001, and the batch size is set to 16.

[0011] As a preferred solution, in step S8, morphological de-aliasing and similarity de-aliasing are used to remove pseudo spots from the pest and disease areas of the remote sensing image data, wherein: Morphological de-aliasing uses morphological characteristic factors to remove pseudo-image spots. Morphological characteristic factors include image spot area, narrow length, and compactness. The threshold of morphological characteristic factors is set according to the ground features of pine wood nematode-infected trees, and targets with values ​​greater than the threshold are removed. Similarity de-aliasing uses spectral feature factors and texture feature factors to remove artifacts. The remote sensing image data to be detected and the grid probability map of the pest and disease range are loaded. The spectral feature factor threshold and texture feature factor threshold of pine wilt diseased trees with different degrees of pests and diseases are calculated using the ground features of pine wilt diseased trees that have just become ill, are sick, or have died in the remote sensing image sample data. The spots that differ greatly from the spectral feature factor threshold and texture feature factor threshold are removed.

[0012] As a preferred solution, in step S9, vector thinning and smoothing processing is first performed based on vector thinning of the Douglas-Peucker algorithm to simplify the vector data by removing unnecessary points while maintaining the main features, and then the data points are fitted based on the Bezier curve to generate a smooth curve.

[0013] The present invention also provides a remote sensing satellite-based pine wood nematode disease monitoring system, comprising: The label collection module is used to obtain remote sensing image data collected by remote sensing satellites, mark pine wilt diseased trees in the remote sensing image data, and give different labels according to the degree of disease and insect pests of pine wilt diseased trees. The degree of disease and insect pests of pine wilt diseased trees is divided into newly infected, medium infected and dead, and obtain pine wilt diseased tree label data containing pine wilt diseased tree information; Sample preprocessing module, used to preprocess remote sensing image data; The sample making module is used to cut the remote sensing image data processed by the sample preprocessing module and the pine wood nematode diseased tree label data obtained by the label acquisition module to obtain tile sample data and form a sample library. The tile sample data includes image tiles and label image tiles. The sample management module is used to manage the sample library, store, query, select, edit and display the pine wood nematode disease samples produced by the sample production module, and provide support for subsequent model training; The model training module is used to build a pine wilt-infected wood extraction model and train the model. It uses pine wilt-infected wood samples in the sample library as input, supports custom configuration of training framework, network structure, and training parameters to carry out cluster training, and supports real-time monitoring and display of training progress and model accuracy. The model testing module is used to test and evaluate the model generated by the training iteration to evaluate the effect and generalization ability of the model; Model management module, used to manage, package and deploy the trained pine wood nematode disease extraction model; The model prediction module is used to obtain the remote sensing image data to be detected, and use the trained pine wood nematode diseased tree extraction model to extract the diseased and insect-infested trees from the remote sensing image data to be detected, and obtain a raster probability map of the disease and insect pest range; and combine the image characteristics of the remote sensing image data and the raster probability map of the disease and insect pest range to remove pseudo-spots from the disease and insect pest area of ​​the remote sensing image data to obtain a raster result; and based on the obtained raster result, the result is vectorized, and maximum and minimum analysis, small facet removal and vector thinning and smoothing processing are performed to obtain the final disease and insect pest spots at different disease stages.

[0014] In addition, the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned remote sensing satellite-based pine wood nematode diseased wood monitoring method is implemented.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention assigns labels of newly infected, diseased and dead according to the degree of disease and insect pests of pine wood nematode diseased trees. The model is trained by tile sample data containing the above-mentioned different disease degree labels, and diseased trees with different disease and insect pests degrees can be identified with high efficiency. It avoids selecting different feature combinations and designing different monitoring methods for different diseased diseased tree scenes. The obtained pine wood nematode disease degree is more accurate, providing clearer guidance for subsequent diseased tree removal operations.

[0016] It also integrates the technical means of remote sensing information extraction and monitoring with deep learning technical means, and uses image enhancement processing methods such as radiation enhancement and resolution enhancement to expand the prominence of graphic information on the range of pine wilt diseased trees. At the same time, it uses a full range of pseudo-spot removal methods to greatly improve the accuracy of pine wilt diseased tree monitoring.

[0017] In the present invention, for the monitoring scenario of pine wood nematode diseased trees, with deep learning technology as the core, a convenient and efficient pine wood nematode diseased tree monitoring system covering the entire business process is constructed. The system provides capabilities such as sample data production and management, model training iteration, model testing, model management, and model prediction, and realizes the full-process business opening of pine wood nematode diseased tree extraction model training and application. At the same time, the training process can monitor and display the training progress and model accuracy in real time, and the model test results can be visualized and compared and analyzed, fully supporting the verification of model effects and generalization capabilities, and providing reliable model support for the extraction of pest-infected tree elements. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The present invention is a flowchart of a method for monitoring trees infected with pine wood nematodes based on remote sensing satellites according to an embodiment of the present invention.

[0019] Figure 2 It is an information flow diagram of a method for monitoring pine wood nematode diseased trees based on remote sensing satellites according to an embodiment of the present invention.

[0020] Figure 3 It is a structural diagram of the backbone network of the pine wood nematode diseased wood extraction model according to an embodiment of the present invention.

[0021] Figure 4 It is a training flow chart of a pine wood nematode diseased wood extraction model according to an embodiment of the present invention.

[0022] Figure 5The present invention is a block diagram of the principle of a remote sensing satellite-based pine wood nematode disease monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0024] Embodiment 1 like Figures 1 to 4 As shown, the method for monitoring pine wood nematode diseased trees based on remote sensing satellites in a preferred embodiment of the present invention comprises the following steps: S1: Acquire remote sensing image data collected by remote sensing satellites and pre-process the remote sensing image data; S2: Label the pine wilt diseased trees in the remote sensing image data, and give different labels according to the degree of disease and insect damage of the pine wilt diseased trees. The degree of disease and insect damage of the pine wilt diseased trees is divided into newly infected, diseased and dead, and obtain the pine wilt diseased tree label data containing the information of pine wilt diseased trees; S3: Crop the remote sensing image data processed in step S1 and the pine wood nematode diseased tree label data obtained in step S2 to obtain tile sample data, where the tile sample data includes image tiles and label image tiles. In this embodiment, the remote sensing image sample data is cropped into image tiles of 512×512 pixels, and the pine wood nematode diseased tree label data is cropped into label image tiles of 512×512 pixels. S4: dividing the tile sample data into a training tile sample set and a verification tile sample set, wherein the training tile sample set is an accurate and representative sample set, and the verification tile sample set is a sample set containing typical sample types; S5: Construct a pine wilt diseased wood extraction model, which is a deep convolutional neural network model; specifically, the pine wilt diseased wood extraction model is a parallel attention network that combines convolution and Transformer; S6: Iteratively train the pine wood nematode diseased wood extraction model using the training tile sample set, and evaluate the pine wood nematode diseased wood extraction model using the validation tile sample set to obtain the trained pine wood nematode diseased wood extraction model; S7: acquiring remote sensing image data to be detected, extracting diseased and insect-infested trees from the remote sensing image data to be detected using the trained pine wood nematode disease-infested tree extraction model, and obtaining a grid probability map of the disease and insect pest range; S8: Combine the image characteristics of remote sensing image data and the grid probability map of pest and disease range, remove pseudo spots from the pest and disease area of ​​remote sensing image data, and obtain a grid result; S9: Based on the raster result obtained in step S8, the result is vectorized, and primary and secondary analysis, small face removal and vector thinning and smoothing processing are performed to obtain the final map containing pests and diseases in different disease stages.

[0025] The present invention gives labels of newly-infected, diseased and dead according to the degree of disease and insect pests of pine wood nematode diseased trees, and trains the model through tile sample data containing the above-mentioned different disease degree labels, so that diseased trees with different disease and insect pest degrees can be identified, with high efficiency, avoiding the selection of different feature combinations and the design of different monitoring methods for different diseased diseased tree scenes, and the obtained pine wood nematode disease degree is more accurate, providing clearer guidance for subsequent diseased tree removal operations. And realize the integration of technical means of remote sensing information extraction and monitoring with deep learning technical means, and use image enhancement processing means such as radiation enhancement and resolution enhancement to expand the significance of the graphic information of the range of pine wood nematode diseased trees, and use a full range of pseudo-image spot removal methods to greatly improve the accuracy of pine wood nematode diseased tree monitoring.

[0026] Specifically, in step S1, the remote sensing image data uses the image data collected by Beijing No. 3N satellite, and the resolution of the image data collected by Beijing No. 3N satellite is 0.3. And when marking the pine wood nematode diseased trees in the remote sensing image sample data, the pine wood nematode diseased tree seed points are outlined by gesture traction. The seed point refers to the initial pixel point used to start the image filling algorithm in image processing. The Graph cut algorithm is used to achieve multi-scale segmentation around the seed point, and a weighted graph of the similarity between each surrounding pixel point and the seed point is established, and the feature difference threshold is calculated to obtain the range label of the pine wood nematode diseased tree. Specifically, in distinguishing the degree of pests and diseases of pine wood nematode diseased trees, the label of the diseased tree area that has just become ill is set to 1, and the label of the diseased tree area in the diseased state is set to 2; the diseased and dead diseased tree area is set to 3, and the rest of the area is set to 0.

[0027] In step S2, the remote sensing image data is homogenized using the spatial reference template of the remote sensing satellite, so that the remote sensing image data has uniform tone, moderate contrast, close to natural true color, and rich spectral information; A spatial reference template is a feature class or layer that is used as a template to set the values ​​of a spatial reference.

[0028] Among them, the uniform light and color include: performing multi-scale feature statistics on the remote sensing image data based on the spatial reference template of the remote sensing satellite, and constructing the original eigenvalue space and the target eigenvalue space respectively; then performing Wallis transformation processing based on the original eigenvalue space and the target eigenvalue space to obtain the target eigenvalue corresponding to each pixel of the image; finally, the processed remote sensing image sample data is obtained through color mode conversion.

[0029] In addition, this embodiment also performs band processing on the remote sensing image data before uniform light and color.

[0030] Specifically, the Wallis transform is processed through the Wallis filter. The Wallis filter is a linear transformation for the local part of the image, so that the grayscale mean and variance of different images or different positions of the image tend to similar values, that is, the contrast is increased in areas with small image contrast, and the contrast is reduced in areas with large image contrast. The mean of the image represents the hue and brightness, while the variance represents the range of variation of the pixel value. The transition between different images or different positions of the image should be continuous, with consistent hue, brightness and contrast, consistent grayscale dynamic range, and the desired mean and variance should be consistent. The principle of the Wallis filter is to reduce the differences in mean and variance between images to achieve consistent image color. The formula for the change of grayscale value of an image processed by the Wallis transform is:

[0031] In the formula, is the gray value of the original image; is the gray value of the resulting image after Wallis transformation; is the local mean of the original image; is the local variance of the original image; is the target value of the local mean of the result image; is the target value of the local variance of the result image; is the expansion constant of the image variance; is the brightness coefficient of the image.

[0032] In step S4, a deep learning training system is used to divide all tile sample data into a training tile sample set and a verification tile sample set in proportion or interactively, and the ratio is generally 7:3. The typical sample types contained in the verification tile sample set include the types of newly infected pine wood nematode diseased trees, the types of diseased pine wood nematode diseased trees, and the types of dead pine wood nematode diseased trees. In addition, on remote sensing images, newly infected and diseased pine wood nematode diseased trees are easily confused with trees in the flowering period, and the characteristics of dead pine wood nematode diseased trees are easily confused with bare land, resulting in mis-extraction. Therefore, in the process of making tile samples, positive and negative samples are usually distinguished, and the image area containing easily confused objects is cut as a negative sample.

[0033] In step S5, the pine wood nematode diseased wood extraction model is based on the encoder-decoder structure, and the MPT network (Multi-Level Parallel Transformer, multi-level parallel attention network) is designed. The encoder part designs the backbone network of CNN+Transformer, and the Transformer part adopts a local parallel method, which makes the model have the advantages of convolution and maintains the advantages of Self-Attention. While enhancing the feature extraction capability, the model efficiency is also optimized; the decoder part uses UPerHead, which is composed of the detection head of the pyramid pooling module PPM (Pyramid Pooling Module) and the top-down feature pyramid network FPN (Feature Pyramid Network) decoder. The encoded features first pass through the PPM detection head, and then perform feature fusion at each layer of FPN, and predict the fused multi-scale features to complete the construction of the pine wood nematode diseased wood extraction model.

[0034] like Figure 3 As shown, the backbone network of the pine wood nematode diseased wood extraction model of this embodiment includes a dimensionality reduction sampling module, a first encoder architecture, a second encoder architecture, and a third encoder architecture connected in sequence. Among them, the first encoder architecture, the second encoder architecture, and the third encoder architecture all include a parallel multi-head attention module and a multi-scale feature aggregation module. The parallel multi-head attention module mainly adopts a PTB (parallel transformer block) module, and the multi-scale feature extraction module mainly adopts a pyramid convolution module PRM (Pyramid Reduction Module). Specifically, in the parallel multi-head attention module, the H×W image is input into the pyramid convolution module to obtain three feature maps of sizes H×W, H / 2×W / 2, and H / 4×W / 4. The three feature maps are each input into a multi-layer attention mechanism layer based on a sliding window to output an output feature map. The three output feature maps, the input H×W image, and the three feature maps of sizes H×W, H / 2×W / 2, and H / 4×W / 4 obtained by convolution of the pyramid convolution module are input into the feature fusion of the multi-layer perceptron for feature fusion.

[0035] Further, in step S6, the image obtained in step S4 is subjected to radiation enhancement processing and resolution enhancement processing, wherein the radiation enhancement processing includes brightness, hue, saturation, contrast, noise disturbance, and image blur, and the resolution enhancement processing includes random scaling, random horizontal flipping, and random vertical flipping; the training tile sample set obtained after the radiation enhancement processing and the resolution enhancement processing is used to train the pine wood nematode diseased wood extraction model; During training, the pine wilt diseased wood extraction model is input into the pine wilt diseased wood extraction model, and the binary cross entropy loss function is adopted. The training parameters are set for iterative training. When the training reaches the iterative stop condition, the trained pine wilt diseased wood extraction model is obtained and encapsulated. The set training parameters include learning rate, learning rate decay, and batch size. The initial learning rate is set to 0.001, the learning rate decay is set to 0.000001, and the batch size is set to 16.

[0036] like Figure 4 As shown, the training process of the pine wood nematode disease extraction model in this embodiment is: (1) Forward propagation: the input data is passed through the neural network to obtain the prediction result; (2) Loss calculation: the loss value is calculated based on the prediction result and the true label; (3) Backward propagation: the network parameters are updated according to the loss value, usually using the gradient descent method; (4) Parameter update: the network parameters are adjusted according to the gradient information to reduce the loss value.

[0037] Specifically, in step S8, morphological de-aliasing and similarity de-aliasing are used to remove pseudo spots from the pest and disease areas of the remote sensing image data, wherein: Morphological de-aliasing uses morphological characteristic factors to remove pseudo-image spots. Morphological characteristic factors include image spot area, narrow length, and compactness. The threshold of the morphological characteristic factor is set according to the ground feature of pine wood nematode infected trees, and the targets with a value greater than the threshold are removed. Among them, compactness refers to the ratio of the square of the perimeter to the area. In this embodiment, the thresholds are set according to the ground feature characteristics: the area is 0.8 , the narrow length is 2, the compactness is 0.4, and objects larger than the threshold are removed during anti-aliasing.

[0038] Similarity de-aliasing uses spectral feature factors and texture feature factors to remove artifacts. The remote sensing image data to be detected and the grid probability map of the pest and disease range are loaded. The spectral feature factor threshold and texture feature factor threshold of pine wilt diseased trees with different degrees of pests and diseases are calculated using the ground features of pine wilt diseased trees that have just become ill, are sick, or have died in the remote sensing image sample data. The spots that differ greatly from the spectral feature factor threshold and texture feature factor threshold are removed.

[0039] Among them, the spectral characteristic factor is the maximum value, minimum value and average value of the red light band, green light band and blue light band of all n pixels in the extracted image range; Texture features use the angular second moment (ASM), also known as energy, to calculate the uniformity of image grayscale distribution and texture coarseness, which is used to distinguish pest and disease areas from other ground objects. The texture feature calculation formula is:

[0040] Among them, i and j represent rows and columns respectively, and N represents the number of rows and columns. is a pixel ( ).

[0041] In step S9, the major / minority analysis is mainly used to remove small patches, thereby improving the accuracy of the patch results. Face removal is to remove patches whose area is smaller than the set threshold. Vector thinning and smoothing is first based on the Douglas-Peucker Algorithm vector thinning, which simplifies the vector data by removing unnecessary points while maintaining the main features, and then fits the data points based on the Bézier Curves to generate a smooth curve. Vector thinning and smoothing makes the vector result data display effect better.

[0042] Weight-based regularization: In the process of vectorization of raster results, by controlling the weight value in the vector graphics fitting error, combined with the optimal rotation angle, the number of edges of the regularized graphics is fitted and output, and the optimal fitting result is output when the fitting error is minimized; Tolerance-based regularization: In the process of vectorizing raster results, the optimal rotation angle is found to ensure that the number of fitting edges is minimized and the tolerance limits of the fitted graphics and the source graphics are met at the same time, and the fitting result with the smallest fitting error is output.

[0043] Embodiment 2 like Figure 5 As shown, an embodiment of the present invention provides a pine wood nematode disease monitoring system based on remote sensing satellites, comprising: The label collection module is used to obtain remote sensing image data collected by remote sensing satellites, mark pine wilt diseased trees in the remote sensing image data, and give different labels according to the degree of disease and insect pests of pine wilt diseased trees. The degree of disease and insect pests of pine wilt diseased trees is divided into newly infected, medium infected and dead, and obtain pine wilt diseased tree label data containing pine wilt diseased tree information; Sample preprocessing module, used to preprocess remote sensing image data; The sample making module is used to cut the remote sensing image data processed by the sample preprocessing module and the pine wood nematode diseased tree label data obtained by the label acquisition module to obtain tile sample data and form a sample library. The tile sample data includes image tiles and label image tiles. The sample management module is used to manage the sample library, store, query, select, edit and display the pine wood nematode disease samples produced by the sample production module, and provide support for subsequent model training; The model training module is used to build a pine wilt-infected wood extraction model and train the model. It uses pine wilt-infected wood samples in the sample library as input, supports custom configuration of training framework, network structure, and training parameters to carry out cluster training, and supports real-time monitoring and display of training progress and model accuracy. The model testing module is used to test and evaluate the model generated by the training iteration to evaluate the effect and generalization ability of the model; Model management module, used to manage, package and deploy the trained pine wood nematode disease extraction model; The model prediction module is used to obtain the remote sensing image data to be detected, and use the trained pine wood nematode diseased tree extraction model to extract the diseased and insect-infested trees from the remote sensing image data to be detected, and obtain a raster probability map of the disease and insect pest range; and combine the image characteristics of the remote sensing image data and the raster probability map of the disease and insect pest range to remove pseudo-spots from the disease and insect pest area of ​​the remote sensing image data to obtain a raster result; and based on the obtained raster result, the result is vectorized, and maximum and minimum analysis, small facet removal and vector thinning and smoothing processing are performed to obtain the final disease and insect pest spots at different disease stages.

[0044] In this embodiment, for the monitoring scenario of pine wood nematode diseased trees, with deep learning technology as the core, a convenient and efficient pine wood nematode diseased tree monitoring system covering the entire business process is constructed. The system provides capabilities such as sample data production and management, model training iteration, model testing, model management, and model prediction, and realizes the full-process business of training and application of pine wood nematode diseased tree extraction model. At the same time, the training process can monitor and display the training progress and model accuracy in real time, and the model test results can be visualized and compared and analyzed, fully supporting the verification of model effects and generalization capabilities, and providing reliable model support for the extraction of pest-infected tree elements.

[0045] Embodiment 3 An embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the remote sensing satellite-based pine wood nematode disease monitoring method of embodiment 1 is implemented.

[0046] In summary, the embodiment of the present invention provides a pine nematode diseased wood monitoring method based on remote sensing satellites, which gives labels of newly-diseased, diseased and dead according to the degree of pests and diseases of pine nematode diseased wood, and trains the model by tile sample data containing the above-mentioned different lesion degree labels, and can identify diseased wood with different degrees of pests and diseases, with high efficiency, avoiding the selection of different feature combinations and design of different monitoring methods for different lesion diseased wood scenes, and the obtained pine nematode lesion degree is more accurate, providing clearer guidance for subsequent diseased wood removal operations. And realize the integration of remote sensing information extraction and monitoring technical means and deep learning technical means, and use image enhancement processing means such as radiation enhancement and resolution enhancement to expand the significance of pine nematode diseased wood range graphic information, and use a full range of pseudo-image spot removal methods to greatly improve the accuracy of pine nematode diseased wood monitoring. And provide a system and storage medium for implementing the above method.

[0047] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention. These improvements and substitutions should also be regarded as the scope of protection of the present invention.

Claims

1. A method for monitoring trees infected with pine wood nematodes based on remote sensing satellites, characterized in that: The steps include: S1: Acquire remote sensing image data collected by remote sensing satellites and pre-process the remote sensing image data; S2: Label the pine wilt diseased trees in the remote sensing image data, and give different labels according to the degree of disease and insect damage of the pine wilt diseased trees. The degree of disease and insect damage of the pine wilt diseased trees is divided into newly infected, diseased and dead, and obtain the pine wilt diseased tree label data containing the information of pine wilt diseased trees; S3: cutting the remote sensing image data processed in step S1 and the pine wood nematode diseased tree label data obtained in step S2 to obtain tile sample data, where the tile sample data includes image tiles and label image tiles; S4: dividing the tile sample data into a training tile sample set and a verification tile sample set, wherein the training tile sample set is an accurate and representative sample set, and the verification tile sample set is a sample set containing typical sample types; S5: Construct a pine wood nematode diseased wood extraction model, which is a deep convolutional neural network model; S6: Iteratively train the pine wood nematode diseased wood extraction model using the training tile sample set, and evaluate the pine wood nematode diseased wood extraction model using the validation tile sample set to obtain the trained pine wood nematode diseased wood extraction model; S7: acquiring remote sensing image data to be detected, extracting diseased and insect-infested trees from the remote sensing image data to be detected using the trained pine wood nematode disease-infested tree extraction model, and obtaining a grid probability map of the disease and insect pest range; S8: Combine the image characteristics of remote sensing image data and the grid probability map of pest and disease range, remove pseudo spots from the pest and disease area of ​​remote sensing image data, and obtain a grid result; S9: Based on the raster result obtained in step S8, the result is vectorized, and primary and secondary analysis, small face removal and vector thinning and smoothing processing are performed to obtain the final map containing pests and diseases in different disease stages.

2. The method for monitoring pine wood nematode diseased trees based on remote sensing satellites according to claim 1, characterized in that: In step S1, the remote sensing image data adopts the image data collected by Beijing-3N satellite.

3. The method for monitoring pine wood nematode diseased trees based on remote sensing satellites according to claim 1, characterized in that: In step S1, the seed point of the pine wilt diseased tree is outlined by gesture traction, and the Graph cut algorithm is used to achieve multi-scale segmentation around the seed point, and a weighted graph of the similarity between the surrounding pixel points and the seed point is established, and the feature difference threshold is calculated to obtain the range label of the pine wilt diseased tree.

4. The method for monitoring pine wood nematode diseased trees based on remote sensing satellites according to claim 1, characterized in that: In step S2, the remote sensing image data is homogenized using the spatial reference template of the remote sensing satellite, so that the remote sensing image data has uniform tone, moderate contrast, close to natural true color, and rich spectral information; Among them, the uniform light and color include: performing multi-scale feature statistics on the remote sensing image data based on the spatial reference template of the remote sensing satellite, and constructing the original eigenvalue space and the target eigenvalue space respectively; then performing Wallis transformation processing based on the original eigenvalue space and the target eigenvalue space to obtain the target eigenvalue corresponding to each pixel of the image; finally, the processed remote sensing image sample data is obtained through color mode conversion.

5. The method for monitoring pine wood nematode diseased trees based on remote sensing satellites according to claim 1, characterized in that: In step S5, the pine wood nematode diseased wood extraction model is based on the encoder-decoder structure and designs the MPT network. The encoder part designs the CNN+Transformer backbone network, and the Transformer part adopts a local parallel method to give the model the advantages of convolution while maintaining the advantages of Self-Attention. While enhancing the feature extraction capability, the model efficiency is also optimized. The decoder part uses UPerHead, which is composed of the detection head of the pyramid pooling module PPM and the top-down feature pyramid network FPN decoder.

6. The method for monitoring pine wood nematode diseased trees based on remote sensing satellites according to claim 1, characterized in that: In step S6, the training tile sample set obtained in step S4 is subjected to radiation enhancement processing and resolution enhancement processing, wherein the radiation enhancement processing includes brightness, hue, saturation, contrast, noise disturbance, and image blur, and the resolution enhancement processing includes random scaling, random horizontal flipping, and random vertical flipping; the pine wood nematode diseased wood extraction model is trained using the training tile sample set obtained after the radiation enhancement processing and the resolution enhancement processing; During training, the pine wilt diseased wood extraction model is input into the pine wilt diseased wood extraction model, and the binary cross entropy loss function is adopted. The training parameters are set for iterative training. When the training reaches the iterative stop condition, the trained pine wilt diseased wood extraction model is obtained and encapsulated. The set training parameters include learning rate, learning rate decay, and batch size. The initial learning rate is set to 0.001, the learning rate decay is set to 0.000001, and the batch size is set to 16.

7. The method for monitoring pine wood nematode diseased trees based on remote sensing satellites according to claim 1, characterized in that: In step S8, morphological de-aliasing and similarity de-aliasing are used to remove pseudo spots from the pest and disease areas of the remote sensing image data, wherein: Morphological de-aliasing uses morphological characteristic factors to remove pseudo-image spots. Morphological characteristic factors include image spot area, narrow length, and compactness. The threshold of morphological characteristic factors is set according to the ground features of pine wood nematode-infected trees, and targets with values ​​greater than the threshold are removed. Similarity de-aliasing uses spectral feature factors and texture feature factors to remove artifacts. The remote sensing image data to be detected and the grid probability map of the pest and disease range are loaded. The spectral feature factor threshold and texture feature factor threshold of pine wilt diseased trees with different degrees of pests and diseases are calculated using the ground features of pine wilt diseased trees that have just become ill, are sick, or have died in the remote sensing image sample data. The spots that differ greatly from the spectral feature factor threshold and texture feature factor threshold are removed.

8. The method for monitoring pine wood nematode diseased trees based on remote sensing satellites according to claim 1, characterized in that: In step S9, vector thinning and smoothing processing is first performed based on vector thinning of the Douglas-Peucker algorithm to simplify the vector data by removing unnecessary points while maintaining the main features, and then the data points are fitted based on the Bezier curve to generate a smooth curve.

9. A pine wood nematode disease-infected tree monitoring system based on remote sensing satellites, characterized in that: include: The label collection module is used to obtain remote sensing image data collected by remote sensing satellites, mark pine wilt diseased trees in the remote sensing image data, and give different labels according to the degree of disease and insect pests of pine wilt diseased trees. The degree of disease and insect pests of pine wilt diseased trees is divided into newly infected, medium infected and dead, and obtain pine wilt diseased tree label data containing pine wilt diseased tree information; Sample preprocessing module, used to preprocess remote sensing image data; The sample making module is used to cut the remote sensing image data processed by the sample preprocessing module and the pine wood nematode diseased tree label data obtained by the label acquisition module to obtain tile sample data and form a sample library. The tile sample data includes image tiles and label image tiles. The sample management module is used to manage the sample library, store, query, select, edit and display the pine wood nematode disease samples produced by the sample production module, and provide support for subsequent model training; The model training module is used to build a pine wilt-infected wood extraction model and train the model. It uses pine wilt-infected wood samples in the sample library as input, supports custom configuration of training framework, network structure, and training parameters to carry out cluster training, and supports real-time monitoring and display of training progress and model accuracy. The model testing module is used to test and evaluate the model generated by the training iteration to evaluate the effect and generalization ability of the model; Model management module, used to manage, package and deploy the trained pine wood nematode disease extraction model; The model prediction module is used to obtain the remote sensing image data to be detected, and use the trained pine wood nematode diseased tree extraction model to extract the diseased and insect-infested trees from the remote sensing image data to be detected, and obtain a raster probability map of the disease and insect pest range; and combine the image characteristics of the remote sensing image data and the raster probability map of the disease and insect pest range to remove pseudo-spots from the disease and insect pest area of ​​the remote sensing image data to obtain a raster result; and based on the obtained raster result, the result is vectorized, and maximum and minimum analysis, small facet removal and vector thinning and smoothing processing are performed to obtain the final disease and insect pest spots at different disease stages.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the remote sensing satellite-based method for monitoring pine wood nematode diseased trees as described in any one of claims 1 to 8 is implemented.

Images