Forest intelligent disease and pest inspection system based on disease feature extraction

Abstract

The application discloses a forest intelligent disease and pest inspection system based on disease feature extraction, comprising: a registration module, which performs space-time registration on multi-period multi-modal image data to obtain a time sequence alignment image sequence; a matrix construction module, which constructs a multi-modal time sequence evolution matrix based on the first derivative of the feature parameters in the sequence; a vector generation module, which generates a disease evolution trajectory vector using a time convolution network; a calculation and clustering module, which generates a candidate unknown disease class using density peak clustering when the Mahalanobis distance between the trajectory vector and a known disease centroid vector is greater than a dynamic rejection threshold; and a verification and update module, which calculates a propagation confidence index in combination with spatial distribution topology data and micro-meteorological vector data, and saves to a disease knowledge base when the verification condition is met. Thus, the limitations of static recognition are overcome, and autonomous discovery, dynamic tracking and multi-dimensional objective verification of unknown diseases in forests are realized.

Description

Technical Field

[0001] This application relates to the technical field of intelligent forestry monitoring, and in particular to an intelligent forest pest and disease inspection system based on disease feature extraction. Background Technology

[0002] Pests and diseases are important factors affecting tree growth and forest ecological security. Existing forest pest and disease inspection techniques usually use visible light, multispectral, or thermal infrared images to identify diseases, but most of them are based on static features obtained from a single inspection for classification and judgment, lacking the ability to continuously track and analyze the development process of diseases.

[0003] Meanwhile, existing technologies primarily target the identification of known diseases, lacking effective mechanisms for detecting and identifying newly emerging disease types or pathogen variations, which can easily lead to misidentification. Furthermore, existing systems typically lack comprehensive verification methods that combine transmission patterns, host information, and environmental factors, making it difficult to continuously update disease knowledge and resulting in insufficient ability to detect unknown diseases and long-term adaptability. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, one objective of this application is to propose an intelligent forest pest and disease inspection system based on disease feature extraction. By performing spatiotemporal registration and temporal evolution modeling on multimodal images from multiple inspection cycles, the system can extract dynamic change features during the disease development process and improve the accuracy of disease identification.

[0006] To achieve the above objectives, the first aspect of this application proposes an intelligent forest pest and disease inspection system based on disease feature extraction, characterized in that it includes:

[0007] The registration module is used to receive multimodal image data of the target area in multiple inspection cycles, perform spatiotemporal registration processing on the multimodal image data, and obtain a time-aligned image sequence.

[0008] The matrix construction module is used to extract feature parameters from the temporally aligned image sequence and construct a multimodal temporal evolution matrix of the target area based on the first derivative of the feature parameters between multiple inspection cycles.

[0009] The vector generation module is used to input the multimodal temporal evolution matrix into a temporal convolutional network for temporal dependency capture processing, and generate disease evolution trajectory vectors.

[0010] The calculation module is used to obtain the centroid vector of the known disease trajectory cluster from the disease knowledge base and calculate the Mahalanobis distance between the disease evolution trajectory vector and the centroid vector.

[0011] The clustering module is used to mark the disease evolution trajectory vector as an abnormal feature when the Mahalanobis distance is greater than a preset dynamic rejection threshold, and to perform high-dimensional feature clustering on the abnormal feature using the density peak clustering algorithm to generate candidate unknown disease categories.

[0012] The verification module is used to calculate the propagation credibility index based on the spatial distribution topology data and micro-meteorological vector data corresponding to the candidate unknown disease categories;

[0013] The update module is used to save the disease evolution trajectory vector of the candidate unknown disease category to the disease knowledge base when the propagation credibility index meets the preset periodic verification conditions.

[0014] In addition, the intelligent forest pest and disease inspection system based on disease feature extraction proposed in this application may also have the following additional technical features:

[0015] In one embodiment of this application, the registration module includes an intrinsic parameter data acquisition unit and a deep network registration unit;

[0016] The deep network registration unit specifically includes:

[0017] Deep semantic key points are extracted from single-cycle fused images by deploying a self-supervised feature point extraction network;

[0018] A graph neural network model is used to capture the spatial topological relationship of the deep semantic key points between images in different inspection cycles, and the key point matching probability matrix is ​​calculated and output.

[0019] The corresponding matching point pairs are determined based on the key point matching probability matrix, and the non-rigid space mapping parameters are obtained by solving the corresponding matching point pairs.

[0020] Based on the non-rigid spatial mapping parameters, spatial pixel alignment mapping is performed on the single-cycle fused images of multiple inspection cycles to generate the temporally aligned image sequence.

[0021] In one embodiment of this application, the matrix construction module includes a rate conversion unit and a timing splicing unit;

[0022] The rate conversion unit is used to divide the characteristic difference between two adjacent inspection cycles by the time interval parameter to calculate the area expansion rate. Temperature diffusion rate Texture degradation rate and color migration rate , as a dynamic parameter vector :

[0023] ;

[0024] The temporal splicing unit is used to process the static parameter vector corresponding to the feature parameters. The dynamic parameter vector is concatenated with the dynamic parameter vector in time series order to generate the multimodal temporal evolution matrix: .

[0025] In one embodiment of this application, the vector generation module includes a temporally dependent feature capture unit and a fully connected dimension mapping unit;

[0026] The time-dependent feature capture unit captures the multimodal time-series evolution matrix. Perform specific dilated causal convolutional network operations to output temporal feature representation data. The formula is:

[0027] ;

[0028] In the formula, For the set nonlinear activation function, These are the network convolution kernel weight parameters. For network bias parameters, The dilated causal convolution operation is performed; the fully connected dimension mapping unit is used to represent the temporal feature data. Perform fully connected layer computational mapping processing to generate the disease evolution trajectory vector with a fixed dimension length in the system. .

[0029] In one embodiment of this application, the calculation module calculates the disease evolution trajectory vector. With the centroid vector of a known cluster of disease trajectories Mahalanobis distance The formula is:

[0030] ;

[0031] In the formula The covariance matrix is ​​the corresponding category; the clustering module extracts the standard deviation of the intra-cluster distance and constructs the preset dynamic rejection threshold for each of the known disease trajectory clusters. The formula is: ;

[0032] In the formula The expected parameter of the intra-cluster Mahalanobis distance is... The variance parameter of the Mahalanobis distance within the cluster. The set standard confidence level coefficient.

[0033] In one embodiment of this application, the clustering module includes a local density calculation unit and an anomaly screening unit;

[0034] The local density calculation unit calculates the corresponding abnormal features. Spatial local clustering density index :

[0035] ;

[0036] In the formula, This is the preset cutoff distance parameter;

[0037] The anomaly screening unit calculates the relative distance index for each of the anomaly features. :

[0038] ;

[0039] The system calculates the comprehensive clustering evaluation index. ;

[0040] The comprehensive clustering evaluation index was selected. The corresponding node that exceeds the set indicator threshold is designated as the candidate unknown disease category.

[0041] In one embodiment of this application, the system further includes a graph verification module located between the clustering module and the verification module, the graph verification module including a basic rule interception unit and a graph deep inference unit:

[0042] The basic rule interception unit is used to input the host tree species entity attribute data of all target areas into a preset hash hard rule exclusion association matrix. If the system determines that there is an absolute biological exclusion rule, the candidate unknown disease category is directly removed from the system.

[0043] The graph deep reasoning unit is used to perform deep relationship inference calculations on the candidate unknown disease categories that have been filtered and retained by the basic rule interception unit, and to output the calculation results to trigger the execution of the verification module.

[0044] In one embodiment of this application, the verification module includes a direction calculation unit, a feature fusion unit, and a tracking unit;

[0045] The direction calculation unit calculates the overall spread direction vector along the time series evolution based on the spatial distribution topology data;

[0046] The feature fusion unit extracts the expansion rate of the overall spread direction vector. The relative humidity parameter in the micro-meteorological vector data and spatial clustering density index Calculate the disease transmission potential index ;

[0047] In the formula, Preset weighting coefficients;

[0048] The tracking unit will measure the disease transmission potential index. As the credibility index of the propagation, the index is continuously extracted in multiple consecutive inspection cycles. If the preset threshold is met continuously, it is determined that the preset cycle verification condition is met.

[0049] In one embodiment of this application, the system further includes a disease knowledge base construction module located at the front end of the verification module;

[0050] The disease knowledge base construction module is used for:

[0051] Obtain the standard time-series aligned image training sequence corresponding to historical known disease sample data;

[0052] The matrix construction module and the vector generation module are invoked to generate and extract a set of disease evolution trajectory vectors from the historical known disease sample data.

[0053] Based on the determined disease category labels, high-dimensional feature space clustering calculation is performed on the disease evolution trajectory vector set, and the centroid vectors and corresponding category covariance matrices of each category are calculated and output respectively.

[0054] The obtained centroid vector and category covariance matrix are initially written into the disease knowledge base as the basis for subsequent rejection discrimination comparison data.

[0055] The intelligent forest pest and disease inspection system based on disease feature extraction proposed in this application:

[0056] 1. By extracting the first derivatives of feature parameters within multiple inspection cycles and constructing a multimodal temporal evolution matrix, a disease evolution trajectory vector is generated. This overcomes the limitation of existing technologies that rely solely on static features obtained from a single inspection for classification and establishes the ability to continuously track and analyze the disease development process.

[0057] 2. By calculating the Mahalanobis distance between the disease evolution trajectory vector and the centroid vector of known diseases, and combining a dynamic rejection threshold and density peak clustering algorithm, high-dimensional clustering analysis is performed on anomalous features. An anomaly detection and independent discrimination mechanism for newly emerging disease types or pathogen variations is constructed, effectively avoiding the misidentification problem that the system is prone to when facing unknown diseases.

[0058] 3. By introducing a graph rule verification based on host information, the propagation reliability index is calculated after fusing spatial distribution topology data with micrometeorological environmental factors. Upon successful verification, the evolution trajectory vector of the new disease is synchronously updated to the disease knowledge base. This addresses the shortcomings of existing systems that lack comprehensive verification methods combining propagation patterns, host information, and environmental factors, achieving objective cross-verification of suspected diseases.

[0059] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0060] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0061] Figure 1 This is an overall block diagram of the intelligent forest pest and disease inspection system based on disease feature extraction proposed in this application;

[0062] Figure 2 This is a structural block diagram of the registration module of the intelligent forest pest and disease inspection system based on disease feature extraction in this application;

[0063] Figure 3 This is a structural block diagram of the matrix construction and vector generation module of the intelligent forest pest and disease inspection system based on disease feature extraction in this application;

[0064] Figure 4 This is a structural block diagram of the clustering module and the map verification module of the intelligent forest pest and disease inspection system based on disease feature extraction in this application;

[0065] Figure 5 This is a block diagram of the verification and updating module of the intelligent forest pest and disease inspection system based on disease feature extraction in this application. Detailed Implementation

[0066] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0067] The following describes, with reference to the accompanying drawings, an intelligent forest pest and disease inspection system based on disease feature extraction, according to an embodiment of this application.

[0068] like Figure 1As shown in this embodiment, a forest intelligent pest and disease inspection system based on disease feature extraction is presented. The system can be deployed and operated based on a cloud-edge-device collaborative physical architecture: the edge-side UAV is equipped with optical sensors to collect the multimodal image data, the edge-side base station executes the registration module to achieve efficient alignment and save transmission bandwidth, and the cloud-based central server centrally runs the complex network simulation from the matrix construction module to the update module.

[0069] Specifically, the data interaction logic in the cloud-edge-device collaborative physical architecture is as follows: after the UAV on the edge acquires the original image, it performs preprocessing such as distortion correction and packages it into a unified format data frame; the base station on the edge receives the data frame through a wireless network and runs a registration algorithm using a lightweight inference engine, and then uploads the time-aligned image sequence after removing redundant backgrounds to the cloud central server, thereby significantly reducing network communication overhead and improving the real-time response of the system.

[0070] The system may include:

[0071] The registration module is used to receive multimodal image data of the target area in multiple inspection cycles, perform spatiotemporal registration processing on the multimodal image data, and obtain a time-aligned image sequence.

[0072] like Figure 2 As shown, in one embodiment of this application, the registration module includes an intrinsic parameter data acquisition unit and a deep network registration unit;

[0073] The deep network registration unit specifically includes:

[0074] Deep semantic key points in a single-cycle fused image are extracted by deploying a self-supervised feature point extraction network. It should be noted that the self-supervised feature point extraction network adopts a VGG convolution backbone with a feature pyramid structure and outputs a dense high-dimensional feature map through multi-scale feature fusion.

[0075] A graph neural network model is used to capture the spatial topological relationship of the deep semantic key points between images in different inspection cycles, and the key point matching probability matrix is ​​calculated and output.

[0076] It should be noted that the graph neural network model obtains global context information through graph structure reasoning and outputs the key point matching probability matrix, thereby accurately filtering out mismatched points.

[0077] Specifically, the graph neural network model includes a multi-layer self-attention mechanism and a cross-attention mechanism. It iteratively updates the key point features and uses the Sinkhorn algorithm to optimize the similarity matrix until the probability distribution difference between two adjacent iterations is less than a preset minimum convergence precision (e.g., 1e-5). Finally, the key point matching probability matrix is ​​output.

[0078] The corresponding matching point pairs are determined based on the key point matching probability matrix, and the non-rigid space mapping parameters are obtained by solving the corresponding matching point pairs.

[0079] Based on the non-rigid spatial mapping parameters, spatial pixel alignment mapping is performed on the single-cycle fused images of multiple inspection cycles to generate the temporally aligned image sequence.

[0080] like Figure 3 As shown, the matrix construction module is used to extract feature parameters from the temporally aligned image sequence and construct a multimodal temporal evolution matrix of the target area based on the first derivative of the feature parameters between multiple inspection cycles.

[0081] It should be noted that the purpose of the matrix construction module is to completely eliminate interference from static physical contamination such as bird droppings and mud splashes, and to distinguish real, living diseases with biological dynamic characteristics.

[0082] In one embodiment of this application, the matrix construction module includes a rate conversion unit and a timing splicing unit;

[0083] The rate conversion unit is used to divide the characteristic difference between two adjacent inspection cycles by the time interval parameter to calculate the area expansion rate. Temperature diffusion rate Texture degradation rate and color migration rate , as a dynamic parameter vector :

[0084] ;

[0085] It should be noted that the area is obtained by multiplying the total number of pixels in the target area of ​​the image semantic segmentation by the actual physical area of ​​a single pixel; the temperature is obtained by the absolute temperature difference between the highest temperature in the infrared thermal imaging data and the average temperature of the background; and the time interval parameter is calibrated based on the actual absolute time difference between two inspections (e.g., 72 hours).

[0086] The temporal splicing unit is used to process the static parameter vector corresponding to the feature parameters. The dynamic parameter vector is concatenated with the dynamic parameter vector in time series order to generate the multimodal temporal evolution matrix:

[0087] .

[0088] In this embodiment, the multidimensional matrix transforms simple visual features into a temporal sequence of physical dimensions.

[0089] In one embodiment of this application, the vector generation module includes a temporally dependent feature capture unit and a fully connected dimension mapping unit;

[0090] The time-dependent feature capture unit captures the multimodal time-series evolution matrix. Perform specific dilated causal convolutional network operations to output temporal feature representation data. The formula is:

[0091] ;

[0092] In the formula, For the set nonlinear activation function, These are the network convolution kernel weight parameters. For network bias parameters, The extended causal convolution operation is performed by expanding the receptive field through an exponentially increased hole rate. For example, multiple hidden layers in the temporal convolutional network are configured with an increased expansion factor (e.g., expansion factors of 1, 2, 4, and 8, respectively). This not only ensures the causal dependence of time series information but also captures pathological evolution cycles spanning several months with low computational complexity.

[0093] Specifically, for the training process of this model, a joint loss function is used to optimize and update the network parameters. The joint loss function is composed of a weighted sum of classification cross-entropy loss and center loss. It is solved iteratively through backpropagation algorithm and Adam optimizer until the loss value is lower than the preset accuracy limit, thereby forcing the network to learn high-dimensional feature representations that are compact within classes and dispersed between classes.

[0094] The fully connected dimension mapping unit is used to represent the temporal feature data. Perform fully connected layer computational mapping processing to generate the disease evolution trajectory vector with a fixed dimension length in the system. The disease evolution trajectory vector is a unique identifier in high-dimensional space that can highly characterize the complete life cycle features of a specific disease.

[0095] The vector generation module is used to input the multimodal temporal evolution matrix into a temporal convolutional network for temporal dependency capture processing, and generate disease evolution trajectory vectors.

[0096] The calculation module is used to obtain the centroid vector of the known disease trajectory cluster from the disease knowledge base and calculate the Mahalanobis distance between the disease evolution trajectory vector and the centroid vector.

[0097] It should be noted that the technical effect of the computing module is to realize the abnormal feature rejection mechanism in the open set environment. Since the physical dimensions of features of different dimensions (such as temperature distribution and expansion area) are different and there is multicollinearity, the conventional Euclidean distance cannot accurately measure the classification boundary.

[0098] In one embodiment of this application, the calculation module calculates the disease evolution trajectory vector. With the The centroid vector of each known disease trajectory cluster Mahalanobis distance The formula is:

[0099] ;

[0100] In the formula, The covariance matrix is ​​the corresponding category; the clustering module extracts the standard deviation of the intra-cluster distance and constructs the preset dynamic rejection threshold for each of the known disease trajectory clusters. The formula is: ;

[0101] In the formula The expected parameter of the intra-cluster Mahalanobis distance is... The variance parameter of the Mahalanobis distance within the cluster. The set standard confidence level coefficient.

[0102] For example, the standard confidence coefficient is set based on prior knowledge of the Gaussian distribution, and its preferred value range is 2.5-3.5 (e.g., This ensures that known samples that are normally moved fall within the rejection boundary with a very high probability.

[0103] In one embodiment, the dimension length of the disease evolution trajectory vector can be set to 128, 256, or 512 dimensions to balance feature representation capability and computational complexity.

[0104] By combining the above mechanism with the morphological distribution variance of each known category to adaptively define the rejection boundary, when the calculated distance exceeds the corresponding threshold, the system does not force it to be misclassified, but precisely triggers the downstream unknown disease exploration mechanism.

[0105] like Figure 4 As shown, the clustering module is used to mark the disease evolution trajectory vector as an abnormal feature when the Mahalanobis distance is greater than a preset dynamic rejection threshold, and to perform high-dimensional feature clustering on the abnormal feature using the density peak clustering algorithm to generate candidate unknown disease categories.

[0106] It can be understood that, in this embodiment, the clustering module uses density peak clustering to discover new forest diseases without prior labels.

[0107] In one embodiment of this application, the clustering module includes a local density calculation unit and an anomaly screening unit;

[0108] The local density calculation unit calculates the corresponding abnormal features. Spatial local clustering density index :

[0109] ;

[0110] In the formula, The preset cutoff distance parameter is used as the basis for setting the preset cutoff distance parameter. It should be noted that the Euclidean distance between the sample points is sorted in ascending order, and the distance value in the 1%-2% range is selected as the preset cutoff distance parameter to ensure that each sample point has an appropriate number of neighboring nodes on average.

[0111] The anomaly screening unit calculates the relative distance index for each of the anomaly features. :

[0112] ;

[0113] The system calculates the comprehensive clustering evaluation index. ;

[0114] The comprehensive clustering evaluation index was selected. The corresponding node that exceeds the set indicator threshold is designated as the candidate unknown disease category.

[0115] For example, the threshold of the set index can be set based on the Elbow Method, which involves sorting the comprehensive clustering evaluation index of all abnormal features in descending order, calculating the descending slope of adjacent indexes, and taking the value corresponding to the abrupt change point with the largest slope change as the threshold of the set index.

[0116] This method can automatically find products with high cohesion. (High value) and far from other high-density centers ( Feature nodes with large values ​​are used as cluster centers to filter out a large number of isolated noise points and effectively prevent the generation of pseudo-unknown categories.

[0117] In one embodiment of this application, the system further includes a graph verification module located between the clustering module and the verification module, the graph verification module including a basic rule interception unit and a graph deep inference unit:

[0118] It should be noted that the spectrum verification module significantly reduces the system's inference delay and blind calculation by establishing a two-layer verification structure.

[0119] The basic rule interception unit is used to input the host tree species entity attribute data of all target areas into a preset hash hard rule exclusion association matrix. If the system determines that there is an absolute biological exclusion rule, the candidate unknown disease category is directly removed from the system.

[0120] Specifically, the basic rule interception unit establishes a fast initial screening mechanism with a complexity of O(1).

[0121] For example, the preset hash hard rule exclusion association matrix uses a key-value pair structure for data storage and high-speed retrieval. It uses the preset hash hard rule exclusion association matrix to query forestry hard rules (for example, a candidate category exhibits biological characteristics that specifically stain coniferous trees, but the host tree species entity attribute data of the current target area is confirmed as "broadleaf poplar"). Once a match is found, it is regarded as feature drift and directly blocked.

[0122] For the remaining categories that have not been excluded, the map depth inference unit combines the structural features of the map nodes for each category, such as altitude distribution and forest stand density.

[0123] The graph deep reasoning unit is used to perform deep relationship inference calculations on the candidate unknown disease categories that have been filtered and retained by the basic rule interception unit, and to output the calculation results to trigger the execution of the verification module.

[0124] Specifically, the graph deep reasoning unit uses a knowledge graph embedding algorithm (such as the TransE algorithm) to map entities and relations into a continuous low-dimensional vector space. By calculating the distance score between the head entity vector, the relation vector, and the tail entity vector, and after normalization by the Softmax function, the output probability score normalized to the [0,1] interval is used as the calculation and inference result, ensuring that the subsequent system verification process has solid biological support.

[0125] like Figure 5 As shown, the verification module is used to calculate the propagation credibility index based on the spatial distribution topology data and micro-meteorological vector data corresponding to the candidate unknown disease category;

[0126] In one embodiment of this application, the verification module includes a direction calculation unit, a feature fusion unit, and a tracking unit;

[0127] The direction calculation unit calculates the overall spread direction vector along the time series evolution based on the spatial distribution topology data;

[0128] It is understood that after the direction calculation unit extracts the overall spread direction vector from the target point distribution, the feature fusion unit associates it with meteorological medium parameters.

[0129] The feature fusion unit extracts the expansion rate of the overall spread direction vector. The relative humidity parameter in the micro-meteorological vector data and spatial clustering density index Calculate the disease transmission potential index ;

[0130] In the formula, Preset weighting coefficients;

[0131] It should be noted that the preset weighting coefficients satisfy... Mathematical constraints, (for example, The expansion rate, relative humidity parameter, and spatial aggregation density index are normalized before fusion calculation.

[0132] As one possibility, the preset weight coefficients can be obtained through supervised learning training using historical disease sample data, or initialized based on the experience of forestry experts, and iteratively optimized through cross-validation.

[0133] The tracking unit will measure the disease transmission potential index. As the credibility index of the propagation, the index is continuously extracted in multiple consecutive inspection cycles. If the preset threshold is met continuously, it is determined that the preset cycle verification condition is met.

[0134] The disease transmission potential index constructs an evidence chain that cross-verifies between the image domain and the meteorological domain. To further reduce the false alarm rate, the tracking unit performs time-axis verification: only when all indicators of the candidate target show a stable epidemiological expansion trend (i.e., continuously meet the single preset threshold) within the aforementioned multiple consecutive inspection cycles (e.g., continuously tracking for 3 inspection days) can the preset cycle verification condition be confirmed as achieved.

[0135] For example, the single preset threshold can be set to an empirical value of 0.75 to 0.85 based on the prior knowledge of industry experts; the preset cycle verification condition is specifically set to be greater than or equal to 3 consecutive inspection cycles, thereby eliminating environmental artifacts caused by short-term drift of the thermal imager or temporary drought.

[0136] In one embodiment of this application, the system further includes a disease knowledge base construction module located at the front end of the verification module;

[0137] The disease knowledge base construction module is responsible for the initialization of the system and the establishment of standard baselines before system operation. During the system cold start phase, historical known disease sample data of confirmed typical forestry diseases (such as pine wilt disease) are collected, and a complete registration and time series extraction workflow is executed to construct a standardized and reliable standard time-series aligned image training sequence. Subsequently, the system statistically analyzes the distribution characteristics of each cluster in the feature space.

[0138] The disease knowledge base construction module is used for:

[0139] Obtain the standard time-series aligned image training sequence corresponding to historical known disease sample data;

[0140] The matrix construction module and the vector generation module are invoked to generate and extract a set of disease evolution trajectory vectors from the historical known disease sample data.

[0141] Based on the determined disease category labels, high-dimensional feature space clustering calculation is performed on the disease evolution trajectory vector set, and the centroid vectors and corresponding category covariance matrices of each category are calculated and output respectively.

[0142] The obtained centroid vector and category covariance matrix are initially written into the disease knowledge base as the basis for subsequent rejection discrimination comparison data.

[0143] Specifically, the high-dimensional feature space clustering calculation uses a Gaussian mixture model (GMM) combined with the expectation-maximization (EM) algorithm for optimization. The expectation-maximization algorithm alternately executes the E-step to calculate the posterior probability of a sample belonging to each Gaussian component and the M-step to update the mean and covariance. When the increment of the log-likelihood function between two adjacent iterations is less than the preset convergence precision (e.g., 1e-4), it is determined that the convergence condition is met and the iteration stops, thereby outputting the optimal centroid vector and the category covariance matrix. This result can accurately describe the cluster center boundary and feature dispersion of a specific known disease.

[0144] After being saved to the disease knowledge base, it provides a high-confidence comparison anchor point for the calculation module, enabling the system to not only have the benchmark identification capability of closed set known diseases, but also to lay a data foundation for the continuous evolution of the system and automatic collection of new disease categories.

[0145] After adding new candidate unknown disease categories to the disease knowledge base, the system recalculates the centroid vector and covariance matrix of the corresponding category and adopts an incremental update strategy to maintain the stability of the knowledge base distribution.

[0146] The update module is used to save the disease evolution trajectory vector of the candidate unknown disease category to the disease knowledge base when the propagation credibility index meets the preset periodic verification conditions.

[0147] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0148] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A forest intelligent pest and disease inspection system based on disease feature extraction, characterized in that, include: The registration module is used to receive multimodal image data of the target area in multiple inspection cycles, perform spatiotemporal registration processing on the multimodal image data, and obtain a time-aligned image sequence. The matrix construction module is used to extract feature parameters from the temporally aligned image sequence and construct a multimodal temporal evolution matrix of the target area based on the first derivative of the feature parameters between multiple inspection cycles. The vector generation module is used to input the multimodal temporal evolution matrix into a temporal convolutional network for temporal dependency capture processing, and generate disease evolution trajectory vectors. The calculation module is used to obtain the centroid vector of the known disease trajectory cluster from the disease knowledge base and calculate the Mahalanobis distance between the disease evolution trajectory vector and the centroid vector. The clustering module is used to mark the disease evolution trajectory vector as an abnormal feature when the Mahalanobis distance is greater than a preset dynamic rejection threshold, and to perform high-dimensional feature clustering on the abnormal feature using the density peak clustering algorithm to generate candidate unknown disease categories. The verification module is used to calculate the propagation credibility index based on the spatial distribution topology data and micro-meteorological vector data corresponding to the candidate unknown disease categories; The update module is used to save the disease evolution trajectory vector of the candidate unknown disease category to the disease knowledge base when the propagation credibility index meets the preset periodic verification conditions.

2. The intelligent forest pest and disease inspection system based on disease feature extraction according to claim 1, characterized in that, The registration module includes an intrinsic parameter data acquisition unit and a deep network registration unit; The deep network registration unit specifically includes: Deep semantic key points are extracted from single-cycle fused images by deploying a self-supervised feature point extraction network; A graph neural network model is used to capture the spatial topological relationship of the deep semantic key points between images in different inspection cycles, and the key point matching probability matrix is ​​calculated and output. The corresponding matching point pairs are determined based on the key point matching probability matrix, and the non-rigid space mapping parameters are obtained by solving the corresponding matching point pairs. Based on the non-rigid spatial mapping parameters, spatial pixel alignment mapping is performed on the single-cycle fused images of multiple inspection cycles to generate the temporally aligned image sequence.

3. The intelligent forest pest and disease inspection system based on disease feature extraction according to claim 1, characterized in that, The matrix construction module includes a rate conversion unit and a timing splicing unit; The rate conversion unit is used to divide the characteristic difference between two adjacent inspection cycles by the time interval parameter to calculate the area expansion rate. Temperature diffusion rate Texture degradation rate and color migration rate , as a dynamic parameter vector : ； The temporal splicing unit is used to process the static parameter vector corresponding to the feature parameters. The dynamic parameter vector is concatenated with the dynamic parameter vector in time series order to generate the multimodal temporal evolution matrix: 。 4. The intelligent forest pest and disease inspection system based on disease feature extraction according to claim 1, characterized in that, The vector generation module includes a temporally dependent feature capture unit and a fully connected dimension mapping unit; The time-dependent feature capture unit captures the multimodal time-series evolution matrix. Perform specific dilated causal convolutional network operations to output temporal feature representation data. The formula is: ； In the formula, For the set nonlinear activation function, These are the network convolution kernel weight parameters. For network bias parameters, The dilated causal convolution operation is performed; the fully connected dimension mapping unit is used to represent the temporal feature data. Perform fully connected layer computational mapping processing to generate the disease evolution trajectory vector with a fixed dimension length in the system. .

5. The intelligent forest pest and disease inspection system based on disease feature extraction according to claim 1, characterized in that, The calculation module calculates the disease evolution trajectory vector. With the The centroid vector of a known cluster of disease trajectories Mahalanobis distance The formula is: ； In the formula The covariance matrix is ​​the corresponding category; the clustering module extracts the standard deviation of the intra-cluster distance and constructs the preset dynamic rejection threshold for each of the known disease trajectory clusters. The formula is: ; In the formula The expected parameter of the intra-cluster Mahalanobis distance is... The variance parameter of the Mahalanobis distance within the cluster. The set standard confidence level coefficient.

6. The intelligent forest pest and disease inspection system based on disease feature extraction according to claim 1, characterized in that, The clustering module includes a local density calculation unit and an anomaly screening unit; The local density calculation unit calculates the corresponding abnormal features. Spatial local clustering density index : ； In the formula, This is the preset cutoff distance parameter; The anomaly screening unit calculates the relative distance index for each of the anomaly features. : ； The system calculates the comprehensive clustering evaluation index. ; The comprehensive clustering evaluation index was selected. The corresponding node that exceeds the set indicator threshold is designated as the candidate unknown disease category.

7. The intelligent forest pest and disease inspection system based on disease feature extraction according to claim 1, characterized in that, The system further includes a graph verification module located between the clustering module and the verification module. The graph verification module includes a basic rule interception unit and a graph deep inference unit. The basic rule interception unit is used to input the host tree species entity attribute data of all target areas into a preset hash hard rule exclusion association matrix. If the system determines that there is an absolute biological exclusion rule, the candidate unknown disease category is directly removed from the system. The graph deep reasoning unit is used to perform deep relationship inference calculations on the candidate unknown disease categories that have been filtered and retained by the basic rule interception unit, and to output the calculation results to trigger the execution of the verification module.

8. The intelligent forest pest and disease inspection system based on disease feature extraction according to claim 1, characterized in that, The verification module includes a direction calculation unit, a feature fusion unit, and a tracking unit; The direction calculation unit calculates the overall spread direction vector along the time series evolution based on the spatial distribution topology data; The feature fusion unit extracts the expansion rate of the overall spread direction vector. The relative humidity parameter in the micro-meteorological vector data and spatial clustering density index Calculate the disease transmission potential index ; In the formula, Preset weighting coefficients; The tracking unit will measure the disease transmission potential index. As the credibility index of the propagation, the index is continuously extracted in multiple consecutive inspection cycles. If the preset threshold is met continuously, it is determined that the preset cycle verification condition is met.

9. The intelligent forest pest and disease inspection system based on disease feature extraction according to claim 1, characterized in that, The system also includes a disease knowledge base construction module located at the front end of the verification module; The disease knowledge base construction module is used for: Obtain the standard time-series aligned image training sequence corresponding to historical known disease sample data; The matrix construction module and the vector generation module are invoked to generate and extract a set of disease evolution trajectory vectors from the historical known disease sample data. Based on the determined disease category labels, high-dimensional feature space clustering calculation is performed on the disease evolution trajectory vector set, and the centroid vectors and corresponding category covariance matrices of each category are calculated and output respectively. The obtained centroid vector and category covariance matrix are initially written into the disease knowledge base as the basis for subsequent rejection discrimination comparison data.

Images