Forestry pest intelligent identification method and system based on remote sensing data

Abstract

The application discloses a forestry disease and pest intelligent identification method and system based on remote sensing data, relates to the technical field of forestry disease and pest monitoring, and comprises the following steps: acquiring target forest area multi-source remote sensing images and preprocessing the remote sensing images to obtain a benchmark image set; extracting a vegetation index feature, fusing texture and color features to construct a multi-dimensional forest health state feature tensor; inputting the feature tensor into a space-time graph neural network inference model, aggregating neighborhood pixels and time evolution information, and outputting a per-pixel disease and pest occurrence probability distribution map; performing space morphological processing on the probability distribution map, identifying a suspected disease area boundary, and counting geometric morphological features; inverting a suspected disease area initial physiological state parameter; combining historical health data to construct a disease pathogen diffusion dynamics state inversion equation, and solving the equation to obtain a propagation rate and direction field; and deducing a future disease and pest risk prediction map. The method improves identification accuracy and realizes early prediction of disease and pest risks.

Description

Technical Field

[0001] This invention belongs to the field of forestry pest and disease monitoring technology, specifically a method and system for intelligent identification of forestry pests and diseases based on remote sensing data. Background Technology

[0002] Forest pests and diseases are a significant factor hindering the healthy development of forestry. Remote sensing technology, due to its wide coverage and high monitoring efficiency, has been widely applied in the field of forest pest and disease identification. Current technologies for remote sensing identification of forest pests and diseases largely rely on single vegetation index features, classifying remote sensing images using traditional machine learning models to identify suspected diseased areas. Some technologies supplement this with simple spatial texture analysis, but a systematic feature fusion framework has not yet been established.

[0003] Existing technical solutions have obvious shortcomings. The feature extraction dimension is too limited, and relying solely on vegetation indices is insufficient to comprehensively characterize the health status of forests. They cannot capture the subtle physiological changes during the early stages of pest and disease infection, resulting in low identification accuracy and a high risk of misjudgment. Furthermore, existing technologies can only identify the current pest and disease outbreak areas, but cannot infer the physiological state of the forest in the early stages of infection, nor can they analyze the spread patterns of pathogens. This makes it difficult to predict future pest and disease risks, and the solutions can only passively respond to pests and diseases that have already occurred, failing to achieve early prevention and control.

[0004] How to improve the accuracy of pest and disease identification through multi-dimensional feature fusion and quantify the probability of pest and disease occurrence on a pixel-by-pixel basis; how to trace the physiological state of forests in the early stage of disease infection, analyze the dynamic characteristics of pathogen spread, and thus achieve accurate prediction of future pest and disease risks have become urgent problems to be solved in the field of forest pest and disease remote sensing identification. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes an intelligent identification method for forestry pests and diseases based on remote sensing data, including: Multi-source remote sensing image data of the target forest area were acquired, and a baseline image set was obtained after preprocessing. Vegetation index features are extracted from a preprocessed set of benchmark images, and texture and color features are fused to construct a multidimensional forest health status feature tensor. The forest health status feature tensor is input into the inference calculation model based on spatiotemporal graph neural network. The inference calculation model based on spatiotemporal graph neural network outputs a pixel-by-pixel probability distribution map of pest and disease occurrence by aggregating neighborhood pixel information and temporal evolution information. Spatial morphological processing is performed on the probability distribution map of the disease and pest occurrence to identify the specific boundaries of suspected disease-infected areas, and the geometric morphological characteristics of the suspected disease-infected areas are statistically analyzed. Based on the geometric morphological features of the suspected disease-infected area, the original physiological state indicators of the suspected disease-infected area in the early stage of infection are inverted from the multidimensional forest health status feature tensor, resulting in a set of inverted physiological state parameters. Using the inverted physiological state parameters and combined with remote sensing data from historical healthy periods, a state inversion equation for pathogen diffusion dynamics is constructed, and the state inversion equation is solved to obtain the pathogen's propagation rate and orientation field. Based on the obtained propagation rate and direction field, the probability distribution map of the occurrence of pests and diseases is extrapolated forward along the time axis to generate a risk prediction map of pests and diseases in the future growth cycle.

[0006] Furthermore, the acquisition of multi-source remote sensing image data of the target forest area, after preprocessing to obtain a reference image set, includes: Acquire multi-source remote sensing image data of the target forest area, including high spatial resolution optical images, multispectral images, and shortwave infrared images; Radiometric calibration and atmospheric correction are performed on the multi-source remote sensing image data to eliminate sensor errors and atmospheric scattering interference, resulting in a preprocessed reference image set. Vegetation index features are extracted from the preprocessed reference image set, and texture and color features are fused to construct a multidimensional forest health status feature tensor. The process of performing radiometric calibration and atmospheric correction on the multi-source remote sensing image data eliminates sensor-specific errors and atmospheric scattering interference, resulting in a preprocessed reference image set, including: The metadata information in the header file of the multi-source remote sensing image data is read to obtain the sensor's gain and bias parameters, and the original digital quantization value is linearly transformed based on the gain and bias parameters to obtain the apparent radiance. Using the aerosol optical thickness product provided by the medium resolution imaging spectrometer, atmospheric scattering removal is performed on the apparent radiance to calculate the true surface reflectance. For the high spatial resolution optical image, orthorectification is performed to eliminate geometric distortion caused by terrain undulations and ensure the correspondence between pixel positions and geographic coordinates; Band registration is performed on the multispectral image and the shortwave infrared image, and image data from different sources are resampled to the same spatial resolution grid coordinate system; All processed image data are matched in the time dimension, and images with cloud cover below a preset threshold are selected as valid data and combined to form the preprocessed reference image set.

[0007] Furthermore, vegetation index features are extracted from the preprocessed reference image set, and texture and color features are fused to construct a multidimensional forest health status feature tensor, including: The reflectance ratio of the near-infrared band to the red band in the preprocessed reference image set is calculated to obtain the normalized vegetation index, and the enhanced vegetation index and the soil-regulated vegetation index are calculated simultaneously. Gray-level co-occurrence matrix analysis was performed on the high spatial resolution optical image to extract the second moment of the angle, contrast and entropy as the texture features to characterize the degree of damage to the blade surface structure. The preprocessed reference image set is converted to the HSV color space, and the statistical histograms of the hue and saturation components are extracted as the color features to capture the phenomenon of plants losing their green color and turning yellow. The calculated vegetation index features, texture features, and color features are concatenated at the pixel level, and latitude and longitude coordinate information is added as location features to form the multidimensional forest health status feature tensor.

[0008] Furthermore, the forest health status feature tensor is input into a spatiotemporal graph neural network-based inference calculation model. This model, by aggregating neighborhood pixel information and temporal evolution information, outputs a pixel-by-pixel probability distribution map of pest and disease occurrence, including: Using each pixel in the multidimensional forest health status feature tensor as a graph node, and using the spatial distance and feature similarity between pixels as edge weights, a spatial graph structure is constructed. By using the sliding time window technique, the spatial graph structures of multiple consecutive time points are stacked to form a spatiotemporal cube data structure; The inference computation model based on spatiotemporal graph neural network uses graph convolution operations to aggregate the feature information of the central pixel and its first-order neighboring pixels, and update the state representation of the central pixel. After the graph convolution operation, a gated recurrent unit is introduced to encode the state representation on the time series to capture the spread trend of pests and diseases over time. The node state representation output at the final time step is mapped to a value between 0 and 1 through a fully connected layer and an activation function. This value represents the probability of pest and disease occurrence per pixel.

[0009] Furthermore, the probability distribution map of the pest and disease occurrence is subjected to spatial morphological processing to identify the specific boundaries of suspected disease-infected areas, and the geometric morphological characteristics of the suspected disease-infected areas are statistically analyzed, including: A probability threshold is set, and the probability distribution map of the occurrence of pests and diseases is binarized. Pixels with probability values ​​higher than the probability threshold are marked as candidate pixels for disease infection. The binarized image is subjected to an opening operation to remove isolated noise points, and then a closing operation is performed to fill the holes inside the lesions, resulting in connected lesion regions. The contour lines of each connected lesion region are extracted using an edge detection algorithm, and the length and area of ​​the contour lines are calculated to obtain the geometric morphological features; Calculate the aspect ratio and compactness index of the lesion shape to distinguish between strip-shaped lesions caused by wind propagation and cluster-shaped lesions caused by insect propagation; Connected lesion regions with similar geometric features are merged to form the final suspected diseased region, and their vector boundaries are recorded.

[0010] Furthermore, based on the geometric morphological characteristics of the suspected diseased area, the original physiological state indicators of the suspected diseased area in the early stage of infection are inverted from the multidimensional forest health state feature tensor, resulting in a set of inverted physiological state parameters, including: Starting from the geometric center pixel of the suspected diseased area, backtracking along the historical timeline, extract the multidimensional forest health status feature tensor data of the geometric center pixel in the same historical time period; By analyzing the differences in the multispectral response curves of the suspected disease-infected areas at the early stage of infection and at the present stage, the specific bands most sensitive to pest and disease stress were identified. Using a radiative transfer model, with the current vegetation canopy reflectance as input, the model iteratively adjusts the mesophyll structure parameters and chlorophyll content parameters to minimize the residual between the model-simulated reflectance and the actual observed reflectance. When the radiative transfer model converges, the obtained mesophyll structure parameters and chlorophyll content parameters are the inverted physiological state parameters, representing the physiological state of the plant in the early stage of infection.

[0011] Furthermore, using the retrieved physiological state parameters and remote sensing data from historical healthy periods, a state inversion equation regarding pathogen diffusion dynamics is constructed, and the state inversion equation is solved to obtain the pathogen's propagation rate and orientation field, including: The forest area is divided into regular grid units, and the inverted physiological state parameters are used as the initial susceptibility values ​​for each grid unit, while the average vegetation index during historical healthy periods is used as the initial resilience value. Based on cellular automata theory, the state inversion equation is defined, and the state transition probability in the state inversion equation is jointly determined by the disease and pest occurrence probability distribution map and the neighboring infection state of the current grid cell. By using the maximum likelihood estimation method, the diffusion coefficient and wind direction influence factor in the state inversion equation are adjusted so that the infection range simulated by the model and the actual boundary of the suspected infected area achieve the best fit. After the model parameters are determined, the state inversion equation is run to record the spatial displacement of infected pixels per unit time, which is then quantified as the propagation rate, and the orientation field is obtained through vector synthesis calculation.

[0012] Furthermore, based on the obtained propagation rate and direction field, the probability distribution map of pest and disease occurrence is extrapolated forward along the time axis to generate a pest and disease risk prediction map for future growth cycles, including: The propagation rate and the direction field are discretized into eight vector components in each direction, which correspond to the diffusion tendency in different directions. Using the disease and pest occurrence probability distribution map as the initial infection source layer, the infection probability is spatially allocated and accumulated according to the diffusion tendency in each direction within each time step. Considering the differences in resistance among different tree species, the dominant tree species of each grid unit are read from the forest stand map, and the cumulative infection probability is weighted and corrected by the tree species resistance coefficient. Repeat the diffusion and correction process of the time step until the cumulative time reaches a complete growth cycle. The output cumulative probability layer is the pest and disease risk prediction map.

[0013] Furthermore, it also includes: The disease and pest risk prediction map is classified into levels, and the regions with different risk levels are associated with specific tree species types by combining the forest stand map; By integrating the boundaries of the suspected infected areas, the inverted physiological state parameters, the transmission rate and direction field, and the pest and disease risk prediction map, a structured intelligent diagnosis and prediction report of forestry pests and diseases is generated. The process of classifying the pest and disease risk prediction map into risk levels and mapping different risk levels to specific tree species in conjunction with the forest stand map includes: Extract the risk probability values ​​of all pixels in the pest and disease risk prediction map, and perform statistical analysis to determine the mean and standard deviation of the probability distribution; Based on the quantile principle of the normal distribution, the pest and disease risk prediction map is divided into three levels: low-risk area, medium-risk area and high-risk area. Spatial overlay analysis was performed on the small-compartment zoning data and tree species composition data in the forest stand map and the pest and disease risk prediction map. The planting area percentage of each major tree species within each risk level area was statistically analyzed to create a table showing the correspondence between tree species and risk levels. For dominant tree species in high-risk areas, their historical damage records should be further investigated as a basis for key monitoring.

[0014] Furthermore, the present invention also includes a forestry pest and disease intelligent identification system based on remote sensing data. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the forestry pest and disease intelligent identification method based on remote sensing data as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Vegetation index features are extracted from the preprocessed baseline image set of the target forest area. Texture and color features are fused to construct a multidimensional forest health status feature tensor. This feature tensor is input into an inference computation model based on a spatiotemporal graph neural network. The model aggregates neighborhood pixel information and temporal evolution information to output a pixel-by-pixel probability distribution map of pest and disease occurrence. Multidimensional feature fusion breaks through the limitations of single feature representation, enabling comprehensive capture of subtle changes in forest health status, especially the physiological and morphological differences during the early infection of pests and diseases. The spatiotemporal graph neural network can simultaneously mine the spatial correlation and temporal evolution patterns of pixels, avoiding the shortcomings of traditional models that can only identify macroscopic diseased areas and cannot accurately locate pixel-by-pixel risks. This achieves refined and probabilistic pest and disease identification, reducing misjudgments and omissions.

[0016] Based on the geometric morphological features of suspected diseased areas obtained after spatial morphological processing, the original physiological state indicators of the area at the initial stage of infection are inverted from the multidimensional forest health state feature tensor. Using the inverted physiological state parameters combined with remote sensing data from historical health periods, a state inversion equation for pathogen diffusion dynamics is constructed and solved to obtain the pathogen's propagation rate and orientation field. Then, based on the propagation rate and orientation field, the probability distribution map of pest and disease occurrence is extrapolated forward along the time axis to generate a pest and disease risk prediction map for future growth cycles. This scheme can trace the forest physiological state at the initial stage of pest and disease infection, clarify the initial causes of pest and disease occurrence, and accurately capture the diffusion patterns of pathogens. It extends from current pest and disease identification to future risk prediction, breaking through the limitations of traditional technologies that can only passively identify pests and cannot prevent them in advance. It achieves full-process coverage of pest and disease monitoring, making prevention and control work more targeted and forward-looking. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the intelligent identification method for forestry pests and diseases based on remote sensing data described in this invention. Figure 2 Flowchart for constructing the forest health status characteristic tensor; Figure 3 Heat map showing the probability of forest pests and diseases; Figure 4 A comparative analysis chart of reflectance correction for multi-source remote sensing images; Figure 5Heatmap for spatial correlation analysis of forestry pest and disease risks and dominant tree species. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 Multi-source remote sensing image data of the target forest area was acquired and preprocessed to obtain a baseline image set. Vegetation index features were extracted from the preprocessed baseline image set, and texture and color features were fused to construct a multi-dimensional forest health status feature tensor. This forest health status feature tensor was input into an inference computation model based on a spatiotemporal graph neural network. This model aggregates neighborhood pixel information and temporal evolution information to output a pixel-by-pixel probability distribution map of pest and disease occurrence. Spatial morphological processing was performed on this probability distribution map to identify the specific boundaries of suspected disease-affected areas and statistically analyze the geometric morphological features of these areas. Based on the geometric morphological features of the suspected disease-affected areas, the original physiological state indicators of these areas in the early stages of infection were inverted from the multi-dimensional forest health status feature tensor, yielding a set of inverted physiological state parameters. Using these inverted physiological state parameters, combined with remote sensing data from historical health periods, a state inversion equation regarding pathogen diffusion dynamics was constructed, and the state inversion equation was solved to obtain the pathogen's propagation rate and orientation field. Based on the obtained propagation rate and direction field, the probability distribution map of pest and disease occurrence is extrapolated forward along the time axis to generate a pest and disease risk prediction map for future growth cycles.

[0020] In one embodiment of the present invention, multi-source remote sensing image data of the target forest area is acquired. This multi-source remote sensing image data includes high spatial resolution optical images, multispectral images, and shortwave infrared images. Radiometric calibration and atmospheric correction processing are performed on the multi-source remote sensing image data to eliminate sensor-specific errors and atmospheric scattering interference, resulting in a preprocessed reference image set. Vegetation index features are extracted from the preprocessed reference image set, and texture and color features are fused to construct a multidimensional forest health status feature tensor.

[0021] Radiometric calibration and atmospheric correction processing includes reading metadata information from the header file of multi-source remote sensing image data, obtaining the sensor's gain and bias parameters, and performing a linear transformation on the raw digital quantized values ​​based on the gain and bias parameters to obtain the apparent radiance. In some embodiments, the linear transformation relationship can be expressed as: in: The wavelength at the entrance pupil of the sensor is represented by The apparent radiance of the band, The original digital quantization values ​​representing the same frequency band. This represents the gain coefficient for that frequency band. This represents the bias coefficient for that wavelength band. Using aerosol optical thickness products provided by a medium-resolution imaging spectrometer, atmospheric scattering is removed from the apparent radiance to calculate the true surface reflectance. Optionally, the atmospheric scattering removal process employs an atmospheric correction algorithm based on a radiative transfer model, such as the 6S model or the FLAASH model, taking apparent radiance and atmospheric parameters such as aerosol optical thickness, water vapor content, and ozone content as inputs to solve for the true surface reflectance.

[0022] For high spatial resolution optical imagery, orthorectification is performed to eliminate geometric distortions caused by terrain undulations and ensure the correspondence between pixel positions and geographic coordinates. The orthorectification process requires digital elevation model data and the orbital and attitude parameters of the sensor during imaging. By constructing a rigorous imaging geometric model or rational function model, projection difference correction is applied to each pixel. In essence, orthorectification corrects the image onto a unified orthophoto projection plane. Band registration is then performed on multispectral and shortwave infrared imagery, resampling image data from different sources to a raster coordinate system with the same spatial resolution. The band registration process uses the high spatial resolution optical imagery as a geometric reference. Uniformly distributed control points are selected on the multispectral and shortwave infrared images, and a polynomial transformation model is established to resample each band of the multispectral and shortwave infrared images to the same spatial resolution and projection coordinate system as the optical imagery. Resampling methods can include nearest neighbor, bilinear interpolation, or cubic convolution. All processed image data are matched along the time dimension, and images with cloud cover below a preset threshold are selected as valid data and combined into a preprocessed reference image set. In some embodiments, matching along the time dimension means uniformly converting the acquisition time of all images to Julian days and ensuring that the image time series within the same growing season has a consistent observation period. Optionally, the preset cloud cover threshold is set to 10%, and a cloud detection algorithm automatically identifies and removes pixels covered by clouds and cloud shadows, retaining image frames with no cloud or very little cloud cover for inclusion in the reference image set.

[0023] In one embodiment of the present invention, see [reference] Figure 2 The reflectance ratio of the near-infrared band to the red band in the preprocessed reference image set is calculated to obtain the normalized vegetation index (NVI). Simultaneously, the enhanced vegetation index (EDI) and the soil-regulated vegetation index (SRI) are calculated. In some embodiments, the formula for calculating the NVI is as follows: ,in The surface reflectance representing the near-infrared band, The surface reflectance, representing the red light band, is used in the calculation of the enhanced vegetation index, which incorporates an atmospheric correction coefficient and a soil adjustment factor. The calculation of the soil-adjusted vegetation index further considers the influence of soil background brightness. Gray-level co-occurrence matrix (GLCM) analysis is performed on high spatial resolution optical images, extracting angular second moments, contrast, and entropy as texture features to characterize the degree of damage to leaf surface structures. The GLCM is calculated within a specified moving window, for example, set to 5 pixels by 5 pixels. The angular second moment reflects the uniformity of the image's gray-level distribution, contrast measures the degree of local variation in the image, and entropy indicates the randomness of the texture.

[0024] In practice, the preprocessed reference image set is converted to the HSV color space, and statistical histograms of the hue and saturation components are extracted as color features to capture the phenomenon of plants losing their green color and turning yellow. The images are converted from the original RGB color space to the HSV color space according to the standard color space conversion formula. The hue component represents the dominant color category of the pixel, and the saturation component represents the purity of the color. The statistical histograms statistically analyze the distribution frequency of hue and saturation values ​​at the pixel level. In some embodiments, the extraction of texture features and color features can be performed in parallel. Texture features are extracted from the panchromatic band of the high spatial resolution optical image or the fused visible light band, while color features are extracted from the red, green, and blue bands of the preprocessed reference image set.

[0025] It is understandable that the calculated vegetation index features, texture features, and color features differ in feature dimensions and numerical ranges, requiring standardization. Optionally, Z-score standardization is used, ensuring that the mean of each feature is 0 and the variance is 1. In practice, the standardized vegetation index features, texture features, and color features are concatenated at the pixel level, and latitude and longitude coordinates are added as location features to form a multidimensional forest health status feature tensor. The location features consist of the geodetic coordinates of the center point of each pixel, with longitude and latitude transformed into Cartesian coordinates consistent with the image. Ultimately, each pixel corresponds to a feature vector, and the feature vectors of all pixels are spatially organized according to the arrangement of the original image, thus forming a three-dimensional forest health status feature tensor. Its two spatial dimensions correspond to the number of rows and columns of the image, while the feature dimension corresponds to the total number of features after concatenation.

[0026] In one embodiment of the present invention, each pixel in the multidimensional forest health state feature tensor is used as a graph node, and the spatial distance and feature similarity between pixels are used as edge weights to construct a spatial graph structure. In the spatial graph structure, each graph node contains all feature values ​​in the corresponding forest health state feature tensor. The calculation of edge weights comprehensively considers spatial adjacency and Euclidean distance in the feature space. The calculation can be expressed as: in: and These are nodes and nodes eigenvectors, Represents Euclidean distance. It is a scale parameter that controls the rate of feature similarity decay. It is a node With nodes The spatial Euclidean distance between them It is a predefined spatial neighborhood radius. This is an indicator function; its value is 1 when the condition within the parentheses is met, and 0 otherwise. Using a sliding time window technique, the spatial graph structures of multiple consecutive time points are stacked to form a spatiotemporal cube data structure. In some embodiments, the length of the sliding time window is set to include 5 consecutive observation phases, with each phase spaced 16 days apart. The spatiotemporal cube data structure is organized in memory as a five-dimensional tensor, with dimensions representing the time step, image height, image width, number of feature channels, and adjacency relationships.

[0027] In practical implementation, the inference computing model based on spatiotemporal graph neural networks uses graph convolution operations to aggregate the feature information of the center pixel and its first-order neighboring pixels, updating the state representation of the center pixel. Graph convolution operations are performed independently on the graph structure at each time step, and the aggregation function uses either mean aggregation or weighted aggregation based on attention weights. After the graph convolution operations, a gated recurrent unit is introduced to encode the state representation on the time series, capturing the spread trend of pests and diseases over time. It can be understood that the input of the gated recurrent unit is the state of all nodes updated by graph convolution at the current time step, and its hidden states convey the pest and disease spread information from historical time steps. The node state representation output at the final time step is mapped to a value between 0 and 1 through a fully connected layer and an activation function; this value represents the probability of pest and disease occurrence per pixel. Optionally, when training the inference computing model based on spatiotemporal graph neural networks, a binary cross-entropy loss function is used, and supervised learning is performed using labeled historical pest and disease occurrence area data.

[0028] In practice, the spatial morphological processing of the probability distribution map of pest and disease occurrence involves setting a probability threshold, binarizing the map, and marking pixels with probability values ​​higher than the threshold as candidate infected pixels. The probability threshold is determined by analyzing the probability distribution histogram of historical data; for example, a threshold of 0.65 is set. An opening operation is performed on the binarized image to remove isolated noise points. The opening operation first performs an erosion operation followed by a dilation operation, using a circular structuring element with a radius of 3 pixels. A closing operation is then performed to fill the voids inside the lesions. This closing operation first performs a dilation operation followed by an erosion operation, using the same structuring element to obtain connected lesion regions. An edge detection algorithm is used to extract the contour lines of each connected lesion region, and the length and area of ​​the contour lines are calculated to obtain geometric morphological features. The edge detection algorithm uses the Canny operator; the contour line length is obtained by summing the Euclidean distances between adjacent pixels on the contour, and the area is obtained by counting the number of pixels contained within the contour line. The aspect ratio and compactness index of lesion shapes were calculated to distinguish between strip-shaped lesions caused by wind propagation and cluster-shaped lesions caused by insect propagation. The aspect ratio is the ratio of the longer side to the shorter side of the smallest bounding rectangle, and the compactness index is calculated using the following formula: Connected lesion regions with similar geometric features are merged to form the final suspected infected areas, and their vector boundaries are recorded. In some embodiments, the merging condition is that the distance between the centroids of multiple connected lesion regions is less than a set threshold, and the absolute value of the difference between their aspect ratio and compactness index is less than the corresponding tolerance threshold. Optionally, the vector boundaries are stored in the vector layer of the geographic information system in the form of polygonal rings.

[0029] See Figure 3 This is a heatmap showing the probability distribution of forest pests and diseases within a target forest area, pixel by pixel. The color gradient ranges from light yellow (0.75) to dark red (1.00), with darker colors indicating a higher probability of pest and disease occurrence at that pixel location. Large areas in the map represent medium-to-high probabilities (0.85-1.00), while only a few areas have low probabilities. The probability distribution exhibits patchy clusters without a clear single center, consistent with the random nature of the natural spread of forest pests and diseases. The scattered distribution of high-probability areas suggests that pests and diseases may spread through multiple points via wind or insect vectors, while low-probability areas are few in number, mostly isolated pixels, possibly indicating healthy forest areas or noise points. This provides forestry departments with accurate pest and disease risk location, guiding targeted monitoring, pesticide spraying, or manual intervention to reduce losses from disease spread.

[0030] In one embodiment of the present invention, based on the geometric morphological characteristics of the suspected disease-affected area, the original physiological state indicators of the suspected disease-affected area in the early stage of infection are inverted from the multidimensional forest health state feature tensor, resulting in a set of inverted physiological state parameters. Specifically, the inversion process starts from the geometric center pixel of the suspected disease-affected area and traces back along the historical time axis, extracting the multidimensional forest health state feature tensor data of the geometric center pixel at the same historical time phase. The time length of the tracing spans a complete phenological cycle. By analyzing the differences in the multispectral response curves of the suspected disease-affected area in the early stage of infection and the current period, the specific bands most sensitive to pest and disease stress are identified. It can be understood that the differences in the multispectral response curves are quantified by calculating the root mean square error of the reflectance in the blue, green, red, near-infrared, and short-wave infrared bands of the two periods, and the band with the largest root mean square error is identified as the most sensitive band. Using a radiative transfer model with the current vegetation canopy reflectance as input, the model iteratively adjusts the mesophyll structure and chlorophyll content parameters to minimize the residual between the simulated and observed reflectance. The optimization process aims to minimize the objective function. in: Representative model simulates reflectivity Compared with actual observed reflectance In all The sum of squared residuals across each band is iteratively adjusted using the Levenberg-Marquardt algorithm. When the radiative transfer model converges, the resulting mesophyll structure parameters and chlorophyll content parameters are the inverted physiological state parameters, representing the plant's physiological state in the early stages of infection. In some embodiments, the PROSAIL model is used as the radiative transfer model.

[0031] In practice, the retrieved physiological state parameters are used in conjunction with remote sensing data from historical healthy periods to construct a state inversion equation regarding pathogen diffusion dynamics. Solving this equation yields the pathogen's transmission rate and orientation field. Specifically, the forest area is divided into regular grid cells. The retrieved physiological state parameters serve as the initial susceptibility value for each grid cell, while the average vegetation index from historical healthy periods is used as the initial resilience value. Based on cellular automata theory, a state inversion equation is defined. The state transition probability in this equation is jointly determined by the disease and pest occurrence probability distribution map and the infection status of the current grid cell's neighboring areas. Using maximum likelihood estimation, the diffusion coefficient and wind direction influence factor in the state inversion equation are adjusted to achieve an optimal fit between the model-simulated infection range and the actual suspected infection area boundary. Refer to Table 1; the parameter fitting process can be illustrated using examples in Table 1.

[0032] Table 1: Parameter Fitting Process of State Inversion Equation Table 1 shows a comparison between the simulation results and actual observation results of the state inversion equation under different combinations of diffusion coefficients and wind direction influence factors. The simulated infected area is the total area of ​​the infected region predicted by the model, while the actual infected area is the total area of ​​the suspected infected region identified by remote sensing. Boundary overlap is used to measure the degree of spatial agreement between the simulated boundary and the actual boundary; the closer its value is to 1, the better the fit. Based on maximum likelihood estimation, the parameter combination that maximizes the boundary overlap is selected as the final solution.

[0033] After the model parameters are determined, the state inversion equation is run to record the spatial displacement of infected pixels per unit time, quantify it as a propagation rate, and obtain the orientation field through vector synthesis calculation. In some embodiments, the state inversion equation is run to simulate the daily diffusion process over the next 30 days, recording the displacement vector of newly infected pixels relative to their infection source each day, averaging all displacement vectors to obtain the daily average propagation rate, and statistically analyzing the direction of the displacement vectors, using the prevailing wind direction as the main direction of the orientation field. Optionally, the unit of time can be defined as one day, the unit of propagation rate is meters per day, and the orientation field is represented by angles, for example, 0 degrees represents true north. It can be understood that the vector synthesis calculation uses the method of calculating the average orientation angle of all displacement vectors and the magnitude of the synthesized vector.

[0034] See Figure 4 This is a comparative analysis chart of reflectance correction in multi-source remote sensing images. It shows the comparison curves of the original and corrected reflectance of multi-source remote sensing images in different bands. Its core purpose is to demonstrate the effect of radiometric calibration and atmospheric correction in remote sensing data preprocessing, a crucial step in feature extraction for intelligent identification of forestry pests and diseases. The reflectance in the visible light band is generally low, with a significant difference between the original and corrected curves. This band is sensitive to vegetation chlorophyll content. The near-infrared band shows the first peak in reflectance and is the core band for monitoring vegetation health. The short-wave infrared band exhibits the greatest reflectance fluctuations, with multiple peaks. It is sensitive to vegetation moisture and lignin content and can help determine the degree of pest and disease infection. The corrected reflectance is generally lower than the original reflectance, consistent with the physical logic of atmospheric correction. Atmospheric scattering enhances the apparent value of surface reflectance, making the corrected reflectance closer to the true surface reflectance. The difference is more significant in the short-wave infrared band, indicating that atmospheric interference has a more pronounced impact on short-wave infrared reflectance, making correction more valuable for improving the accuracy of data in this band.

[0035] In one embodiment of the present invention, based on the obtained propagation rate and direction field, the probability distribution map of pest and disease occurrence is extrapolated forward along the time axis to generate a pest and disease risk prediction map for future growth cycles. The propagation rate and direction field are discretized into eight directional vector components, each corresponding to a different diffusion tendency: north, northeast, east, southeast, south, southwest, west, and northwest. The probability distribution map of pest and disease occurrence is used as the initial infection source layer. Within each time step, the infection probability is spatially allocated and accumulated according to the diffusion tendency of each direction. The accumulation process of the infection probability can be expressed as follows: in: Represents the time step Located in the space grid The cumulative infection probability at the location Representing the The diffusion tendency weights for each azimuth are obtained by normalizing the magnitudes of the vector components for that azimuth. Representing the The row and column offsets corresponding to each location. Considering the differences in resistance among different tree species, the dominant tree species for each grid cell are read from the forest stand map, and the cumulative infection probability is weighted and corrected using the tree species resistance coefficient. The correction formula is as follows: ,in This is the corrected infection probability. It is a grid cell Superior tree species The resistance coefficient, resistance coefficient It is a value between 0 and 1 obtained from historical data statistics. The diffusion and correction process of the time step is repeated until the cumulative time reaches a complete growth cycle. The output cumulative probability layer is the pest and disease risk prediction map. It can be understood that a complete growth cycle is set to, for example, 120 days.

[0036] In practice, the pest and disease risk prediction map is classified into different levels, and the different risk levels are mapped to specific tree species in conjunction with the forest stand map. The risk probability values ​​of all pixels in the pest and disease risk prediction map are extracted and statistically analyzed to determine the mean and standard deviation of the probability distribution. In some embodiments, the statistical analysis calculates the arithmetic mean of the risk probability values ​​of all pixels within the entire prediction map area. and sample standard deviation Based on the quantile principle of the normal distribution, the pest and disease risk prediction map is divided into three levels: low-risk, medium-risk, and high-risk areas. Optionally, the division rule is defined as follows: risk probability values ​​less than... The area is a low-risk zone, with a risk probability value between [missing information]. and The area between these two points is a medium-risk area, with a risk probability value greater than [missing value]. The area is designated as a high-risk zone. Spatial overlay analysis was performed on the forest stand map's sub-compartment zoning data, tree species composition data, and pest and disease risk prediction map. This overlay analysis was achieved through spatial connection operations within a geographic information system, linking the attributes of the risk level raster data with the attributes of the sub-compartment polygon vector data.

[0037] The planting area percentage of each major tree species within each risk level area is statistically analyzed to create a table mapping tree species to risk levels. This mapping can be understood as follows: the planting area percentage is calculated by accumulating the area of ​​all pixels within a specific risk level area where the dominant tree species is a particular species, and then dividing that area by the total area of ​​that risk level area. For dominant tree species in high-risk areas, their historical damage records are further queried as a basis for key monitoring. Optionally, the tree species-risk level mapping table is stored in a structured table format, with table fields including at least the risk level, dominant tree species, planting area, area percentage, and historical damage record identifier.

[0038] See Figure 5 This is a heatmap showing the spatial correlation between forestry pest and disease risk and dominant tree species. It visually presents the distribution of the planting area of ​​three dominant tree species in areas with different risk levels, and is a core visualization result of the "risk level-tree species association mapping" in intelligent identification of forestry pests and diseases. The darker the color, the higher the planting area of ​​that tree species in the corresponding risk level area. Poplar accounts for 100% of the area in high-risk areas, making it the absolute dominant tree species in high-risk areas. Combined with historical data, it can be inferred that poplar has the weakest resistance to pests and diseases and is the core susceptible tree species for pest and disease infection. Poplar should be the key monitoring target in high-risk areas, and priority should be given to intervention measures such as chemical control and artificial tending. Chinese fir has the highest area in low-risk areas (~60%), only ~25% in medium-risk areas, and is not distributed in high-risk areas, showing strong resistance to pests and diseases, and is a relatively safe tree species in forest areas. Pine trees account for approximately 25% of low-risk areas, approximately 50% of medium-risk areas, and are not found in high-risk areas. Their resistance is between that of Chinese fir and poplar, making them the dominant tree species in medium-risk areas.

[0039] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for intelligent identification of forestry pests and diseases based on remote sensing data, characterized in that, include: Multi-source remote sensing image data of the target forest area were acquired, and a baseline image set was obtained after preprocessing. Vegetation index features are extracted from a preprocessed set of benchmark images, and texture and color features are fused to construct a multidimensional forest health status feature tensor. The forest health status feature tensor is input into a spatiotemporal graph neural network-based inference model. This model aggregates neighborhood pixel information and temporal evolution information to output a pixel-by-pixel probability distribution map of pest and disease occurrence, including: Using each pixel in the multidimensional forest health status feature tensor as a graph node, and using the spatial distance and feature similarity between pixels as edge weights, a spatial graph structure is constructed. By using the sliding time window technique, the spatial graph structures of multiple consecutive time points are stacked to form a spatiotemporal cube data structure; The inference computation model based on spatiotemporal graph neural network uses graph convolution operations to aggregate the feature information of the central pixel and its first-order neighboring pixels, and update the state representation of the central pixel. After the graph convolution operation, a gated recurrent unit is introduced to encode the state representation on the time series to capture the spread trend of pests and diseases over time. The node state representation output at the final time step is mapped to a value between 0 and 1 through a fully connected layer and an activation function. This value represents the probability of pest and disease occurrence per pixel. Spatial morphological processing is performed on the probability distribution map of the disease and pest occurrence to identify the specific boundaries of suspected disease-infected areas, and the geometric morphological characteristics of the suspected disease-infected areas are statistically analyzed. Based on the geometric morphological features of the suspected disease-infected area, the original physiological state indicators of the suspected disease-infected area in the early stage of infection are inverted from the multidimensional forest health status feature tensor, resulting in a set of inverted physiological state parameters. Using the retrieved physiological state parameters and remote sensing data from historical healthy periods, a state inversion equation for pathogen diffusion dynamics is constructed, and the equation is solved to obtain the pathogen's propagation rate and orientation field, including: The forest area is divided into regular grid units, and the inverted physiological state parameters are used as the initial susceptibility values ​​for each grid unit, while the average vegetation index during historical healthy periods is used as the initial resilience value. Based on cellular automata theory, the state inversion equation is defined, and the state transition probability in the state inversion equation is jointly determined by the disease and pest occurrence probability distribution map and the neighboring infection state of the current grid cell. By using the maximum likelihood estimation method, the diffusion coefficient and wind direction influence factor in the state inversion equation are adjusted so that the infection range simulated by the model and the actual boundary of the suspected infected area achieve the best fit. After the model parameters are determined, the state inversion equation is run to record the spatial displacement of the infected pixels per unit time, which is quantified as the propagation rate, and the direction field is obtained by vector synthesis calculation. Based on the obtained propagation rate and direction field, the probability distribution map of the occurrence of pests and diseases is extrapolated forward along the time axis to generate a risk prediction map of pests and diseases in the future growth cycle.

2. The intelligent identification method for forestry pests and diseases based on remote sensing data as described in claim 1, characterized in that, The acquisition of multi-source remote sensing image data of the target forest area, after preprocessing, yields a reference image set, including: Acquire multi-source remote sensing image data of the target forest area, including high spatial resolution optical images, multispectral images, and shortwave infrared images; Radiometric calibration and atmospheric correction are performed on the multi-source remote sensing image data to eliminate sensor errors and atmospheric scattering interference, resulting in a preprocessed reference image set. Vegetation index features are extracted from the preprocessed reference image set, and texture and color features are fused to construct a multidimensional forest health status feature tensor. The process of performing radiometric calibration and atmospheric correction on the multi-source remote sensing image data eliminates sensor-specific errors and atmospheric scattering interference, resulting in a preprocessed reference image set, including: The metadata information in the header file of the multi-source remote sensing image data is read to obtain the sensor's gain and bias parameters, and the original digital quantization value is linearly transformed based on the gain and bias parameters to obtain the apparent radiance. Using the aerosol optical thickness product provided by the medium resolution imaging spectrometer, atmospheric scattering removal is performed on the apparent radiance to calculate the true surface reflectance. For the high spatial resolution optical image, orthorectification is performed to eliminate geometric distortion caused by terrain undulations and ensure the correspondence between pixel positions and geographic coordinates; Band registration is performed on the multispectral image and the shortwave infrared image, and image data from different sources are resampled to the same spatial resolution grid coordinate system; All processed image data are matched in the time dimension, and images with cloud cover below a preset threshold are selected as valid data and combined to form the preprocessed reference image set.

3. The intelligent identification method for forestry pests and diseases based on remote sensing data as described in claim 2, characterized in that, Vegetation index features are extracted from the preprocessed reference image set, and texture and color features are fused to construct a multidimensional forest health status feature tensor, including: The reflectance ratio of the near-infrared band to the red band in the preprocessed reference image set is calculated to obtain the normalized vegetation index, and the enhanced vegetation index and the soil-regulated vegetation index are calculated simultaneously. Gray-level co-occurrence matrix analysis was performed on the high spatial resolution optical image to extract the second moment of the angle, contrast and entropy as the texture features to characterize the degree of damage to the blade surface structure. The preprocessed reference image set is converted to the HSV color space, and the statistical histograms of the hue and saturation components are extracted as the color features to capture the phenomenon of plants losing their green color and turning yellow. The calculated vegetation index features, texture features, and color features are concatenated at the pixel level, and latitude and longitude coordinate information is added as location features to form the multidimensional forest health status feature tensor.

4. The intelligent identification method for forestry pests and diseases based on remote sensing data as described in claim 3, characterized in that, Spatial morphological processing was performed on the probability distribution map of the pest and disease occurrence to identify the specific boundaries of suspected infected areas, and the geometric morphological characteristics of the suspected infected areas were statistically analyzed, including: A probability threshold is set, and the probability distribution map of the occurrence of pests and diseases is binarized. Pixels with probability values ​​higher than the probability threshold are marked as candidate pixels for disease infection. The binarized image is subjected to an opening operation to remove isolated noise points, and then a closing operation is performed to fill the holes inside the lesions, resulting in connected lesion regions. The contour lines of each connected lesion region are extracted using an edge detection algorithm, and the length and area of ​​the contour lines are calculated to obtain the geometric morphological features; Calculate the aspect ratio and compactness index of the lesion shape to distinguish between strip-shaped lesions caused by wind propagation and cluster-shaped lesions caused by insect propagation; Connected lesion regions with similar geometric features are merged to form the final suspected diseased region, and their vector boundaries are recorded.

5. The intelligent identification method for forestry pests and diseases based on remote sensing data as described in claim 4, characterized in that, Based on the geometric morphological characteristics of the suspected disease-affected area, the original physiological state indicators of the suspected disease-affected area in the early stage of infection are inverted from the multidimensional forest health status feature tensor, resulting in a set of inverted physiological state parameters, including: Starting from the geometric center pixel of the suspected diseased area, backtracking along the historical timeline, extract the multidimensional forest health status feature tensor data of the geometric center pixel in the same historical time period; Analyze the differences in multispectral response curves of the suspected disease-infected areas at the early stage of infection and at the current stage to identify the specific bands most sensitive to pest and disease stress. Using a radiative transfer model, with the current vegetation canopy reflectance as input, the model iteratively adjusts the mesophyll structure parameters and chlorophyll content parameters to minimize the residual between the model-simulated reflectance and the actual observed reflectance. When the radiative transfer model converges, the obtained mesophyll structure parameters and chlorophyll content parameters are the inverted physiological state parameters, representing the physiological state of the plant in the early stage of infection.

6. The intelligent identification method for forestry pests and diseases based on remote sensing data as described in claim 5, characterized in that, Based on the obtained propagation rate and direction field, the probability distribution map of pest and disease occurrence is extrapolated forward along the time axis to generate a pest and disease risk prediction map for future growth cycles, including: The propagation rate and the direction field are discretized into eight vector components in each direction, which correspond to the diffusion tendency in different directions. Using the disease and pest occurrence probability distribution map as the initial infection source layer, the infection probability is spatially allocated and accumulated according to the diffusion tendency in each direction within each time step. Considering the differences in resistance among different tree species, the dominant tree species of each grid unit are read from the forest stand map, and the cumulative infection probability is weighted and corrected by the tree species resistance coefficient. Repeat the diffusion and correction process of the time step until the cumulative time reaches a complete growth cycle. The output cumulative probability layer is the pest and disease risk prediction map.

7. The intelligent identification method for forestry pests and diseases based on remote sensing data as described in claim 6, characterized in that, Also includes: The disease and pest risk prediction map is classified into levels, and the regions with different risk levels are associated with specific tree species types by combining the forest stand map; By integrating the boundaries of the suspected infected areas, the inverted physiological state parameters, the transmission rate and direction field, and the pest and disease risk prediction map, a structured intelligent diagnosis and prediction report of forestry pests and diseases is generated. The process of classifying the pest and disease risk prediction map into risk levels and mapping different risk levels to specific tree species in conjunction with the forest stand map includes: Extract the risk probability values ​​of all pixels in the pest and disease risk prediction map, and perform statistical analysis to determine the mean and standard deviation of the probability distribution; Based on the quantile principle of the normal distribution, the pest and disease risk prediction map is divided into three levels: low-risk area, medium-risk area and high-risk area. Spatial overlay analysis was performed on the small-compartment zoning data and tree species composition data in the forest stand map and the pest and disease risk prediction map. The planting area percentage of each major tree species within each risk level area was statistically analyzed to create a table showing the correspondence between tree species and risk levels. For dominant tree species in high-risk areas, their historical damage records should be further investigated as a basis for key monitoring.

8. A forestry pest and disease intelligent identification system based on remote sensing data, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent identification method for forestry pests and diseases based on remote sensing data as described in any one of claims 1 to 7.

Images