A forestry tree species classification and identification method and system based on image recognition

Abstract

The application discloses a forestry tree species classification and identification method and system based on image recognition and relates to the technical field of image processing. The continuum removal method is used to extract the spectral fingerprint parameters of the hyperspectral image to be detected, and the component coupling ratio is calculated. Based on the component coupling ratio, the biochemical structural heterogeneity features are obtained, and the biochemical content feature vector is obtained by combination. The adaptive neighborhood structure is used to check the biochemical content feature vector to obtain the spatial spectral features by smoothing processing. Based on the sample set, the target mask area is identified and the pseudo-background spectrum is filled. The spatial spectral features are used to solve the filter coefficient vector. The response polarization score map is output by combining the preset exclusive content fingerprint library. The features of the map and the biochemical content feature vector are fused to obtain the white spruce and blue spruce feature point cloud. The white spruce and blue spruce feature point cloud is used to generate the discriminant distance matrix by combining the second-order K nearest neighbor algorithm. Based on the discriminant distance matrix, the class assignment and the spatial consistency correction are performed, and the white spruce and blue spruce classification map is output.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for classifying and identifying forestry tree species based on image recognition. Background Technology

[0002] In recent years, high-precision tree species classification has become fundamental to forest resource monitoring. *Picea alba* and *Picea chinensis*, as closely related species within the genus *Picea*, face significant challenges in traditional remote sensing identification: First, their spectral curves highly overlap, exhibiting severe heterospectral differences and heterospectral similarities, making it difficult for algorithms based on apparent spectra to achieve essential distinction; second, uneven lighting and topographical shadows in forest areas easily lead to feature drift, and existing methods lack physiological mechanism constraints, resulting in poor identification stability; finally, insufficient utilization of spatial information, with traditional filtering easily causing edge blurring and severe salt-and-pepper noise in the classification images. Therefore, deeply coupling biochemical content features with spatial topology to solve the high-precision identification of closely related tree species in complex backgrounds is a critical problem that urgently needs to be addressed.

[0003] Currently, Chinese invention patent application CN202111447482.1 discloses an intelligent detection method for afforestation quality based on UAV remote sensing, which includes the following steps: (S1) selecting several ground control points; (S2) mounting a visible light camera and a hyperspectral imager on a UAV to collect data on newly afforested forest land from the air; (S3) using a convolutional neural network to identify "killed seedlings" from the visible light images of the forest land data; (S4) stitching together the hyperspectral images from the forest land data to form a forest land image based on the control point location information; (S5) using a pre-trained deep learning model to classify tree species in the forest land image based on the size, shape, and spectral characteristics of the seedlings, and calculating the quantity configuration of each tree species; (S6) calculating each evaluation index based on the forest land image and the quantity configuration of each tree species. However, the related technology ignores the phenomena of different spectra for the same species and the same spectra for different species, as well as the spatial coherence of forest stand distribution. Summary of the Invention

[0004] The technical problem solved by this invention is that the spectral curves of *Phyllostachys edulis* and *Phyllostachys bambusoids* are extremely similar in the visible-near-infrared band, exhibiting severe heterospectral differences and heterogeneous spectral similarities. Related technologies struggle to capture the subtle differences in the proportions of trace biochemical components such as waxes, polyphenols, and lignin. Furthermore, the pixel-by-pixel classification of related technologies ignores the spatial continuity of forest stand distribution, resulting in numerous misclassification points at target boundaries in the final classification map.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] A method for classifying and identifying forestry tree species based on image recognition includes the following steps:

[0007] Step S1: Process the hyperspectral image to be detected using the continuum removal method, extract spectral fingerprint parameters, lignin index and cellulose absorption index, calculate the component coupling ratio using the lignin index and cellulose absorption index, and calculate the biochemical structural heterogeneity characteristics based on the component coupling ratio. Combine the spectral fingerprint parameters, component coupling ratio and biochemical structural heterogeneity characteristics to obtain the biochemical content feature vector.

[0008] Step S2: The biochemical content feature vector is smoothed using an adaptive neighborhood structure kernel to obtain spatial spectral features. Based on the acquired sample set, the target mask region is identified and pseudo-background spectral filling is performed to generate a background autocorrelation matrix. The filter coefficient vector is solved using the minimum output energy criterion through the background autocorrelation matrix and spatial spectral features. The response polarization score map is output in combination with the preset exclusive content fingerprint database.

[0009] Step S3: The response polarization score map and the biochemical content feature vector are fused to obtain the feature point cloud of white and green foliage. The feature point cloud of white and green foliage is then combined with second-order... The nearest neighbor algorithm generates a discriminant distance matrix, and performs category assignment and spatial consistency correction based on the discriminant distance matrix, outputting a white and green classification map.

[0010] Preferably, step S1 specifically includes:

[0011] Step S11: Obtain the hyperspectral image to be detected, and normalize the hyperspectral image to be detected using the continuum removal method to obtain the spectral curve;

[0012] Step S12: Extract the wax scattering intercept at 450 nm and the polyphenol absorption depth at 1660 nm from the spectral curve to obtain the spectral fingerprint parameters.

[0013] Step S13: Based on the spectral fingerprint parameters, calculate the lignin index in the 1750nm band and the cellulose absorption index in the 2100nm band, and normalize them. Calculate the component coupling ratio using the normalized lignin index and the normalized cellulose absorption index, and obtain the component coupling ratio distribution map.

[0014] Step S14: Use a preset sliding window to traverse the component coupling ratio distribution map, and obtain the biochemical structural heterogeneity map by statistically analyzing the biochemical structural heterogeneity characteristics of the component content in each sliding window.

[0015] The biochemical structural heterogeneity features include spatial gradient, discrete coefficient, and information entropy;

[0016] Step S15: Stack the spectral fingerprint parameters, component coupling ratio distribution map and biochemical structural heterogeneity map into a matrix, perform numerical mapping using a preset exclusive content fingerprint library, and combine the feature values ​​corresponding to the map at each pixel location into the biochemical content feature vector of that pixel.

[0017] The preset exclusive content fingerprint database includes the content mean vector of the target tree species, the content covariance matrix of the target tree species, and the effective content range of the target tree species.

[0018] Preferably, the formula for calculating the component coupling ratio is:

[0019] ;

[0020] in, The component coupling ratio, The normalized lignin index, This is the normalized cellulose absorption index.

[0021] Preferably, step S2 specifically includes:

[0022] Step S21: Smooth the biochemical content feature vector using an adaptive neighborhood structure kernel and extract spatial spectral features;

[0023] Step S22: Obtain the sample set, calculate the white target guidance vector and the green target guidance vector based on the sample set, identify the target mask region and perform pseudo-background spectral filling, calculate the covariance distribution, and generate the background autocorrelation matrix;

[0024] The sample set includes standard hyperspectral data of white sage and standard hyperspectral data of green sage;

[0025] Step S23: Based on the minimum output energy criterion, the filter coefficient vector is solved using the background autocorrelation matrix, and the response polarization score map is obtained by combining the spatial spectral features and biochemical content feature vectors. The response polarization score map includes the green sap response map and the white sap response map.

[0026] Preferably, step S21 specifically includes:

[0027] Based on the preset structural kernel window, the cosine distance between the center pixel and its neighboring pixels within the preset structural kernel window is calculated, and the normalized Euclidean distance between the center pixel and its neighboring pixels on the biochemical content feature vector within the preset structural kernel window is also calculated. A Gaussian function is then used to map the cosine distance and the normalized Euclidean distance into joint weight coefficients. The calculation expression is as follows:

[0028] ;

[0029] in, These are the joint weighting coefficients. Center pixel With neighboring pixels The cosine distance between them For normalized Euclidean distance, and These are the spectral bandwidth and the content bandwidth, respectively.

[0030] The spatial spectral features are obtained by summing and averaging the biochemical content feature vectors of all pixels within the preset structural kernel window according to the joint weight coefficients.

[0031] Preferably, in step S22, the specific processing logic for generating the background autocorrelation matrix includes:

[0032] The characteristic mean values ​​of wax scattering intercept, polyphenol absorption depth and component coupling ratio of the standard hyperspectral of white pine sample and green pine sample were calculated respectively, and used as the target guiding vectors for white pine and green pine.

[0033] Initial energy detection is performed based on the white and green target guiding vectors to identify potential target regions and generate target masks. Biochemical content feature vectors of the background region within the target mask neighborhood are extracted. The arithmetic mean of the biochemical content feature vectors of the background region is calculated by averaging the biochemical content feature vectors of the background region. The original feature values ​​in the target mask coverage area are replaced with the biochemical content mean vector to obtain a pseudo-background image. The covariance distribution is calculated using the spatial spectral features of the pseudo-background image to generate the background autocorrelation matrix.

[0034] Preferably, step S23 specifically includes:

[0035] Based on the criterion of minimizing output energy, the filter coefficient vector is solved, and the calculation expression is as follows:

[0036] ;

[0037] in, Let m be the filter coefficient vector of the m-th target tree species. Guide vector for the target tree species. Let R be the transpose of the target tree species' guiding vector, and let R be the background autocorrelation matrix. It is the inverse of the background autocorrelation matrix. To normalize the denominator, Index for target tree species categories, When =1, it represents white. When =2, it represents green moss;

[0038] The initial response score is calculated using the filter coefficient vector and the background autocorrelation matrix. The calculation expression is as follows:

[0039] ;

[0040] in, Let be the transpose of the filter coefficient vector for the m-th target tree species. Let x be the spatial spectral feature vector of pixel x. The initial response score;

[0041] The prior probability factor of the content is calculated using the mean vector of the content and the biochemical content feature vector of the target tree species. The calculation expression is as follows:

[0042] ;

[0043] in, The content prior probability factor, The biochemical content feature vector of pixel x, Let be the vector of the average content of the m-th target tree species. Let be the inverse matrix of the covariance matrix of the content of the m-th target tree species;

[0044] The response polarization score is calculated by combining the filter coefficient vector and the prior probability factor of the content. The calculation expression is as follows:

[0045] ;

[0046] in, In response to polarization scores, The effective range for content;

[0047] Obtain the response polarization score map based on the response polarization score corresponding to each pixel.

[0048] Preferably, in step S3, the processing logic for generating the discriminant distance matrix specifically includes:

[0049] The response polarization score map and the biochemical content feature vector are normalized, and the normalized response polarization score map and the biochemical content feature vector are fused to obtain the feature point cloud of white and green foliage.

[0050] Searching for each pixel in the feature point cloud of white and green haze using Euclidean distance. 1. ... 1. Neighboring points, construct a second order The set of nearest neighbor representative points;

[0051] Statistical analysis of shared second order of any adjacent pixel pairs The number of nearest neighbors is calculated, and the Mahalanobis distance between adjacent pixel pairs on the biochemical content feature vector is used as the content deviation.

[0052] Will share second order The nearest neighbor count plus 1 is used as the denominator, and the content deviation is used as the numerator to perform a decay operation, generating a discrimination distance matrix. The calculation expression is as follows:

[0053] ;

[0054] in, To determine the distance matrix, The deviation in content between pixel i and pixel j. Let be the inverse matrix of the covariance matrix of the content of the m-th target tree species. To share the number of second-order nearest neighbors, and These are the biochemical content feature vectors of pixels i and j, respectively.

[0055] Preferably, in step S3, performing category assignment and spatial consistency correction based on the discriminant distance matrix specifically includes:

[0056] By discriminating the distance matrix, the local density and relative distance of each pixel are calculated, and the weight index is obtained by multiplying the local density and relative distance.

[0057] The pixels with the largest weight index, whose response polarization score in the green pine response map is greater than the first high value threshold, and whose biochemical content feature vector is located within the effective content range of the target tree species are selected as the initial green pine class center.

[0058] The initial white elm class center was selected based on the largest weight index, the response polarization score in the white elm response map being greater than the second highest threshold, and the biochemical content feature vector being located within the effective content range of the target tree species.

[0059] After normalizing based on the number of shared second-order nearest neighbors, the shared nearest neighbor similarity is obtained. A joint similarity is then constructed, and its calculation expression is as follows:

[0060] ;

[0061] in, For joint similarity, To share the similarity of nearest neighbors, The content prior probability factor, In response to polarization scores, As the first preset weight, As the second preset weight, The third preset weight;

[0062] Based on the maximum similarity criterion, each pixel to be classified is assigned to the category of the class center with the highest joint similarity and given the corresponding initial category label;

[0063] A spatial window is established centered on each pixel to be classified. The initial class labels of all neighboring pixels within the spatial window are counted. It is then determined whether the initial class label of the central pixel is consistent with the initial class label that appears most frequently among the neighboring pixels.

[0064] If there is a discrepancy, and the proportion of the most frequently occurring initial category label exceeds the preset consistency threshold, then the initial category label of the center pixel is replaced with the most frequently occurring initial category label according to the local voting principle to obtain the final category label;

[0065] The final category labels of all the pixels are spatially mapped according to the original geographic coordinates to output a classification map of white and green irises.

[0066] A forestry tree species classification and identification system based on image recognition includes a feature extraction module, a processing module, and a classification module.

[0067] The feature extraction module is used to process the hyperspectral image to be detected using the continuum removal method, extract spectral fingerprint parameters, lignin index and cellulose absorption index, calculate the component coupling ratio using the lignin index and cellulose absorption index, and calculate the biochemical structural heterogeneity characteristics based on the component coupling ratio to obtain the component coupling ratio distribution map and the biochemical structural heterogeneity map, which are combined with the spectral fingerprint parameters to obtain the biochemical content feature vector.

[0068] The processing module is used to smooth the biochemical content feature vector using an adaptive neighborhood structure kernel to obtain spatial spectral features. Based on the acquired sample set, the target mask region is identified and pseudo-background spectral filling is performed to generate a background autocorrelation matrix. Through the background autocorrelation matrix and spatial spectral features, the filter coefficient vector is solved using the minimum output energy criterion. Combined with a preset exclusive content fingerprint database, the response polarization score map is output.

[0069] The classification module is used to fuse the response polarization score map with the biochemical content feature vector to obtain the feature point cloud of white and green foliage. This feature point cloud is then combined with second-order... The nearest neighbor algorithm generates a discriminant distance matrix, and performs category assignment and spatial consistency correction based on the discriminant distance matrix, outputting a white and green classification map.

[0070] The beneficial effects of this invention are as follows: By introducing biochemical mechanism constraints and spatial topological discrimination, this invention improves the subdivision and identification accuracy of closely related heterogeneous tree species. Addressing the pain point of highly overlapping spectra between *Pterocarya stenoptera* and *Pterocarya acutissima*, this invention utilizes biochemical content characteristics sensitive to leaf surface waxes and polyphenol metabolites to fundamentally separate interclass distances based on physiological structure, reducing the misclassification rate of closely related tree species. Simultaneously, a dual robust mechanism of component coupling ratio and effective content range is constructed, using ratio calculations to offset light fluctuations and hard-gating to automatically filter background noise exceeding physiological constants, enhancing anti-interference capabilities in complex forest stand environments. An adaptive structural kernel achieves high-performance edge-preserving smoothing, suppressing canopy texture noise while effectively preventing edge blurring through biochemical component mutation identification. Finally, by combining a shared second-order k-nearest neighbor topology and discriminant distance matrix, an assignment logic of accelerated intra-class aggregation and forced inter-class blocking is established, along with spatial consistency correction, solving the salt-and-pepper noise problem in traditional classification maps and outputting high-precision forestry classification results with clear boundaries and spatial coherence. Attached Figure Description

[0071] Figure 1 A basic flowchart of a forestry tree species classification and identification method based on image recognition is provided in one embodiment of the present invention;

[0072] Figure 2 This is a schematic diagram of the basic process of a forestry tree species classification and identification system based on image recognition, provided as an embodiment of the present invention. Detailed Implementation

[0073] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0074] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for classifying and identifying forestry tree species based on image recognition is provided, comprising the following steps:

[0075] Step S1: Process the hyperspectral image to be detected using the continuum removal method, extract spectral fingerprint parameters, lignin index and cellulose absorption index, calculate the component coupling ratio using the lignin index and cellulose absorption index, and calculate the biochemical structural heterogeneity characteristics based on the component coupling ratio. Combine the spectral fingerprint parameters, component coupling ratio and biochemical structural heterogeneity characteristics to obtain the biochemical content feature vector.

[0076] Step S2: The biochemical content feature vector is smoothed using an adaptive neighborhood structure kernel to obtain spatial spectral features. Based on the acquired sample set, the target mask region is identified and pseudo-background spectral filling is performed to generate a background autocorrelation matrix. The filter coefficient vector is solved using the minimum output energy criterion through the background autocorrelation matrix and spatial spectral features. The response polarization score map is output in combination with the preset exclusive content fingerprint database.

[0077] Step S3: The response polarization score map and the biochemical content feature vector are fused to obtain the feature point cloud of white and green foliage. The feature point cloud of white and green foliage is then combined with second-order... The nearest neighbor algorithm generates a discriminant distance matrix, and performs category assignment and spatial consistency correction based on the discriminant distance matrix, outputting a white and green classification map.

[0078] This invention addresses the challenge of spectral overlap among tree species, which leads to heterogeneous species exhibiting similar spectra. Instead of relying on fluctuating apparent spectra, it extracts fingerprints of wax and polyphenol content that are sensitive to physiological differences, establishing a discrimination criterion from a microscopic biochemical perspective, significantly reducing the false positive rate. Through the synergistic constraint of component coupling ratios and effective content ranges, this scheme uses ratio calculations to offset ambient light interference and employs physiological constants as hard gating to automatically remove background noise, enhancing the algorithm's robustness in complex forest stands. Particularly noteworthy is its innovative construction of a joint biochemical consistency adaptive structure kernel, performing smoothing only between pixels with similar components, thus suppressing canopy texture noise while preserving accurate tree species distribution boundaries. By sharing a second-order nearest neighbor topology, content features are mapped to a high-dimensional point cloud space. A discriminative distance matrix is ​​used to achieve category assignment based on a biochemical manifold, and spatial consistency correction addresses the fragmentation problem caused by traditional pixel-by-pixel classification, outputting a classification map of *Phyllostachys edulis* and *Phyllostachys aurea* that combines biochemical mechanisms and spatial semantics.

[0079] In a specific embodiment, step S1 specifically includes:

[0080] Step S11: Obtain the hyperspectral image to be detected, and normalize the hyperspectral image to be detected using the continuum removal method to obtain the spectral curve;

[0081] Step S12: Extract the wax scattering intercept at 450 nm and the polyphenol absorption depth at 1660 nm from the spectral curve to obtain the spectral fingerprint parameters.

[0082] Step S13: Based on the spectral fingerprint parameters, calculate the lignin index in the 1750nm band and the cellulose absorption index in the 2100nm band, and normalize them. Calculate the component coupling ratio using the normalized lignin index and the normalized cellulose absorption index, and obtain the component coupling ratio distribution map.

[0083] Specifically, the continuum removal method is used to fit the hyperspectral image to be detected, generating a spectral envelope. The hyperspectral image value is divided by the corresponding value on the envelope to obtain a spectral curve between 0 and 1. At the 450 nm band of the spectral curve, its reflectance intercept is extracted; this reflectance intercept reflects the scattering intensity of the wax layer. At the 1660 nm absorption valley of the spectral curve, the polyphenol absorption depth at this point is calculated using the following expression:

[0084] DF = 1 - Fs;

[0085] Where DF is the polyphenol absorption depth and Fs is the reflectance;

[0086] The lignin index, representing the degree of lignification, was calculated using the reflectance at 1750 nm; the cellulose absorption index was calculated using the absorption characteristics at 2100 nm.

[0087] It should be noted that this scheme selects 450nm as the calculation point for the waxy scattering intercept, rather than the red light window of the general vegetation index, because the waxy layer exhibits extremely strong specificity for incoherent scattering of short-wavelength light. Meanwhile, there are significant differences between *Picea spp.* and *Picea acutissima* in the accumulation of polyphenolic stress-resistance metabolites. The 1660nm band is located at the boundary between cellulose absorption and lignin characteristic peaks, and can sensitively capture changes in the slope of the spectral envelope caused by differences in polyphenol content. Compared to conventional wideband scanning, this scheme successfully differentiates the spatial distance between the characteristics of two visually easily confused spruce species through specific spectral fingerprint parameters.

[0088] Step S14: Use a preset sliding window to traverse the component coupling ratio distribution map, and obtain the biochemical structural heterogeneity map by statistically analyzing the biochemical structural heterogeneity characteristics of the component content in each sliding window.

[0089] The biochemical structural heterogeneity features include spatial gradient, discrete coefficient, and information entropy;

[0090] Step S15: Stack the spectral fingerprint parameters, component coupling ratio distribution map and biochemical structural heterogeneity map into a matrix, perform numerical mapping using a preset exclusive content fingerprint library, and combine the feature values ​​corresponding to the map at each pixel location into the biochemical content feature vector of that pixel.

[0091] The preset exclusive content fingerprint database includes the content mean vector of the target tree species, the content covariance matrix of the target tree species, and the effective content range of the target tree species.

[0092] Specifically, using the preset size A sliding window is used. The maximum rate of change in the component coupling ratio between the central pixel and its neighboring pixels within the sliding window is calculated as the spatial gradient. The ratio of the standard deviation to the mean of the component coupling ratio within the sliding window is calculated as the dispersion coefficient. The probability distribution density of the component coupling ratio within the sliding window is statistically analyzed. The spatial gradient, dispersion coefficient, and information entropy are stacked as matrices in the vertical direction. The spatial gradient reflects the degree of abrupt change in biochemical components in space. Due to the different optical shading caused by the needle arrangement and branching structure of *Phyllostachys nigra* and *Phyllostachys aurea*, the spatial transition rates of their component contents differ significantly. The dispersion coefficient is used to eliminate the influence of the average content background and purely describe the fluctuation amplitude of biochemical component distribution. It can accurately capture the subtle perturbations caused by physiological metabolic heterogeneity within the forest canopy.

[0093] Information entropy characterizes the disorder or complexity of the spatial arrangement of biochemical components. The higher the information entropy, the more complex the biochemical structure of the tree species within the current spatial range, which is an important nonlinear characteristic for distinguishing between two types of spruce.

[0094] Numerical mapping using a pre-defined, proprietary content fingerprint database specifically includes:

[0095] The spectral fingerprint parameter map, component coupling ratio distribution map, and biochemical structural heterogeneity map are scaled and normalized to eliminate numerical deviations caused by different dimensions among wax scattering intercept, polyphenol absorption depth, component coupling ratio, and spatial structural statistics. The obtained apparent spectral feature values ​​are compared with the mean content vector in a dedicated biochemical content fingerprint database. A preset mapping function converts spectral absorption and scattering intensity into quantitative indicators of biochemical components characterizing plant physiological states. Hard constraint calibration is performed on the mapping results using the effective content range. If the feature value of a pixel exceeds the normal physiological range for *Phyllostachys edulis* or *Phyllostachys pubescens*, the value is determined to be environmental noise interference. The calibrated spectral fingerprint parameters, component coupling ratio, and biochemical structural heterogeneity features are stacked in a multi-channel matrix according to pixel spatial location to form the biochemical content feature vector for that pixel.

[0096] The preset mapping function adopts linear proportional mapping logic, which calculates the offset ratio of the apparent spectral feature value relative to the content mean vector in the fingerprint database, and proportionally scales the spectral magnitude to the content standard range defined by the fingerprint database, thereby achieving normalization calibration of the feature dimension.

[0097] In a specific embodiment, the expression for calculating the component coupling ratio is as follows:

[0098] ;

[0099] in, The component coupling ratio, The normalized lignin index, This is the normalized cellulose absorption index.

[0100] In a specific embodiment, step S2 specifically includes:

[0101] Step S21: Smooth the biochemical content feature vector using an adaptive neighborhood structure kernel and extract spatial spectral features;

[0102] Step S22: Obtain the sample set, calculate the white target guidance vector and the green target guidance vector based on the sample set, identify the target mask region and perform pseudo-background spectral filling, calculate the covariance distribution, and generate the background autocorrelation matrix;

[0103] The sample set includes standard hyperspectral data of white sage and standard hyperspectral data of green sage;

[0104] Step S23: Based on the minimum output energy criterion, the filter coefficient vector is solved using the background autocorrelation matrix, and the response polarization score map is obtained by combining the spatial spectral features and biochemical content feature vectors. The response polarization score map includes the green sap response map and the white sap response map.

[0105] In a specific embodiment, step S21 specifically includes:

[0106] Based on the preset structural kernel window, the cosine distance between the center pixel and its neighboring pixels within the preset structural kernel window is calculated, and the normalized Euclidean distance between the center pixel and its neighboring pixels on the biochemical content feature vector within the preset structural kernel window is also calculated. A Gaussian function is then used to map the cosine distance and the normalized Euclidean distance into joint weight coefficients. The calculation expression is as follows:

[0107] ;

[0108] in, These are the joint weighting coefficients. Center pixel With neighboring pixels The cosine distance between them For normalized Euclidean distance, and These are the spectral bandwidth and the content bandwidth, respectively.

[0109] The spatial spectral features are obtained by summing and averaging the biochemical content feature vectors of all pixels within the preset structural kernel window according to the joint weight coefficients.

[0110] Specifically, the preset structure kernel window is Pixel matrix.

[0111] Calculating the cosine distance between the center pixel and its neighboring pixels within the preset kernel window specifically includes:

[0112] Extract the observation vector of the center pixel of the preset structural kernel window and the observation vector of any neighboring pixel, and calculate the dot product and magnitude of these two vectors. Divide the dot product by the product of the magnitudes to obtain the cosine similarity. Finally, subtract this similarity from 1 to obtain the cosine distance. This distance measures the difference in spectral curve shape between the center pixel and its neighboring pixels. The smaller the cosine distance, the more consistent the spectral characteristics of the two pixels, and the higher the probability that they belong to the same type of substance.

[0113] The calculation of the normalized Euclidean distance between the center pixel and its neighboring pixels within the preset structural kernel window on the biochemical content feature vector specifically includes:

[0114] The biochemical content feature vector of the center pixel is subtracted from the biochemical content feature vectors of its neighboring pixels in each feature dimension, and the difference is squared. The square root of the sum of the squared differences in all dimensions is then taken to obtain the original Euclidean distance. This original Euclidean distance is then scaled by the maximum value of the Euclidean distances of all neighboring pixels within a preset structural kernel window to obtain the normalized Euclidean distance. This distance reflects the absolute difference in biochemical composition between the center pixel and its neighboring pixels; the smaller the distance, the closer their biochemical structures are.

[0115] The process of summing and averaging the biochemical content feature vectors of all pixels within the preset structural kernel window according to the joint weight coefficients specifically includes:

[0116] The system multiplies the biochemical content feature vector of each pixel within the preset structural kernel window with its corresponding joint weight coefficient to obtain the weighted feature vector of that pixel. The magnitude of the joint weight coefficient determines the contribution of the neighboring pixels to the feature correction of the central pixel: pixels highly similar to the central pixel in spectral shape and biochemical composition are given higher weights, and their feature information is largely preserved; conversely, pixels with large differences have weights close to zero, and their feature information is suppressed. Subsequently, the system sums the weighted feature vectors of all pixels within the window, and divides the sum by the sum of all joint weight coefficients within the window to complete the normalized average calculation.

[0117] Spectral bandwidth controls the sensitivity of weights to differences in spectral shape. When the cosine distance is constant, a smaller spectral bandwidth causes the weights to decay rapidly, enhancing the ability to distinguish subtle spectral changes and helping to extract purer structural features.

[0118] Content bandwidth is used to adjust the tolerance of weights to differences in biochemical components. Since there are natural physiological fluctuations in the biochemical content between different pixels, the content bandwidth setting allows for a high weighting coefficient to be maintained even with a certain range of content deviations.

[0119] In a specific embodiment, the specific processing logic for generating the background autocorrelation matrix in step S22 includes:

[0120] The characteristic mean values ​​of wax scattering intercept, polyphenol absorption depth and component coupling ratio of the standard hyperspectral of white pine sample and green pine sample were calculated respectively, and used as the target guiding vectors for white pine and green pine.

[0121] Initial energy detection is performed based on the white and green target guiding vectors to identify potential target regions and generate target masks. Biochemical content feature vectors of the background region within the target mask neighborhood are extracted. The arithmetic mean of the biochemical content feature vectors of the background region is calculated by averaging the biochemical content feature vectors of the background region. The original feature values ​​in the target mask coverage area are replaced with the biochemical content mean vector to obtain a pseudo-background image. The covariance distribution is calculated using the spatial spectral features of the pseudo-background image to generate the background autocorrelation matrix.

[0122] Specifically, a sample set including standard hyperspectral data of *Phyllostachys edulis* and *Phyllostachys pubescens* samples is obtained. Based on this sample set, the mean values ​​of features for *Phyllostachys edulis* and *Phyllostachys pubescens* in the dimensions of wax scattering intercept, polyphenol absorption depth, and component coupling ratio are calculated respectively, thereby constructing target guiding vectors for *Phyllostachys edulis* and *Phyllostachys pubescens*. Subsequently, the target guiding vectors are used to perform initial energy detection on the entire image to identify candidate pixel regions that are highly similar to the biochemical characteristics of the target tree species, generating target mask regions. Within the coverage area of ​​the target mask, the biochemical content feature vectors of background pixels in the neighborhood of the target mask are extracted. The mean biochemical content vector of the background region is calculated by arithmetic mean, and the original feature values ​​in the mask coverage area are replaced with the mean biochemical content vector to construct a pseudo-background image without the target. The spatial spectral feature vectors corresponding to each pixel position in the pseudo-background image are extracted, and their covariance distribution in the spatial and feature dimensions is calculated to generate a background autocorrelation matrix reflecting the environmental noise characteristics of the current forest stand, thereby providing accurate background statistical constraints for solving the filter coefficients under the minimum output energy criterion.

[0123] High similarity refers to calculating the cosine similarity between the biochemical content feature vector of a pixel and the target guiding vector, and determining whether it is greater than the initial screening threshold;

[0124] If the cosine similarity is greater than the initial screening threshold, the pixel is determined to have a high degree of consistency with the target tree species in the biochemical feature space, and is thus identified as a potential target pixel.

[0125] The initial screening threshold is a pre-set value based on the statistical characteristics of the distribution of target tree species in the sample set.

[0126] The main purpose of the pseudo-background image is to eliminate the interference of the target tree species signal on the background statistical features and prevent the target self-inhibition effect during the background autocorrelation matrix solution. By extracting the mean vector of biochemical content in the background region within the target mask neighborhood for filling, this invention can construct a covariance distribution that only contains the statistical characteristics of background noise. This ensures that the filter coefficient vector, while minimizing energy output, can suppress only background interference, thereby maximizing the feature contrast between the target and the background during the polarization response stage and improving the recognition accuracy of closely related tree species in complex backgrounds.

[0127] In a specific embodiment, step S23 specifically includes:

[0128] Based on the criterion of minimizing output energy, the filter coefficient vector is solved, and the calculation expression is as follows:

[0129] ;

[0130] in, Let m be the filter coefficient vector of the m-th target tree species. Guide vector for the target tree species. Let R be the transpose of the target tree species' guiding vector, and let R be the background autocorrelation matrix. It is the inverse of the background autocorrelation matrix. To normalize the denominator, Index for target tree species categories, When =1, it represents white. When =2, it represents green moss;

[0131] The initial response score is calculated using the filter coefficient vector and the background autocorrelation matrix. The calculation expression is as follows:

[0132] ;

[0133] in, Let be the transpose of the filter coefficient vector for the m-th target tree species. Let x be the spatial spectral feature vector of pixel x. The initial response score;

[0134] The prior probability factor of the content is calculated using the mean vector of the content and the biochemical content feature vector of the target tree species. The calculation expression is as follows:

[0135] ;

[0136] in, The content prior probability factor, The biochemical content feature vector of pixel x, Let be the vector of the average content of the m-th target tree species. Let be the inverse matrix of the covariance matrix of the content of the m-th target tree species;

[0137] The response polarization score is calculated by combining the filter coefficient vector and the prior probability factor of the content. The calculation expression is as follows:

[0138] ;

[0139] in, In response to polarization scores, The effective range for content;

[0140] Obtain the response polarization score map based on the response polarization score corresponding to each pixel.

[0141] Specifically, under the constraint of minimizing output energy, this scheme maps spatial spectral features to the direction of the target tree species guide vector, which can adaptively suppress complex background noise such as shadows, bare ground and weeds in the forest area. By calculating the initial response score, it maximizes the numerical difference between white and green trees and the background environment, ensuring the sensitivity of the detection operator in low signal-to-noise ratio environments.

[0142] Traditional spectral identification is susceptible to spectral overlap caused by environmental fluctuations. This invention introduces a prior probability factor for content, combining the mean vector and covariance matrix of content from a dedicated biochemical content fingerprint database to calculate the initial response. Even when spectral features are highly similar, if the biochemical composition of a pixel does not conform to the physiological fingerprint characteristics of *Picea spp.* or *Picea acutissima*, its score will be significantly reduced, effectively solving the common technical challenge of distinguishing between different spruce species.

[0143] This invention establishes a stepped response polarization score calculation logic. It utilizes hard gating to remove outlier pixels within the effective interval and soft correction using prior probability to enhance target confidence. This results in the final generated response polarization score map where the target region's score is pushed towards 1, while non-target regions are suppressed to near 0.

[0144] In a specific embodiment, step S3, the processing logic for generating the discriminant distance matrix specifically includes:

[0145] The response polarization score map and the biochemical content feature vector are normalized, and the normalized response polarization score map and the biochemical content feature vector are fused to obtain the feature point cloud of white and green foliage.

[0146] Searching for each pixel in the feature point cloud of white and green haze using Euclidean distance. 1. ... 1. Neighboring points, construct a second order The set of nearest neighbor representative points;

[0147] Statistical analysis of shared second order of any adjacent pixel pairs The number of nearest neighbors is calculated, and the Mahalanobis distance between adjacent pixel pairs on the biochemical content feature vector is used as the content deviation.

[0148] Will share second order The nearest neighbor count plus 1 is used as the denominator, and the content deviation is used as the numerator to perform a decay operation, generating a discrimination distance matrix. The calculation expression is as follows:

[0149] ;

[0150] in, To determine the distance matrix, The deviation in content between pixel i and pixel j. Let be the inverse matrix of the covariance matrix of the content of the m-th target tree species. To share the number of second-order nearest neighbors, and These are the biochemical content feature vectors of pixels i and j, respectively.

[0151] Specifically, this scheme constructs a discrimination mechanism that accelerates intra-class aggregation and forces inter-class blocking. For pixel pairs belonging to the same tree species, their content deviation is small and they share many nearest neighbors, resulting in a very small discrimination distance, which is conducive to forming clear density peaks. However, for edge pixels located at the boundary between white and green trees, even if the spatial distance is close, the discrimination distance will increase exponentially due to the large deviation in biochemical content and the sharp decrease in the number of shared nearest neighbors. This mechanism forms a clear inter-class gap in the feature space, solving the classification ambiguity problem caused by different spectra of the same species and the same spectra of different species.

[0152] In a specific embodiment, step S3, which involves performing category assignment and spatial consistency correction based on the discriminant distance matrix, specifically includes:

[0153] By discriminating the distance matrix, the local density and relative distance of each pixel are calculated, and the weight index is obtained by multiplying the local density and relative distance.

[0154] Specifically, calculating the local density and relative distance of each pixel using the discriminant distance matrix includes:

[0155] By discriminating the distance matrix, the number of neighboring pixels of a pixel within a preset cutoff distance range is counted. The higher the local density, the more likely the pixel is located in the core distribution area of ​​that tree species in the biochemical feature space.

[0156] Calculate the minimum discriminant distance between a pixel and all pixels with a local density higher than its own. If the local density of a pixel is the global maximum, then the relative distance is defined as the discriminant distance between it and the farthest pixel.

[0157] The preset cutoff distance range is the one that ranks first in the total. The distance value at that location.

[0158] The pixels with the largest weight index, whose response polarization score in the green pine response map is greater than the first high value threshold, and whose biochemical content feature vector is located within the effective content range of the target tree species are selected as the initial green pine class center.

[0159] The initial white elm class center was selected based on the largest weight index, the response polarization score in the white elm response map being greater than the second highest threshold, and the biochemical content feature vector being located within the effective content range of the target tree species.

[0160] After normalizing based on the number of shared second-order nearest neighbors, the shared nearest neighbor similarity is obtained. A joint similarity is then constructed, and its calculation expression is as follows:

[0161] ;

[0162] in, For joint similarity, To share the similarity of nearest neighbors, The content prior probability factor, In response to polarization scores, As the first preset weight, As the second preset weight, The third preset weight;

[0163] Specifically, the sum of the first preset weight, the second preset weight, and the third preset weight is 1, where the first preset weight is 0.3, the second preset weight is 0.5, and the third preset weight is 0.2.

[0164] The first preset weight is used for spatial topological constraints, which can correct the single-point discrimination result by utilizing the consistency of neighboring pixels. This can effectively suppress noise in the classification map, making the recognition result more consistent with the actual distribution pattern of the forest stand in geographic space.

[0165] The second preset weight is assigned a value of 0.5 to ensure that classification decisions are always guided by physiological mechanisms. This effectively solves the problem of "different spectra for the same organism," and even when uneven illumination causes spectral distortion, it can still make accurate judgments based on biochemical properties, thus improving the biological reliability of classification.

[0166] The third preset weight is used to further widen the score gap between the target tree species and the background in complex environments. Although its weight is relatively small, when coupled with the second preset weight, it plays a crucial role in background filtering, improving the signal-to-noise ratio of the algorithm in heterogeneous environments.

[0167] Based on the maximum similarity criterion, each pixel to be classified is assigned to the category of the class center with the highest joint similarity and given the corresponding initial category label;

[0168] A spatial window is established centered on each pixel to be classified. The initial class labels of all neighboring pixels within the spatial window are counted. It is then determined whether the initial class label of the central pixel is consistent with the initial class label that appears most frequently among the neighboring pixels.

[0169] If there is a discrepancy, and the proportion of the most frequently occurring initial category label exceeds the preset consistency threshold, then the initial category label of the center pixel is replaced with the most frequently occurring initial category label according to the local voting principle to obtain the final category label;

[0170] The final category labels of all the pixels are spatially mapped according to the original geographic coordinates to output a classification map of white and green irises.

[0171] Specifically, when the spatial window moves to a certain position in the image, the pixel to be classified that is at the geometric center of the window is defined as the center pixel.

[0172] Example 2, refer to Figure 2 This is another embodiment of the present invention, which differs from the first embodiment in that it provides a forestry tree species classification and identification system based on image recognition, including a feature extraction module, a processing module, and a classification module:

[0173] The feature extraction module is used to process the hyperspectral image to be detected using the continuum removal method, extract spectral fingerprint parameters, lignin index and cellulose absorption index, calculate the component coupling ratio using the lignin index and cellulose absorption index, and calculate the biochemical structural heterogeneity characteristics based on the component coupling ratio to obtain the component coupling ratio distribution map and the biochemical structural heterogeneity map, which are combined with the spectral fingerprint parameters to obtain the biochemical content feature vector.

[0174] The processing module is used to smooth the biochemical content feature vector using an adaptive neighborhood structure kernel to obtain spatial spectral features. Based on the acquired sample set, the target mask region is identified and pseudo-background spectral filling is performed to generate a background autocorrelation matrix. Through the background autocorrelation matrix and spatial spectral features, the filter coefficient vector is solved using the minimum output energy criterion. Combined with a preset exclusive content fingerprint database, the response polarization score map is output.

[0175] The classification module is used to fuse the response polarization score map with the biochemical content feature vector to obtain the feature point cloud of white and green foliage. This feature point cloud is then combined with second-order... The nearest neighbor algorithm generates a discriminant distance matrix, and performs category assignment and spatial consistency correction based on the discriminant distance matrix, outputting a white and green classification map.

[0176] This invention addresses the challenge of spectral overlap among tree species, which leads to heterogeneous species exhibiting similar spectra. Instead of relying on fluctuating apparent spectra, it extracts fingerprints of wax and polyphenol content that are sensitive to physiological differences, establishing a discrimination criterion from a microscopic biochemical perspective, significantly reducing the false positive rate. Through the synergistic constraint of component coupling ratios and effective content ranges, this scheme uses ratio calculations to offset ambient light interference and employs physiological constants as hard gating to automatically remove background noise, enhancing the algorithm's robustness in complex forest stands. Particularly noteworthy is its innovative construction of a joint biochemical consistency adaptive structure kernel, performing smoothing only between pixels with similar components, thus suppressing canopy texture noise while preserving accurate tree species distribution boundaries. By sharing a second-order nearest neighbor topology, content features are mapped to a high-dimensional point cloud space. A discriminative distance matrix is ​​used to achieve category assignment based on a biochemical manifold, and spatial consistency correction addresses the fragmentation problem caused by traditional pixel-by-pixel classification, outputting a classification map of *Phyllostachys edulis* and *Phyllostachys aurea* that combines biochemical mechanisms and spatial semantics.

[0177] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0178] It should be noted that the above embodiments are only used to illustrate the technical solutions 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 solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for classifying and identifying forestry tree species based on image recognition, characterized in that, Includes the following steps: Step S1: Process the hyperspectral image to be detected using the continuum removal method, extract spectral fingerprint parameters, lignin index and cellulose absorption index, calculate the component coupling ratio using the lignin index and cellulose absorption index, and calculate the biochemical structural heterogeneity characteristics based on the component coupling ratio to obtain the component coupling ratio distribution map and the biochemical structural heterogeneity map, and combine them with the spectral fingerprint parameters to obtain the biochemical content feature vector. The hyperspectral image to be detected is acquired, and the hyperspectral image to be detected is normalized using the continuum removal method to obtain the spectral curve; The wax scattering intercept at 450 nm and the polyphenol absorption depth at 1660 nm were extracted from the spectral curve to obtain the spectral fingerprint parameters. At the 450nm band of the spectral curve, the reflectance intercept is extracted and used as the wax scattering intercept. At the 1660 nm absorption valley of the spectral curve, the polyphenol absorption depth at the 1660 nm absorption valley is calculated using the following expression: ； in, The depth of polyphenol absorption. Reflectivity; Based on spectral fingerprint parameters, the lignin index in the 1750 nm band and the cellulose absorption index in the 2100 nm band were calculated and normalized. The component coupling ratio was calculated by normalizing the lignin index and the cellulose absorption index, and the component coupling ratio distribution map was obtained. The lignin index was calculated using the reflectance at 1750 nm. The cellulose absorption index was calculated using the absorption characteristics in the 2100nm band. The formula for calculating the component coupling ratio is: ； in, The component coupling ratio, The normalized lignin index, Normalized cellulose absorption index; A preset sliding window is used to traverse the component coupling ratio distribution map, and the biochemical structural heterogeneity map is obtained by statistically analyzing the biochemical structural heterogeneity characteristics of the component content within each sliding window. Using the preset size The sliding window is used to calculate the maximum rate of change of the component coupling ratio between the center pixel and the neighboring pixels within the sliding window as the spatial gradient. The ratio of the standard deviation to the mean of the component coupling ratio within the sliding window is calculated as the discrete coefficient. The probability distribution density of the component coupling ratio within the sliding window is statistically analyzed. The spatial gradient, discrete coefficient, and information entropy are stacked as matrices in the vertical direction. The biochemical structural heterogeneity features include spatial gradient, discrete coefficient, and information entropy; Step S2: The biochemical content feature vector is smoothed using an adaptive neighborhood structure kernel to obtain spatial spectral features. Based on the acquired sample set, the target mask region is identified and pseudo-background spectral filling is performed to generate a background autocorrelation matrix. The filter coefficient vector is solved using the minimum output energy criterion through the background autocorrelation matrix and spatial spectral features. The response polarization score map is output in combination with the preset exclusive content fingerprint database. The preset exclusive content fingerprint database includes the content mean vector of the target tree species, the content covariance matrix of the target tree species, and the effective content range of the target tree species. Based on the preset structural kernel window, the cosine distance between the center pixel and its neighboring pixels within the preset structural kernel window is calculated, and the normalized Euclidean distance between the center pixel and its neighboring pixels on the biochemical content feature vector within the preset structural kernel window is also calculated. A Gaussian function is then used to map the cosine distance and the normalized Euclidean distance into joint weight coefficients. The calculation expression is as follows: ； in, These are the joint weighting coefficients. Center pixel With neighboring pixels The cosine distance between them For normalized Euclidean distance, and These are the spectral bandwidth and the content bandwidth, respectively. The spatial spectral features are obtained by summing and averaging the biochemical content feature vectors of all pixels within the preset structural kernel window according to the joint weight coefficients. The specific processing logic for generating the background autocorrelation matrix includes: The characteristic mean values ​​of wax scattering intercept, polyphenol absorption depth and component coupling ratio of the standard hyperspectral of white pine sample and green pine sample were calculated respectively, and used as the target guiding vectors for white pine and green pine. Initial energy detection is performed based on the white and green target guiding vectors to identify potential target regions and generate target masks. Biochemical content feature vectors of background regions within the target mask neighborhood are extracted. The arithmetic mean of biochemical content feature vectors of background regions is calculated by averaging the biochemical content feature vectors of background regions. The original feature values ​​within the target mask coverage area are replaced by the biochemical content mean vector to obtain a pseudo-background image. The covariance distribution is calculated using the spatial spectral features of the pseudo-background image to generate a background autocorrelation matrix. The prior probability factor of the content is calculated using the mean vector of the content and the biochemical content feature vector of the target tree species. The calculation expression is as follows: ； in, The content prior probability factor, The biochemical content feature vector of pixel x, Let be the vector of the average content of the m-th target tree species. Let be the inverse matrix of the covariance matrix of the content of the m-th target tree species; The initial response score is calculated using the filter coefficient vector and the background autocorrelation matrix. The calculation expression is as follows: ； in, Let be the transpose of the filter coefficient vector for the m-th target tree species. Let x be the spatial spectral feature vector of pixel x. The initial response score; The response polarization score is calculated by combining the filter coefficient vector and the prior probability factor of the content. The calculation expression is as follows: ； in, In response to polarization scores, Effective range of content Obtain the response polarization score map based on the response polarization score corresponding to each pixel; Step S3: The response polarization score map and the biochemical content feature vector are fused to obtain the feature point cloud of white and green foliage. The feature point cloud of white and green foliage is then combined with second-order... The nearest neighbor algorithm generates a discriminant distance matrix, and performs category assignment and spatial consistency correction based on the discriminant distance matrix, outputting a white and green classification map; The specific processing logic for generating the discriminant distance matrix includes: The response polarization score map and the biochemical content feature vector are normalized, and the normalized response polarization score map and the biochemical content feature vector are fused to obtain the feature point cloud of white and green foliage. Searching for each pixel in the feature point cloud of white and green haze using Euclidean distance.

1. ...

1. Neighboring points, construct a second order The set of nearest neighbor representative points; Statistical analysis of shared second order of any adjacent pixel pairs The number of nearest neighbors is calculated, and the Mahalanobis distance between adjacent pixel pairs on the biochemical content feature vector is used as the content deviation. Will share second order The nearest neighbor count plus 1 is used as the denominator, and the content deviation is used as the numerator to perform a decay operation, generating a discrimination distance matrix. The calculation expression is as follows: ； in, To determine the distance matrix, The deviation in content between pixel i and pixel j. Let be the inverse matrix of the covariance matrix of the content of the m-th target tree species. To share the number of second-order nearest neighbors, and These are the biochemical content feature vectors of pixels i and j, respectively.

2. The forestry tree species classification and identification method based on image recognition as described in claim 1, characterized in that, In step S1, the spectral fingerprint parameters, component coupling ratio distribution map and biochemical structural heterogeneity map are stacked into a matrix, and numerical mapping is performed using a preset exclusive content fingerprint library. The feature values ​​corresponding to the map at each pixel location are combined to form the biochemical content feature vector of that pixel.

3. The forestry tree species classification and identification method based on image recognition as described in claim 2, characterized in that, Step S2 specifically includes: Step S21: Smooth the biochemical content feature vector using an adaptive neighborhood structure kernel and extract spatial spectral features; Step S22: Obtain the sample set, calculate the white target guidance vector and the green target guidance vector based on the sample set, identify the target mask region and perform pseudo-background spectral filling, calculate the covariance distribution, and generate the background autocorrelation matrix; The sample set includes standard hyperspectral data of white sage and standard hyperspectral data of green sage; Step S23: Based on the minimum output energy criterion, the filter coefficient vector is solved using the background autocorrelation matrix, and the response polarization score map is obtained by combining the spatial spectral features and biochemical content feature vectors. The response polarization score map includes the green sap response map and the white sap response map.

4. The forestry tree species classification and identification method based on image recognition as described in claim 3, characterized in that, Based on the criterion of minimizing output energy, the filter coefficient vector is solved, and the calculation expression is as follows: ； in, Let m be the filter coefficient vector of the m-th target tree species. Guide vector for the target tree species. Let R be the transpose of the target tree species' guiding vector, and let R be the background autocorrelation matrix. It is the inverse of the background autocorrelation matrix. To normalize the denominator, Index for target tree species categories, When =1, it represents white. When =2, it represents green moss.

5. The forestry tree species classification and identification method based on image recognition as described in claim 4, characterized in that, In step S3, the specific steps of performing category assignment and spatial consistency correction based on the discriminant distance matrix include: By discriminating the distance matrix, the local density and relative distance of each pixel are calculated, and the weight index is obtained by multiplying the local density and relative distance. The pixels with the largest weight index, whose response polarization score in the green pine response map is greater than the first high value threshold, and whose biochemical content feature vector is located within the effective content range of the target tree species are selected as the initial green pine class center. The initial white elm class center was selected based on the largest weight index, the response polarization score in the white elm response map being greater than the second highest threshold, and the biochemical content feature vector being located within the effective content range of the target tree species. After normalizing based on the number of shared second-order nearest neighbors, the shared nearest neighbor similarity is obtained. A joint similarity is then constructed, and its calculation expression is as follows: ； in, For joint similarity, To share the similarity of nearest neighbors, The content prior probability factor, In response to polarization scores, As the first preset weight, As the second preset weight, The third preset weight; Based on the maximum similarity criterion, each pixel to be classified is assigned to the category of the class center with the highest joint similarity and given the corresponding initial category label; A spatial window is established centered on each pixel to be classified. The initial class labels of all neighboring pixels within the spatial window are counted. It is then determined whether the initial class label of the central pixel is consistent with the initial class label that appears most frequently among the neighboring pixels. If there is a discrepancy, and the proportion of the most frequently occurring initial category label exceeds the preset consistency threshold, then the initial category label of the center pixel is replaced with the most frequently occurring initial category label according to the local voting principle to obtain the final category label; The final category labels of all the pixels are spatially mapped according to the original geographic coordinates to output a classification map of white and green irises.

6. A forestry tree species classification and identification system based on image recognition, characterized in that, It includes a feature extraction module, a processing module, and a classification module: The feature extraction module is used to process the hyperspectral image to be detected using the continuum removal method, extract spectral fingerprint parameters, lignin index and cellulose absorption index, calculate the component coupling ratio using the lignin index and cellulose absorption index, and calculate the biochemical structural heterogeneity characteristics based on the component coupling ratio to obtain the component coupling ratio distribution map and the biochemical structural heterogeneity map, which are combined with the spectral fingerprint parameters to obtain the biochemical content feature vector. The hyperspectral image to be detected is acquired, and the hyperspectral image to be detected is normalized using the continuum removal method to obtain the spectral curve; The wax scattering intercept at 450 nm and the polyphenol absorption depth at 1660 nm were extracted from the spectral curve to obtain the spectral fingerprint parameters. At the 450nm band of the spectral curve, the reflectance intercept is extracted and used as the wax scattering intercept. At the 1660 nm absorption valley of the spectral curve, the polyphenol absorption depth at the 1660 nm absorption valley is calculated using the following expression: ； in, The depth of polyphenol absorption. Reflectivity; Based on spectral fingerprint parameters, the lignin index in the 1750 nm band and the cellulose absorption index in the 2100 nm band were calculated and normalized. The component coupling ratio was calculated by normalizing the lignin index and the cellulose absorption index, and the component coupling ratio distribution map was obtained. The lignin index was calculated using the reflectance at 1750 nm. The cellulose absorption index was calculated using the absorption characteristics in the 2100nm band. The formula for calculating the component coupling ratio is: ； in, The component coupling ratio, The normalized lignin index, Normalized cellulose absorption index; A preset sliding window is used to traverse the component coupling ratio distribution map, and the biochemical structural heterogeneity map is obtained by statistically analyzing the biochemical structural heterogeneity characteristics of the component content within each sliding window. Using the preset size The sliding window is used to calculate the maximum rate of change of the component coupling ratio between the center pixel and the neighboring pixels within the sliding window as the spatial gradient. The ratio of the standard deviation to the mean of the component coupling ratio within the sliding window is calculated as the discrete coefficient. The probability distribution density of the component coupling ratio within the sliding window is statistically analyzed. The spatial gradient, discrete coefficient, and information entropy are stacked as matrices in the vertical direction. The biochemical structural heterogeneity features include spatial gradient, discrete coefficient, and information entropy; The processing module is used to smooth the biochemical content feature vector using an adaptive neighborhood structure kernel to obtain spatial spectral features. Based on the acquired sample set, the target mask region is identified and pseudo-background spectral filling is performed to generate a background autocorrelation matrix. Through the background autocorrelation matrix and spatial spectral features, the filter coefficient vector is solved using the minimum output energy criterion. Combined with a preset exclusive content fingerprint database, the response polarization score map is output. The preset exclusive content fingerprint database includes the content mean vector of the target tree species, the content covariance matrix of the target tree species, and the effective content range of the target tree species. Based on the preset structural kernel window, the cosine distance between the center pixel and its neighboring pixels within the preset structural kernel window is calculated, and the normalized Euclidean distance between the center pixel and its neighboring pixels on the biochemical content feature vector within the preset structural kernel window is also calculated. A Gaussian function is then used to map the cosine distance and the normalized Euclidean distance into joint weight coefficients. The calculation expression is as follows: ； in, These are the joint weighting coefficients. Center pixel With neighboring pixels The cosine distance between them For normalized Euclidean distance, and These are the spectral bandwidth and the content bandwidth, respectively. The spatial spectral features are obtained by summing and averaging the biochemical content feature vectors of all pixels within the preset structural kernel window according to the joint weight coefficients. The specific processing logic for generating the background autocorrelation matrix includes: The characteristic mean values ​​of wax scattering intercept, polyphenol absorption depth and component coupling ratio of the standard hyperspectral of white pine sample and green pine sample were calculated respectively, and used as the target guiding vectors for white pine and green pine. Initial energy detection is performed based on the white and green target guiding vectors to identify potential target regions and generate target masks. Biochemical content feature vectors of background regions within the target mask neighborhood are extracted. The arithmetic mean of biochemical content feature vectors of background regions is calculated by averaging the biochemical content feature vectors of background regions. The original feature values ​​within the target mask coverage area are replaced by the biochemical content mean vector to obtain a pseudo-background image. The covariance distribution is calculated using the spatial spectral features of the pseudo-background image to generate a background autocorrelation matrix. The prior probability factor of the content is calculated using the mean vector of the content and the biochemical content feature vector of the target tree species. The calculation expression is as follows: ； in, The content prior probability factor, The biochemical content feature vector of pixel x, Let be the vector of the average content of the m-th target tree species. Let be the inverse matrix of the covariance matrix of the content of the m-th target tree species; The initial response score is calculated using the filter coefficient vector and the background autocorrelation matrix. The calculation expression is as follows: ； in, Let be the transpose of the filter coefficient vector for the m-th target tree species. Let x be the spatial spectral feature vector of pixel x. The initial response score; The response polarization score is calculated by combining the filter coefficient vector and the prior probability factor of the content. The calculation expression is as follows: ； in, In response to polarization scores, The effective range for content; Obtain the response polarization score map based on the response polarization score corresponding to each pixel; The classification module is used to fuse the response polarization score map with the biochemical content feature vector to obtain the feature point cloud of white and green foliage. This feature point cloud is then combined with second-order... The nearest neighbor algorithm generates a discriminant distance matrix, and performs category assignment and spatial consistency correction based on the discriminant distance matrix, outputting a white and green classification map; The specific processing logic for generating the discriminant distance matrix includes: The response polarization score map and the biochemical content feature vector are normalized, and the normalized response polarization score map and the biochemical content feature vector are fused to obtain the feature point cloud of white and green foliage. Searching for each pixel in the feature point cloud of white and green haze using Euclidean distance.

1. ...

1. Neighboring points, construct a second order The set of nearest neighbor representative points; Statistical analysis of shared second order of any adjacent pixel pairs The number of nearest neighbors is calculated, and the Mahalanobis distance between adjacent pixel pairs on the biochemical content feature vector is used as the content deviation. Will share second order The nearest neighbor count plus 1 is used as the denominator, and the content deviation is used as the numerator to perform a decay operation, generating a discrimination distance matrix. The calculation expression is as follows: ； in, To determine the distance matrix, The deviation in content between pixel i and pixel j. Let be the inverse matrix of the covariance matrix of the content of the m-th target tree species. To share the number of second-order nearest neighbors, and These are the biochemical content feature vectors of pixels i and j, respectively.

Images