A tree age recognition system and method based on leaf hyperspectral data

Abstract

A tree age recognition system and method based on leaf hyperspectral data belong to the technical field of forestry investigation. In order to solve the problem of non-invasive tree age recognition, the present application constructs a tree branch segmentation module to segment the pixel spectral features of the tree crown leaf hyperspectral image; an abnormal sample elimination module is constructed to eliminate abnormal samples from the leaf hyperspectral image data; a feature extraction module is constructed to extract features from the pure leaf hyperspectral image data, first calculate the mutual information of the features and the true value of the tree age, then use the mutual information combined with the partial least squares method to screen the features, input the screened features into the random forest regression model for further extraction, and obtain the core features of the pure leaf; a leaf age recognition model module is constructed to recognize the tree age by using the constructed leaf age recognition model for the core features of the pure leaf, and obtain the tree age prediction value. The present application realizes non-destructive detection of tree age.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of forestry investigation, and particularly relates to a tree age identification system and method based on leaf hyperspectral data. BACKGROUND

[0002] Existing tree age identification methods are mostly lossy and contact type identification methods, which directly or indirectly identify tree age through tree rings, such as C14 identification, drilling wood to obtain tree core to identify tree rings, tree disc to identify tree rings, CT scanning tree rings, etc. These methods cannot obtain the tree age of each tree in the forest from above the forest canopy in a large range, for a long time, with high efficiency and without damage. SUMMARY

[0003] The problem to be solved by the present application is to identify tree age without damage, and a tree age identification system and method based on leaf hyperspectral data are provided.

[0004] To achieve the above-mentioned purpose, the present application realizes the following technical scheme:

[0005] A tree age identification method based on leaf hyperspectral data, comprising the following steps:

[0006] S1. A hyperspectral camera is used to collect a tree crown leaf hyperspectral image, and pixel spectral features of the tree crown leaf hyperspectral image are obtained after preprocessing;

[0007] S2. A tree leaf branch segmentation module is constructed, the pixel spectral features of the tree crown leaf hyperspectral image obtained in step S1 are subjected to tree leaf branch segmentation, a support vector machine method is used to define a classification rule, and target optimization is performed, the pixel spectral features of the tree crown leaf hyperspectral image are subjected to mask cropping, and leaf hyperspectral image data is obtained;

[0008] S3. An abnormal sample elimination module is constructed, the leaf hyperspectral image data obtained in step S2 is subjected to abnormal sample elimination, an isolated forest method is used to calculate the abnormal score of the pixels in the leaf hyperspectral image, abnormal pixels are eliminated, and pure leaf hyperspectral image data is obtained;

[0009] S4. A feature extraction module is constructed, the pure leaf hyperspectral image data obtained in step S3 is subjected to feature extraction, mutual information between the features and the true value of the tree age is calculated first, then the mutual information is combined with a partial least squares method for feature screening, the screened features are input into a random forest regression model for further extraction, and core features of the pure leaf are obtained;

[0010] S5. A leaf identification tree age model module is constructed, the core features of the pure leaf obtained in step S4 are subjected to tree age identification by using the constructed leaf identification tree age model, and a tree age prediction value is obtained.

[0011] Further, the pre-processing process in step S1 includes spectral scanning, black and white balance correction, geometric correction, and obtaining the pixel spectral features of the tree crown leaf hyperspectral image as follows: ,

[0012] wherein, represents the spectral vector of the mth row, nth column pixel in the hyperspectral image, that is, the combination of the reflectivity of the pixel at all wavebands, used to represent the spectral features of the material at this position being a leaf or a branch; represents the reflectivity value of the mth row, nth column pixel at the wavelength band, ; K represents the total number of wavebands of the hyperspectral image; represents the K-dimensional real number space in which the spectral vector is located, that is, the spectral features of each pixel can be regarded as a point in the K-dimensional space.

[0013] Further, step S2 is based on the spectral difference between leaves and branches being linearly separable, and a support vector machine method is used to define the classification rule;

[0014] The hyperplane in the K-dimensional space is set as follows:

[0015]

[0016] wherein, , is the weight vector of the hyperplane, representing the importance of each waveband to to the classification;is the bias term, used to adjust the position of the hyperplane in the K-dimensional space; the classification decision rule is defined as follows:

[0017]

[0018] wherein, is the predicted class of the mth row, nth column pixel, wherein when is determined as a leaf, is determined as a branch; then the optimal function is used to segment all sample pixel points, thereby obtaining the mask of the leaves, and the hyperspectral image is cropped according to the mask to obtain the hyperspectral image data of the leaves.

[0019] Further, the specific implementation method of step S3 includes the following steps:

[0020] S3.1. Set the pixel set of the leaf hyperspectral image obtained in step S2 wherein, represents the spectral vector of the pth pixel in the leaf area, ;

[0021] S3.2. The constructed isolation forest is composed of individual isolation trees, denoted as , where is the total number of isolation trees, determined by cross-validation, taking 100-200; the determination threshold of the anomaly score is denoted as , the optimal threshold based on 500+ group sample statistics, adjusted according to tree species;

[0022] S3.3. For each pixel spectral vector in step S3.1 , the anomaly score is calculated using the path length algorithm of the isolation forest , is the anomaly score function, taking the value range ;

[0023] According to the anomaly score, if , it is determined that the pixel is an abnormal pixel and is removed; if , the pixel is retained, and finally the pure leaf hyperspectral image data after removing the abnormal pixels is obtained :

[0024] .

[0025] Further, the specific implementation method of step S4 includes the following steps:

[0026] S4.1. Calculate the mutual information of each hyperspectral band reflectance in , which is used to measure the linear / nonlinear correlation of two variables, denoted as , where represents the true value of the sample age; according to the descending order of the mutual information value, select the top bands with the largest mutual information to form the feature matrix after mutual information screening :

[0027]

[0028] where, represents the reflectance feature vector of the i-th hyperspectral band, and when the total number of samples is Q, , , is the number of features;

[0029] S4.2. Based on the obtained in step S4.1, a partial least squares regression PLS model is constructed as follows:

[0030]

[0031] in, , This is the weight matrix of the PLS model, representing the degree of influence of each screening feature on tree age prediction; , The residual matrix of the model reflects the deviation between the model's predicted values ​​and the actual values, and satisfies... ; Calculate the determination coefficient of the PLS model Used to measure the goodness of fit of a model, with a value range of The formula is as follows:

[0032]

[0033] in, Represents the total number of samples. For the first The true tree age of each sample. For use Constructing the PLS model for the first Predicted tree age for each sample The sample mean of the true tree age;

[0034] Curvature of the determination coefficients of the PLS model The formula is as follows:

[0035]

[0036] in, for Number of features The first derivative, for Number of features The second derivative; the number of optimal features For curvature The value of the number of features corresponding to the maximum value, i.e. Finally selected The optimized feature matrix consists of several features. ;

[0037] S4.3. Based on the results obtained in step S4.2 The data is fed into a random forest regression model for deep feature extraction: through feature splitting and voting ensemble of multiple decision trees, the key features that contribute most to tree age identification are selected, and finally the core feature matrix of pure leaves is obtained. .

[0038] Furthermore, the specific implementation method of step S5 includes the following steps:

[0039] S5.1. Core feature matrix of pure leaves obtained in step S4 , build an integrated leaf node recognition age model; through round iterative training to generate decision trees, the final age prediction value of the model for the first sample is obtained by weighted average of the prediction values of the decision trees, and the prediction function is as follows:

[0040]

[0041] , wherein , is the core feature vector of the first sample, is the prediction function of the first decision tree, is the total number of integrated decision trees, is the integrated prediction function of the model, is the function space formed by all possible decision trees;

[0042] S5.2. To avoid model overfitting, a target function with L2 regularization is constructed, and a decision tree is learned by minimizing the target function in each round of iteration , and the formula of the target function is as follows:

[0043]

[0044] , wherein is the model parameter of the first round of iteration, is the mean square loss function; is the cumulative prediction value of the first sample by the previous decision trees, is the regularization term of the decision tree ;

[0045]

[0046]

[0047] , wherein is the leaf node number regularization hyperparameter, is the total number of leaf nodes of the decision tree , is the weight regularization hyperparameter, is the output weight of the first leaf node;

[0048] S5.3. To simplify the optimization process of the target function, the in step S5.2 is replaced by ​​The second-order Taylor expansion is performed at the t-th iteration and high-order small items are ignored to obtain an approximate optimization objective of the t-th iteration :

[0049]

[0050] wherein, is a loss function The first-order gradient statistics of , is the second-order gradient statistics, and is defined as follows:

[0051]

[0052] ;

[0053] For the t-th leaf node of the decision tree , the calculation formula of the optimal output weight is as follows:

[0054] wherein,

[0055] is the sample index set belonging to the t-th leaf node; the split feature and the split threshold of the decision tree are determined by maximizing the gain , and the gain calculation formula is as follows:

[0056]

[0057] wherein, is the sample set before splitting, and are the sample sets of the left and right child nodes after splitting respectively.

[0058] A tree age identification system based on leaf hyperspectral data, comprising a leaf branch segmentation module, an abnormal sample elimination module, a feature extraction module, and a leaf identification tree age model module, wherein the leaf branch segmentation module is sequentially connected to the abnormal sample elimination module, the feature extraction module, and the leaf identification tree age model module.

[0059] The leaf branch segmentation module: the input is a preprocessed tree crown hyperspectral image, a leaf mask is generated through SVM classification, the image is masked and cropped, and the output is leaf hyperspectral image data ;

[0060] The abnormal sample elimination module: the input is , an anomaly score is calculated through an isolation forest, and abnormal pixels are eliminated, and the output is pure leaf hyperspectral image data ;​​

[0061] The feature extraction module: input is , first through mutual information screening features to get , then through random forest depth extraction, output pure leaf core feature matrix ;

[0062] The leaf recognition age model module: input is , through GBDT integrated model to predict the age, output age prediction value .

[0063] The beneficial effects of the application are:

[0064] The tree age recognition method based on leaf hyperspectral data can recognize the tree age non-destructively, quickly and non-contact from the tree canopy by collecting the leaf reflectance hyperspectral image. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 The flow chart of the tree age recognition method based on leaf hyperspectral data. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application, that is, the described specific embodiments are only a part of the embodiments of the present application, but not all the specific embodiments. The components of the specific embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations, and the present application can also have other embodiments.

[0067] Therefore, the detailed description of the specific embodiments of the present application provided in the drawings below is not intended to limit the scope of the claimed present application, but only represents selected specific embodiments of the present application. Based on the specific embodiments of the present application, all other specific embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0068] In order to further understand the invention content, characteristics and effects of the present application, the following specific embodiments are exemplified and described in detail as follows in combination with the drawings:

[0069] Example 1:

[0070] A tree age recognition method based on leaf hyperspectral data, comprising the following steps:

[0071] S1. Collecting tree crown leaf hyperspectral images by using a hyperspectral camera, and obtaining pixel spectral features of the tree crown leaf hyperspectral images after preprocessing;

[0072] Further, the preprocessing process in step S1 includes spectral scanning, black and white balance correction, and geometric correction, and the pixel spectral features of the tree crown leaf hyperspectral images are as follows: ,

[0073] Wherein, represents the spectral vector of the mth row and nth column pixel in the hyperspectral image, that is, the combination of reflectivity of the pixel at all wavebands, used to represent the spectral features of the material at this position being a leaf or a branch; represents the reflectivity value of the mth row and nth column pixel at the wavelength band, ; K represents the total number of wavebands of the hyperspectral image; represents the K-dimensional real number space in which the spectral vector is located, that is, the spectral features of each pixel can be regarded as a point in the K-dimensional space.

[0074] S2. Constructing a leaf and branch segmentation module, performing leaf and branch segmentation on the pixel spectral features of the tree crown leaf hyperspectral images obtained in step S1, defining a classification rule by using a support vector machine method, and performing target optimization, and performing mask cropping on the pixel spectral features of the tree crown leaf hyperspectral images to obtain leaf hyperspectral image data;

[0075] Further, step S2 regards the spectral difference between leaves and branches as linearly separable, and defines a classification rule by using a support vector machine method;

[0076] The hyperplane in the K-dimensional space is set as follows:

[0077]

[0078] Wherein, is the weight vector of the hyperplane, representing the importance of each waveband to classification; is the bias term, used to adjust the position of the hyperplane in the K-dimensional space; and the classification decision rule is defined as follows:

[0079]

[0080] Wherein, is the predicted category of the mth row and nth column pixel, wherein when , it is determined as a leaf. The tree branch is determined, and the pixel points of all samples are segmented based on an optimal function to obtain a mask of the tree leaves. The hyperspectral image is cropped according to the mask to obtain the hyperspectral image data of the tree leaves.

[0081] S3. An abnormal sample elimination module is constructed to eliminate abnormal samples from the hyperspectral image data of the tree leaves obtained in step S2. An isolation forest method is used to calculate the abnormal score of each pixel in the hyperspectral image of the tree leaves, and abnormal pixels are eliminated to obtain the hyperspectral image data of pure tree leaves.

[0082] Further, the specific implementation method of step S3 includes the following steps:

[0083] S3.1. A pixel set of the hyperspectral image of the tree leaves obtained in step S2 is set , wherein represents the spectral vector of the pth pixel in the tree leaf region, .

[0084] S3.2. The isolation forest constructed by trees is denoted as , wherein is the total number of isolation trees, which is determined through cross-validation and is 100-200; the determination threshold of the abnormal score is denoted as , the optimal threshold is based on 500+ groups of samples, and is adjusted according to the tree species;

[0085] S3.3. The abnormal score of each pixel spectral vector in step S3.1 is calculated by using the path length algorithm of the isolation forest . is the abnormal score function, and the value range is . The greater the value, the more likely the pixel is abnormal, such as a shadow or a noise point.

[0086] According to the abnormal score, if , the pixel is determined to be an abnormal pixel and is eliminated; if , the pixel is retained, and the hyperspectral image data of the pure tree leaves after the abnormal pixels are eliminated is finally obtained:

[0087] .

[0088] S4. A feature extraction module is constructed to extract features from the hyperspectral image data of the pure tree leaves obtained in step S3. First, the mutual information between the features and the true value of the tree age is calculated, and then the mutual information is combined with the partial least squares method to perform feature screening. The screened features are input into a random forest regression model for further extraction to obtain the core features of the pure tree leaves. ​

[0089] Further, the specific implementation method of step S4 includes the following steps:

[0090] S4.1. Calculation The mutual information of each hyperspectral band reflectance is used to measure the linear / nonlinear correlation of two variables, denoted as , wherein represents the true value of the sample tree age; according to the descending order of the mutual information value, the first mutual information maximum band features are selected to form the feature matrix after mutual information screening :

[0091]

[0092] wherein, represents the reflectance feature vector of the th hyperspectral band, and when the total number of samples is Q, , is the number of features;

[0093] S4.2. Based on the obtained in step S4.1, a partial least squares regression (PLS) model is constructed as follows:

[0094]

[0095] wherein, , is the weight matrix of the PLS model, representing the influence degree of each screened feature on the tree age prediction; , is the residual matrix of the model, reflecting the deviation of the model predicted value from the true value, satisfying ; the determination coefficient of the PLS model is calculated, which is used to measure the goodness of fit of the model, and the value range is , and the closer the value is to 1, the better the fitting is, and the formula is as follows:

[0096]

[0097] wherein, represents the total number of samples, is the true value of the tree age of the th sample, is the predicted value of the tree age of the th sample using the PLS model constructed by , and is the sample mean of the tree age true value;

[0098] To determine the optimal number of features and avoid overfitting caused by too many features and underfitting caused by too few features, the curvature of the determination coefficient of the PLS model is calculated. The point of maximum curvature corresponds to the "inflection point" of the feature number, i.e., the optimal number of features, as shown in the following formula:

[0099]

[0100] in, for Number of features The first derivative reflects the rate of change of goodness of fit with the number of features. for Number of features The second derivative reflects the curvature of the rate of change; a larger curvature indicates a more significant improvement in the goodness of fit from increasing the number of features. The optimal number of features... For curvature The value of the number of features corresponding to the maximum value, i.e. Finally selected The optimized feature matrix consists of several features. ;

[0101] S4.3. Based on the results obtained in step S4.2 The data is input into a random forest regression model for deep feature extraction: through feature splitting and voting ensemble of multiple decision trees, each tree is trained based on a random feature subset to reduce the risk of overfitting, and key features that contribute most to tree age identification, such as the reflectance ratio of a specific wavelength and the depth of the absorption valley, are selected, ultimately yielding the core feature matrix of pure leaves. .

[0102] S5. Construct a leaf identification tree age model module. Use the constructed leaf identification tree age model to identify the tree age based on the core features of the pure leaves obtained in step S4, and obtain the predicted tree age value.

[0103] Furthermore, the specific implementation method of step S5 includes the following steps:

[0104] S5.1. Core feature matrix of pure leaves obtained in step S4 An integrated leaf-based tree age identification model was constructed using a gradient boosting decision tree and the GBDT framework; through... Round of iterative training generation decision tree, model for the first The final tree age prediction value for each sample is determined by The predictions from the decision trees are obtained by weighted averaging. Weighted averaging reduces the error of a single tree. The prediction function is as follows:

[0105]

[0106] in, , For the first The core feature vector of each sample For the first The prediction function of a decision tree takes core features as input and outputs the predicted tree age. The total number of ensemble decision trees (usually taken as...) , This is the ensemble prediction function for the model. , Let be the function space consisting of all possible decision trees;

[0107] S5.2. To avoid overfitting, an objective function with L2 regularization is constructed. In each iteration, the decision tree is learned by minimizing this objective function. The objective function formula is as follows:

[0108]

[0109] in, For the first The model parameters for each iteration include the splitting features, threshold, and leaf node weights of the decision tree. Let be the mean squared loss function, ( This is used to measure the error between the actual and predicted tree age. For the front Round decision tree for the first The cumulative predicted value of each sample. For decision trees The regularization term is used to control the complexity of the tree;

[0110]

[0111]

[0112] in, The number of leaf nodes is a regularization hyperparameter. A larger value further limits the number of leaf nodes, preventing the tree from becoming too deep. For decision trees The total number of leaf nodes, For weight regularization hyperparameters, A larger value restricts the absolute value of the leaf node weights, avoiding extreme predictions. For the first The output weight of each leaf node, that is, the prediction contribution of that node to the sample;

[0113] S5.3. To simplify the optimization process of the objective function, step S5.2 is modified as follows: exist A second-order Taylor expansion is performed at the t-th stage, ignoring higher-order terms. The Taylor expansion approximates the nonlinear loss function as a quadratic function, which facilitates the solution and yields the approximate optimization objective for the t-th round. :

[0114]

[0115] in, For loss function ,right The first-order gradient statistic reflects the direction of error change. This is a second-order gradient statistic that reflects the curvature of the error change, and its definition is as follows:

[0116]

[0117] ;

[0118] For decision trees The leaf nodes, with their optimal output weights To minimize local losses, the calculation formula is as follows:

[0119]

[0120] in, For belonging to the first The set of sample indices for each leaf node, i.e., all core feature vectors. Samples falling into this leaf node The set constituted; the splitting features and splitting threshold of the decision tree are maximized by... It is confirmed that a larger gain indicates a greater reduction in model loss after splitting. The gain calculation formula is as follows:

[0121]

[0122] in, The sample set before splitting, and These are the sample sets of the left and right child nodes after the split, respectively.

[0123] Example 2:

[0124] A tree age identification system based on leaf hyperspectral data includes a leaf and branch segmentation module, an abnormal sample removal module, a feature extraction module, and a leaf identification tree age model module. The leaf and branch segmentation module is sequentially connected to the abnormal sample removal module, the feature extraction module, and the leaf identification tree age model module.

[0125] The leaf and branch segmentation module takes a preprocessed hyperspectral image of the tree canopy (pixel spectral features) as input. ), generate a leaf mask through SVM classification, mask crop the image, and output the hyperspectral image data of the leaf ;

[0126] The abnormal sample elimination module: input is , calculates an abnormal score through an isolation forest , eliminates abnormal pixels, and outputs the hyperspectral image data of the pure leaf ;

[0127] The feature extraction module: input is , first filters features through mutual information to obtain , and then extracts in depth through a random forest, and outputs the core feature matrix of the pure leaf ;

[0128] The leaf recognition age model module: input is , performs age prediction through a GBDT integrated model, and outputs the age prediction value .

[0129] It should be noted that the relational terms, such as "first" and "second", and the like, are used solely to distinguish one from another entity or action, without necessarily requiring or implying any actual relationship or order between or among such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises... a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0130] Although the present application has been described with reference to specific implementations, various modifications and / or changes in form and details can be made thereto by those skilled in the art without departing from the scope of the application. In particular, the features of the specific implementations disclosed herein can be combined in any manner without structural conflicts, and the combinations of features are not exhaustively described in the present specification only for the purpose of brevity and resource saving. Therefore, the present application is not limited to the specific implementations disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A tree age identification method based on leaf hyperspectral data, characterized in that, The method comprises the following steps: S1. Collecting a tree crown leaf hyperspectral image by using a hyperspectral camera, and obtaining pixel spectral features of the tree crown leaf hyperspectral image after preprocessing; S2. Constructing a leaf branch segmentation module, segmenting the pixel spectral features of the tree crown leaf hyperspectral image obtained in step S1, defining a classification rule by using a support vector machine method, and performing target optimization, and performing mask cropping on the pixel spectral features of the tree crown leaf hyperspectral image to obtain hyperspectral image data of leaves; S3. Constructing an abnormal sample elimination module, eliminating abnormal samples from the hyperspectral image data of leaves obtained in step S2, calculating an anomaly score of pixels in the hyperspectral image of leaves by using an isolation forest method, eliminating abnormal pixels, and obtaining pure leaf hyperspectral image data; S4. Constructing a feature extraction module, extracting features from the pure leaf hyperspectral image data obtained in step S3, first calculating mutual information of the features and a true value of tree age, then screening the features by using the mutual information combined with a partial least squares method, inputting the screened features into a random forest regression model for further extraction, and obtaining core features of pure leaves; The specific implementation method of step S4 comprises the following steps: S4.

1. Calculation The mutual information of each hyperspectral band reflectance is used to measure the linear / nonlinear correlation of two variables, denoted as where represents the true value of the sample tree age; according to the descending order of the mutual information value, the top mutual information maximum band features are selected to form the mutual information filtered feature matrix : ； wherein, represents the reflectance feature vector of the i-th hyperspectral band, and the total number of samples is Q, , is the number of features.​ S4.

2. Based on the results from step S4.1 A partial least squares regression, PLS, model was constructed as follows: ； wherein, , is the weight matrix of the PLS model, representing the influence degree of each screening feature on the tree age prediction; , is the residual matrix of the model, reflecting the deviation of the model prediction value from the true value, satisfying ; the determination coefficient of the PLS model is calculated , used to measure the goodness of fit of the model, the value range , and the formula is as follows: ； wherein, denotes the total number of samples, is the true value of the tree age for the th sample, is the predicted value of the tree age for the th sample using the PLS model constructed using is the predicted value of the tree age for the th sample using the PLS model constructed using is the sample mean of the true values of the tree age. Curvature of determination coefficient of a PLS model , as follows: ； in, for Number of features The first derivative, for Number of features The second derivative; the number of optimal features For curvature The value of the number of features corresponding to the maximum value, i.e. Finally selected The optimized feature matrix consists of these features. ; S4.

3. Based on the obtained in step S4.2 Input into the random forest regression model for deep feature extraction: Through the feature splitting and voting integration of multiple decision trees, the key features with the greatest contribution to tree age identification are screened out, and finally the core feature matrix of pure tree leaves is obtained ; S5. Constructing a leaf recognition tree age model module, recognizing tree age by using the constructed leaf recognition tree age model for the core features of pure leaves obtained in step S4, and obtaining a tree age prediction value. 2.The tree age recognition method based on tree leaf hyperspectral data according to claim 1, characterized in that, The process of pre-processing in step S1 includes spectral scanning, black and white balance correction, geometric correction, and the pixel spectral features of the tree crown leaf hyperspectral image are obtained as: , wherein, represents the spectral vector of the mth row, nth column pixel in the hyperspectral image, i.e. the combination of reflectance of the pixel at all wavebands, which is used to characterize the spectral feature of the corresponding location of the material being a leaf or a branch; represents the reflectance value of the mth row, nth column pixel at the wavelength band, ; K represents the total number of wavebands of the hyperspectral image; represents the K-dimensional real number space where the spectral vector is located, i.e. the spectral feature of each pixel can be regarded as a point in the K-dimensional space.

3. The tree age identification method based on tree leaf hyperspectral data according to claim 1 or 2, characterized in that, Step S2 regards the spectral difference between leaves and branches as linearly separable, and defines a classification rule by using a support vector machine method; The hyperplane in the K-dimensional space is set as the following formula: ； where, , is the weight vector of the hyperplane, representing the importance of each band to the importance of the classification; is the bias term, used to adjust the position of the hyperplane in K-dimensional space; the classification decision rule is defined as follows: ； wherein, a prediction class of a pixel in the mth row and the nth column, wherein when a tree leaf is determined, a tree branch is determined; then the pixels of all samples are segmented based on the optimal function, so as to obtain a mask of the tree leaf, and the hyperspectral is cropped according to the mask to obtain hyperspectral image data of the tree leaf.

4. The tree age identification method based on tree leaf hyperspectral data according to claim 3, characterized in that, The specific implementation method of step S3 comprises the following steps: S3.

1. Set the pixel set of the hyperspectral image of the tree leaves obtained in step S2 wherein, denotes the spectral vector of the pth pixel within the tree leaf region, ; S3.

2. The constructed Isolation Forest is composed of a number of Isolation Trees, denoted as , where is the total number of Isolation Trees, determined by cross-validation, taking 100~200; the determination threshold of the anomaly score is denoted as , the optimal threshold based on 500+ groups of sample statistics, adjusted according to the tree species; S3.

3. For each pixel spectral vector in step S3.1 , compute its anomaly score using the path length algorithm of Isolation Forest , is the anomaly score function, taking values in the range ; According to the abnormal score, if , the pixel is determined as an abnormal pixel and is removed; if , the pixel is retained, and finally, the hyperspectral image data of the pure leaves after removing the abnormal pixels is obtained : 。 5. The tree age identification method based on leaf hyperspectral data according to claim 4, characterized in that, The specific implementation method of step S5 comprises the following steps: S5.

1. Constructing the core feature matrix of the pure leaves based on the result of step S4 , and building an integrated leaf recognition age model; through round iterative training, generating decision trees, and the final age prediction value of the model for the first sample is obtained by weighted average of the prediction values of the decision trees, and the prediction function is as follows: ； wherein, , is the core feature vector of the th sample, is the prediction function of the th decision tree, is the total number of integrated decision trees, is the integrated prediction function of the model, is the function space consisting of all possible decision trees; S5.

2. To avoid overfitting of the model, a target function with L2 regularization is constructed, and a decision tree is learned by minimizing the target function in each iteration The formula of the target function is as follows: ； wherein, is the th iteration of the model parameters, is a mean square loss function; is the prediction of the th decision tree for the th sample, is a regularization term for the decision tree . ； ； wherein, is the number of leaf nodes regularization hyperparameter, is the total number of leaf nodes of the decision tree is the weight regularization hyperparameter, is the output weight of the th leaf node;​ S5.

3. To simplify the optimization process of the objective function, the following approximation is made for the step S5.2 At the t-th iteration, the second-order Taylor expansion is performed at and the high-order small terms are ignored, obtaining the approximate optimization objective of the t-th iteration : ； wherein is a loss function for the first order gradient statistics of is a second order gradient statistics defined as follows:​ ； ； For the decision tree of the first leaf node, the formula for calculating the optimal output weight is as follows: ； wherein, is a set of sample indices belonging to the th leaf node; the split feature and split threshold of the decision tree are determined by maximizing the gain , which is calculated as follows: ； wherein, is the set of samples before splitting, and are the sets of samples for the left and right child nodes after splitting, respectively.

6. A tree age identification system based on the tree age identification method based on leaf hyperspectral data according to any one of claims 1-5, characterized in that, The method comprises a leaf branch segmentation module, an abnormal sample elimination module, a feature extraction module, and a leaf recognition tree age model module, wherein the leaf branch segmentation module is sequentially connected to the abnormal sample elimination module, the feature extraction module, and the leaf recognition tree age model module. The leaf tree branch segmentation module: the input is the pre-processed tree crown hyperspectral image, the leaf mask is generated through SVM classification, the image is masked and cut, and the hyperspectral image data of the leaf is output ; The abnormal sample elimination module: input is , calculate abnormal score through isolated forest and eliminate abnormal pixels, output pure tree leaf hyperspectral image data ; The feature extraction module: the input is , first screen features by mutual information to get , and then extract by random forest depth, output the core feature matrix of pure leaves ; The leaf recognition tree age model module: the input is , the tree age is predicted through the GBDT integrated model, and a tree age prediction value is output .

Images