Method and system for measuring carbon reserves from single tree to regional forest based on hyperspectral image

Abstract

The invention discloses a method and a system for measuring carbon reserves from a single tree to a regional forest based on a hyperspectral image, which can effectively improve the precision and efficiency of forest carbon reserves, and the method comprises the following steps: obtaining hyperspectral data and field survey data; performing spectral feature analysis to determine a target tree species; extracting a crown breadth range of the crown based on a multi-scale segmentation algorithm of edge detection; based on the hyperspectral remote sensing data, classifying the target tree species by using a machine learning algorithm; selecting a tree species biomass equation based on the tree species category of the target area, and calculating the biomass of the overground part; based on the calculation result of the biomass equation and the carbon conversion coefficient, utilizing a deep learning algorithm to perform inversion of the aboveground carbon reserve of the target area; and on the basis of the above-ground carbon reserve inversion result, estimating the above-ground carbon reserve of the target area to obtain the above-ground carbon reserve spatial distribution of the vegetation in the target area.

Description

Technical Field

[0001] The present invention relates to the intersection of forest carbon stock measurement and computer technology, and in particular to a method and system for measuring forest carbon stock from a single tree to a region based on hyperspectral images. Background Art

[0002] Forest carbon pools are the main carbon pools of terrestrial ecosystems. Accurately estimating forest carbon storage has always been a difficult and painful point in academia. At present, many methods have been applied to forest carbon storage estimation, including biomass conversion factor method, carbon flux method, process model method, remote sensing model method, etc. The application of these methods has greatly promoted the estimation of regional forest carbon storage. However, the construction (or verification) of such methods or models requires dozens or hundreds of ground sample plots, and the diameter at breast height and tree height of each tree in the sample plot need to be measured, which is a huge workload. At the same time, due to differences in spatial resolution, such as 30m, 250m and 1000m, it is difficult to directly estimate the carbon storage of individual trees at the regional scale, resulting in overestimation or underestimation of forest carbon storage, limiting the application of the above methods in production practice, and having limited guiding role in forest logging and management, which seriously affects the sustainable management of forests. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for measuring carbon stocks from individual trees to regional forests based on hyperspectral images, in response to the problems existing in the prior art, which can effectively improve the accuracy and efficiency of estimating carbon stocks from individual trees to regional forests, thereby greatly reducing the workload in the field, and can be used to guide the design and management of forest felling operations.

[0004] In order to achieve the above objectives, the technical solutions adopted by the present invention include the following aspects.

[0005] A method for measuring forest carbon stocks from single trees to regional forests based on hyperspectral images, which includes:

[0006] Select the target area, obtain hyperspectral data and field survey data, and perform data preprocessing;

[0007] Spectral information of each tree species in the target area extracted from the hyperspectral image is analyzed for spectral characteristics, sensitive bands are identified, and target tree species are determined;

[0008] The three types of variables, original band, vegetation index and texture feature, were screened, and some variables were selected for classification;

[0009] The crown width range of the tree crown is extracted based on the multi-scale segmentation algorithm based on edge detection;

[0010] Based on the selected variables, machine learning algorithms were used to classify the target tree species based on hyperspectral remote sensing data;

[0011] Based on the tree species in the target area, select the corresponding tree species biomass equation to calculate the aboveground biomass;

[0012] Based on the calculation results of the biomass equation and the carbon conversion coefficient, the deep learning algorithm is used to invert the aboveground carbon storage in the target area;

[0013] Based on the inversion results of aboveground carbon storage, a quantitative estimation of aboveground carbon sinks in the target area is carried out to obtain the spatial distribution data of aboveground carbon storage in vegetation in the target area.

[0014] A system for measuring forest carbon stocks from single trees to regions based on hyperspectral images, comprising at least one processor and a memory connected to the at least one processor in communication; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above method.

[0015] In summary, due to the adoption of the above technical solution, the present invention has at least the following beneficial effects:

[0016] Based on hyperspectral images, single tree segmentation and tree species classification are carried out to realize the calculation of carbon storage of individual trees, breaking away from the traditional method's dependence on sample plots and breast diameter, and providing a new technical path and method for accurate estimation of carbon storage of individual trees and regions;

[0017] A new model for estimating carbon storage of individual trees has been constructed, which can be extended to different regions. This model has achieved high-precision calculation of forest carbon storage from individual trees to regional scales, and has enabled the estimation of forest carbon storage from individual trees to regional scales, so as to guide the design of forest management and logging operations.

[0018] Based on individual tree samples of different diameters and species, tree species classification was performed, with an accuracy of 81%. On this basis, a single tree carbon storage prediction model for any tree species was constructed, with an accuracy of 84%, reducing the workload of field measurements to less than 20% of existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic flow chart of a method for measuring forest carbon stocks from single trees to regional areas based on hyperspectral images according to an exemplary embodiment of the present invention.

[0020] Figure 2 are original spectrograms of 24 target tree species according to an exemplary embodiment of the present invention.

[0021] Figure 3 is a first-order derivative graph of original spectra of 24 target tree species according to an exemplary embodiment of the present invention.

[0022] Figure 4is an envelope-removed spectrum analysis graph of 24 target tree species according to an exemplary embodiment of the present invention.

[0023] Figure 5 4 is a diagram of single tree crown width extraction according to an exemplary embodiment of the present invention.

[0024] Figure 6 is a schematic diagram of accuracy evaluation of model inversion results according to an exemplary embodiment of the present invention.

[0025] Figure 7 Schematic diagram of single tree aboveground mass by model inversion according to an exemplary embodiment of the present invention.

[0026] Figure 8 Schematic diagram of aboveground carbon storage inversion using a model according to an exemplary embodiment of the present invention.

[0027] Fig. 9 It is a schematic diagram of the structure of a system for measuring carbon storage from single trees to regional forests based on hyperspectral images according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments to make the purpose, technical solutions and advantages of the present invention more clearly understood. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0029] Figure 1 A method for measuring forest carbon stocks from single trees to regional forests based on hyperspectral images according to an exemplary embodiment of the present invention is shown. The method of this embodiment mainly includes:

[0030] Data preparation: select the target area, obtain hyperspectral data and field survey data, and perform data preprocessing;

[0031] Spectral feature analysis: Spectral information extracted from hyperspectral images of various tree species is analyzed by original spectrum, first-order derivative, envelope elimination, etc., to identify sensitive bands and determine target tree species;

[0032] Variable selection: The vegetation index and texture features calculated from the original band and sensitive band are screened, and some variables are selected for classification;

[0033] Single tree segmentation: A multi-scale segmentation algorithm based on edge detection is used to extract the crown width range of the tree crown and achieve single tree segmentation;

[0034] Tree species classification: Based on the selected variables, machine learning algorithms such as random forest and support vector machine are used to classify the target tree species in hyperspectral remote sensing data;

[0035] Construction of biomass equation: Based on the previous tree species classification, the biomass allometric growth equation of tree species suitable for the target area is selected to calculate the single tree biomass of the target tree species;

[0036] Inversion of aboveground carbon storage based on deep learning: Based on the calculation results of the tree species biomass allometric growth equation and the carbon conversion coefficient, the aboveground carbon storage of the target area is inverted using a deep learning algorithm;

[0037] Measurement of carbon stocks from single trees to regional forests: Based on the above-ground carbon stock inversion results, the above-ground carbon stocks from single trees to regional forests in the target area are measured, and the measurement data results can be intuitively presented through mapping and other forms.

[0038] In the data preparation step, hyperspectral data is obtained. Specifically, a UAV equipped with a Gaiasky-mini 3-VN hyperspectral imaging system can be used to take multiple aerial photos of the target area to obtain hyperspectral image data with a spectral range of 400-1000nm and a spectral resolution of 5nm. The acquired image data is preprocessed by lens calibration, radiation calibration, atmospheric correction, reflectivity correction, splicing, and color uniformity to obtain the entire UAV hyperspectral data of the target area.

[0039] Among them, obtaining field survey data includes: conducting field surveys on tree species in the target area many times, observing the canopy characteristics of each tree on the spot, selecting trees whose canopies can be directly photographed by drone hyperspectral sensors, and using a total station to measure coordinates and record the names of the tree species.

[0040] Preprocessing and descriptive statistical analysis of field survey data were performed, including identification of typical tree species, calculation of the standard deviation of average DBH size and tree species distribution, etc., in order to understand the basic trend and variability of the data. On this basis, the corresponding band values ​​of these typical tree species were extracted from the drone hyperspectral images according to their coordinates, and the spectral curves of each type of tree species were further analyzed, and points with differences greater than the threshold were removed to determine the final samples. The sample target tree species are shown in Table 1, with a total of 309 samples. Among them, the largest sample size is cypress, with a total of 221 samples.

[0041] Table 1 Target tree species

[0042]

[0043] In the spectral feature analysis step, the original spectrum analysis is first performed. Taking the 24 target tree species in Table 1 as an example, the original spectra are as follows: Figure 2As shown; in the visible light range of 400-680nm, the change is small, there is a small reflection peak located at 550nm green light, and there is a small reflection valley at 670nm red light on its right side. This feature is due to the strong absorption of chlorophyll on blue and red light, and strong reflection on green light; in the red edge range of 680-740nm, the spectral reflectance of each tree species shows a sharp rise, but there are large differences in the range of change, which is mainly due to the concentration of chlorophyll a and b, changes in leaf cell structure and plant canopy structure; and in the near-infrared range of 740-1000nm, its reflectance reaches the maximum value near 900nm and then begins to decline, and presents a more obvious reflection valley at 950nm, which is due to the absorption of water. Therefore, through the characteristic analysis of the original spectra of different tree species, the 550, 670, 900 and 950nm bands, namely B57, B101, B185, and B203 bands, are selected as characteristic bands.

[0044] Hyperspectral preprocessing derivative algorithms include first-order derivatives, second-order derivatives, etc. Taking the 24 target tree species in Table 1 as an example, the first-order derivatives are as follows: Figure 3 As shown: in the visible light range of 400-670nm, the spectral reflectance of each tree species changes slightly. Due to the strong absorption of blue and red light and strong reflection of green light by chlorophyll, there are large increases and decreases respectively; in the range of 670-760nm, due to the influence of chlorophyll a and b concentrations, changes in leaf cell structure and plant canopy structure, the spectral reflectance of each tree species changes greatly and reaches the maximum near 740nm. It is precisely because of the large change rate in the red edge range that the red edge band is often used as an effective means of vegetation identification and detection; in the range of 760-1000nm, the spectral reflectance of each tree species tends to be stable, and due to the influence of water absorption, it shows a downward trend.

[0045] By analyzing the original spectra of different tree species ( Figure 2 ) and its first-order derivative ( Figure 3 ) and selected 740nm, i.e. B127 band, as the characteristic band.

[0046] When the spectral curves are similar, it is not convenient to directly extract the spectral features from them, so the spectral curves can be further processed to highlight the spectral features. The envelope elimination method can effectively highlight the absorption, reflection and emission features of the spectral curves and normalize them to a consistent spectral background, which is conducive to comparing the characteristic values ​​with other spectral curves. The algorithm execution steps include:

[0047] (1) Obtain all the maximum points on the spectrum curve, i.e., the "peak" value points, by taking the derivative, and then compare the sizes to obtain the maximum value point among the maximum value points;

[0048] (2) Take the maximum value point as one endpoint of the envelope line, calculate the slope of the line connecting this point and each maximum value in the long-wave direction (the direction of wavelength growth), take the point with the maximum slope as the next endpoint of the envelope line, and then use this point as the starting point to repeat until the last point;

[0049] (3) Take the maximum value point as one endpoint of the envelope selection line, perform similar calculations towards the short wave (the direction of decreasing wavelength), take the point with the minimum slope as the next endpoint, and then use this point as the starting point to repeat until the starting point on the curve;

[0050] (4) By connecting all endpoints along the long-wave direction, the envelope of the curve can be formed. By subtracting the reflectance value of the corresponding band on the envelope from the actual spectral reflectance, the normalized value after the envelope elimination method can be obtained.

[0051] The spectral analysis of the 24 target tree species in Table 1 using envelope removal found that: within the range of 400-670nm, its value showed a decreasing trend, similar to the original spectrum. Due to the strong absorption of blue and red light and strong reflection of green light by chlorophyll, it showed a peak at 550nm, and valleys appeared at 490 and 670nm on both sides; within the 670-760nm spectrum range, the values ​​of each tree species showed a sharp increase; and within the range of 760-1000nm, its value changed little, and only a small valley shape was formed near 950nm due to the influence of water absorption, such as Figure 4 Therefore, according to the results of spectral analysis of 24 tree species by envelope removal method, 490, 550, 670 and 950nm, namely B34, B57, B101 and B203 bands were selected as characteristic bands.

[0052] By performing spectral analysis on 24 target tree species using original spectra, first-order derivatives, and envelope removal, B34, B57, B101, B127, B185, and B203 were finally selected as sensitive bands that can highlight the spectral characteristics of the target tree species.

[0053] The variables used for target tree species classification in the present invention include 224 bands of the collected UAV hyperspectral remote sensing data, and the original band information is shown in Table 2.

[0054] Table 2 Original band information

[0055]

[0056] In order to highlight the characteristics of different tree species and improve the classification accuracy, in addition to the original bands, the six sensitive bands of B34, B57, B101, B127, B185 and B203 obtained by spectral feature analysis can be combined with the normalized vegetation index (NDVI), enhanced vegetation index (EVI), simple ratio vegetation index (SR), soil adjusted vegetation index (SAVI), normalized chlorophyll index (NDchl) and other original bands to construct different vegetation indices, a total of 44, as shown in Table 3.

[0057] Table 3 Vegetation index

[0058]

[0059] In addition, considering that the vegetation index may be saturated, eight texture features (see Table 4) including mean, variance, correlation, hom, dissimilarity, contrast, information entropy and second-order moment can be selected as predictive variables for classification. The texture features are calculated using the B101 red band, with the grayscale set to 32, the window size 5×5, and the direction (0, 1).

[0060] Table 4 Texture features

[0061]

[0062] Since the calculation of 276 variables, including 224 original bands, 44 vegetation indices and 8 texture features, for the classification of target tree species is very large, these three types of variables can be screened first. Variable screening can use the vifcor function in the usdm package of the R language to select variables based on the variance inflation factor.

[0063] The steps of single tree segmentation include: first, using a total station to measure the longitude and latitude coordinates of the trees as the accurate position of the single trees, and performing single tree position detection, that is, taking the crown vertices found as the spatial position of the single trees; then taking the extracted crown vertices as the marking points, performing single tree crown segmentation; using a multi-scale segmentation algorithm based on edge detection to segment the corrected hyperspectral image, setting a grayscale threshold to eliminate background information, and thus extracting the crown width of the single tree; based on the crown width of the single tree, extracting relevant environmental variables according to arbitrary polygons for tree species classification and single tree carbon storage prediction, the single tree segmentation results are as follows: Figure 5 shown.

[0064] In the process of tree species classification, the random forest algorithm randomly extracts multiple samples from the training data set through self-service sampling technology and builds a decision tree for each sample. During the training process, these decision trees will randomly select a certain proportion of features from the feature set without replacement, and find the best split point to separate the nodes based on these features. This method is not only suitable for processing numerical and categorical data, but also effectively solves regression and classification problems. Random forests are fast to train and can usually generate models with high generalization capabilities, thereby reducing the risk of overfitting to a certain extent. This makes random forests an efficient and versatile machine learning algorithm.

[0065] In the tree species classification step, the support vector machine algorithm can also be used. The support vector machine is a supervised learning model and related learning algorithm for analyzing data in classification and regression analysis. Specifically, it distinguishes different categories by mapping instance points in space. This mapping creates a wide interval, making the boundaries between categories clearly visible. When a new instance appears, the support vector machine maps it to the same space and determines its possible category based on its position. The support vector machine can not only perform linear classification, but also nonlinear classification through kernel function technology, which means that it can map the input data to a higher-dimensional feature space to find patterns that may not be intuitively recognizable in the original space. It is particularly effective when the data set is small because it can reduce overfitting and improve the generalization ability of the model, making it a powerful tool for dealing with complex classification problems.

[0066] The present invention adopts the cross-validation method to evaluate the classification effect according to the overall accuracy and Kappa coefficient, wherein the overall classification accuracy refers to the probability that the classification result is consistent with the test data type for each random sample. .

[0067] The Kappa coefficient is a multivariate discrete method for evaluating the classification accuracy and error matrix of remote sensing images. The overall classification accuracy takes into account the number of correctly classified pixels in the diagonal direction, while the Kappa coefficient also takes into account various missed and misclassified pixels outside the diagonal. The calculation formula of the Kappa coefficient based on the confusion matrix is ​​as follows:

[0068]

[0069] Among them, Po is the overall accuracy, and Pe is the product of the actual and predicted numbers divided by the square of the total number of samples.

[0070] In an exemplary embodiment of the present invention, the classification effects of the random forest and support vector machine algorithms are compared under the three prediction variable conditions of original band, vegetation index plus texture feature, and original band plus vegetation index and texture feature. The performance of the random forest and support vector machine algorithms is shown in the following table:

[0071] Table 5 Tree species classification results

[0072]

[0073] Based on the original bands, five bands, B2, B104, B127, B185 and B224, were selected as feature variables for classification. The overall accuracy of the random forest was 0.75, and the Kappa coefficient was 0.52. For the support vector machine, seven bands, B2, B57, B101, B127, B185, B203 and B224, were selected as feature variables for classification. The overall accuracy of the support vector machine algorithm was 0.73, and the Kappa coefficient was 0.50.

[0074] Based on vegetation index and texture characteristics, 13 indices including ARI2, DWSI4, REP, RVSI, VIgreen, SR_1, PWI_1, PSRI_1, CRI1_1, PRI_1, mean, var, and cor were selected for classification. The overall accuracy of the random forest algorithm was 0.79, and the Kappa coefficient was 0.53. For the support vector machine, 11 indices including VIgreen, ARI2, CRI2, DWSI4, PRI, PSRI, REP, RVSI, hom, var and cor were selected as feature variables for classification. The overall accuracy of the support vector machine algorithm was 0.77, and the Kappa coefficient was 0.52.

[0075] From all the variables, a total of 14 feature variables including the B224 band, 10 indexes such as ARI1 and 3 texture features such as mean were selected for classification. The overall accuracy of the classification result of the random forest algorithm was 0.80, and the Kappa coefficient was 0.55. For the support vector machine, a total of 15 feature variables including 2 bands such as B104, 10 indexes such as CRI1_1 and 3 texture features such as hom were selected for classification. The overall accuracy of the support vector machine algorithm was 0.79, and the Kappa coefficient was 0.54.

[0076] Therefore, in a preferred embodiment, the classification effect of using vegetation index and texture features as prediction variables is better than that of original bands (higher overall accuracy and Kappa coefficient), and the combination of original bands, vegetation index and texture features can achieve the highest classification accuracy.

[0077] In a preferred embodiment, the drone remote sensing image is segmented with a scale of 6, a shape factor of 0.8, and a compact factor of 0.1. After the image segmentation is completed, the crown of the sample plot is extracted using the nearest neighbor classification method. Representative samples are selected as classification indicators, and a threshold model is constructed for classification. The crown and segmentation misclassification and redundant parts are corrected, and the overly fragmented crown objects are merged to complete the crown width extraction of all sample plots.

[0078] The specific steps for constructing the biomass equation include: based on the tree species categories in the target area, selecting the tree species biomass allometric growth equation suitable for the target area, and for each type of tree species, dividing it into four parts: trunk, branch, leaf and root according to the organs for calculation, and adding the biomass of the three parts, namely trunk, branch and leaf, as the biomass of the aboveground part.

[0079] The aboveground carbon storage inversion step based on deep learning can specifically use the Multilayer Perceptron (MLP) model to invert the aboveground carbon storage in the target area. Among them, for the environmental variables input into the model, 224 original bands, 45 vegetation indices and 8 texture features, a total of 277 environmental variables, can be selected for analysis.

[0080] Since the 227 environmental variables are very computationally intensive for the classification of target tree species, and there is a high degree of collinearity between environmental variables, these three types of variables can be screened first. Variable screening can use the vifcor function in the usdm package of the R language to select the variance inflation factor, and the correlation coefficient threshold is set to 0.9.

[0081] In a preferred embodiment, the variables screened out can be further ranked by importance, and 15 environmental variables including RVSI, Var, DWSI4, PWI, ARI2, GNDVI_1, Band2, VIgreen, ARVI, PRI_1, REP, Hom, Band224, PSRI_1 and Mean are selected to participate in the inversion of the model.

[0082] According to the characteristics of the model and data, an MLP model with three layers of neurons can be constructed. The model parameters are adjusted as follows: the number of neurons in the first layer, the second layer, and the third layer are set to 1000, 800, and 500 respectively, the neuron dropout rate of the three layers is set to 0.15, and the batch_size of the number of samples selected for one training is set to 32.

[0083] In a preferred embodiment, the model can be run to perform aboveground biomass inversion based on the selected 15 environmental variables and model parameter settings, and the accuracy of the model inversion results can be evaluated. The accuracy evaluation results and the aboveground biomass inversion results are as follows: Figure 6 , Figure 7 As shown. Through cross-validation of field data, the preliminary accuracy reached 84%. The final determination coefficient R of the model 2 The average value of the aboveground biomass of a single tree is 0.84, and the root mean square error (RMSE) is 73.82 kg. The correlation between the measured aboveground biomass of a single tree and the predicted aboveground biomass of a single tree is strong, and the overall accuracy of the model is high.

[0084] The aboveground carbon storage in the target area was inverted based on the model parameters with the highest accuracy, and the model inversion obtained the predicted data of aboveground carbon storage of individual trees in the target area.

[0085] On this basis, based on the predicted data of aboveground biomass of individual trees in the target area, the aboveground carbon storage of vegetation in the target area was calculated using a carbon content coefficient of 0.5 to obtain the spatial distribution data of aboveground carbon storage of vegetation in the target area, such as Figure 8 As shown, the accuracy and efficiency of forest carbon stock estimation are effectively improved.

[0086] Fig. 9 A system for measuring forest carbon stocks from single tree to regional scales based on hyperspectral images according to an exemplary embodiment of the present invention is shown, namely, an electronic device 310 (such as a computer server with a program execution function), which includes at least one processor 311, a power supply 314, and a memory 312 and an input-output interface 313 that are communicatively connected to the at least one processor 311; the memory 312 stores instructions that can be executed by the at least one processor 311, and the instructions are executed by the at least one processor 311 so that the at least one processor 311 can execute the method disclosed in any of the aforementioned embodiments; the input-output interface 313 may include a display, a keyboard, a mouse, and a USB interface for inputting and outputting data; the power supply 314 is used to provide power to the electronic device 310.

[0087] Those skilled in the art can understand that: all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM), disks or optical disks, etc. Various media that can store program codes.

[0088] When the above-mentioned integrated unit of the present invention is implemented in the form of a software functional unit and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0089] The above is only a detailed description of the specific implementation of the present invention, rather than a limitation of the present invention. Various substitutions, modifications and improvements made by those skilled in the relevant art without departing from the principle and scope of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for measuring forest carbon stocks from single trees to regional forests based on hyperspectral images, characterized in that: include: Select the target area, obtain hyperspectral data and field survey data, and perform data preprocessing; Spectral information of each tree species in the target area extracted from the hyperspectral image is analyzed for spectral characteristics, sensitive bands are identified, and target tree species are determined; Screen the three types of variables: original band, vegetation index, and texture characteristics, and select some variables for classification; The crown width range of the tree crown is extracted based on the multi-scale segmentation algorithm based on edge detection; Based on the selected variables, machine learning algorithms were used to classify the target tree species based on hyperspectral remote sensing data; Based on the tree species in the target area, select the corresponding tree species biomass equation to calculate the aboveground biomass; Based on the calculation results of the biomass equation and the carbon conversion coefficient, the deep learning algorithm is used to invert the aboveground carbon storage in the target area; Based on the inversion results of aboveground carbon storage, a quantitative estimation of aboveground carbon sinks in the target area is carried out to obtain the spatial distribution data of aboveground carbon storage in vegetation in the target area.

2. The method according to claim 1, characterized in that The obtaining of hyperspectral data comprises: A UAV equipped with the Gaiasky-mini 3-VN hyperspectral imaging system was used to take multiple aerial photos of the target area to obtain hyperspectral image data with a spectral range of 400-1000nm and a spectral resolution of 5nm. The acquired image data was subjected to lens calibration, radiation calibration, atmospheric correction, reflectivity correction, stitching, and uniform color preprocessing to obtain the entire hyperspectral data of the target area.

3. The method according to claim 1, characterized in that The spectral feature analysis includes: The original spectrum, first-order derivative, and envelope removal were used for spectral analysis, and B34, B57, B101, B127, B185, and B203 were selected as sensitive bands to highlight the spectral characteristics of the target tree species.

4. The method according to claim 3, characterized in that The method further comprises: Based on the spectral characteristic analysis, 6 sensitive bands were obtained and combined with the normalized difference vegetation index NDVI, enhanced vegetation index EVI, simple ratio vegetation index SR, soil adjusted vegetation index SAVI and normalized chlorophyll index NDchl to construct 44 vegetation indices.

5. The method according to claim 4, characterized in that The method of using a machine learning algorithm to classify target tree species based on hyperspectral remote sensing data includes: The classification effects of random forest and support vector machine algorithms were compared under three prediction variable conditions: original band, vegetation index plus texture features, and original band plus vegetation index and texture features. The algorithm and classification result data with the highest overall accuracy and Kappa coefficient were selected.

6. The method according to claim 5, characterized in that The method further includes: resampling the spatial resolution of the hyperspectral image from 0.18m to 0.5m, 1m, 1.5m, 2m, 2.5m and 3m respectively, and classifying using two algorithms, random forest and support vector machine, respectively, and selecting the algorithm and classification result data with the highest overall accuracy and Kappa coefficient.

7. The method according to claim 5, characterized in that The method comprises: selecting a total of 14 characteristic variables including the B224 band, 10 vegetation indices, and 3 texture features, and classifying target tree species using a random forest algorithm.

8. The method according to claim 6, characterized in that The crown width extraction method includes: firstly, using the latitude and longitude coordinates of the field survey sample points as the accurate position of the single tree, detecting the position of the single tree, and using the searched crown vertex as the spatial position of the single tree; then using the extracted crown vertex as the mark point to segment the crown of the single tree; using the multi-scale segmentation algorithm based on edge detection to segment the calibrated high spatial resolution image, setting the gray threshold to eliminate background information, and thus extracting the crown width of the single tree; Among them, the multi-scale segmentation algorithm parameters are set as scale 6, shape factor 0.8, and compact factor 0.

1.

9. The method according to claim 1, characterized in that: The inversion of aboveground carbon storage in the target area using deep learning includes: The multi-layer perceptron (MLP) model was used to invert the aboveground carbon storage in the target area. The MLP model parameters included: the number of neurons in the first, second and third layers were set to 1000, 800 and 500 respectively, the neuron dropout rate of the three layers was set to 0.15, and the batch_size of the number of samples selected for one training was set to 32.

10. A system for measuring forest carbon stocks from single trees to regions based on hyperspectral images, characterized in that: It includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method described in any one of claims 1 to 9.

Images