A rapid assessment method for forest carbon storage
By combining multispectral images and spectral band classification networks and combining neighborhood area information for correction, the problem of low evaluation accuracy of forest carbon storage in the prior art is solved, and the rapid and accurate evaluation of forest carbon storage is achieved.
Patent Information
- Application Number
- CN202411389475.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-08
- Publication Date
- 2024-12-20
- Estimated Expiration
- 2044-10-08
AI Technical Summary
The existing forest land carbon storage evaluation method based on remote sensing data relies on a single vegetation index and ignores the role of different spectral bands, resulting in low accuracy in carbon storage evaluation.
Multispectral images are used to calculate NDVI values, and spectral information is classified through spectral band classification network, and vegetation luxuriant values in neighboring areas are corrected to estimate forest carbon storage. The specific steps include collecting multi-spectral images, calculating NDVI values, splitting the spectral information into blue, green, red and near-infrared bands, performing matrix element differential enhancement processing, and using the spectral band classification network to classify vegetation dense values, and passing the above , lower, left and right neighborhood regional corrections, and finally estimate the carbon reserves in forest land.
The accuracy of forest land carbon storage evaluation is improved, and the classification accuracy of vegetation dense values is enhanced through the fusion and correction of multi-spectral information, and the rapid and accurate evaluation of forest land carbon storage is achieved.
Smart Images
Figure CN118918390B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of image processing, and in particular to a method for rapid assessment of forestland carbon reserves. Background Art
[0002] Traditional forest carbon stock assessment methods usually require a lot of ground surveys and laboratory analysis, which is not only time-consuming and labor-intensive, but also difficult to achieve large-scale and efficient monitoring. Therefore, developing a fast and accurate forest carbon stock assessment method is of great practical significance.
[0003] In recent years, the development of remote sensing technology has provided new possibilities for the rapid assessment of forest carbon stocks. By acquiring multispectral images, large-scale forest information can be obtained non-contactly, providing basic data for the rapid assessment of forest carbon stocks. The existing forest carbon stock assessment method based on remote sensing data relies on a single vegetation index to assess forest carbon stocks. Although this method is simple and easy to implement, it often ignores the role and influence of different spectral bands, and there is a problem of low accuracy in carbon stock assessment. Summary of the invention
[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for rapid assessment of forest carbon reserves, which solves the problem of low accuracy of carbon reserve assessment in the prior art.
[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for rapid assessment of forest carbon reserves, comprising the following steps:
[0006] S1, collect multispectral images of forest land, calculate the NDVI value of each pixel point, and form an NDVI image;
[0007] S2, taking each local area on the multispectral image as the central area;
[0008] S3, according to the spectral information on the central area and the NDVI value of the same position on the NDVI image, the spectral information is classified using a spectral band classification network to obtain the vegetation density value of the central area;
[0009] S4, correcting the vegetation density value of the central area according to the classification values of the upper neighboring area, the lower neighboring area, the left neighboring area, and the right neighboring area of the central area to obtain a vegetation density correction value of the central area;
[0010] S5. Estimate the forest carbon storage based on the vegetation density correction value of each central area on the multispectral image.
[0011] Furthermore, the formula for calculating the NDVI value of each pixel in S1 is:
[0012] , where NDVI is the NDVI value of the pixel, NIR is the reflectivity of the near-infrared band of the pixel, and Red is the reflectivity of the red light band.
[0013] Furthermore, S3 includes the following sub-steps:
[0014] S31, splitting the spectral information of the central area into a blue light band, a green light band, a red light band and a near infrared band, respectively forming a blue light band matrix, a green light band matrix, a red light band matrix and a near infrared band matrix;
[0015] S32, on the NDVI image, taking the NDVI value at the same position as the central area, and performing matrix element difference enhancement processing to form an NDVI attention matrix;
[0016] S33. Use the spectral band classification network to process the blue light band matrix, green light band matrix, red light band matrix and near infrared band matrix, and obtain the vegetation density value of the central area based on the adjustment of the NDVI attention matrix.
[0017] The beneficial effect of the above-mentioned further scheme is as follows: the present invention splits the spectral information of a central area, separates it, constitutes matrices of each band, takes the NDVI value at the same position as the central area, and the taken NDVI values constitute a matrix. The arrangement of the elements in the matrix is consistent with that on the NDVI image, and the matrix element difference is enhanced to highlight the effective value to obtain the NDVI attention matrix. The spectral band classification network is then used to process the blue light band matrix, the green light band matrix, the red light band matrix and the near-infrared band matrix. Based on the element attention provided by the NDVI attention matrix, the classification accuracy of the vegetation density value is improved.
[0018] Furthermore, the specific process of performing matrix element difference enhancement processing in S32 to form the NDVI attention matrix is as follows: each NDVI value is enhanced:
[0019] , where z j is the jth enhanced NDVI value, e is a natural constant, z NDVI,j Each enhanced NDVI value is normalized to the jth NDVI value at the same location as the central area: , multiple Construct the NDVI attention matrix Z, is the jth attention value in the NDVI attention matrix, max{z j} for multiple z j Take the maximum value, j is a positive integer.
[0020] The beneficial effect of the above further solution is: the present invention uses an exponential function to enhance each NDVI value taken out, and then uses max{z j The normalization process enhances the difference between NDVI values. The larger the NDVI value, the larger the attention value, and the smaller the NDVI value, the smaller the attention value. This makes it easier to apply different attention values to the elements at corresponding positions in the feature fusion matrix output by the spectral information fusion layer. The NDVI value can reflect the density of vegetation. Therefore, the NDVI attention matrix is constructed with the help of the NDVI value, and attention is paid to each element in the feature fusion matrix output by the spectral information fusion layer to increase the attention of important elements.
[0021] Further, the spectral band classification network in S33 includes: a first Conv layer, a second Conv layer, a third Conv layer, a fourth Conv layer, a spectral information fusion layer, a feature enhancement unit, a fifth Conv layer, an upsampling layer, a three-channel feature extraction unit and a fully connected layer;
[0022] The input end of the first Conv layer is used to input a blue light band matrix; the input end of the second Conv layer is used to input a green light band matrix; the input end of the third Conv layer is used to input a red light band matrix; the input end of the fourth Conv layer is used to input a near infrared band matrix;
[0023] The input end of the spectral information fusion layer is respectively connected to the output ends of the first Conv layer, the second Conv layer, the third Conv layer and the fourth Conv layer, and its output end is connected to the first input end of the feature enhancement unit; the second input end of the feature enhancement unit is used to input the NDVI attention matrix, and its output end is connected to the input end of the fifth Conv layer; the input end of the upsampling layer is connected to the output end of the fifth Conv layer; the input end of the three-channel feature extraction unit is connected to the output end of the upsampling layer, and its output end is connected to the input end of the fully connected layer; the output end of the fully connected layer serves as the output end of the spectral band classification network.
[0024] The beneficial effect of the above further solution is that the convolution kernel sizes of the first Conv layer, the second Conv layer, the third Conv layer and the fourth Conv layer are all , in each band matrix The size of the matrix does not change after the Conv layer of different sizes. The features of each band are fused at the spectral information fusion layer. The NDVI attention matrix is used to enhance the important features of the feature fusion matrix output by the spectral information fusion layer. The fifth Conv layer is used for feature extraction again. The upsampling layer is used to increase the size of the output features of the fifth Conv layer to improve the generalization ability, so that the three-channel feature extraction unit can extract different features. The fully connected layer is combined with the three-channel feature extraction unit to extract different features to obtain the vegetation density value.
[0025] Furthermore, the expression of the spectral information fusion layer is:
[0026] , , where H is the feature fusion matrix output by the spectral information fusion layer, X i is the i-th input feature matrix, the first input feature matrix is the output of the first Conv layer, the second input feature matrix is the output of the second Conv layer, the third input feature matrix is the output of the third Conv layer, and the fourth input feature matrix is the output of the fourth Conv layer. i is the i-th weighted feature matrix, The corresponding elements of the two matrices are multiplied, Y 1 is the first weighted feature matrix, Y 2 is the second weighted feature matrix, Y 3 is the third weighted feature matrix, Y 4 is the fourth weighted feature matrix, w F,i is the i-th input feature matrix X i The weight matrix, b F,i is the i-th input feature matrix X i The bias matrix of , i is a positive integer.
[0027] The beneficial effect of the above further solution is: in the spectral information fusion layer, X i The weight matrix and bias matrix are assigned, and then the corresponding elements of the two matrices are multiplied, so that the position of each eigenvalue does not change, which is convenient for the processing of the feature enhancement unit.
[0028] Furthermore, the expression of the feature enhancement unit is: , where R is the enhancement matrix output by the feature enhancement unit, H is the feature fusion matrix output by the spectral information fusion layer, and Z is the NDVI attention matrix. Multiply the corresponding elements of two matrices.
[0029] Further, the three-channel feature extraction unit includes: a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer and a Concat layer;
[0030] The input end of the first maximum pooling layer, the input end of the second maximum pooling layer, and the input end of the third maximum pooling layer are all used as the input end of the three-channel feature extraction unit; the input end of the Concat layer is connected to the output end of the first maximum pooling layer, the output end of the second maximum pooling layer, and the output end of the third maximum pooling layer, respectively, and the output end thereof is used as the output end of the three-channel feature extraction unit;
[0031] The pooling kernel size of the first maximum pooling layer , the pooling kernel size of the second maximum pooling layer , the pooling kernel size of the third maximum pooling layer .
[0032] The beneficial effect of the above further solution is: the maximum pooling layer using three different pooling kernels extracts different eigenvalues, which makes it easier for the fully connected layer to obtain more eigenvalues.
[0033] Furthermore, the correction formula in S4 is:
[0034] ,in, is the vegetation density correction value in the central area, s o is the vegetation density value of the central area, s up is the vegetation density value of the upper neighborhood area, s down is the vegetation density value of the lower neighborhood area, s r is the vegetation density value of the right neighboring area, s l is the vegetation density value of the left neighborhood area.
[0035] The beneficial effect of the above further scheme is that the present invention takes into account the neighborhood information around the central area, corrects the vegetation density value of the central area, and obtains a more accurate vegetation density correction value. Finally, based on these corrected vegetation density values, a rapid and accurate estimation of forest land carbon storage is achieved.
[0036] Furthermore, the formula for estimating forest carbon storage in S5 is:
[0037] , where y is the forest land carbon storage, is the vegetation density correction value of the nth central area on the multispectral image, N is the number of central areas on the multispectral image, n is a positive integer, and θ is the proportional coefficient of carbon storage conversion.
[0038] The beneficial effects of the present invention are as follows: the present invention calculates the NDVI value of each pixel point to form an NDVI image of the same size as the multispectral image, combines the spectral information on the central area and the NDVI value at the same position on the NDVI image, and adopts a spectral band classification network for classification to obtain the vegetation density value of the central area, so that the spectral band classification network combines all spectral information and pays attention to the NDVI value at the same position, thereby improving the classification accuracy of the vegetation density value, and then corrects the central area in combination with the upper, lower, left and right corrections, thereby further improving the classification accuracy of the vegetation density value, thereby improving the accuracy of forest carbon storage assessment. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 This is a flowchart of a rapid assessment method for forest land carbon stocks;
[0040] Figure 2 This is a schematic diagram of the structure of the spectral band classification network;
[0041] Figure 3 Schematic diagram of the structure of the three-channel feature extraction unit. DETAILED DESCRIPTION
[0042] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0043] like Figure 1 As shown, a rapid assessment method for forest carbon storage includes the following steps:
[0044] S1, collect multispectral images of forest land, calculate the NDVI value of each pixel point, and form an NDVI image;
[0045] S2, taking each local area on the multispectral image as the central area;
[0046] S3, according to the spectral information on the central area and the NDVI value of the same position on the NDVI image, the spectral information is classified using a spectral band classification network to obtain the vegetation density value of the central area;
[0047] S4, correcting the vegetation density value of the central area according to the classification values of the upper neighboring area, the lower neighboring area, the left neighboring area, and the right neighboring area of the central area to obtain a vegetation density correction value of the central area;
[0048] S5. Estimate the forest carbon storage based on the vegetation density correction value of each central area on the multispectral image.
[0049] The present invention is equivalent to dividing the multispectral image into multiple local areas, taking each local area as the central area, that is, extracting the spectral information of each local area. In this embodiment, the size of the local area is set to .
[0050] After step S3, it is equivalent to obtaining the vegetation density value of each local area in the multispectral image. Therefore, the vegetation density value of the central area is corrected according to the vegetation density values of the local areas above, below, left and right of the central area.
[0051] The formula for calculating the NDVI value of each pixel in S1 is:
[0052] , where NDVI is the NDVI value of the pixel, NIR is the reflectivity of the near-infrared band of the pixel, and Red is the reflectivity of the red light band.
[0053] The S3 comprises the following sub-steps:
[0054] S31, splitting the spectral information of the central area into a blue light band, a green light band, a red light band and a near infrared band, respectively forming a blue light band matrix, a green light band matrix, a red light band matrix and a near infrared band matrix;
[0055] S32, on the NDVI image, taking the NDVI value at the same position as the central area, and performing matrix element difference enhancement processing to form an NDVI attention matrix;
[0056] S33. Use the spectral band classification network to process the blue light band matrix, green light band matrix, red light band matrix and near infrared band matrix, and obtain the vegetation density value of the central area based on the adjustment of the NDVI attention matrix.
[0057] The present invention splits the spectral information of a central area to separate it and form matrices of each band. The NDVI value at the same position as the central area is taken to form a matrix. The arrangement of elements in the matrix is consistent with that on the NDVI image. The matrix element difference is enhanced to highlight the effective value to obtain the NDVI attention matrix. The spectral band classification network is then used to process the blue light band matrix, the green light band matrix, the red light band matrix and the near-infrared band matrix. Based on the element attention provided by the NDVI attention matrix, the classification accuracy of the vegetation density value is improved.
[0058] The arrangement of each reflectance value in the blue light band matrix, the green light band matrix, the red light band matrix, and the near infrared band matrix is the same as the arrangement of each reflectance value in the central area.
[0059] The specific process of performing matrix element difference enhancement processing in S32 to form the NDVI attention matrix is as follows: each NDVI value is enhanced: , where z j is the jth enhanced NDVI value, e is a natural constant, z NDVI,j Each enhanced NDVI value is normalized to the jth NDVI value at the same location as the central area: , multiple Construct the NDVI attention matrix Z, is the jth attention value in the NDVI attention matrix, max{z j} for multiple z j Take the maximum value, j is a positive integer.
[0060] The present invention uses an exponential function to enhance each NDVI value obtained, and then uses max{z j The normalization process enhances the difference between NDVI values. The larger the NDVI value, the larger the attention value, and the smaller the NDVI value, the smaller the attention value. This makes it easier to apply different attention values to the elements at corresponding positions in the feature fusion matrix output by the spectral information fusion layer. The NDVI value can reflect the density of vegetation. Therefore, the NDVI attention matrix is constructed with the help of the NDVI value, and attention is paid to each element in the feature fusion matrix output by the spectral information fusion layer to increase the attention of important elements.
[0061] like Figure 2 As shown, the spectral band classification network in S33 includes: a first Conv layer, a second Conv layer, a third Conv layer, a fourth Conv layer, a spectral information fusion layer, a feature enhancement unit, a fifth Conv layer, an upsampling layer, a three-channel feature extraction unit and a fully connected layer;
[0062] The input end of the first Conv layer is used to input a blue light band matrix; the input end of the second Conv layer is used to input a green light band matrix; the input end of the third Conv layer is used to input a red light band matrix; the input end of the fourth Conv layer is used to input a near infrared band matrix;
[0063] The input end of the spectral information fusion layer is respectively connected to the output ends of the first Conv layer, the second Conv layer, the third Conv layer and the fourth Conv layer, and its output end is connected to the first input end of the feature enhancement unit; the second input end of the feature enhancement unit is used to input the NDVI attention matrix, and its output end is connected to the input end of the fifth Conv layer; the input end of the upsampling layer is connected to the output end of the fifth Conv layer; the input end of the three-channel feature extraction unit is connected to the output end of the upsampling layer, and its output end is connected to the input end of the fully connected layer; the output end of the fully connected layer serves as the output end of the spectral band classification network.
[0064] The convolution kernel sizes of the first Conv layer, the second Conv layer, the third Conv layer, and the fourth Conv layer are all , in each band matrix The size of the matrix does not change after the Conv layer of different sizes. The features of each band are fused at the spectral information fusion layer. The NDVI attention matrix is used to enhance the important features of the feature fusion matrix output by the spectral information fusion layer. The fifth Conv layer is used for feature extraction again. The upsampling layer is used to increase the size of the output features of the fifth Conv layer to improve the generalization ability, so that the three-channel feature extraction unit can extract different features. The fully connected layer is combined with the three-channel feature extraction unit to extract different features to obtain the vegetation density value.
[0065] The expression of the spectral information fusion layer is:
[0066] , , where H is the feature fusion matrix output by the spectral information fusion layer, X i is the i-th input feature matrix, the first input feature matrix is the output of the first Conv layer, the second input feature matrix is the output of the second Conv layer, the third input feature matrix is the output of the third Conv layer, and the fourth input feature matrix is the output of the fourth Conv layer. i is the i-th weighted feature matrix, The corresponding elements of the two matrices are multiplied, Y 1 is the first weighted feature matrix, Y 2 is the second weighted feature matrix, Y 3 is the third weighted feature matrix, Y 4 is the fourth weighted feature matrix, w F,i is the i-th input feature matrix X i The weight matrix, b F,i is the i-th input feature matrix X i The bias matrix of , i is a positive integer.
[0067] In the spectral information fusion layer, X i The weight matrix and bias matrix are assigned, and then the corresponding elements of the two matrices are multiplied, so that the position of each eigenvalue does not change, which is convenient for the processing of the feature enhancement unit.
[0068] The expression of the feature enhancement unit is: , where R is the enhancement matrix output by the feature enhancement unit, H is the feature fusion matrix output by the spectral information fusion layer, and Z is the NDVI attention matrix. Multiply the corresponding elements of two matrices.
[0069] like Figure 3 As shown, the three-channel feature extraction unit includes: a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer and a Concat layer;
[0070] The input end of the first maximum pooling layer, the input end of the second maximum pooling layer, and the input end of the third maximum pooling layer are all used as the input end of the three-channel feature extraction unit; the input end of the Concat layer is connected to the output end of the first maximum pooling layer, the output end of the second maximum pooling layer, and the output end of the third maximum pooling layer, respectively, and the output end thereof is used as the output end of the three-channel feature extraction unit;
[0071] The pooling kernel size of the first maximum pooling layer , the pooling kernel size of the second maximum pooling layer , the pooling kernel size of the third maximum pooling layer .
[0072] The maximum pooling layer using three different pooling kernels extracts different eigenvalues, making it easier for the fully connected layer to obtain more eigenvalues.
[0073] The correction formula in S4 is:
[0074] ,in, is the vegetation density correction value in the central area, s o is the vegetation density value of the central area, s up is the vegetation density value of the upper neighborhood area, s down is the vegetation density value of the lower neighborhood area, s r is the vegetation density value of the right neighboring area, s l is the vegetation density value of the left neighborhood area.
[0075] The present invention takes into account the neighborhood information around the central area, corrects the vegetation density value of the central area, and obtains a more accurate vegetation density correction value. Finally, based on these corrected vegetation density values, a rapid and accurate estimation of forestland carbon storage is achieved.
[0076] The formula for estimating forest carbon storage in S5 is:
[0077] , where y is the forest land carbon storage, is the vegetation density correction value of the nth central area on the multispectral image, N is the number of central areas on the multispectral image, n is a positive integer, and θ is the proportional coefficient of carbon storage conversion.
[0078] In this embodiment, the proportionality coefficient of carbon storage conversion is specifically set based on experiments or experience.
[0079] The present invention calculates the NDVI value of each pixel point to form an NDVI image of the same size as the multispectral image, combines the spectral information on the central area and the NDVI value at the same position on the NDVI image, and adopts a spectral band classification network for classification to obtain the vegetation density value of the central area. The spectral band classification network combines all spectral information and pays attention to the NDVI value at the same position to improve the classification accuracy of the vegetation density value. Then, the central area is corrected in combination with the upper, lower, left and right corrections to further improve the classification accuracy of the vegetation density value, thereby improving the accuracy of forest carbon storage assessment.
[0080] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A rapid assessment method for forest carbon storage, It is characterized in that The following steps are involved: S1, collect multispectral images of forest land, calculate the NDVI value of each pixel point, and form an NDVI image; S2, taking each local area on the multispectral image as the central area; S3, according to the spectral information on the central area and the NDVI value of the same position on the NDVI image, the spectral information is classified using a spectral band classification network to obtain the vegetation density value of the central area; S4, correcting the vegetation density value of the central area according to the classification values of the upper neighboring area, the lower neighboring area, the left neighboring area, and the right neighboring area of the central area to obtain a vegetation density correction value of the central area; S5. Estimate the forest carbon storage based on the vegetation density correction value of each central area on the multispectral image; The S3 comprises the following sub-steps: S31, splitting the spectral information of the central area into a blue light band, a green light band, a red light band and a near infrared band, respectively forming a blue light band matrix, a green light band matrix, a red light band matrix and a near infrared band matrix; S32, on the NDVI image, taking the NDVI value at the same position as the central area, and performing matrix element difference enhancement processing to form an NDVI attention matrix; S33. Use the spectral band classification network to process the blue light band matrix, green light band matrix, red light band matrix and near infrared band matrix, and obtain the vegetation density value of the central area based on the adjustment of the NDVI attention matrix.
2. The method for rapid assessment of forest carbon reserves according to claim 1, It is characterized in that The formula for calculating the NDVI value of each pixel in S1 is: , where NDVI is the NDVI value of the pixel, NIR is the reflectivity of the near-infrared band of the pixel, and Red is the reflectivity of the red light band.
3. The method for rapid assessment of forest carbon reserves according to claim 1, It is characterized in that The specific process of performing matrix element difference enhancement processing in S32 to form the NDVI attention matrix is as follows: each NDVI value is enhanced: , where z j is the jth enhanced NDVI value, e is a natural constant, z NDVI,j Each enhanced NDVI value is normalized to the jth NDVI value at the same location as the central area: , multiple Construct the NDVI attention matrix Z, is the jth attention value in the NDVI attention matrix, max{z j } for multiple z j Take the maximum value, j is a positive integer.
4. The method for rapid assessment of forest carbon reserves according to claim 1, It is characterized in that The spectral band classification network in S33 includes: a first Conv layer, a second Conv layer, a third Conv layer, a fourth Conv layer, a spectral information fusion layer, a feature enhancement unit, a fifth Conv layer, an upsampling layer, a three-channel feature extraction unit and a fully connected layer; The input end of the first Conv layer is used to input a blue light band matrix; the input end of the second Conv layer is used to input a green light band matrix; the input end of the third Conv layer is used to input a red light band matrix; the input end of the fourth Conv layer is used to input a near infrared band matrix; The input end of the spectral information fusion layer is respectively connected to the output ends of the first Conv layer, the second Conv layer, the third Conv layer and the fourth Conv layer, and its output end is connected to the first input end of the feature enhancement unit; the second input end of the feature enhancement unit is used to input the NDVI attention matrix, and its output end is connected to the input end of the fifth Conv layer; the input end of the upsampling layer is connected to the output end of the fifth Conv layer; the input end of the three-channel feature extraction unit is connected to the output end of the upsampling layer, and its output end is connected to the input end of the fully connected layer; the output end of the fully connected layer serves as the output end of the spectral band classification network.
5. The method for rapid assessment of forest carbon reserves according to claim 4, It is characterized in that The expression of the spectral information fusion layer is: , , where H is the feature fusion matrix output by the spectral information fusion layer, X i is the i-th input feature matrix, the first input feature matrix is the output of the first Conv layer, the second input feature matrix is the output of the second Conv layer, the third input feature matrix is the output of the third Conv layer, and the fourth input feature matrix is the output of the fourth Conv layer. i is the i-th weighted feature matrix, The corresponding elements of the two matrices are multiplied, Y 1 is the first weighted feature matrix, Y 2 is the second weighted feature matrix, Y 3 is the third weighted feature matrix, Y 4 is the fourth weighted feature matrix, w F,i is the i-th input feature matrix X i The weight matrix, b F,i is the i-th input feature matrix X i The bias matrix of , i is a positive integer.
6. The method for rapid assessment of forest carbon reserves according to claim 4, It is characterized in that The expression of the feature enhancement unit is: , where R is the enhancement matrix output by the feature enhancement unit, H is the feature fusion matrix output by the spectral information fusion layer, and Z is the NDVI attention matrix. Multiply the corresponding elements of two matrices.
7. The method for rapid assessment of forest carbon reserves according to claim 4, It is characterized in that The three-channel feature extraction unit includes: a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer and a Concat layer; The input end of the first maximum pooling layer, the input end of the second maximum pooling layer, and the input end of the third maximum pooling layer are all used as the input end of the three-channel feature extraction unit; the input end of the Concat layer is connected to the output end of the first maximum pooling layer, the output end of the second maximum pooling layer, and the output end of the third maximum pooling layer, respectively, and the output end thereof is used as the output end of the three-channel feature extraction unit; The pooling kernel size of the first maximum pooling layer , the pooling kernel size of the second maximum pooling layer , the pooling kernel size of the third maximum pooling layer .
8. The method for rapid assessment of forest carbon reserves according to claim 1, It is characterized in that The correction formula in S4 is: ,in, is the vegetation density correction value in the central area, s o is the vegetation density value of the central area, s up is the vegetation density value of the upper neighborhood area, s down is the vegetation density value of the lower neighborhood area, s r is the vegetation density value of the right neighboring area, s l is the vegetation density value of the left neighborhood area.
9. The method for rapid assessment of forest carbon reserves according to claim 1, It is characterized in that The formula for estimating forest carbon storage in S5 is: , where y is the forest land carbon storage, is the vegetation density correction value of the nth central area on the multispectral image, N is the number of central areas on the multispectral image, n is a positive integer, and θ is the proportional coefficient of carbon storage conversion.