Forest accumulation prediction method and system

Abstract

The invention discloses a forest accumulation prediction method and system, and relates to the technical field of image data processing, and the method comprises the steps: obtaining a target region SAR image through an L-band synthetic aperture radar, and generating a polarization interference coherence matrix; combining a digital elevation model to extract a terrain gradient and a local incident angle, and calculating a polarization azimuth angle offset; performing model polarization decomposition based on the coherence matrix and the offset, and estimating a pure body decoherence coefficient and body scattering intensity; introducing a three-layer S-RVoG model in combination with the interference fringe pattern to invert the forest canopy height and the extinction coefficient, and constructing a multi-dimensional feature vector; and inputting the result into a pre-training model, and outputting a forest stock prediction result, thereby achieving the technical effects of reducing the terrain influence and improving the prediction precision.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a method and system for predicting forest stock volume. Background Technology

[0002] Forest stock volume is a core indicator for measuring the abundance of forest resources, assessing the quality of the forest ecological environment, and calculating carbon storage. Traditional forest stock volume surveys rely on ground quadrat measurements and forestry data tables. Although these methods offer high accuracy, they are time-consuming, labor-intensive, and unable to cover large areas of remote forest regions, thus failing to meet the urgent need for accurate acquisition of forest stock volume over large areas.

[0003] With the development of remote sensing technology, large-scale forest parameter inversion using satellite data has become a research hotspot. Synthetic Aperture Radar (SAR) has shown unique advantages in forest volume monitoring due to its all-weather, all-day imaging capabilities and penetration into the vertical structure of forests.

[0004] Existing methods for predicting forest stock volume using synthetic aperture radar (SAR) include the backscattering coefficient method, the coherence coefficient method, and the interferometric phase method. However, these methods are susceptible to the influence of surface water content and roughness. They also exhibit severe saturation phenomena in areas with high stock volume and do not consider topography and the tree trunk layer, which affects the accuracy of the prediction. Summary of the Invention

[0005] This invention provides a method and system for predicting forest stock volume, which solves the technical problems of large terrain interference, incomplete physical mechanism description and reliance on single features in the prior art, and achieves the technical effect of reducing the impact of terrain and improving prediction accuracy.

[0006] In a first aspect, the present invention provides a method for predicting forest stock volume, wherein the method for predicting forest stock volume includes: SAR image data of the target area is acquired using L-band synthetic aperture radar, and a polarimetric interferometric coherence matrix is ​​generated based on the SAR image data. Obtain the digital elevation model of the target area, extract the terrain slope and local incident angle based on the digital elevation model, and calculate the polarization azimuth offset based on the terrain slope. Based on the polarization interference coherence matrix and the polarization azimuth offset, a model-based polarization decomposition method is used to estimate the pure volume decoherence coefficient and the volume scattering intensity. Combining the pure volume decoherence coefficient, the local incident angle, and the interferometric fringe pattern generated from the SAR image data, a three-layer S-RVoG model is introduced to perform a three-stage inversion to obtain the forest canopy height and extinction coefficient of the target area, and a multidimensional feature vector is constructed. The multidimensional feature vector is input into a pre-trained forest stock volume estimation model to generate forest stock volume prediction results.

[0007] In one feasible implementation, SAR image data of the target area is acquired using L-band synthetic aperture radar, and a polarimetric interferometric coherence matrix is ​​generated based on the SAR image data, including: Raw SAR image data is acquired using L-band synthetic aperture radar; The original SAR image data is finely registered, and the registered image is then subjected to spectral filtering. Interference coherence calculation is performed based on the spectral filtering results to generate an interference fringe pattern and a coherence coefficient pattern. Based on the interference fringe pattern and the coherence coefficient pattern, the polarization interference coherence matrix is ​​constructed.

[0008] In one feasible implementation, a digital elevation model of the target area is obtained; based on the digital elevation model, the terrain slope and local incident angle are extracted; and the polarization azimuth offset is calculated based on the terrain slope, including: Extract the terrain slope and aspect values ​​of each pixel from the digital elevation model; By combining the imaging geometry parameters of the synthetic aperture radar with the terrain slope value and terrain aspect value, the local incident angle of each pixel is calculated; The polarization azimuth offset caused by the terrain is calculated based on the terrain slope value, terrain aspect value and local incident angle.

[0009] In one feasible implementation, based on the polarization interferometry coherence matrix and the polarization azimuth offset, a model-based polarization decomposition method is used to estimate the pure volume decoherence coefficient and the volume scattering intensity, including: Based on the polarization interference coherence matrix, the contributions of the three scattering mechanisms—surface scattering, secondary scattering, and volume scattering—are calculated respectively. Based on the aforementioned contribution, the polarization coherence matrix is ​​expressed as the first linear combination of the scattering matrices corresponding to the three scattering mechanisms: surface scattering, secondary scattering, and volume scattering. Based on the aforementioned contribution, the polarization interference matrix is ​​decomposed into a second linear combination of the polarization interference matrices corresponding to the three scattering mechanisms: surface scattering, secondary scattering, and volume scattering. The first linear combination and the second linear combination are combined to form a model matrix for describing the target region; Based on the model matrix, the best fit is performed with the goal of minimizing the residual common factor between the polarization interference coherence matrix and the model matrix, and the volume decoherence coefficient and the volume scattering intensity are obtained by solving.

[0010] In one feasible implementation, the three-layer S-RVoG model includes: A surface scattering model, which is used to characterize Bragg scattering on a rough surface of the Earth. A secondary scattering model, which is used to characterize secondary scattering between the ground surface and the tree trunk; A volume scattering model is used to characterize random volume scattering in the forest canopy.

[0011] In one feasible implementation, the pure volume decoherence coefficient, the local incident angle, and the interferometric fringe pattern generated from the SAR image data are combined and introduced into a three-layer S-RVoG model for three-stage inversion to obtain the forest canopy height and extinction coefficient of the target area, and a multi-dimensional feature vector is constructed, including: Calculate the local incident angle, and based on the local incident angle, calculate several complex coherence coefficients corresponding to different polarization modes, and then fit a straight line on the complex plane; To remove vegetation bias, the two points obtained by the intersection of the fitted straight line in the first step and the unit circle of the complex plane are used. The phase angle corresponding to one of these points is selected as the surface phase based on the principle of maximum vegetation bias. Using the surface phase, the known vertical wavenumber, the pure volume decoherence coefficient, and the local incident angle, combined with the physical equations of the three-layer S-RVoG model, a two-dimensional lookup table is established for the pure volume decoherence coefficient, the forest canopy height, and the extinction coefficient, and the forest canopy height is obtained by looking up the table. A multidimensional feature vector is constructed based on the forest canopy height, the volume scattering intensity, the local incident angle, and the tree species classification information of the target area.

[0012] In one feasible implementation, the multidimensional feature vector is input into a pre-trained forest stock volume estimation model to generate a forest stock volume prediction result. Prior to this, the process further includes: Obtain a sample dataset, which contains multiple samples, each sample being composed of the paradigm of the multidimensional feature vector and the corresponding measured accumulation data of the ground quadrat; The sample dataset is divided into a training sample set and a validation sample set; Using the multidimensional feature vectors in the training sample set as input and the corresponding measured stock volume data of ground quadrats as output, the random forest regression model is trained and its performance is optimized until the prediction accuracy on the validation sample set meets a preset threshold, thus establishing the forest stock volume estimation model.

[0013] In one feasible implementation, the prediction accuracy includes the coefficient of determination and the root mean square error.

[0014] In one feasible implementation, a two-dimensional lookup table is established, relating the pure volume decoherence coefficient to the forest canopy height and the extinction coefficient, including: By combining historical data and quadrat data, the range of forest canopy height and extinction coefficient in the target area were obtained; Within the range of forest canopy height and the range of extinction coefficient, grid sampling is performed with a preset step size; The gridded sampling results are traversed and substituted into the physical equations of the three-layer S-RVoG model to calculate the theoretical pure volume decoherence coefficient; The theoretical pure volume decoherence coefficients and the gridded sampling results are stored in a structured manner to form a two-dimensional lookup table in the form of a two-dimensional table.

[0015] Secondly, the present invention also provides a forest stock volume prediction system, wherein the forest stock volume prediction system includes: The SAR data acquisition module is used to acquire SAR image data of the target area through L-band synthetic aperture radar, and generate a polarimetric interferometric coherence matrix based on the SAR image data. The terrain parameter extraction module is used to obtain the digital elevation model of the target area, extract the terrain slope and local incident angle based on the digital elevation model, and calculate the polarization azimuth offset based on the terrain slope. The polarization decomposition estimation module is used to estimate the pure volume decoherence coefficient and volume scattering intensity based on the polarization interference coherence matrix and the polarization azimuth offset using a model-based polarization decomposition method. The three-stage inversion module is used to combine the pure volume decoherence coefficient, the local incident angle, and the interferometric fringe pattern generated from the SAR image data, and introduce them into a three-layer S-RVoG model for three-stage inversion to obtain the forest canopy height and extinction coefficient of the target area, and construct a multi-dimensional feature vector. The forest stock volume prediction module is used to input the multidimensional feature vector into a pre-trained forest stock volume estimation model to generate forest stock volume prediction results.

[0016] Beneficial Effects: This invention discloses a method and system for predicting forest stock volume. It utilizes L-band synthetic aperture radar to acquire SAR image data of a target area and constructs a polarimetric interferometric coherence matrix based on the SAR image data. It then acquires a digital elevation model corresponding to the target area, extracts terrain slope and local incident angle information from it, and calculates the polarimetric azimuth offset parameter based on the terrain slope. Based on the polarimetric interferometric coherence matrix and the polarimetric azimuth offset parameter, a model-driven polarimetric decomposition method is used to estimate the pure volume decoherence coefficient and volume scattering intensity. Finally, it fuses the pure volume decoherence... The coefficients, local incident angles, and interferometric fringe patterns generated from the SAR image data are introduced into a three-layer S-RVoG model for staged inversion processing to obtain the forest canopy height and extinction coefficient of the target area, and a multidimensional feature vector is constructed accordingly. The multidimensional feature vector is then input into a pre-trained forest volume estimation model to output the corresponding forest volume prediction result. The forest volume prediction method and system disclosed in this invention solves the technical problems of large terrain interference, incomplete physical mechanism characterization, and dependence on a single feature, and achieves the technical effects of reducing terrain influence and improving prediction accuracy. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the forest stock prediction method of the present invention; Figure 2 This is a schematic diagram of the element flow in the forest stock prediction method of the present invention; Figure 3 The data provided are the average tree height data of an exemplary quadrat in the forest stock prediction method of this invention. Figure 4 This is an exemplary tree height inversion result of the forest stock prediction method of the present invention; Figure 5 This is a schematic diagram of the forest stock prediction system of the present invention.

[0018] The components represented by each number in the attached figure are described as follows: SAR data acquisition module 11, terrain parameter extraction module 12, polarization decomposition estimation module 13, three-stage inversion module 14, and volume prediction module 15. Detailed Implementation

[0019] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0020] Example 1, as Figure 1 and Figure 2 This is a flowchart illustrating the forest stock prediction method of the present invention, wherein the forest stock prediction method includes: S100: Acquire SAR image data of the target area using L-band synthetic aperture radar, and generate a polarimetric interferometric coherence matrix based on the SAR image data.

[0021] Specifically, L-band synthetic aperture radar refers to a microwave remote sensing system with an operating frequency of approximately 1 to 2 GHz and a wavelength of approximately 15 to 30 cm. Its longer wavelength gives it strong vegetation penetration capabilities, enabling it to acquire scattering information from the forest canopy and some dry structures. SAR image data refers to complex image data formed by compressing the synthetic aperture radar echo signal in the range and azimuth directions. Each pixel contains amplitude and phase information.

[0022] For example, the data acquisition scheme for a certain area is as follows: LiDAR UAV, flight altitude 120m, flight speed 2m / s, overlap rate 50%, point cloud density 62 points / m. 2 .

[0023] In some embodiments, SAR image data of a target area is acquired using L-band synthetic aperture radar, and a polarimetric interferometric coherence matrix is ​​generated based on the SAR image data, including: Raw SAR image data is acquired using L-band synthetic aperture radar; the raw SAR image data is finely registered, and the registered image is spectral filtered; based on the spectral filtering results, interferometric coherence calculation is performed to generate an interferometric fringe pattern and a coherence coefficient pattern, and based on the interferometric fringe pattern and the coherence coefficient pattern, the polarization interferometric coherence matrix is ​​constructed.

[0024] Specifically, polarization refers to the polarization mode of radar electromagnetic wave transmission and reception, such as HH, HV, VH, and VV polarization channels; interferometry is a technique that uses the phase difference between two or more images of the same area to obtain spatial structure information; the coherence coefficient is a complex correlation coefficient used to characterize the degree of correlation between two complex images, with a value range of 0 to 1; the polarization interferometric coherence matrix is ​​the result of matrix representation of the interferometric coherence relationship between different polarization channels, which usually contains complex coherence coefficient elements under each polarization combination, and is used to characterize the forest body scattering structure and vertical layer information.

[0025] Specifically, the polarimetric interferometric coherence matrix (PMC) is a 6×6 complex matrix composed of polarimetric interferometric information from two full-polarimetric SAR images, serving as the data foundation for forest height inversion. This PMC contains two 3×3 sub-matrices: the polarimetric coherence matrix T11 of the main image and the polarimetric coherence matrix T22 of the auxiliary image, as well as a matrix Ω describing the polarimetric interferometric coherence between the two. Each element of the T6 matrix contains amplitude and phase information under a specific polarization combination, providing complete physical quantity inputs for subsequent model-based polarization decomposition and three-stage inversion.

[0026] Specifically, firstly, raw SAR image data covering the target area is acquired using L-band synthetic aperture radar. The long-wave characteristics of the L-band give it strong penetration capabilities into the forest canopy, enabling better detection of the canopy's internal structure. Secondly, the primary and secondary images are precisely registered using methods based on correlation coefficients or maximum spectrum estimation to ensure accurate alignment of pixel positions for the same ground feature in both images, which is the foundation of interferometric processing. Then, the registered images undergo spectral filtering, such as using the Goldstein filtering method, extracting common spectral bandwidths in the azimuth and range directions. The filtering window size is typically set to 32×32 to 64×64 pixels to balance noise suppression and spatial resolution loss, thereby suppressing speckle noise and improving the signal-to-noise ratio. Finally, interferometric coherence calculations are performed based on the filtered interferogram, generating interferometric fringe patterns and coherence coefficient maps to characterize phase differences and coherence levels, respectively.

[0027] Preferably, the acquired interferometric fringe patterns and coherence coefficient patterns undergo flat-ground phase removal and terrain-removal phase processing to reduce the interferometric phase difference caused by the different distances between the two antennas and the ground elements, as well as the phase difference caused by terrain influence. Finally, the polarization scattering matrices of the main and auxiliary images are converted into Pauli basis vectors. By calculating the outer product of these vectors and statistically averaging them, the polarization interferometric coherence matrix is ​​finally constructed.

[0028] The generated polarimetric interferometric coherence matrix can effectively fuse polarimetric and interferometric information, preserving the radar signal's sensitivity to the shape and orientation of forest scatterers while also including height information of the canopy's vertical distribution.

[0029] Through the above process, the conversion from raw SAR echo signals to structured polarimetric interferometric features was achieved, providing high-quality data input for subsequent terrain correction, scattering mechanism separation, and forest height inversion. The selection of the L-band ensured effective penetration of the forest canopy, while precise registration and spectral filtering significantly improved the quality and reliability of the interferometric phase. The construction of the polarimetric interferometric coherence matrix fully preserved the joint information of polarization and interferometry, enabling subsequent steps to distinguish different scattering mechanisms and quantify their vertical distribution characteristics, thus laying a data foundation for high-precision prediction of forest stock volume.

[0030] S200: Obtain the digital elevation model of the target area, extract the terrain slope and local incident angle based on the digital elevation model, and calculate the polarization azimuth offset based on the terrain slope.

[0031] Specifically, the digital elevation model is a raster dataset representing the spatial distribution of surface elevation; terrain slope is the steepness and direction of the surface inclination; local incident angle is the angle between the radar beam and the surface normal, which is different from the incident angle of radar ground observation. It integrates the influence of radar geometry and terrain undulation, and directly determines the actual penetration path and backscattering characteristics of the radar beam on the slope; polarization azimuth offset is the rotation angle of the radar polarization direction caused by the terrain slope.

[0032] Specifically, the digital elevation model is used to provide basic data for terrain analysis. The terrain slope and local incident angle are the core parameters describing the geometric relationship between the radar and the ground surface, while the polarization azimuth offset is the key correction quantity for modifying the polarization scattering model.

[0033] In some embodiments, a digital elevation model of the target area is obtained, and based on the digital elevation model, the terrain slope and local incident angle are extracted, and the polarization azimuth offset is calculated based on the terrain slope, including: The terrain slope and aspect values ​​of each pixel are extracted from the digital elevation model; the local incidence angle of each pixel is calculated by combining the imaging geometry parameters of the synthetic aperture radar with the terrain slope and aspect values; and the polarization azimuth offset caused by the terrain is calculated based on the terrain slope, aspect values ​​and local incidence angle.

[0034] Specifically, the terrain slope value is the steepness of the surface inclination, and the terrain aspect value is the direction of the surface inclination. Together with the local incident angle and polarization azimuth offset, they describe the geometric relationship between radar signals and terrain, providing key inputs for terrain correction in polarimetric interferometric SAR processing.

[0035] Specifically, firstly, the terrain slope and aspect values ​​for each pixel are extracted from the digital elevation model. This can be achieved by calculating the partial derivative of the elevation, such as using a difference method to calculate the slope angle and azimuth angle, where the slope angle is obtained through the gradient modulus and the aspect angle is determined through the gradient direction. Then, combining the imaging geometry parameters of the synthetic aperture radar (satellite orbital altitude, radar incident angle) with the aforementioned terrain slope and aspect values, the local incident angle for each pixel is calculated. This calculation is based on the dot product of the radar line-of-sight direction and the surface normal vector. For L-band data, the incident angle typically ranges from 10° to 35°. For example, satellite orbital parameters and imaging time information can be used to calculate the local incident angle through coordinate transformation. Finally, based on the terrain slope value, aspect value, and local incident angle, the polarization azimuth offset caused by the terrain is calculated by substituting into the geometric formula. This geometric formula is based on the geometric principle of polarization direction rotation. For example, the polarization azimuth offset Δψ=arctan[(sinα·sinφs) / (sinα·cosφs·cosθ+cosα·sinθ)], where α is the slope angle, φs is the aspect angle, and θ is the local incident angle.

[0036] Through the above process, the extracted terrain slope and aspect values ​​provide the basis for calculating the local incident angle. The calculation of the local incident angle quantifies the geometric relationship between the radar signal and the ground surface, ensuring the accuracy of the radar beam incident angle. The calculation of the polarization azimuth offset directly compensates for the polarization direction rotation caused by the terrain, enabling subsequent model-based polarization decomposition to more accurately separate surface scattering and secondary scattering components, thereby improving the estimation accuracy of the pure volume decoherence coefficient. This enhances its applicability in complex terrain areas.

[0037] S300: Based on the polarization interference coherence matrix and the polarization azimuth offset, the pure volume decoherence coefficient and volume scattering intensity are estimated using a model-based polarization decomposition method.

[0038] Specifically, the model-based polarization decomposition method represents the observed polarization interference coherence matrix as a linear combination of several scattering mechanisms with clear physical meaning (surface scattering, volume scattering, and secondary scattering), and solves the contribution parameters of each mechanism through optimization algorithms. The core is a mathematical framework that includes polarization azimuth offset correction.

[0039] Specifically, the pure volume decoherence coefficient is the interferometric coherence attenuation caused solely by random volume scattering from the forest canopy, excluding the effects of surface scattering, trunk scattering, and topographic decoherence, and is directly related to forest vertical structure parameters. Volume scattering intensity characterizes the energy contribution of the forest canopy in polarimetric scattering, reflecting comprehensive information on stand density, biomass, and canopy water content. The application of polarimetric azimuth offset is manifested as rotation compensation of the polarimetric interferometric coherence matrix. Through unitary transformation, the observation matrix is ​​converted from the inclined topographic coordinate system to the horizontal reference coordinate system, eliminating the polarimetric basis rotation effect caused by topography and ensuring the physical consistency of the decomposition results.

[0040] In some embodiments, based on the polarization interferometry coherence matrix and the polarization azimuth offset, a model-based polarization decomposition method is used to estimate the pure volume decoherence coefficient and the volume scattering intensity, including: Based on the polarization interference coherence matrix, the contributions of surface scattering, secondary scattering, and volume scattering are calculated respectively. Combining these contributions, the polarization coherence matrix is ​​expressed as a first linear combination of the scattering matrices corresponding to the three scattering mechanisms. Based on these contributions, the polarization interference matrix is ​​decomposed into a second linear combination of the polarization interference matrices corresponding to the three scattering mechanisms. The first linear combination and the second linear combination are combined to form a model matrix describing the target region. Based on the model matrix, a best fit is performed with the goal of minimizing the residual factor between the polarization interference coherence matrix and the model matrix, and the volume decoherence coefficient and the volume scattering intensity are obtained.

[0041] Specifically, the contribution refers to the weight of each scattering mechanism in the total polarization scattering energy, satisfying the normalization constraint condition; the scattering matrix is ​​a 3×3 Hermitian matrix describing the polarization characteristics of a specific scattering mechanism. The surface scattering matrix characterizes the Bragg scattering characteristics of a rough surface, the secondary scattering matrix characterizes the scattering characteristics of the surface-trunk dihedral structure, and the volume scattering matrix characterizes the scattering characteristics of the canopy random orientation dipole cloud.

[0042] Specifically, the model matrix is ​​the integrated model framework that combines the decomposed polarization coherence matrix and the decomposed polarization interference matrix. It is the decomposition result and is used to fully describe the electromagnetic response of the target region in the polarization-interference joint domain. The residual norm minimization refers to the optimization criterion that minimizes the difference in Frobenius norm between the observation matrix and the model matrix. The optimal model parameters are solved through an iterative algorithm.

[0043] Specifically, firstly, the contributions of the three scattering mechanisms are calculated based on the polarimetric interferometry coherence matrix: using polarimetric decomposition techniques, such as Cloude decomposition or Freeman-Durden decomposition, the power contributions of surface scattering, secondary scattering, and volume scattering are extracted from the polarimetric coherence matrix T to obtain the weighting coefficients of each mechanism. These coefficients reflect the relative importance of each scattering mechanism in the total echo.

[0044] Secondly, a first linear combination is constructed, which combines the rotated corrected surface scattering matrix. Secondary scattering matrix Volume scattering matrix With weighting coefficients By performing a linear combination, we obtain the overall polarization coherence matrix model: .in, This is the polarization azimuth offset, used to correct the polarization direction of the surface and secondary scattering.

[0045] Next, a second linear combination is constructed, and a decorrelation coefficient is introduced. And assuming that all interference decoherence is independent of polarization mode, the polarization interference matrix is... It can be decomposed into a linear combination of the polarization interference matrices corresponding to the three scattering mechanisms: .here This is the pure volume decoherence coefficient to be determined, which is directly related to the forest canopy structure.

[0046] Then, the first linear combination and the second linear combination are integrated into a complete polarization interferometric coherence matrix model: Finally, the best fit is achieved by minimizing the residuals; in other words, the best fit is obtained using the observed polarization interferometry coherence matrix. With model matrix To minimize the difference, construct the objective function. The Frobenius norm is used to iteratively solve for the optimal pure volume decoherence coefficient through a nonlinear optimization algorithm. Volume scattering intensity (by Volume scattering matrix The power is jointly determined by the scattering mechanism, and the solution term consists of three volume decoherence terms related to the scattering mechanism. .

[0047] Through the model-based polarization decomposition described above, a physical interpretation and quantitative estimation of polarization interferometry information are achieved. The introduction of polarization azimuth offset enables the surface scattering and secondary scattering models to adapt to complex terrain, significantly improving the physical rationality and accuracy of the decomposition results. By constructing and optimizing a complete model matrix incorporating three scattering mechanisms, this method can directly and robustly estimate the pure volume decoherence coefficient. This method, which estimates volume scattering intensity, avoids the error accumulation that may be introduced by step-by-step estimation in traditional methods. The estimated... It is directly related to forest canopy height and density, while volume scattering intensity is related to canopy biomass, providing high-quality and reliable input features for subsequent forest canopy height inversion and volume prediction.

[0048] S400: Combining the pure volume decoherence coefficient, the local incident angle, and the interferometric fringe pattern generated from the SAR image data, a three-layer S-RVoG model is introduced to perform a three-stage inversion to obtain the forest canopy height and extinction coefficient of the target area, and a multidimensional feature vector is constructed.

[0049] Specifically, the three-layer S-RVoG model is an extended stochastic surface coverage model that adds a tree trunk scattering layer and introduces terrain slope correction on the basis of the classic two-layer model, thereby more accurately describing the scattering mechanism of forest canopy under complex terrain.

[0050] The three-stage inversion method is a forest height inversion method based on geometrical optics and a three-layer S-RVoG model. It separates the land surface and canopy phases by analyzing the distribution of polarization coherence in the complex plane. Among them, forest canopy height is the core parameter of forest vertical structure, extinction coefficient characterizes the attenuation of radar waves in the canopy, interferometric fringe pattern is a phase difference image generated by SAR interferometry, which includes the contributions of terrain phase and vegetation phase, and multidimensional feature vector is structured data used as input for machine learning models.

[0051] In some embodiments, the three-layer S-RVoG model includes: A surface scattering model is used to characterize Bragg scattering on a rough surface; a secondary scattering model is used to characterize secondary scattering between the surface and the tree trunk; and a volume scattering model is used to characterize random volume scattering from the forest canopy.

[0052] Specifically, the three-layer S-RVoG model is an improved random volume scattering model for mountainous forest scenes. Based on the traditional RVOG model, it introduces a terrain slope correction factor and divides the vertical structure of the forest into a three-layer architecture: surface scattering layer (rough surface), secondary scattering layer (surface-trunk interaction), and volume scattering layer (canopy random medium).

[0053] The surface scattering model is used to characterize Bragg scattering on rough surfaces. In the model, it can be represented as a coherence matrix, with its energy primarily concentrated in the co-polarized channels (HH and VV), and influenced by polarization azimuth offset.

[0054] The secondary scattering model is used to characterize the dihedral scattering between the tree trunk and the ground surface. Radar waves first illuminate the tree trunk, then reflect off the ground, and finally propagate back. It is also represented by a coherence matrix, with its energy mainly concentrated in the same polarization channel, and its phase center located at the ground surface, similarly affected by polarization azimuth offset.

[0055] Among them, the volume scattering model is used to characterize random volume scattering within the forest canopy (leaves, branches, etc.). This is the main contributor to the forest canopy, and its coherence matrix describes the response of randomly distributed scatterers within the canopy to radar waves, with the phase center located inside the canopy.

[0056] In some embodiments, the pure volume decoherence coefficient, the local incident angle, and the interferometric fringe pattern generated from the SAR image data are combined and introduced into a three-layer S-RVoG model for three-stage inversion to obtain the forest canopy height and extinction coefficient of the target area, and a multidimensional feature vector is constructed, including: Calculate the local incident angle and, based on the local incident angle, calculate several complex coherence coefficients corresponding to different polarization modes, and fit a straight line on the complex plane. Remove vegetation bias by using the two points obtained from the intersection of the fitted line from the first step and the unit circle of the complex plane, and select the phase angle corresponding to one of these points as the surface phase according to the principle of maximum vegetation bias. Using the surface phase, the known vertical wavenumber, the volume decoherence coefficient, and the local incident angle, combined with the physical equations of the three-layer S-RVoG model, establish a two-dimensional lookup table for the volume decoherence coefficient, the forest canopy height, and the extinction coefficient, and obtain the forest canopy height from the table. Based on the forest canopy height, the volume scattering intensity, the local incident angle, and the tree species classification information of the target area, construct a multidimensional feature vector.

[0057] Specifically, the implementation process of three-stage inversion and construction of multi-dimensional feature vectors based on the three-layer S-RVoG model includes the following three stages: First, in the straight-line fitting stage, the local incident angle is obtained, and based on this, a series of complex coherence coefficients for different polarizations (HH, HV, VV, etc.) are calculated. These complex coherence coefficients are complex numbers containing amplitude and phase information. The complex coherence coefficients for all polarizations are plotted within the unit circle of the complex plane. Due to noise and model errors, these points are not perfectly collinear, but they will be distributed within a narrow linear region. The overall least squares method can be used to fit these points, minimizing the perpendicular distance from each polarization complex coherence coefficient to the straight line, resulting in a straight line with a goodness of fit R. 2 A value greater than 0.95 typically indicates that the straight line model is valid. This straight line theoretically connects the surface phase point and the canopy phase point.

[0058] Then, after removing the vegetation bias stage, the fitted straight line intersects the unit circle in the complex plane at two points (corresponding to the possible surface phase and canopy phase, respectively). Based on the "maximum vegetation bias principle," one of these points is selected as the surface phase point. This principle is based on physical understanding that surface scattering typically has higher coherence (closer to the edge of the unit circle) and its phase is relatively stable. Therefore, the point with higher coherence among the intersections with the unit circle, or the point that, according to the model, is closer to surface characteristics, is selected; its corresponding phase angle is the surface phase.

[0059] For example, taking a coniferous forest pixel with a slope of 25° as an example, HH polarization γ=0.65·exp(i0.85), HV polarization γ=0.48·exp(i1.12), VV polarization γ=0.58·exp(i0.92), the linear fitting yields a=0.82 and b=0.15, the phases of the intersection points with the unit circle are 0.78 radians and 1.35 radians respectively, and the corresponding |γv| are 0.38 and 0.62 respectively. Based on the principle of maximum vegetation deviation, the surface phase = 1.35 radians is selected.

[0060] Subsequently, in the parameter estimation and lookup table stage, a two-dimensional lookup table is pre-constructed using known vertical wavenumbers (calculated from SAR system geometric parameters), surface phase, volumetric decoherence coefficient, and local incidence angle, combined with a three-layer S-RVoG model (which establishes a functional relationship between the volumetric decoherence coefficient and forest canopy height and extinction coefficient). This two-dimensional lookup table is discretely sampled within the possible range of forest canopy height and extinction coefficient values, with each sampled result corresponding to a theoretically calculated volumetric decoherence coefficient value. Finally, by matching and interpolating the actual estimated volumetric decoherence coefficient values ​​with the theoretical values ​​in the two-dimensional lookup table, the corresponding forest canopy height and extinction coefficient can be retrieved.

[0061] Furthermore, based on the inverted forest canopy height and extinction coefficient, as well as the volume scattering intensity and local incident angle obtained in previous steps, and combined with the tree species classification information of the target area (which can be pre-classified through polarization decomposition features such as entropy and anisotropy, or obtained from external data), feature fusion is performed to construct a feature vector containing multi-dimensional information such as forest height, extinction coefficient, volume scattering intensity, local incident angle, and tree species classification. This vector comprehensively describes the vertical structure, density, geometric characteristics, and type of the forest, providing input for subsequent machine learning modeling.

[0062] Through the above process, the three-layer S-RVoG model decomposes the data using three scattering mechanisms, providing a solid physical foundation for the three-stage inversion. The first stage uses geometric properties to separate the surface and canopy phases; the second stage determines the surface phase using physical criteria; and the third stage utilizes a pre-computed lookup table to achieve efficient and accurate conversion from polarimetric interferometry parameters to forest structure parameters. The resulting multidimensional feature vector integrates various remote sensing information, laying a crucial data foundation for high-precision prediction of forest stock volume.

[0063] In some implementations, a two-dimensional lookup table is established, relating the pure volume decoherence coefficient to the forest canopy height and the extinction coefficient, including: By combining historical data and quadrat data, the range of forest canopy height and extinction coefficient in the target area are obtained; within the range of forest canopy height and the range of extinction coefficient, grid sampling is performed with a preset step size; the grid sampling results are traversed, and the physical equations of the three-layer S-RVoG model are used to calculate the theoretical pure volume decoherence coefficient; multiple theoretical pure volume decoherence coefficients and grid sampling results are structured and stored to form a two-dimensional lookup table in the form of a two-dimensional table.

[0064] Specifically, historical data refers to previous forest resource survey data, remote sensing inversion products, or ecological observation records of the target area, used to define reasonable ranges for forest structure parameters; sample plot data consists of observational data of typical forest stands obtained from ground measurements or airborne lidar, including parameters such as canopy height, leaf area index, and stand density, providing prior constraints for setting the lookup table range; the physical equations of the three-layer S-RVoG model are mathematical expressions describing the quantitative relationship between the pure volume decoherence coefficient and forest canopy height, extinction coefficient, and local incident angle, derived based on the theory of electromagnetic wave radiative transfer in random media; structured storage organizes the calculation results in a two-dimensional table format with row and column indexes, supporting fast keyword-based retrieval and interpolation queries.

[0065] Specifically, firstly, by combining historical data and quadrat data, the range of forest canopy height and extinction coefficient for the target area is obtained. For example, for the Qinling Mountains region, the range of forest canopy height might be set to 5 to 35 meters, and the range of extinction coefficient might be set to 0.1 Np / m to 1.0 Np / m. Secondly, within the determined range of forest canopy height and extinction coefficient, grid sampling is performed with a preset step size. For example, the height step size can be set to 0.5 meters, and the extinction coefficient step size can be set to 0.01 Np / m, thereby generating a two-dimensional grid, where each node corresponds to a combination of (height, extinction coefficient). Then, the grid sampling results are traversed, and each combination of (height, extinction coefficient) is substituted into the physical equation of the three-layer S-RVoG model for calculation. This physical equation describes the relationship between the volume decoherence coefficient and (height, extinction coefficient), and the expression includes other known parameters, such as surface phase, vertical wavenumber, local incident angle, and terrain slope-related parameters.

[0066] For example, the calculation formula is shown below: ; in, The theoretical pure volume decoherence coefficient, whose modulus represents the coherence intensity of volume scattering, is directly related to the forest canopy height. and extinction coefficient Related, The imaginary unit, It is a height correction factor, dimensionless, and is a terrain correction parameter introduced in the three-layer S-RVoG model. It is used to adjust the projection relationship of the canopy height on the sloping surface, so that the model is more in line with the actual geometry. It is the local incident angle, measured in radians, and refers to the angle between the radar beam and the normal to the Earth's surface. The vertical wavenumber, expressed in rad / m, represents the sensitivity of radar waves to changes in vertical altitude in interferometric SAR. It is the slope angle of the terrain, measured in radians, which indicates the steepness of the terrain's inclination. This is a correction value for the vertical wavenumber, expressed in rad / m, and may be based on the local angle of incidence. Fine-tuning is made to the terrain slope to more accurately reflect the vertical sensitivity of radar waves in complex terrain. is the base of the natural logarithm.

[0067] Specifically, by traversing all grid points, the corresponding theoretical values ​​can be calculated. Finally, the calculated theoretical pure volume decoherence coefficients and their corresponding gridded sampling results (i.e., h and σ values) are structurally stored to form a two-dimensional lookup table. This table can be a matrix, with row indices corresponding to forest canopy height and column indices corresponding to extinction coefficients. The matrix cells store the calculated theoretical values. Values. The storage format should facilitate quick interpolation and lookup later.

[0068] Specifically, in practice, bilinear interpolation can be used to obtain the inverted forest canopy height and extinction coefficient.

[0069] Through the above process, the established two-dimensional lookup table enables the discretization and pre-calculation of the physical equations of the three-layer S-RVoG model. The combination of historical data and sample plot data ensures the rationality of the lookup table's parameter range and its regional applicability. Gridded sampling and traversal calculation ensure the completeness of the lookup table. Structured storage eliminates the need for repeated complex physical model solutions in subsequent inversion processes; simply matching and interpolating the actual estimated values ​​with the lookup table quickly yields the corresponding forest canopy height and extinction coefficient. This significantly improves the efficiency of the inversion algorithm, reduces computational complexity, and ensures the physical consistency of the inversion results.

[0070] S500: Input the multidimensional feature vector into the pre-trained forest stock volume estimation model to generate forest stock volume prediction results.

[0071] Specifically, the pre-trained forest stock volume estimation model is a regression prediction model built based on machine learning algorithms. It is used to learn the nonlinear mapping relationship between multidimensional feature vectors and forest stock volume through historical sample data. The forest stock volume prediction result is a quantitative estimate of the timber volume per unit area of ​​forest stands in the target region, typically expressed in cubic meters per hectare (m³). 3 The unit is / ha).

[0072] Specifically, the implementation process includes two stages: model training and prediction. First, model training is conducted: based on concurrently conducted ground quadrat survey data, the measured volume of each quadrat is used as the target variable. Simultaneously, the multidimensional feature vector generated in the previous steps (including forest canopy height) is extracted from the corresponding area of ​​that quadrat. Extinction coefficient The system uses features such as volume scattering intensity, local incident angle, and tree species classification as input features. Then, a random forest regression algorithm is used for training. Bootstrap sampling is used to randomly sample multiple subsets from the training data, and a decision tree is built for each subset. Finally, the average of the predictions from all decision trees is taken as the final output. During training, out-of-bag (OOB) data (samples remaining after each decision tree construction) is used to estimate the generalization error and optimize key parameters (number of trees). and the number of features considered for each tree Once training is complete, a stable forest stock estimation model is obtained.

[0073] Secondly, prediction is performed: the multi-dimensional feature vector corresponding to each pixel in the study area is input into a pre-trained forest stock estimation model. The model uses its internal decision tree combination to reason and outputs the predicted forest stock value for that pixel. Finally, the predicted values ​​of all pixels are combined to generate a forest stock distribution map covering the target area, thus completing the forest stock prediction.

[0074] In some embodiments, the multidimensional feature vector is input into a pre-trained forest stock volume estimation model to generate a forest stock volume prediction result. Prior to this, the method further includes: A sample dataset is obtained, which contains multiple samples, each sample being constructed based on the paradigm of the multidimensional feature vector and the corresponding measured volume data of the ground quadrat; the sample dataset is divided into a training sample set and a validation sample set; the random forest regression model is trained using the multidimensional feature vector in the training sample set as input and the corresponding measured volume data of the ground quadrat as output, and the model performance is optimized until the prediction accuracy on the validation sample set meets a preset threshold, thereby establishing the forest volume estimation model.

[0075] In some embodiments, the prediction accuracy includes the coefficient of determination and the root mean square error.

[0076] Specifically, the sample dataset is the basic data resource for model training. It consists of a pair of multidimensional feature vectors and ground-measured timber volume data. The paradigm of the multidimensional feature vectors refers to the standardized representation of the feature vectors, including the numerical encoding of forest canopy height, extinction coefficient, volume scattering intensity, local incident angle, and tree species classification information. The ground-measured timber volume data is the measured value of timber volume per unit area obtained through standard forestry survey methods, which serves as the ground truth label for model training.

[0077] Specifically, the division of the training and validation sample sets follows general principles, employing random sampling or stratified spatial sampling, with a ratio of 7:3 or 8:2 to ensure the representativeness of the two types of samples in both feature space and geographic space. The coefficient of determination (R²) 2 The value is used to measure the model's ability to explain data variation, ranging from 0 to 1. The closer to 1, the higher the goodness of fit. The root mean square error (RMSE) measures the average deviation between the predicted and measured values. The unit is the same as the volume. The smaller the value, the higher the accuracy. The preset threshold is an acceptance criterion set according to the accuracy standards of forestry surveys or application requirements.

[0078] Specifically, the first step is to acquire a sample dataset. For example, 40 ground quadrats of 30m x 30m are set up, and the diameter at breast height (DBH) and tree height of each tree are measured using the angle gauge control measuring method. Figure 3 (As shown), refer to the local one-yuan timber volume table to calculate the volume of a single timber, and sum them up to obtain the measured value of the sample plot volume, which ranges from 65-320m³. 3 / ha. Then, for each sample plot center corresponding to a SAR pixel, the multidimensional feature vector generated in the previous steps is extracted, forming 40 paired samples. Next, stratified random sampling is used, according to accumulation level (low <120, medium 120-200, high >200m). 3 / ha) stratification, with 70% randomly selected as the training set (28 groups) and 30% as the validation set (12 groups) within each stratum, to ensure that the proportion of samples in the training set and validation set is consistent between the two strata.

[0079] Furthermore, the model parameters of the random forest regression model are initialized, using the training set feature vectors as input and the measured accumulation as output. 200 decision trees are fitted, with each tree constructed using Bootstrap resampling to build a training subset. When splitting nodes, 4 dimensions are randomly selected from the 5-dimensional features, with minimizing the mean squared error (MSE) as the splitting criterion. Hyperparameter optimization is performed: 5-fold cross-validation is used, with the validation set R... 2 The goal is to maximize the optimal combination of parameters. For example, based on... Figure 4 The accuracy of the tree height inversion results was verified. The average canopy height in the sample plot area was 15.5732m, the average measured height was 13.2716m, and the RMSE result was 2.4633m, as shown in Table 1 below: Table 1. Exemplary Forest Tree Height Inversion Accuracy Analysis Table

[0080] Through the above process, the data-driven construction and rigorous validation of the forest stock estimation model were achieved, ensuring the reliability and generalization ability of the model from training to deployment. The dual evaluation of the coefficient of determination and root mean square error provides a comprehensive measure of accuracy, thereby supporting the technical effects of reducing the impact of terrain and improving prediction accuracy at the model construction and validation levels.

[0081] In summary, the forest stock prediction method provided by this invention has the following technical effects: SAR image data of the target area is acquired using L-band synthetic aperture radar, and a polarimetric interferometric coherence matrix is ​​constructed based on the SAR image data. A digital elevation model (DEM) corresponding to the target area is obtained, from which terrain slope and local incident angle information are extracted, and the polarimetric azimuth offset parameter is calculated based on the terrain slope. Based on the polarimetric interferometric coherence matrix and the polarimetric azimuth offset parameter, a model-driven polarimetric decomposition method is used to estimate the volume decoherence coefficient and volume scattering intensity. The volume decoherence coefficient, local incident angle, and interferometric fringe pattern generated from the SAR image data are fused and introduced into a three-layer S-RVoG model for staged inversion processing to obtain the forest canopy height and extinction coefficient of the target area, and a multidimensional feature vector is constructed accordingly. The multidimensional feature vector is input into a pre-trained forest volume estimation model, and the corresponding forest volume prediction result is output, thereby achieving the technical effect of reducing the influence of terrain and improving prediction accuracy.

[0082] Example 2, as Figure 5 This is a schematic diagram of the forest stock prediction system of the present invention. For example, Figure 1 The flowchart of the forest stock prediction method of the present invention can be seen as follows: Figure 5 The structure shown is implemented.

[0083] Based on the same concept as the forest stock prediction method in the above embodiments, the present invention also provides a forest stock prediction system comprising: The SAR data acquisition module 11 is used to acquire SAR image data of the target area through L-band synthetic aperture radar, and generate a polarimetric interferometric coherence matrix based on the SAR image data.

[0084] The terrain parameter extraction module 12 is used to obtain the digital elevation model of the target area, extract the terrain slope and local incident angle based on the digital elevation model, and calculate the polarization azimuth offset based on the terrain slope.

[0085] The polarization decomposition estimation module 13 is used to estimate the pure volume decoherence coefficient and volume scattering intensity based on the polarization interference coherence matrix and the polarization azimuth offset using a model-based polarization decomposition method.

[0086] The three-stage inversion module 14 is used to combine the pure volume decoherence coefficient, the local incident angle, and the interferometric fringe pattern generated from the SAR image data, and introduce them into a three-layer S-RVoG model for three-stage inversion to obtain the forest canopy height and extinction coefficient of the target area, and to construct a multi-dimensional feature vector.

[0087] The forest stock volume prediction module 15 is used to input the multidimensional feature vector into a pre-trained forest stock volume estimation model to generate forest stock volume prediction results.

[0088] In some embodiments, the execution steps of the SAR data acquisition module 11 include: Raw SAR image data is acquired using L-band synthetic aperture radar.

[0089] The original SAR image data is finely registered, and the registered image is then subjected to spectral filtering.

[0090] Interference coherence calculation is performed based on the spectral filtering results to generate an interference fringe pattern and a coherence coefficient pattern. Based on the interference fringe pattern and the coherence coefficient pattern, the polarization interference coherence matrix is ​​constructed.

[0091] In some embodiments, the execution steps of the terrain parameter extraction module 12 include: The terrain slope and aspect values ​​of each pixel are extracted from the digital elevation model.

[0092] By combining the imaging geometry parameters of the synthetic aperture radar with the terrain slope and aspect values, the local incidence angle of each pixel is calculated.

[0093] The polarization azimuth offset caused by the terrain is calculated based on the terrain slope value, terrain aspect value and local incident angle.

[0094] In some embodiments, the execution steps of the polarization decomposition estimation module 13 include: Based on the polarization interference coherence matrix, the contributions of the three scattering mechanisms—surface scattering, secondary scattering, and volume scattering—are calculated respectively.

[0095] Based on the aforementioned contribution, the polarization coherence matrix is ​​expressed as the first linear combination of the scattering matrices corresponding to the three scattering mechanisms: surface scattering, secondary scattering, and volume scattering.

[0096] Based on the aforementioned contribution, the polarization interference matrix is ​​decomposed into a second linear combination of the polarization interference matrices corresponding to the three scattering mechanisms: surface scattering, secondary scattering, and volume scattering.

[0097] The first linear combination and the second linear combination are combined to form a model matrix for describing the target region.

[0098] Based on the model matrix, the volume decoherence coefficient and the volume scattering intensity are obtained by performing an optimal fit between the polarization interference coherence matrix and the model matrix with the goal of minimizing the residual universality.

[0099] In some embodiments, the three-layer S-RVoG model includes: A surface scattering model is used to characterize Bragg scattering on rough surfaces of the Earth.

[0100] A secondary scattering model is used to characterize secondary scattering between the ground surface and the tree trunk.

[0101] A volume scattering model is used to characterize random volume scattering in the forest canopy.

[0102] In some embodiments, the execution steps of the three-stage inversion module 14 include: Calculate the local incident angle, and based on the local incident angle, calculate several complex coherence coefficients corresponding to different polarization modes, and then fit a straight line on the complex plane.

[0103] To remove vegetation bias, the two points obtained by the intersection of the fitted straight line in the first step and the unit circle of the complex plane are used. The phase angle corresponding to one of these points is selected as the surface phase based on the principle of maximum vegetation bias.

[0104] Using the surface phase, the known vertical wavenumber, the volume decoherence coefficient, and the local incident angle, combined with the physical equations of the three-layer S-RVoG model, a two-dimensional lookup table is established for the volume decoherence coefficient, the forest canopy height, and the extinction coefficient, and the forest canopy height is obtained by looking up the table.

[0105] A multidimensional feature vector is constructed based on the forest canopy height, the volume scattering intensity, the local incident angle, and the tree species classification information of the target area.

[0106] In some embodiments, the execution steps of the three-stage inversion module 14 further include: Obtain a sample dataset, which contains multiple samples, each sample being composed of the paradigm of the multidimensional feature vector and the corresponding measured accumulation data of the ground quadrat.

[0107] The sample dataset is divided into a training sample set and a validation sample set.

[0108] Using the multidimensional feature vectors in the training sample set as input and the corresponding measured stock volume data of ground quadrats as output, the random forest regression model is trained and its performance is optimized until the prediction accuracy on the validation sample set meets a preset threshold, thus establishing the forest stock volume estimation model.

[0109] In some embodiments, the prediction accuracy includes the coefficient of determination and the root mean square error.

[0110] In some embodiments, the execution steps of the three-stage inversion module 14 further include: By combining historical data and quadrat data, the range of forest canopy height and extinction coefficient in the target area were obtained.

[0111] Within the range of forest canopy height and the range of extinction coefficient, grid sampling is performed with a preset step size.

[0112] The gridded sampling results are traversed, and the physical equations of the three-layer S-RVoG model are used to calculate the theoretical pure volume decoherence coefficient.

[0113] The theoretical pure volume decoherence coefficients and the gridded sampling results are stored in a structured manner to form a two-dimensional lookup table in the form of a two-dimensional table.

[0114] It should be understood that the focus of the embodiments mentioned in this specification is their difference from other embodiments. The specific embodiments in the aforementioned Embodiment 1 are also applicable to the forest stock prediction system described in Embodiment 2. For the sake of brevity, they will not be elaborated further here.

[0115] It should be understood that the embodiments disclosed in this invention and the above description enable those skilled in the art to implement this invention. However, this invention is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be included within the protection scope of this invention.

Claims

1. A method for predicting forest stock volume, characterized in that, include: SAR image data of the target area is acquired using L-band synthetic aperture radar, and a polarimetric interferometric coherence matrix is ​​generated based on the SAR image data. Obtain the digital elevation model of the target area, extract the terrain slope and local incident angle based on the digital elevation model, and calculate the polarization azimuth offset based on the terrain slope. Based on the polarization interference coherence matrix and the polarization azimuth offset, a model-based polarization decomposition method is used to estimate the pure volume decoherence coefficient and the volume scattering intensity. Combining the pure volume decoherence coefficient, the local incident angle, and the interferometric fringe pattern generated from the SAR image data, a three-layer S-RVoG model is introduced to perform a three-stage inversion to obtain the forest canopy height and extinction coefficient of the target area, and a multidimensional feature vector is constructed. The multidimensional feature vector is input into a pre-trained forest stock volume estimation model to generate forest stock volume prediction results.

2. The forest stock prediction method as described in claim 1, characterized in that, Using L-band synthetic aperture radar, SAR image data of the target area is acquired, and a polarimetric interferometric coherence matrix is ​​generated based on the SAR image data, including: Raw SAR image data is acquired using L-band synthetic aperture radar; The original SAR image data is finely registered, and the registered image is then subjected to spectral filtering. Interference coherence calculation is performed based on the spectral filtering results to generate an interference fringe pattern and a coherence coefficient pattern. Based on the interference fringe pattern and the coherence coefficient pattern, the polarization interference coherence matrix is ​​constructed.

3. The forest stock prediction method as described in claim 1, characterized in that, Obtain a digital elevation model of the target area; extract the terrain slope and local incident angle based on the digital elevation model; and calculate the polarization azimuth offset based on the terrain slope, including: Extract the terrain slope and aspect values ​​of each pixel from the digital elevation model; By combining the imaging geometry parameters of the synthetic aperture radar with the terrain slope value and terrain aspect value, the local incident angle of each pixel is calculated; The polarization azimuth offset caused by the terrain is calculated based on the terrain slope value, terrain aspect value and local incident angle.

4. The forest stock prediction method as described in claim 1, characterized in that, Based on the polarization interferometry coherence matrix and the polarization azimuth offset, a model-based polarization decomposition method is used to estimate the pure volume decoherence coefficient and volume scattering intensity, including: Based on the polarization interference coherence matrix, the contributions of the three scattering mechanisms—surface scattering, secondary scattering, and volume scattering—are calculated respectively. Based on the aforementioned contribution, the polarization coherence matrix is ​​expressed as the first linear combination of the scattering matrices corresponding to the three scattering mechanisms: surface scattering, secondary scattering, and volume scattering. Based on the aforementioned contribution, the polarization interference matrix is ​​decomposed into a second linear combination of the polarization interference matrices corresponding to the three scattering mechanisms: surface scattering, secondary scattering, and volume scattering. The first linear combination and the second linear combination are combined to form a model matrix for describing the target region; Based on the model matrix, the best fit is performed with the goal of minimizing the residual common factor between the polarization interference coherence matrix and the model matrix, and the volume decoherence coefficient and the volume scattering intensity are obtained by solving.

5. The forest stock prediction method as described in claim 1, characterized in that, The three-layer S-RVoG model includes: A surface scattering model, which is used to characterize Bragg scattering on a rough surface of the Earth. A secondary scattering model, which is used to characterize secondary scattering between the ground surface and the tree trunk; A volume scattering model is used to characterize random volume scattering in the forest canopy.

6. The forest stock prediction method as described in claim 5, characterized in that, Combining the pure volume decoherence coefficient, the local incident angle, and the interferometric fringe pattern generated from the SAR image data, a three-layer S-RVoG model is introduced for three-stage inversion to obtain the forest canopy height and extinction coefficient of the target area, and a multi-dimensional feature vector is constructed, including: Calculate the local incident angle, and based on the local incident angle, calculate several complex coherence coefficients corresponding to different polarization modes, and then fit a straight line on the complex plane; To remove vegetation bias, the two points obtained by the intersection of the fitted straight line in the first step and the unit circle of the complex plane are used. The phase angle corresponding to one of these points is selected as the surface phase based on the principle of maximum vegetation bias. Using the surface phase, the known vertical wavenumber, the pure volume decoherence coefficient, and the local incident angle, combined with the physical equations of the three-layer S-RVoG model, a two-dimensional lookup table is established for the pure volume decoherence coefficient, the forest canopy height, and the extinction coefficient, and the forest canopy height is obtained by looking up the table. A multidimensional feature vector is constructed based on the forest canopy height, the volume scattering intensity, the local incident angle, and the tree species classification information of the target area.

7. The forest stock prediction method as described in claim 1, characterized in that, The multidimensional feature vector is input into a pre-trained forest stock volume estimation model to generate a forest stock volume prediction result. Prior to this, the process also includes: Obtain a sample dataset, which contains multiple samples, each sample being composed of the paradigm of the multidimensional feature vector and the corresponding measured accumulation data of the ground quadrat; The sample dataset is divided into a training sample set and a validation sample set; Using the multidimensional feature vectors in the training sample set as input and the corresponding measured stock volume data of ground quadrats as output, the random forest regression model is trained and its performance is optimized until the prediction accuracy on the validation sample set meets a preset threshold, thus establishing the forest stock volume estimation model.

8. The forest stock prediction method as described in claim 7, characterized in that, The prediction accuracy includes the coefficient of determination and the root mean square error.

9. The forest stock prediction method as described in claim 6, characterized in that, Establish a two-dimensional lookup table relating the pure volume decoherence coefficient to the forest canopy height and the extinction coefficient, including: By combining historical data and quadrat data, the range of forest canopy height and extinction coefficient in the target area were obtained; Within the range of forest canopy height and the range of extinction coefficient, grid sampling is performed with a preset step size; The gridded sampling results are traversed and substituted into the physical equations of the three-layer S-RVoG model to calculate the theoretical pure volume decoherence coefficient; The theoretical pure volume decoherence coefficients and the gridded sampling results are stored in a structured manner to form a two-dimensional lookup table in the form of a two-dimensional table.

10. A forest stock volume prediction system, characterized in that, The system for implementing the forest stock prediction method according to any one of claims 1 to 9 comprises: The SAR data acquisition module is used to acquire SAR image data of the target area through L-band synthetic aperture radar, and generate a polarimetric interferometric coherence matrix based on the SAR image data. The terrain parameter extraction module is used to obtain the digital elevation model of the target area, extract the terrain slope and local incident angle based on the digital elevation model, and calculate the polarization azimuth offset based on the terrain slope. The polarization decomposition estimation module is used to estimate the pure volume decoherence coefficient and volume scattering intensity based on the polarization interference coherence matrix and the polarization azimuth offset using a model-based polarization decomposition method. The three-stage inversion module is used to combine the pure volume decoherence coefficient, the local incident angle, and the interferometric fringe pattern generated from the SAR image data, and introduce them into a three-layer S-RVoG model for three-stage inversion to obtain the forest canopy height and extinction coefficient of the target area, and construct a multi-dimensional feature vector. The forest stock volume prediction module is used to input the multidimensional feature vector into a pre-trained forest stock volume estimation model to generate forest stock volume prediction results.

Images