A method for estimating bamboo forest biomass and carbon sink capacity based on image analysis

Abstract

The application discloses a bamboo forest biomass and carbon sink capacity estimation method based on image analysis and belongs to the technical field of bamboo forest biomass estimation. The method comprises the following steps: S1, obtaining multi-source remote sensing data of a target bamboo forest area, wherein the multi-source remote sensing data comprises unmanned aerial vehicle hyperspectral images, satellite-borne laser radar point cloud data and time-series multi-spectral satellite images; S2, pre-processing and feature extraction are performed on the multi-source remote sensing data, and a multi-scale feature data set is constructed, wherein the multi-scale feature data set comprises a spectral feature set, a structural feature set and a time-series feature set, and a spatial scale conversion relationship between the spectral feature set and the structural feature set is established through a feature-level fusion strategy; and S3, a vegetation index time-series curve is reconstructed based on the time-series multi-spectral satellite images, and phenological phase recognition is implemented. On the basis of realizing bamboo forest biomass and carbon sink capacity estimation, the method also breaks through the physical limitation of a single sensor and improves the ecological rationality of biomass inversion and carbon sink capacity conversion.

Description

Technical Field

[0001] This invention relates to the field of bamboo forest biomass estimation technology, and more specifically, to a method for estimating bamboo forest biomass and carbon sink based on image analysis. Background Technology

[0002] Bamboo forest biomass refers to the total weight of all parts of a bamboo forest, including culms, branches, leaves, roots, and rhizomes. It is an important indicator for assessing the carbon sequestration function and ecological value of bamboo forests.1 Its size is affected by various factors such as bamboo species, forest age, site conditions, and density, and can be measured at both the individual culm and stand (unit area) levels.

[0003] Existing publicly available bamboo forest biomass and carbon sink estimation schemes suffer from significant bottlenecks due to their reliance on a single data source. Most schemes depend on a single optical remote sensing image. Since bamboo forests have a highly closed single-layer structure, optical vegetation indices quickly saturate at medium biomass levels, failing to effectively reflect the dynamic changes in high biomass areas. While some schemes incorporate SAR or a single airborne LiDAR, SAR has weak penetration in high-density bamboo thickets, and spaceborne LiDARs such as GEDI / ICESat-2 only provide discrete sampling points, making it difficult to directly generate spatially continuous coverage maps. This results in numerous spatial interpolation errors in large-scale mapping, failing to meet the needs of precision forestry. Summary of the Invention

[0004] In view of the problems existing in the prior art, the purpose of this invention is to provide a method for estimating bamboo forest biomass and carbon sink based on image analysis. This invention not only realizes the estimation of bamboo forest biomass and carbon sink, but also breaks through the physical limitations of a single sensor and improves the ecological rationality of biomass inversion and carbon sink conversion.

[0005] To solve the above problems, the present invention adopts the following technical solution: A method for estimating bamboo forest biomass and carbon sequestration based on image analysis includes: S1. Acquire multi-source remote sensing data of the target bamboo forest area, including UAV hyperspectral imagery, spaceborne lidar point cloud data, and time-series multispectral satellite imagery. S2. Preprocess and extract features from the multi-source remote sensing data to construct a multi-scale feature dataset. The multi-scale feature dataset includes a spectral feature set, a structural feature set, and a temporal feature set. A spatial scale transformation relationship is established between the spectral feature set and the structural feature set through a feature-level fusion strategy. S3. Based on the time-series multispectral satellite imagery, reconstruct the vegetation index time-series curve, perform phenological phase identification, determine the key phenological period of bamboo forest, construct a phenological period adaptive parameter lookup table, and determine the feature fusion weight based on the phenological period adaptive parameter lookup table. Perform dynamic weighted fusion on the multi-scale feature dataset to generate a fusion feature set after phenological period correction. S4. Obtain digital elevation model data of the target bamboo forest area, extract terrain morphology features and terrain shadow coefficient, and construct a terrain parameter feature set; based on the fused feature set and the terrain parameter feature set, use a pre-trained conditional generative adversarial network to implement terrain interference correction and generate standardized spectral and structural features. S5. Input the standardized spectrum and structural features into the pre-trained bamboo forest biomass estimation model, output the bamboo forest aboveground biomass estimation results, and convert the aboveground biomass into carbon sink based on the carbon content of the target bamboo species.

[0006] Compared with the prior art, the advantages of this invention are: (1) This invention achieves three-dimensional complementarity of bamboo canopy biochemical components and vertical structure through the air-space-ground collaboration of UAV hyperspectral, spaceborne LiDAR point cloud and time-series multispectral satellite imagery. To address the problem of spatial discontinuity of discrete sampling points in spaceborne LiDAR, a photon separation algorithm based on density threshold and normalized height calculation are used to accurately extract the relative height quantile features of bamboo canopy. Combined with the continuous narrow band reflectance of hyperspectral 400-1700nm, a competitive adaptive reweighted sampling-continuous projection algorithm is used to accurately reduce the dimensionality of high-dimensional data, eliminate redundant bands, and extract the specific spectral features most sensitive to bamboo chlorophyll, water and lignin. In terms of fusion strategy, a joint matrix local preservation projection and cross-attention mechanism are introduced, which not only achieves adaptive alignment of multi-scale spatial grids, but also suppresses redundant features through dynamic weight allocation. This completely solves the physical contradiction between the rapid saturation of optical signals in single-layer high-density bamboo canopy and the insufficient penetration ability of radar signals, greatly improving the physical representation ability of feature datasets for complex bamboo canopy structures. Secondly, for complex mountainous areas To address the severe spectral distortion caused by topography, this scheme not only calculates slope, aspect, topographic relief, and topographic shadow coefficient (TSC), but also constructs a pre-trained conditional generative adversarial network (CGAN). By introducing topographic residual constraints and adversarial loss functions, it forces the spatial gradient of generated features to approximate flat reference features without topographic interference, fundamentally eliminating texture distortion caused by nonlinear topographic umbra and shadow. This achieves high-fidelity generation of standardized spectral and structural features. Simultaneously, it utilizes satellite sunlight-induced chlorophyll fluorescence (SIF) and vegetation indices to construct a coupled phenological perception index (SV), and reconstructs time-series curves using a dual logistic model. This accurately identifies five phenological phases: shoot emergence, jointing, leaf expansion, stable growth, and dormancy. Based on effective accumulated temperature, it constructs an adaptive parameter lookup table for phenological phases, assigning maximum weight to structural features during shoot emergence and jointing, and emphasizing spectral features during leaf expansion and stable growth. This allows the feature fusion weights to dynamically adjust with the bamboo growth rhythm, effectively eliminating the interference of phenological misalignment and LAI saturation on estimation accuracy, and achieving optimal adaptive configuration of multi-scale features throughout the entire growth cycle.

[0007] (2) This invention uses extreme gradient lifting or heterogeneous stacking ensemble algorithms to invert the average diameter at breast height (DBH), bamboo height, bamboo age, and number of standing bamboo plants. The system automatically retrieves and calls specific parameter lookup tables for large-diameter scattered bamboo, small-diameter scattered bamboo, or clump bamboo based on the dominant bamboo species type in the target area. It matches the corresponding nonlinear allometric growth equation or power function equation containing DBH and bamboo height that introduces bamboo age adjustment factors, accurately characterizing the significant differences in morphological construction strategies and carbon allocation ratios of different bamboo species. A bamboo age threshold Tb is introduced to implement a dynamic correction mechanism for developmental stages. In the developmental stage where the bamboo age is below the maturity threshold Tb, the average growth method is used to correct the aboveground biomass. After the forest stabilizes, a dynamic balance is maintained, which scientifically reflects the carbon accumulation process of rapid growth and dynamic balance in bamboo forests. In addition, it addresses the issue of the well-developed underground rhizome system of bamboo forests and its long-term presence. Leveraging the unique ecological characteristics of bamboo, this scheme breaks away from the traditional assumption of a fixed root-to-rhizome ratio. It constructs a dynamic underground biomass model that grows in segments according to bamboo age, accurately calculating the carbon accumulation of bamboo stumps, roots, and rhizomes. Ultimately, the total biomass is scientifically converted into carbon sinks, avoiding the severe underestimation of underground carbon sinks caused by static estimation. This deep coupling from multi-source remote sensing feature extraction to bamboo species-specific ecological modeling not only significantly improves the absolute accuracy of biomass estimation under different bamboo species and terrain conditions, eliminating the drawback of a sharp drop in accuracy when applied across regions and bamboo species, but also provides a low-cost, high-efficiency, and high-precision replicable technical tool for regional carbon sink inventory compilation, forest carbon sink trading project design, and precise management of bamboo forest ecosystems. This strongly supports the accurate implementation of national carbon neutrality goals and the accurate assessment of global carbon balance. Attached Figure Description

[0008] Figure 1 This is a first-view stereoscopic image of a method for estimating bamboo forest biomass and carbon sequestration based on image analysis according to the present invention. Figure 2 This is a flowchart of step S1 in the image analysis-based method for estimating bamboo forest biomass and carbon sink of the present invention. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0010] Example: Please see Figure 1-2 A method for estimating bamboo forest biomass and carbon sequestration based on image analysis, comprising: S1. Acquire multi-source remote sensing data of the target bamboo forest area. The multi-source remote sensing data includes UAV hyperspectral imagery, spaceborne lidar point cloud data, and time-series multispectral satellite imagery. S2. Preprocess and extract features from multi-source remote sensing data to construct a multi-scale feature dataset. The multi-scale feature dataset includes a spectral feature set, a structural feature set, and a temporal feature set. A spatial scale transformation relationship is established between the spectral feature set and the structural feature set through a feature-level fusion strategy. S3. Based on time-series multispectral satellite imagery, reconstruct the vegetation index time-series curve, implement phenological phase identification, determine the key phenological period of bamboo forest, construct a phenological period adaptive parameter lookup table, determine the feature fusion weight based on the phenological period adaptive parameter lookup table, perform dynamic weighted fusion of multi-scale feature datasets, and generate a fusion feature set after phenological period correction. S4. Obtain digital elevation model data of the target bamboo forest area, extract terrain morphology features and terrain shadow coefficient, and construct a terrain parameter feature set; based on the fused feature set and terrain parameter feature set, use a pre-trained conditional generative adversarial network to implement terrain interference correction and generate standardized spectral and structural features. S5. Input the standardized spectrum and structural features into the pre-trained bamboo forest biomass estimation model, output the bamboo forest aboveground biomass estimation results, and convert the aboveground biomass into carbon sink based on the carbon content of the target bamboo species.

[0011] In a specific embodiment of the present invention, by sequentially acquiring UAV hyperspectral images, spaceborne lidar point cloud data, and time-series multispectral satellite images, complementary acquisition of multi-dimensional and multi-scale spatial information is achieved. During operation, the UAV hyperspectral imagery obtains high-fidelity surface reflectance through rigorous radiometric calibration and dark current correction. The spaceborne lidar uses a density threshold algorithm to accurately separate ground and canopy photons and calculates relative height to characterize vertical structure. The time-series multispectral imagery eliminates terrain shadows through resampling and improved C-correction, and extracts the YCBI index to capture the biennial cycle characteristics of bamboo forests. Finally, the images are unified to UTM projection to achieve spatial registration within one pixel.

[0012] Specifically, S1 includes: UAV hyperspectral image acquisition: A UAV platform equipped with a pushbroom hyperspectral imager with a spectral coverage range of 400-1700nm was selected to acquire continuous narrow-band spectral characteristics of the bamboo forest canopy. The flight path overlap rate was set to ≥80%, the lateral overlap rate to ≥70%, the flight altitude to 50-200m, and the ground speed to 3-8m / s. Before takeoff and after landing, standard reflectors were placed in unobstructed locations to capture calibration images. Completely black images were then captured with the lens cap on for dark current correction. Based on the calibration images, the original digital quantization value (DN) was converted to surface reflectance using the formula... ,in wavelength The surface reflectance is dimensionless. The original digital quantization value of the target pixel is dimensionless. The digital quantization value of dark current is dimensionless. The standard whiteboard digital quantization value is dimensionless. The reflectance of a standard whiteboard is calibrated to be dimensionless; Spaceborne LiDAR Point Cloud Data Acquisition: Acquire spaceborne LiDAR ATL03 or ATL08 data products covering the target bamboo forest area; use a density threshold-based algorithm to separate ground photons and canopy photons, normalize the point cloud, calculate the relative height of the canopy, and establish a vertical structure-specific correlation for the bamboo forest. The formula is as follows: ,in The relative height of the canopy. For the absolute elevation of the photon, For the corresponding position Ground elevation; Temporal multispectral satellite imagery acquisition: Landsat 8 multi-temporal images covering the target bamboo forest area were acquired, with a cloud cover threshold of <20%. SNAP-Sen2Cor software was used for radiometric calibration and atmospheric correction to obtain surface reflectance. A cubic convolution method was used to resample the 20m spatial resolution bands to 10m. An improved C-correction model was used for terrain correction to eliminate the interference of terrain shadows on the spectral characteristics of the bamboo forest. The formula is as follows: ,in The surface reflectance after topographic correction is dimensionless. To ensure the surface reflectance is dimensionless before correction, Let denoted as zenith angle, i as terrain slope angle, and m and b as values ​​based on pixel reflectivity and... The coefficients obtained from linear regression are dimensionless. Based on the biennial growth characteristics of bamboo forests, the annual bamboo forest variation index (YCBI) is extracted to establish time-series-specific correlation characteristics for bamboo forests. The formula is as follows: ,in The bamboo forest change index for year y is dimensionless. The reflectance of the near-infrared band at 842nm in year y is... The reflectance of the red-edge band at 783nm in year y is... The reflectance of the red-edge band at 865nm in year y; Unified spatial reference for multi-source data: The universal transverse Mercator UTM projection system is uniformly selected, and Sentinel-2 data is used as the spatial reference to perform spatial registration of UAV hyperspectral imagery and spaceborne lidar point cloud, with the registration accuracy controlled within 1 pixel.

[0013] Specifically, step S2 includes: Hyperspectral image preprocessing and spectral feature set extraction: Radiometric calibration and FLAASH atmospheric correction were performed on UAV hyperspectral images. Savitzky-Golay filters were used for smoothing and denoising. A competitive adaptive reweighted sampling continuous projection algorithm was introduced to extract bamboo forest canopy-specific spectral feature bands. The hyperspectral normalized vegetation index was calculated using the following formula: ,in The surface reflectance in the 800nm ​​near-infrared band is dimensionless. The surface reflectance in the 670nm red light band is dimensionless. Preprocessing and structural feature set extraction of spaceborne lidar point clouds: Noise filtering and ground point classification are performed on the spaceborne lidar point clouds to generate a digital elevation model (DEM); the relative height of canopy photons is calculated and canopy height quantile features are extracted to construct a multi-scale local spatial feature expression. ,in Represents multi-scale local spatial features. This represents the feature of the i-th intersection channel. This represents the j-th convolution or pooling operation operator for the i-th cross-channel feature. This represents the convolution operation; Temporal multispectral satellite image preprocessing and temporal feature set extraction: A multi-temporal image dataset combining Sentinel-2 and Landsat-8 was acquired. Pixels with values ​​of 4 to 8 were extracted using the Scene Classification Band (SCL) to remove cloud shadows and high-probability clouds. The reflectance values ​​were normalized to the [0,1] interval by dividing by 10000. An asymmetric Gaussian function was used to fit and reconstruct the temporal curve of the vegetation index. Where a is the dimensionless background NDVI baseline value, b is the dimensionless NDVI amplitude, c is the daily sequence corresponding to the NDVI peak value in days, and d is the curve width parameter in days. 2 t represents the day unit in the time series; Spatial scale transformation and feature-level fusion strategy: The spectral feature set of hyperspectral imagery and the structural feature set of spaceborne lidar are resampled to a unified spatial resolution grid; a joint matrix local preservation projection algorithm is used for feature dimensionality reduction and scale transformation; and spectral features, structural features, and temporal features are dynamically fused based on a cross-attention mechanism, expressed as follows: ,in For multi-scale fusion feature sets, and These are spectral and structural feature sets, respectively. and The weight matrix is ​​learned based on channel attention, and Concat represents the cascading operation.

[0014] In a specific embodiment of the present invention, the preferred characteristic band of CARS-SPA in step S2 is specifically defined as follows: First, the competitive adaptive reweighted sampling algorithm CARS, based on Monte Carlo sampling and partial least squares, selects wavelengths with greater importance, iterates multiple times to obtain multiple different wavelength subsets, and selects the wavelength subset with the smallest root mean square error based on cross-validation. Then, the continuous projection algorithm SPA is used to perform secondary effective feature wavelength selection on the feature wavelength initially selected by CARS, and the original spectral matrix is ​​projected to select the maximum projection vector as the feature wavelength. The coefficient of determination R^2 and the root mean square error RMSE are used as the evaluation criteria for feature subsets. ,in Let i be the true value of the i-th sample. Let be the predicted value of the i-th sample, and n be the total number of samples.

[0015] Specifically, the feature-level fusion in step S2 also includes a feature sensitivity ranking constraint based on continuous wavelet projection: The multi-source fusion features are organized into a two-dimensional spectral matrix. The continuous wavelet transform is used to perform multi-scale decomposition of the two-dimensional spectral matrix in the time-frequency domain to capture the texture and edge features of the bamboo canopy, obtain the corresponding wavelet coefficients, and construct a feature chain matrix. ; Calculate the feature sensitivity matrix The mean matrix is ​​obtained by taking the mean of each column of elements. Find the column containing the largest element in the mean matrix. The feature sensitivity matrix quantifies the sensitivity of each wavelet coefficient feature to the target of bamboo forest aboveground biomass inversion; feature chain matrix The wavelet coefficients of the column corresponding to the maximum mean are extracted as the optimal feature combination; Principal Component Analysis (PCA) algorithm is used to reduce the dimensionality of the optimal feature combination, extract principal component features to eliminate collinearity and information redundancy between features, and use the feature vector output after dimensionality reduction as the final fused feature set, which is input into the pre-trained bamboo forest biomass regression estimation model to participate in subsequent aboveground biomass estimation.

[0016] In a specific embodiment of the present invention, the limitations of Landsat-8 and Sentinel-2 collaborative data preprocessing in step S2 include: Bitwise operations are used to extract sharp pixels from cloudless shadows. ; For Landsat-8 data lacking native red-edge bands, a weighted formula is used. Simulated red-edge band reflectivity, where To simulate dimensionless surface reflectance in the red-edge band, The surface reflectance in the near-infrared band is dimensionless. The surface reflectance in the red light band is dimensionless. Feature sensitivity matrix To quantify the response of the multi-scale feature dataset output or bamboo biomass inversion target to small perturbations of each input feature, the Jacobian matrix is ​​specifically defined as the ratio of the difference between the maximum and minimum values ​​of the inversion target observed by a certain feature within its uncertainty range at each parameter expectation value to the observed values ​​when all parameters are fixed at the expectation values, thus forming a matrix that can reflect the uncertainty and sensitivity range of the parameters. Feature sensitivity matrix Derived from the feature chain matrix generated by the continuous wavelet projection algorithm Independent feature sensitivity analysis was performed, and the feature chain matrix was analyzed. It consists of wavelet coefficients that capture texture and edge features after decomposing a two-dimensional spectral matrix fused from multiple sources in the time-frequency domain using wavelet transform; Feature sensitivity matrix The calculation method is as follows: For the regression problem of bamboo forest biomass estimation, the determination coefficient between each wavelet feature in the feature chain matrix and the bamboo forest aboveground biomass inversion target is calculated. The determination coefficient is used as the sensitivity coefficient of that feature, thereby constructing the corresponding feature sensitivity matrix. Then, the feature sensitivity matrix is ​​calculated. The mean of each column's elements is used to obtain the mean matrix. The column containing the largest element in the mean matrix is ​​then identified, and the wavelet coefficient feature combination corresponding to that column is used as the optimal feature combination with the highest sensitivity and input into the subsequent model.

[0017] Specifically, step S3 includes: Vegetation index time-series curve reconstruction and phenological phase identification: Based on acquired time-series multispectral satellite imagery, satellite sunlight-induced chlorophyll fluorescence (SIF) and vegetation indices (VIs) are extracted to construct a coupled bamboo forest phenological perception vegetation index (SV), the formula of which is as follows: ,in The unit of the SIF observation value for the day is , This is the minimum value of the SIF time series. The magnitude of SIF change. The vegetation index observation value for the day is dimensionless. The minimum value of the vegetation index. The SV index was used to enhance the specific expression of phenological characteristics in bamboo forests during bumper and lean years, representing the magnitude of vegetation index changes. Subsequently, a bilogistic model was employed to smooth and reconstruct the SV time series. The fitting function expression is as follows: ,in The fitted SV value for day t is dimensionless, where t is the day number in a year. This represents the winter background baseline value. The amplitude is dimensionless. and These are the curvature parameters (dimensionless) for the rising and falling phases of the growing season, respectively. and The day sequences corresponding to the inflection points of the rising and falling phases are used to perform phenological phase identification and reconstruct the... The second derivative of the curve is calculated, and the rate of change of curvature at each time point is determined. The number of days corresponding to the maximum curvature during the growth and decline phases are taken as the start time (SOS) and end time (EOS) of the bamboo forest growing season. The formula is as follows: ,in The curvature of the curve on day t is dimensionless. and They are respectively The first and second derivatives are defined. , Based on this, bamboo forest phenology is divided into five key phenological phases: shoot emergence period, jointing period, leaf unfolding period, stable growth period, and dormancy period. Constructing an adaptive phenological parameter lookup table: An adaptive phenological parameter lookup table is constructed, defining the effective accumulated temperature during the bamboo forest growing season as the independent variable x. A three-part piecewise logistic function is introduced to simulate the shape equation of vegetation index changes with accumulated temperature. ,in The starting value is dimensionless. and The increments of the endpoint values ​​for the two segments are dimensionless. and The unit for accumulated temperature at the inflection point is ℃ × d. and The growth rate coefficient is in units of , The accumulated temperature threshold for the two-stage switching is in °C × d. A phenological adaptive weight equation is established to determine the fusion weights of multi-scale features as a function of accumulated temperature, and the weights of the spectral feature set are set. Structural feature set weights and time series feature set weights The spectral feature weighting equation is: The structural feature weight equation is: The time series feature weight equation is: ,in , , , , , The fitting coefficients are obtained by least squares optimization based on historical sample plot data. They are dimensionless. x is the effective accumulated temperature corresponding to the current date. A mapping relationship is established between the phenological phase and the corresponding weight coefficient to form a phenological period adaptive parameter lookup table. Dynamic weighted fusion of multi-scale feature datasets: Based on the phenological period adaptive parameter lookup table, the effective accumulated temperature x corresponding to the phenological period of the current phase multi-scale feature dataset is read, the dynamic fusion weight of each feature channel under the current bamboo forest phenological period is calculated, and dynamic weighted fusion is performed. The formula is as follows: ,in The dimensionless fused eigenvalue after phenological period correction at spatial location p is given. , and The spectral, structural, and temporal feature values ​​at spatial location p are dimensionless after standardization. To address the scale differences between adjacent levels of features in multi-scale features, an adaptive gating mechanism is introduced to calibrate the contextual information of the fused features. The formula is as follows: ,in This is the final phenological period-corrected and spatially context-aware fusion feature set. It is the Sigmoid activation function. This is a 1×1 convolution operation; [;] indicates concatenation along the channel dimension. This indicates element-wise multiplication. The adaptive gating mechanism dynamically adjusts the contribution of each channel feature based on the phenological period feature response intensity of the current pixel, thus suppressing background noise redundancy caused by cross-scale feature splicing.

[0018] In a specific embodiment of the present invention, satellite sunlight-induced chlorophyll fluorescence (SIF) and vegetation index are extracted, and a coupled phenological perception index (SV) is constructed to enhance the phenological specificity expression of alternate bearing. The SV time-series curve is reconstructed using a dual logistic model. The curvature change rate is obtained by calculating the second derivative to determine the beginning and end of the growing season. Five phenological phases, such as the bamboo shoot stage and the jointing stage, are divided. A three-segment logistic shape equation with effective accumulated temperature as the independent variable is constructed. The weight equations of each feature set as a function of accumulated temperature are set, dynamic weighted fusion is performed, and an adaptive gating mechanism is introduced to calibrate the context information.

[0019] Specifically, the construction and dynamic weighted fusion of the phenological period adaptive parameter lookup table are limited as follows: During the bamboo shoot emergence and jointing stages The canopy structure changes drastically, setting ; During the leaf unfolding stage and stable growth period of bamboo forest When the canopy leaf area index reaches saturation, the sensitivity of spectral characteristics to changes in biomass decreases, and the setting... ; Temporal feature set weights Maintain a stable decay constraint throughout the year to avoid the interference of time series lag effects on the accuracy of current biomass estimation.

[0020] In a specific embodiment of the present invention, differentiated weight allocation is carried out according to the physiological and ecological characteristics of bamboo forests at different phenological stages: during the shoot emergence and jointing stages, the three-dimensional structure of the canopy undergoes drastic changes, so the structural feature weight is set to be the largest; during the leaf unfolding and stable growth stages, the LAI tends to saturate, and the sensitivity of spectral features decreases, so the spectral and structural weights are set to be equal or slightly better; the temporal feature weight remains stable and decays throughout the year to prevent interference from lag effects.

[0021] Specifically, S4 includes: Digital Elevation Model (DEM) Data Acquisition and Terrain Feature Extraction: Obtain DEM data covering the target bamboo forest area, resample it to the same spatial resolution as the fused feature set, and calculate the slope of the bamboo forest distribution area based on the DEM data. and slope The topographic relief is extracted as a macroscopic topographic feature, and the formula is as follows: ,in Let n be the topographic relief within the neighborhood grid, and n be the number of neighborhood grids. Given the elevation value of the i-th neighboring grid, the surface cutting depth is extracted to reflect the micro-geomorphic erosion characteristics of the bamboo forest, using the formula: ,in The unit for the depth of surface cutting is meters (m). The average elevation of the neighborhood grid is in meters. Calculation of Topographic Shadow Coefficient and Construction of Topographic Parameter Feature Set: Based on multispectral image reflectance and DEM data, a bamboo forest topographic shadow index (SI) is constructed. SI is the dimensionless shading index. , , Dimensionless surface reflectance is used in the coastal band, green band, and near-infrared band. A threshold c is set, and variable coefficients reflecting the intensity of bamboo forest shadows are calculated. , ,in The dimensionless variable coefficient for shadow intensity The maximum SI value for the study area is dimensionless. Based on the geometric relationship between the sun and the terrain, the topographic shading coefficient TSC of the bamboo forest is calculated. Where TSC is the bamboo forest terrain shadow coefficient, The zenith angle of the sun. The slope angle of the terrain. The azimuth of the sun. The slope angle of the terrain. As a variable coefficient for shadow intensity, slope, aspect, topographic relief, surface cutting depth, and topographic shadow coefficient are stacked in a raster pattern to construct a topographic parameter feature set. ; Terrain disturbance correction and normalized feature generation based on conditional generative adversarial networks: Construct a pre-trained conditional generative adversarial network (CGAN) containing a generator G and a discriminator D, and fuse feature sets. With terrain parameter feature set Concatenate along the channel dimension to form the conditional input vector The generator G takes the conditional input vector c and random noise z as input to generate the corrected normalized spectrum and structural features. Construct the generator's overall loss function that includes terrain residual constraints. , ,in To counteract the dimensionless loss, To reconstruct the dimensionless loss, and The weighting coefficients are dimensionless. Represents the spatial gradient operator. Since the flatland reference bamboo forest feature set is dimensionless, the terrain residual constraint term forces the spatial gradient of the generated features to approximate the flatland reference features without terrain interference, eliminating texture distortion caused by terrain. The discriminator D receives the true, shadow-free features. Or generate features And the conditional vector c, output discriminant probability, and adversarial loss function. for Through game-theoretic optimization between the generator and the discriminator, the standardized spectral and structural features generated by the generator are output. This completes the correction of the topographic umbra and shadow effects.

[0022] In a specific embodiment of the present invention, a DEM is acquired and resampled, and the slope, aspect, topographic relief, and surface cutting depth are calculated. Based on multispectral reflectance and the DEM, a shadow index SI and a variable coefficient α are constructed. The topographic shadow coefficient TSC is calculated by combining solar geometry. The fused features and topographic parameters are spliced ​​along the channel as the conditional input of CGAN. The generator outputs the corrected standardized features. An overall loss function containing topographic residual constraints is constructed to force the spatial gradient of the generated features to approximate the flat ground reference features. The final features are output through a game between the discriminator and the generator.

[0023] Specifically, the digital elevation model data obtained in step S4 needs to be projected before use so that it is in the same WGS84 projection coordinate system as the multispectral image, and then resampled using bilinear interpolation. In step S4, the conditional generative adversarial network uses terrain feature pairs extracted from open-source DEM and high-precision DEM as training datasets during the pre-training stage. The terrain feature elements are used as constraints to optimize the network parameters, so as to ensure the generalization ability and terrain reconstruction accuracy in complex mountainous bamboo forest areas.

[0024] In a specific embodiment of the present invention, the DEM is uniformly transformed to the WGS84 projection coordinate system before use, and bilinear interpolation is used for resampling to maintain the continuity of elevation gradient. In the CGAN pre-training stage, terrain feature pairs extracted from open-source DEM and high-precision DEM are used as training datasets, and terrain feature elements are used as constraints to optimize network parameters.

[0025] Specifically, S5 includes: Bamboo forest structural parameter inversion and aboveground biomass estimation: Standardized spectral and structural features are input into a pre-trained bamboo forest biomass estimation model. The model is constructed based on extreme gradient boosting or heterogeneous stacking ensemble learning algorithms and outputs key bamboo forest stand structural parameters, including average diameter at breast height (DBH) D, average bamboo height (H), bamboo age (T), and number of standing bamboo plants per unit area (N). A bamboo forest-specific allometric growth equation for single-plant aboveground biomass is established. For large-diameter, scattered bamboo, a nonlinear equation for a bamboo age regulating factor is introduced. ,in Here, D represents the average diameter at breast height (DBH) of a single bamboo plant (cm), T represents the age of the bamboo plant (years), and a, b, c, d, e, and f are dimensionless, target bamboo species-specific fitting parameters. For large-diameter clumping bamboo or small-diameter scattered bamboo, a power function equation incorporating DBH and bamboo height is used. Where H is the average height of a single bamboo plant in meters, and a', b', and c' are dimensionless, dimensionless, specific fitting parameters for the corresponding bamboo species. The aboveground biomass per unit area of ​​the bamboo forest is calculated based on the number of standing bamboo plants per unit area. ,in The aboveground biomass per unit area of ​​bamboo forest in year t is given by [unit]. N is the number of bamboo plants per unit area. This is a constant for converting kilograms to tons; Biomass dynamic correction during bamboo forest development stages: Based on the characteristics of bamboo forest growth and development stages, a threshold age for bamboo forests to reach maturity and stability is set as follows. In the project bamboo age During the developmental stages, the average growth rate method was used to correct for aboveground biomass. In the project bamboo age > During the stable stage of forest formation, the aboveground biomass maintains a dynamic balance. ,in For the first Annual aboveground biomass per unit area of ​​bamboo forest To determine the aboveground biomass constant per unit area after bamboo forests reach maturity and stability, and to estimate the underground biomass and convert it into total carbon sink, a dynamic growth model of underground biomass with bamboo age is established based on the long-term retention characteristics of the underground portion of bamboo forests. The total biomass of the bamboo forest is then calculated and converted into carbon sink. ,in Carbon sequestration in bamboo forests The carbon content of the target bamboo species. This is the molecular weight constant for converting the mass of carbon to the mass of carbon dioxide.

[0026] In a specific embodiment of the present invention, the dynamic growth model is as follows: when hour, ; when hour, ; when At that time, it was conservatively assumed that the underground biomass would no longer increase; in For the first Annual underground biomass per unit area of ​​bamboo forest The ratio of underground biomass to aboveground biomass in bamboo forests is represented by the rhizome. , where is the growth coefficient of underground biomass at different stages.

[0027] Specifically, in step S5, based on the dominant bamboo species type in the target bamboo forest area, the corresponding bamboo species-specific parameter lookup table is automatically invoked to match the corresponding allometric growth equation coefficients with the mature and stable bamboo age threshold. This is to ensure the specificity and accuracy of biomass estimation for different bamboo species.

[0028] In a specific embodiment of the present invention, in step S5, the system automatically retrieves and calls the corresponding bamboo species-specific parameter lookup table according to the dominant bamboo species type of the target bamboo forest area, and matches the corresponding allometric growth equation coefficients with the mature and stable bamboo age threshold. This ensures the specificity of biomass estimation for different bamboo species.

[0029] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A method for estimating bamboo forest biomass and carbon sequestration based on image analysis, characterized in that, include: S1. Acquire multi-source remote sensing data of the target bamboo forest area, including UAV hyperspectral imagery, spaceborne lidar point cloud data, and time-series multispectral satellite imagery. S2. Preprocess and extract features from the multi-source remote sensing data to construct a multi-scale feature dataset. The multi-scale feature dataset includes a spectral feature set, a structural feature set, and a temporal feature set. A spatial scale transformation relationship is established between the spectral feature set and the structural feature set through a feature-level fusion strategy. S3. Based on the time-series multispectral satellite imagery, reconstruct the vegetation index time-series curve, perform phenological phase identification, determine the key phenological period of bamboo forest, construct a phenological period adaptive parameter lookup table, and determine the feature fusion weight based on the phenological period adaptive parameter lookup table. Perform dynamic weighted fusion on the multi-scale feature dataset to generate a fusion feature set after phenological period correction. S4. Obtain digital elevation model data of the target bamboo forest area, extract terrain morphology features and terrain shadow coefficient, and construct a terrain parameter feature set; based on the fused feature set and the terrain parameter feature set, use a pre-trained conditional generative adversarial network to implement terrain interference correction and generate standardized spectral and structural features. S5. Input the standardized spectrum and structural features into the pre-trained bamboo forest biomass estimation model, output the bamboo forest aboveground biomass estimation results, and convert the aboveground biomass into carbon sink based on the carbon content of the target bamboo species.

2. The method for estimating bamboo forest biomass and carbon sequestration based on image analysis according to claim 1, characterized in that, S1 includes: UAV hyperspectral image acquisition: A UAV platform equipped with a pushbroom hyperspectral imager with a spectral coverage range of 400-1700nm was selected to obtain continuous narrow-band spectral characteristics of the bamboo forest canopy. Before takeoff and after landing, a standard reflector was placed in an unobstructed location to take calibration images. A completely black image was taken with the lens cap on for dark current correction. Based on the calibration images, the original digital quantization value DN was converted into surface reflectance. Spaceborne lidar point cloud data acquisition: acquire spaceborne lidar data covering the target bamboo forest area, use a density threshold-based algorithm to separate ground photons and canopy photons, normalize the point cloud, calculate the relative height of the canopy, and establish a specific correlation of the vertical structure of the bamboo forest. Temporal multispectral satellite imagery acquisition: Landsat 8 multi-temporal images covering the target bamboo forest area were acquired, with a cloud cover threshold of <20%. SNAP-Sen2Cor software was used for radiometric calibration and atmospheric correction to obtain surface reflectance. The 20m spatial resolution band was resampled to 10m using a cubic convolution method. An improved C-correction model was used for topographic correction to eliminate the interference of topographic shadows on the spectral characteristics of the bamboo forest. Based on the biennial growth characteristics of the bamboo forest, the annual moso bamboo forest variation index YCBI was extracted to establish temporal-specific correlation characteristics of the bamboo forest. Unified spatial reference for multi-source data: The universal transverse Mercator UTM projection system is uniformly selected, and Sentinel-2 data is used as the spatial reference to perform spatial registration of UAV hyperspectral imagery and spaceborne lidar point cloud, with the registration accuracy controlled within 1 pixel.

3. The method for estimating bamboo forest biomass and carbon sequestration based on image analysis according to claim 1, characterized in that, Step S2 includes: Hyperspectral image preprocessing and spectral feature set extraction: Radiometric calibration and FLAASH atmospheric correction were performed on UAV hyperspectral images. Savitzky-Golay filters were used for smoothing and noise reduction. A competitive adaptive reweighted sampling continuous projection algorithm was introduced to extract bamboo forest canopy-specific spectral feature bands. Preprocessing and structural feature set extraction of spaceborne lidar point cloud: noise filtering and ground point classification are performed on the spaceborne lidar point cloud to generate a digital elevation model (DEM), and the relative height of canopy photons is calculated and canopy height quantile features are extracted. Temporal multispectral satellite image preprocessing and temporal feature set extraction: A multi-temporal image dataset of Sentinel-2 and Landsat-8 was obtained. Pixels with values ​​of 4 to 8 were extracted using the Scene Classification Band (SCL) to remove cloud shadows and high-probability clouds. The reflectance values ​​were then divided by 10000 and normalized to the [0,1] interval. Spatial scale transformation and feature-level fusion strategy: The spectral feature set of hyperspectral image and the structural feature set of spaceborne lidar are resampled to a unified spatial resolution grid; the joint matrix local preservation projection algorithm is used for feature dimensionality reduction and scale transformation; and the spectral features, structural features and temporal features are dynamically fused based on the cross attention mechanism.

4. The method for estimating bamboo forest biomass and carbon sequestration based on image analysis according to claim 3, characterized in that, The feature-level fusion in step S2 also includes a feature sensitivity ranking constraint based on continuous wavelet projection: The multi-source fusion features are organized into a two-dimensional spectral matrix. The two-dimensional spectral matrix is ​​decomposed in the time-frequency domain using continuous wavelet transform to capture the texture and edge features of the bamboo canopy, obtain the corresponding wavelet coefficients, and construct a feature chain matrix. ; Calculate the feature sensitivity matrix The mean matrix is ​​obtained by taking the mean of each column of elements. Find the column containing the largest element in the mean matrix. The feature sensitivity matrix quantifies the sensitivity of each wavelet coefficient feature to the target of bamboo forest aboveground biomass inversion; feature chain matrix The wavelet coefficients of the column corresponding to the maximum mean are extracted as the optimal feature combination; Principal Component Analysis (PCA) algorithm is used to reduce the dimensionality of the optimal feature combination, extract principal component features to eliminate collinearity and information redundancy between features, and use the feature vector output after dimensionality reduction as the final fused feature set, which is input into the pre-trained bamboo forest biomass regression estimation model to participate in subsequent aboveground biomass estimation.

5. The method for estimating bamboo forest biomass and carbon sequestration based on image analysis according to claim 1, characterized in that, S3 includes: Vegetation index time-series curve reconstruction and phenological phase identification: Based on acquired time-series multispectral satellite imagery, satellite sunlight-induced chlorophyll fluorescence (SIF) and vegetation indices (VIs) were extracted to construct a coupled bamboo forest phenological perception vegetation index (SV). The SV index was used to enhance the specific expression of phenological characteristics in high-yield and low-yield years of bamboo forests. Subsequently, a dual logistic model was used to smooth and reconstruct the SV time series curves. The fitting function expression is as follows: Implement phenological phase recognition to reconstruct The second derivative of the curve is calculated, and the rate of change of curvature at each time point is calculated. The number of days corresponding to the maximum curvature in the growth stage and the decline stage is taken as the start time SOS and the end time EOS of the bamboo forest growing season. Based on this, the phenology of the bamboo forest is divided into five key phenological phases: the shoot emergence period, the jointing period, the leaf unfolding period, the stable growth period, and the dormancy period. Constructing an adaptive phenological parameter lookup table: An adaptive phenological parameter lookup table is constructed, defining the effective accumulated temperature during the bamboo forest growing season as the independent variable x. A three-part piecewise logistic function is introduced to simulate the shape equation of vegetation index changes with accumulated temperature. A phenological adaptive weighting equation is established to determine the fusion weights of multi-scale features as a function of accumulated temperature, and the weights of the spectral feature set are set. Structural feature set weights and time series feature set weights A mapping relationship is established between the phenological phases and the corresponding weight coefficients to form an adaptive parameter lookup table for phenological periods; Dynamic weighted fusion of multi-scale feature datasets: Based on the phenological period adaptive parameter lookup table, the effective accumulated temperature x corresponding to the phenological period of the current phase multi-scale feature dataset is read, the dynamic fusion weight of each feature channel under the current bamboo forest phenological period is calculated, and dynamic weighted fusion is performed. In view of the scale difference between adjacent layers of features in the multi-scale features, an adaptive gating mechanism is introduced to calibrate the context information of the fused features. The adaptive gating mechanism dynamically adjusts the contribution of each channel feature according to the feature response intensity of the current pixel in the phenological period, and suppresses the background noise redundancy caused by cross-scale feature splicing.

6. The method for estimating bamboo forest biomass and carbon sequestration based on image analysis according to claim 5, characterized in that, The construction and dynamic weighted fusion of the phenological period adaptive parameter lookup table are limited to: During the bamboo shoot emergence and jointing stages The canopy structure changes drastically, setting ; During the leaf unfolding stage and stable growth period of bamboo forest When the canopy leaf area index reaches saturation, the sensitivity of spectral characteristics to changes in biomass decreases, and the setting... ; The weight of the time-series feature set Maintain a stable decay constraint throughout the year to avoid the interference of time series lag effects on the accuracy of current biomass estimation.

7. The method for estimating bamboo forest biomass and carbon sequestration based on image analysis according to claim 1, characterized in that, S4 includes: Digital Elevation Model (DEM) Data Acquisition and Terrain Feature Extraction: Obtain DEM data covering the target bamboo forest area, resample it to the same spatial resolution as the fused feature set, and calculate the slope of the bamboo forest distribution area based on the DEM data. and slope The topographic relief was extracted as a macro-topographic feature, and the surface cutting depth was extracted to reflect the micro-topographic erosion characteristics of the bamboo forest. Calculation of Topographic Shadow Coefficient and Construction of Topographic Parameter Feature Set: Based on multispectral image reflectance and DEM data, a bamboo forest topographic shadow index SI is constructed. A threshold c is set, and a variable coefficient reflecting the intensity of bamboo forest shadow is calculated. By combining the geometric relationship between the sun and the terrain, the topographic shadow coefficient (TSC) of the bamboo forest is calculated. The slope, aspect, topographic relief, surface cutting depth, and topographic shadow coefficient are then rasterized and stacked to construct a topographic parameter feature set. ; Terrain interference correction and normalized feature generation based on conditional generative adversarial networks: Construct a pre-trained conditional generative adversarial network (CGAN) containing a generator G and a discriminator D, and integrate the fused feature set... With the terrain parameter feature set Concatenate along the channel dimension to form the conditional input vector The generator G takes a conditional input vector c and random noise z as input to generate the corrected normalized spectrum and structural features. Construct the generator's overall loss function that includes terrain residual constraints. The terrain residual constraint term forces the spatial gradient of the generated features to approximate a flat reference feature without terrain interference, eliminating texture distortion caused by terrain. The discriminator D receives the true, shadow-free features. Or generate features And the conditional vector c, outputting the discriminant probability, with the adversarial loss function being... Through game-theoretic optimization between the generator and the discriminator, the standardized spectral and structural features generated by the generator are output. This completes the correction of the topographic umbra and shadow effects.

8. The method for estimating bamboo forest biomass and carbon sequestration based on image analysis according to claim 7, characterized in that, The digital elevation model data obtained in step S4 needs to be projected before use so that it is in the same WGS84 projection coordinate system as the multispectral image, and then resampled using bilinear interpolation. In step S4, the conditional generative adversarial network uses terrain feature pairs extracted from open-source DEM and high-precision DEM as training datasets during the pre-training stage. The terrain feature elements are used as constraints to optimize network parameters, so as to ensure the generalization ability and terrain reconstruction accuracy in complex mountainous bamboo forest areas.

9. The method for estimating bamboo forest biomass and carbon sequestration based on image analysis according to claim 1, characterized in that, S5 includes: Bamboo forest structural parameter inversion and aboveground biomass estimation: The standardized spectral and structural features are input into a pre-trained bamboo forest biomass estimation model. This model is constructed based on extreme gradient boosting or heterogeneous stacking ensemble learning algorithms and outputs key bamboo forest stand structural parameters, including average diameter at breast height (DBH) D, average bamboo height (H), bamboo age (T), and number of standing bamboo plants per unit area (N). A bamboo forest-specific allometric growth equation for single-plant aboveground biomass is established. For large-diameter, scattered bamboo, a nonlinear equation for a bamboo age regulating factor is introduced. For large-diameter clump-forming bamboo or small-diameter scattered bamboo, a power function equation incorporating both diameter at breast height (DBH) and bamboo height is used. The aboveground biomass per unit area of ​​bamboo forest is calculated based on the number of bamboo plants per unit area. Biomass dynamic correction during bamboo forest development stages: Based on the characteristics of bamboo forest growth and development stages, a threshold age for bamboo forests to reach maturity and stability is set as follows. In the project bamboo age During the developmental stages, the average growth rate method was used to correct for aboveground biomass. In the project bamboo age > During the stable stage of forest formation, the aboveground biomass maintains a dynamic balance. ,in For the first Annual aboveground biomass per unit area of ​​bamboo forest To determine the aboveground biomass constant per unit area after bamboo forests reach maturity and stability, and to estimate the underground biomass and convert the total carbon sink of bamboo forests, a dynamic growth model of underground biomass with bamboo age is established based on the long-term retention characteristics of the underground part of bamboo forests. The total biomass of bamboo forests is then calculated and converted into carbon sink.

10. The method for estimating bamboo forest biomass and carbon sink based on image analysis according to claim 9, characterized in that, In step S5, based on the dominant bamboo species type in the target bamboo forest area, the corresponding bamboo species-specific parameter lookup table is automatically invoked to match the corresponding allometric growth equation coefficients and the mature and stable bamboo age threshold. This is to ensure the specificity and accuracy of biomass estimation for different bamboo species.

Images