Forest aboveground biomass estimation method and device and electronic equipment

Abstract

The application relates to the technical field of forest biomass estimation, in particular to a forest aboveground biomass estimation method and device and electronic equipment, which comprises the following steps: obtaining a plurality of forest aboveground biomass samples of a target area, the samples comprising first samples and second samples; training a global forest aboveground biomass estimation model of the target area by using the first samples, and determining a prediction residual of the second samples based on the trained global model; determining a spatial heterogeneity score of the second samples based on the prediction residual, and spatially dividing the target area based on the score to obtain a plurality of target sub-areas; training a local forest aboveground biomass estimation model based on the first samples of each target sub-area to obtain a local forest aboveground biomass estimation model corresponding to each target sub-area; and estimating the forest aboveground biomass of a to-be-predicted pixel in the target area by using the trained local model. The application can improve the estimation accuracy and spatial expression capability of the forest aboveground biomass in a heterogeneous area.

Description

Technical Field

[0001] This application relates to the field of forest biomass estimation technology, and in particular to a method, apparatus and electronic equipment for estimating forest aboveground biomass. Background Technology

[0002] Forest aboveground biomass is a crucial parameter characterizing the structure, productivity, and carbon storage of forest ecosystems, and it forms an important foundation for forest carbon cycle research, carbon storage assessment, and regional carbon accounting. With the development of remote sensing technology, large-scale forest aboveground biomass estimation based on multi-source remote sensing data has become an important technical means in forest resource monitoring and ecosystem assessment.

[0003] Remote sensing inversion in related technologies often employs a unified machine learning model at the target area scale to establish an estimation relationship between forest ground biomass and remote sensing characteristic variables. However, due to differences in forest site, vegetation structure, and spatial distribution, this estimation relationship may change with spatial location. A single model is insufficient to effectively characterize this spatial variation, which can easily lead to problems such as high estimation errors in local areas, low estimation accuracy, and insufficient spatial consistency of estimation results. Summary of the Invention

[0004] This application provides a method, apparatus, and electronic device for estimating forest aboveground biomass, in order to solve the problems in related technologies where the unified estimation model is difficult to adapt to the spatial heterogeneity of forest aboveground biomass, resulting in high estimation errors and limited estimation accuracy in local areas.

[0005] The first aspect of this application provides a method for estimating forest aboveground biomass, comprising the following steps: acquiring multiple forest aboveground biomass samples of a target area, wherein the forest aboveground biomass samples include a first sample and a second sample, the first sample being used for model training and the second sample being used for spatial heterogeneity partitioning; training a global forest aboveground biomass estimation model at the target area scale using the first sample, and determining the prediction residual of the second sample based on the trained global forest aboveground biomass estimation model; determining the spatial heterogeneity score of each second sample based on the prediction residual, and spatially partitioning the target area based on the spatial heterogeneity score of the second sample to obtain multiple target sub-regions; extracting a first sample from each target sub-region, and training a corresponding local forest aboveground biomass estimation model based on the first sample of each target sub-region to obtain local forest aboveground biomass estimation models corresponding to multiple target sub-regions; partitioning the pixels to be predicted within the target area into the corresponding target sub-regions, and calling the local forest aboveground biomass estimation model corresponding to the target sub-regions to estimate the forest aboveground biomass of the pixels to be predicted.

[0006] Optionally, the target region is spatially partitioned based on the spatial heterogeneity score of the second sample to obtain multiple target sub-regions, including: dividing the target region into regular grids and matching the second sample to the corresponding regular grids to obtain multiple spatial units; calculating the spatial heterogeneity score of the corresponding spatial unit based on the prediction residuals of all second samples within the spatial unit; taking the target region as the region to be partitioned, sorting the spatial heterogeneity scores of multiple spatial units within the region to be partitioned, and sequentially extracting spatial units based on the sorting results to divide the region to be partitioned into two candidate sub-regions; taking the two candidate sub-regions as the binary partitioning results to be tested, and performing a target test on the binary partitioning results to be tested, wherein the target test includes an accuracy improvement test and a significance test; if the target test passes, the binary partitioning results are retained, and the two candidate sub-regions are taken as new regions to be partitioned, and the binary partitioning and target test are recursively executed until a preset stopping condition is met to obtain multiple target sub-regions.

[0007] Optionally, a target test is performed on the binary partitioning results to be tested, including: obtaining the estimated value of the first forest aboveground biomass corresponding to the region to be partitioned before partitioning, and the estimated value of the second forest aboveground biomass corresponding to the candidate sub-regions after partitioning; based on the estimated value of the second forest aboveground biomass, the estimated value of the first forest aboveground biomass, and the corresponding forest aboveground biomass reference value, determining the change value of the prediction accuracy of the second sample of the candidate sub-region before and after partitioning; if the change value of the prediction accuracy does not meet the preset improvement condition, the binary partitioning results are not retained, and the partitioning of the region to be partitioned is stopped; if the change value of the prediction accuracy meets the preset improvement condition, a significance test is performed on the change value of the prediction accuracy, and the binary partitioning results are retained when the significance test passes, otherwise the binary partitioning results are not retained, and the partitioning of the region to be partitioned is stopped.

[0008] Optionally, the formula for calculating the spatial heterogeneity score of the candidate sub-region is: ; in, The spatial heterogeneity score for candidate sub-region S is given. This represents the additional mean offset of the overall prediction residuals within the candidate sub-region relative to the baseline distribution. The sum of the inverse variance-weighted residuals of all second samples within the candidate sub-region S. , For the second sample In spatial location The predicted residual at the location, Spatial location The standard deviation of the baseline residual distribution. The sum of the inverse variance weights of all second samples within the candidate sub-region S. .

[0009] Optionally, the preset stopping conditions include at least one of the following: the number of first samples in the region to be partitioned is less than a preset sample number threshold; the number of spatial units in the region to be partitioned is less than a preset spatial unit number threshold; the change in prediction accuracy corresponding to the binary partitioning result does not meet the preset improvement conditions; the binary partitioning result fails the significance test considering spatial autocorrelation; the current recursive partitioning depth reaches the preset maximum partitioning depth.

[0010] Optionally, each forest ground biomass sample includes feature variables and forest ground biomass reference values. Obtaining multiple forest ground biomass samples in the target area includes: obtaining multi-source remote sensing data of the target area; extracting feature variables of the corresponding forest ground biomass sample from the multi-source remote sensing data; and obtaining forest ground biomass reference values ​​of multiple forest ground biomass samples in the target area.

[0011] Optionally, the feature variables include a first feature variable and a second feature variable. The first feature variable includes one or more of the following: multispectral band features, multispectral band computational features, texture features, radar dual polarization features, and canopy height features. The second feature variable is a diameter at breast height (DBH) prediction feature variable. Extracting the feature variables corresponding to forest ground biomass samples from multi-source remote sensing data includes: resampling the multi-source remote sensing data to the target resolution; spatially matching the resampled multi-source remote sensing data with multiple forest ground biomass samples to determine the central pixel of the sampling point corresponding to the forest ground biomass sample in the multi-source remote sensing data; extracting features from the central pixel and pixels within the target range of the central pixel to obtain the first feature variable; and determining the second feature variable based on the canopy height feature variable and the DBH prediction model.

[0012] Optionally, based on the canopy height feature variable and the pre-built diameter at breast height (DBH) prediction model, the second feature variable is determined, including: obtaining a third sample, wherein the third sample includes observations of canopy height and DBH; constructing a fitting relationship between canopy height and DBH based on the third sample to obtain the DBH prediction model; inputting the canopy height feature variable into the DBH prediction model, and the DBH prediction model outputting the second feature variable.

[0013] A second aspect of this application provides a forest aboveground biomass estimation device, comprising: an acquisition module for acquiring multiple forest aboveground biomass samples of a target area, wherein the forest aboveground biomass samples include a first sample and a second sample, the first sample being used for model training and the second sample for spatial heterogeneity partitioning; a first training module for training a global forest aboveground biomass estimation model at the target area scale using the first sample, and determining the prediction residual of the second sample based on the trained global forest aboveground biomass estimation model; a partitioning module for determining the spatial heterogeneity score of each second sample based on the prediction residual, and spatially partitioning the target area into multiple target sub-regions based on the spatial heterogeneity score of the second sample; a second training module for extracting a first sample from each target sub-region, and training a corresponding local forest aboveground biomass estimation model based on the first sample of each target sub-region, to obtain local forest aboveground biomass estimation models corresponding to multiple target sub-regions; and an estimation module for partitioning the pixels to be predicted within the target area into corresponding target sub-regions, and calling the local forest aboveground biomass estimation model corresponding to the target sub-regions to estimate the forest aboveground biomass of the pixels to be predicted.

[0014] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform the forest aboveground biomass estimation method as described in the above embodiments.

[0015] A fourth aspect of this application provides a computer-readable storage medium having a computer program or instructions stored thereon, which is executed by a processor to perform the forest aboveground biomass estimation method as described above.

[0016] A fifth aspect of this application provides a computer program product, including a computer program or instructions, which, when executed, implement the forest aboveground biomass estimation method as described in the above embodiments.

[0017] Therefore, this application has at least the following beneficial effects: This application embodiment can divide multiple forest ground biomass samples in a target area into a first sample and a second sample. A global forest ground biomass estimation model at the target area scale is trained using the first sample. Based on the trained global forest ground biomass estimation model, the prediction residual of the second sample is determined. Based on the prediction residual, the spatial heterogeneity score of the second sample is determined. The target area is then spatially divided into multiple target sub-regions based on the spatial heterogeneity score. This accurately identifies and divides areas with large estimation errors in the global forest ground biomass estimation model, eliminating the need for subjective manual region division. Furthermore, a corresponding local forest ground biomass estimation model is trained for each target sub-region based on the first sample. The trained local forest ground biomass estimation model is used to estimate the forest ground biomass of the pixels to be predicted within the target area, thereby improving the accuracy and reliability of forest ground biomass estimation. Simultaneously, a local model can be constructed without a large number of samples, significantly reducing modeling computation costs and balancing large-scale measurement efficiency and estimation accuracy. This solves the technical problems in related technologies where unified estimation models are difficult to adapt to the spatial heterogeneity of forest ground biomass, leading to high estimation errors and limited estimation accuracy in local areas.

[0018] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0019] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a forest aboveground biomass estimation method provided according to an embodiment of this application; Figure 2 This is a flowchart of an explicit modeling method for spatial heterogeneity estimation in forest aboveground biomass, provided according to an embodiment of this application. Figure 3 This is a comparison chart of the accuracy of the global model provided according to the embodiments of this application and the model of the method of this application; Figure 4 This is a spatial heterogeneity partitioning result diagram provided according to an embodiment of this application; Figure 5 This is a diagram showing the independent verification results of field quadrats based on the gradient boosting decision tree algorithm provided in the embodiments of this application; Figure 6 This is a diagram showing the results of cross-validation uncertainty analysis based on the gradient boosting decision tree algorithm provided in the embodiments of this application. Figure 7 This is a schematic diagram illustrating the interpretability analysis results provided according to the embodiments of this application; Figure 8This is a schematic diagram of a forest aboveground biomass estimation device provided according to an embodiment of this application; Figure 9 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0020] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0021] Traditional forest aboveground biomass measurement mainly relies on field surveys and destructive sampling methods. While these methods offer high accuracy, they suffer from high costs, high labor intensity, and limited coverage, making them unsuitable for continuous estimation at large regional scales. Therefore, related technologies typically utilize field plot data, airborne or spaceborne lidar derivatives as reference samples, and combine them with optical remote sensing imagery, synthetic aperture radar imagery, canopy height data, and other environmental features to achieve forest aboveground biomass inversion and spatial mapping through regression modeling.

[0022] Machine learning methods, capable of handling complex nonlinear relationships and adapting to various types of tabular feature data, have been widely applied to large-scale forest aboveground biomass estimation tasks. However, due to spatial differences in factors such as forest type, topographic conditions, climate background, canopy structure, and disturbance history, there is often significant spatial heterogeneity and non-stationary relationship between forest aboveground biomass and remote sensing features in different regions. One type of related technology typically relies on the assumption of spatial stationarity, using a single global model for estimation. This makes it difficult to explicitly characterize spatial heterogeneity, easily leading to underfitting in local areas and affecting the accuracy and reliability of large-scale forest aboveground biomass estimation results. Another type of method employs local modeling strategies such as geographically weighted regression. While this can consider spatial non-stationarity to some extent, it usually relies on a large number of field quadrats and involves high model building and computational costs, thus limiting its application at larger regional scales.

[0023] To alleviate the above problems, there are currently methods for hierarchical modeling of the study area according to administrative divisions, vegetation types, or regular grids. However, such methods usually rely on relatively subjective and singular zoning criteria, making it difficult to adaptively identify real heterogeneous regions based on the spatial distribution of model errors.

[0024] To this end, this application provides a method for estimating forest aboveground biomass. In this method, multi-source remote sensing images are acquired and feature variables are extracted; a global baseline biomass estimation model that does not consider spatial heterogeneity is constructed; a model-independent, data-driven method for spatial heterogeneity identification, zoning, and verification is constructed; forest aboveground biomass extrapolation and spatial mapping are achieved based on local model automatic matching; and the zoning results are analyzed for interpretability based on the SHAP (SHapley Additive exPlanations) method.

[0025] Specifically, Figure 1 This is a flowchart illustrating a method for estimating forest aboveground biomass provided in an embodiment of this application.

[0026] like Figure 1 As shown, this method for estimating forest aboveground biomass includes the following steps: In step S101, multiple forest aboveground biomass samples of the target area are obtained, including a first sample and a second sample. The first sample is used for model training, and the second sample is used for spatial heterogeneity partitioning.

[0027] Each forest aboveground biomass sample includes characteristic variables and forest aboveground biomass reference values; the target area can be understood as the area to be studied; the first sample is used for model training and can be called the training sample; the second sample is used to calculate the model prediction residuals and identify spatial heterogeneity and can be called the validation sample.

[0028] The source of the forest aboveground biomass samples in this application embodiment can be the aboveground biomass product GEDI L4A (Global Ecosystem Dynamics Investigation Level 4A Footprint Level Aboveground Biomass Density) based on spaceborne lidar data; this application embodiment can also delineate a portion of the samples as test samples to evaluate the accuracy of the model's estimation results.

[0029] Furthermore, it should be noted that before using samples based on spaceborne lidar data in this application embodiment, quality control and sample screening are required. Specifically, this may include: retaining samples whose quality control parameters meet preset conditions, removing degraded samples, retaining full-power beam observation samples, setting waveform sensitivity thresholds, removing water and urban area samples, and retaining forest samples with forest coverage greater than a preset threshold. The preset threshold can be set according to the actual situation and is not limited thereto.

[0030] In one embodiment of this application, obtaining multiple forest aboveground biomass samples in a target area includes: obtaining multi-source remote sensing data of the target area; extracting feature variables of the corresponding forest aboveground biomass samples from the multi-source remote sensing data; and obtaining reference values ​​of forest aboveground biomass for multiple forest aboveground biomass samples in the target area.

[0031] The multi-source remote sensing data may include multispectral image data Sentinel-2, radar data Sentinel-1, etc., and the radar data may be synthetic aperture radar data; the feature variables include a first feature variable and a second feature variable. The first feature variable includes one or more of the following: multispectral band features, multispectral band computational features, texture features, radar dual polarization features, and canopy height features. The second feature variable is a diameter at breast height (DBH) prediction feature variable.

[0032] It is understood that the embodiments of this application can acquire multi-source remote sensing data of the target area and extract feature variables of the corresponding forest aboveground biomass samples from the multi-source remote sensing data in order to carry out subsequent model training and spatial heterogeneity partitioning.

[0033] In one embodiment of this application, extracting feature variables corresponding to forest ground biomass samples from multi-source remote sensing data includes: resampling the multi-source remote sensing data to a target resolution; spatially matching the resampled multi-source remote sensing data with multiple forest ground biomass samples to determine the central pixel of the sampling point corresponding to the forest ground biomass sample in the multi-source remote sensing data; extracting features from the central pixel and pixels within the target range of the central pixel to obtain a first feature variable; and determining a second feature variable based on the canopy height feature variable and a pre-built diameter at breast height (DBH) prediction model.

[0034] The target resolution can be set according to the actual situation, such as 20m; the target range can be set according to the spatial footprint range and positioning error of the sample data, such as 60m.

[0035] Since remote sensing data from different sources have different spatial resolutions, this application embodiment can resample multi-source remote sensing data to the target resolution to ensure that the pixel size and spatial coordinates of different sources correspond and avoid spatial errors. The resampled multi-source remote sensing data is spatially matched with multiple forest ground biomass samples to determine the central pixel of the sampling point corresponding to the forest ground biomass sample in the multi-source remote sensing data. Features are extracted from the central pixel and the pixels within the target range of the central pixel to obtain the first feature variable. The second feature variable is determined based on the canopy height feature variable and the pre-constructed diameter at breast height prediction model.

[0036] Specifically, the spatial matching process of forest ground biomass samples and multi-source remote sensing data in this application embodiment includes: determining the corresponding central pixel position of the forest ground biomass sample in the multi-source remote sensing data, and extracting feature values ​​within a 60m range of the central pixel as feature variables of the corresponding forest ground biomass sample.

[0037] In one embodiment of this application, determining a second feature variable based on the canopy height feature variable and a pre-built diameter at breast height (DBH) prediction model includes: obtaining a third sample, wherein the third sample includes observed values ​​of canopy height and DBH; constructing a fitting relationship between canopy height and DBH based on the third sample to obtain a DBH prediction model; inputting the canopy height feature variable into the DBH prediction model, and the DBH prediction model outputting the second feature variable.

[0038] The third sample can be derived from field quadrat or single-tree measured data.

[0039] Due to the limitations of remote sensing, it is not possible to directly obtain the diameter at breast height (DBH) of trees. However, there is a corresponding relationship between crown height and DBH. Therefore, in this application embodiment, a third sample including crown height and DBH observations can be obtained. Based on the third sample, a fitting relationship between crown height and DBH is constructed to obtain a DBH prediction model. Based on the crown height feature variable and the DBH prediction model, a second feature variable is determined.

[0040] It should be noted that the embodiments of this application require adjusting the canopy height in the canopy height dataset based on the measured data to the target resolution in order to achieve spatial alignment between the canopy height and the sampling points and remote sensing data, so as to ensure the accuracy of the canopy height and diameter-to-chest fitting inversion results.

[0041] Specifically, the embodiments of this application can construct a fitting relationship between canopy height and diameter at breast height (DBH) based on field quadrat or single tree measured data, estimate the predicted DBH value based on canopy height data and the fitting relationship, and use the predicted DBH value as a feature variable.

[0042] In step S102, a global forest aboveground biomass estimation model at the target area scale is trained using the first sample, and the prediction residual of the second sample is determined based on the trained global forest aboveground biomass estimation model.

[0043] The global forest aboveground biomass estimation model can be constructed using machine learning regression algorithms, including random forest, gradient boosting tree, extreme gradient boosting, lightweight gradient boosting, class boosting, and multilayer perceptron. The choice can be made based on the actual situation without specific limitations. The input of the global forest aboveground biomass estimation model is the feature variable, the training label is the forest aboveground biomass reference value, and the output is the forest aboveground biomass estimation value.

[0044] It is understood that the embodiments of this application use the first sample to train a global forest aboveground biomass estimation model at the target scale, and determine the prediction residual of the second sample based on the trained global forest aboveground biomass estimation model that does not consider spatial heterogeneity, so as to be used for subsequent spatial heterogeneity identification and partitioning.

[0045] In this embodiment, the feature variables of the second sample can be input into the trained global forest aboveground biomass estimation model. The global forest aboveground biomass estimation model outputs the forest aboveground biomass estimation value, and the prediction residual is calculated based on the forest aboveground biomass reference value and the estimated value corresponding to the second sample.

[0046] In step S103, the spatial heterogeneity score corresponding to the sampling point of each second sample is determined based on the prediction residual, and the target region is spatially divided based on the spatial heterogeneity score to obtain multiple target sub-regions.

[0047] It is understood that the embodiments of this application can determine the spatial heterogeneity score corresponding to the sampling point of each second sample based on the prediction residual, and divide the target region into multiple target sub-regions based on the spatial heterogeneity score, so as to accurately identify the region with large global model estimation error.

[0048] In one embodiment of this application, spatial partitioning of a target region based on the spatial heterogeneity score of a second sample yields multiple target sub-regions. This includes: dividing the target region into regular grids and matching the second sample to the corresponding regular grid to obtain multiple spatial units; calculating the spatial heterogeneity score of the corresponding spatial unit based on the prediction residuals of all second samples within the spatial unit; using the target region as the region to be partitioned, sorting the spatial heterogeneity scores of the multiple spatial units within the region to be partitioned, and sequentially extracting spatial units based on the sorting results to divide the region to be partitioned into two candidate sub-regions; using the two candidate sub-regions as the binary partitioning results to be tested, and performing a target test on the binary partitioning results to be tested, wherein the target test includes an accuracy improvement test and a significance test; if the target test passes, retaining the binary partitioning results, and using the two candidate sub-regions as new regions to be partitioned, and recursively performing binary partitioning and target testing until a preset stopping condition is met, resulting in multiple target sub-regions.

[0049] Among them, the significance test is a significance test that takes into account spatial autocorrelation.

[0050] It is understood that, in this embodiment, the target region can first be divided into a regular grid, and the second sample can be matched to the corresponding regular grid to obtain multiple spatial units. The spatial heterogeneity score of the corresponding spatial unit is calculated based on the prediction residuals of all second samples of the spatial unit. Next, the target region is used as the region to be partitioned, and the spatial heterogeneity scores of the multiple spatial units of the region to be partitioned are sorted. Based on the sorting results, spatial units are extracted sequentially to divide the region to be partitioned into two candidate sub-regions. Then, the two candidate sub-regions are used as the binary partitioning results to be detected, and the binary partitioning results to be tested are subjected to target testing. Finally, if the target testing passes, the binary partitioning results are retained, and the two candidate sub-regions are used as new regions to be partitioned. The binary partitioning and target testing are recursively executed until the preset stopping condition is met to obtain multiple target sub-regions. The spatial heterogeneity of the target region is divided by the spatial heterogeneity score of the second sample.

[0051] In one embodiment of this application, a target test is performed on the binary partitioning result to be tested, including: obtaining the first forest aboveground biomass estimate corresponding to the region to be partitioned before partitioning, and the second forest aboveground biomass estimate corresponding to the candidate sub-regions after partitioning; based on the second forest aboveground biomass estimate, the first forest aboveground biomass estimate, and the corresponding forest aboveground biomass reference value, determining the change in prediction accuracy of the second sample of the candidate sub-region before and after partitioning; if the change in prediction accuracy does not meet the preset improvement condition, the binary partitioning result is not retained, and the partitioning of the region to be partitioned is stopped; if the change in prediction accuracy meets the preset improvement condition, a significance test is performed on the change in prediction accuracy, and the binary partitioning result is retained if the significance test passes, otherwise the binary partitioning result is not retained, and the partitioning of the region to be partitioned is stopped.

[0052] The first forest aboveground biomass estimate is the current complete region before further partitioning, using the model corresponding to the region before partitioning. The second forest aboveground biomass estimate is the forest aboveground biomass estimate calculated for each candidate sub-region after temporarily splitting the region to be partitioned into two candidate sub-regions. The significance check is performed by statistical comparison to determine whether there is a substantial difference in the estimation results and overall error before and after partitioning, so as to determine whether the partitioning operation is meaningful. In this application embodiment, the upper-tail paired T-test method can be used. The preset improvement condition can be that the change in prediction accuracy is improved by more than a set threshold, that is, the prediction error is significantly improved.

[0053] It is understood that the embodiments of this application can obtain the first forest aboveground biomass estimate corresponding to the area to be partitioned before partitioning, and the second forest aboveground biomass estimate corresponding to the candidate sub-areas after partitioning. Based on the second forest aboveground biomass estimate, the first forest aboveground biomass estimate, and the corresponding forest aboveground biomass reference value, the change in prediction accuracy of the second sample of the candidate sub-area before and after partitioning is determined. If the change in prediction accuracy does not meet the preset improvement condition, the binary partitioning result is not retained, and the partitioning of the area to be partitioned is stopped. If the change in prediction accuracy meets the preset improvement condition, the change in prediction accuracy is subjected to a significance test, and the binary partitioning result is retained when the significance test passes; otherwise, the binary partitioning result is not retained, and the partitioning of the area to be partitioned is stopped. The change in preset accuracy and significance test are used to intuitively determine whether the regional modeling brings substantial improvement, and to objectively screen effective partitions.

[0054] The significance verification process in this application includes: comparing the sample-level predicted scores of the partitioned model and the global benchmark model using the upper-tail paired T-test method considering spatial autocorrelation on the same validation sample set; grouping the validation samples according to spatial units; estimating the intra-group correlation coefficient (ICC) based on the variation in score differences between groups; and calculating the design effect (DEFF) based on the average group size (m). ; in, Design effect; is the average group size, i.e., the average number of validation samples within a single spatial unit; ICC is the intragroup correlation coefficient, used to measure the intragroup correlation of the difference in predicted scores among samples within the same spatial unit.

[0055] Furthermore, the effective sample size is calculated based on the design effect. and degrees of freedom (DF): , ; in, This represents the total number of samples representing the score differences; is the effective sample size; DF is the degrees of freedom; based on the test results after correction for degrees of freedom, it is determined whether the current candidate partition meets the preset significance condition. If the preset significance condition is met, the current candidate partition result is accepted; otherwise, the current candidate partition result is rejected.

[0056] In one embodiment of this application, the formula for calculating the spatial heterogeneity score of the candidate sub-region is as follows: ； in, Scoring the spatial heterogeneity of candidate sub-region S This represents the additional mean offset of the overall prediction residuals within the candidate sub-region relative to the baseline distribution. The sum of the inverse variance-weighted residuals of all second samples within the candidate sub-region S. , For the second sample In spatial location The predicted residual at the location, Spatial location The standard deviation of the baseline residual distribution. The sum of the inverse variance weights of all second samples within the candidate sub-region S. .

[0057] In one embodiment of this application, the preset stopping conditions include at least one of the following: the number of first samples in the region to be partitioned is less than a preset sample number threshold; the number of spatial units in the region to be partitioned is less than a preset spatial unit number threshold; the change in prediction accuracy corresponding to the binary partitioning result does not meet the preset improvement condition; the binary partitioning result fails the significance test; the current recursive partitioning depth reaches the preset maximum partitioning depth.

[0058] Among them, the preset sample number threshold, preset spatial unit number threshold, preset boosting condition, preset maximum partition depth, and preset conditions can all be set according to the actual situation, and no specific restrictions are imposed on them.

[0059] Specifically, in this application embodiment, when terminating the partitioning of the target region, preset stopping conditions can be set to terminate the partitioning, such as failure to achieve substantial improvement in accuracy, failure to pass the significance test, reaching the preset maximum partitioning depth, or the number of samples in the sub-region being less than the preset sample number threshold.

[0060] The process of spatial heterogeneity identification and partitioning based on predicted residuals in this application includes: 1. Scoring Function Module: Construct a spatial heterogeneity scoring function based on the prediction residuals of the second sample.

[0061] 2. Partition Optimization Module: Based on the spatial heterogeneity scoring function (i.e., the calculation formula for the spatial heterogeneity score of candidate sub-regions).

[0062] 3. Significance Test Module: Performs a significance test considering spatial autocorrelation on the candidate sub-regions, and recursively partitions the current region when the partitioning conditions are met to obtain the final spatial partitioning results.

[0063] 4. Repeat steps 1-3 to recursively optimize and adjust the spatial heterogeneity partitioning until the preset stopping condition is met: the splitting stops and the final partitioning result is obtained when any of the following conditions are met, including no substantial improvement in accuracy, failure to pass the significance test, reaching the deepest stratification number, or the number of samples in the sub-region is less than the preset sample number threshold.

[0064] The scoring function module is constructed based on the difference in the predicted residual distribution within the candidate sub-region under the null and alternative hypotheses. It quantifies the degree of deviation of the sample residuals within the candidate sub-region relative to the baseline residual distribution. Under the null hypothesis, the predicted residuals within the current region follow a Gaussian distribution centered on the baseline expectation. Under the alternative hypothesis, the predicted residuals within the candidate sub-region exhibit an additional mean shift relative to the baseline distribution. A regular grid is overlaid on the study area, and each grid cell serves as the basic spatial unit for the candidate sub-region search. The scoring function can be expressed as: ; in, , , Indicates candidate sub-regions. Indicates the second sample In spatial location The predicted residual at the location, This represents the standard deviation of the baseline residual distribution at that location. This represents the additional mean offset relative to the baseline distribution within the candidate sub-region. According to the maximum likelihood estimation principle, its estimated value is: .

[0065] The scoring function measures the degree of deviation of the residual distribution within a candidate subregion from the baseline distribution. The higher the score, the greater the likelihood that the candidate subregion has spatial heterogeneity.

[0066] The partitioning optimization module includes: based on the additivity of the spatial heterogeneity scoring function on spatial units, sorting spatial units and forming multiple candidate prefix subsets, using a linear time subset scanning method to identify the candidate sub-region with the highest score, and combining the candidate sub-region and its complement to form the two partitioning result of the current region.

[0067] For example, the process of region division in this application embodiment includes: taking target region A as the region to be divided, the target region being the region to be divided before division, corresponding to a global forest aboveground biomass estimation model, determining the prediction residual of the second sample based on the global forest aboveground biomass estimation model, determining the spatial heterogeneity score based on the prediction residual of the sample, and then dividing region A into two candidate sub-regions B and C based on the spatial heterogeneity score. Furthermore, training local forest aboveground biomass estimation models corresponding to regions B and C based on the first samples within regions B and C, and recursively dividing regions B and C after passing target verification. Target verification involves comparing the prediction results of the model before division of region A with the prediction results of the local models after division of regions B and C based on the forest aboveground biomass reference value corresponding to the second sample within region A, determining the change in prediction accuracy before and after division, and judging whether the change in prediction accuracy meets preset improvement and significance conditions.

[0068] Taking region B as an example, the local forest aboveground biomass estimation model corresponding to region B is used as the pre-partition model in the current partitioning process. The prediction residual of the second sample is determined based on the pre-partition model, and the spatial heterogeneity score is determined based on the sample's prediction residual. Then, based on the spatial heterogeneity score, region B is divided into two candidate sub-regions, region D and region E. Furthermore, local forest aboveground biomass estimation models corresponding to regions D and E are trained based on the first samples within regions D and E. After passing target validation in regions D and E, regions D and E are recursively partitioned. Target validation involves comparing the prediction results of the pre-partition model of region B with the post-partition local models corresponding to regions D and E based on the forest aboveground biomass reference value corresponding to the second sample within region B, determining the change in prediction accuracy before and after partitioning, and judging whether this change in prediction accuracy meets the preset improvement and significance conditions.

[0069] In step S104, the first sample of each target sub-region is extracted, and a pre-constructed local forest aboveground biomass estimation model is trained based on the first sample of each target sub-region to obtain local forest aboveground biomass estimation models corresponding to multiple target sub-regions.

[0070] The local forest aboveground biomass estimation model is constructed in the same way as the global forest aboveground biomass estimation model, and the input and output variables are also the same.

[0071] It is understood that, in the embodiments of this application, the first sample of each target sub-region can be extracted, and a local forest aboveground biomass estimation model corresponding to each target sub-region can be obtained based on the training of a pre-built local forest aboveground biomass estimation model.

[0072] In step S105, the pixels to be predicted within the target area are divided into corresponding target sub-regions, and the local forest aboveground biomass estimation model corresponding to the target sub-region is called to estimate the forest aboveground biomass of the pixels to be predicted.

[0073] It is understood that, in the embodiments of this application, the pixels to be predicted can be divided into corresponding target sub-regions; the local forest aboveground biomass estimation model corresponding to the target sub-region is called to estimate the forest aboveground biomass of the pixels to be predicted, so as to improve the accuracy of the forest aboveground biomass estimation of the pixels to be predicted.

[0074] This application embodiment can divide multiple forest ground biomass samples in a target area into a first sample and a second sample. The first sample is used to train a global forest ground biomass estimation model at the target scale. Based on the trained global forest ground biomass estimation model, the prediction residual of the second sample is determined. Based on the prediction residual, the spatial heterogeneity score of the second sample is determined. The target area is spatially divided into multiple target sub-regions based on the spatial heterogeneity score, thereby accurately identifying and dividing areas with large estimation errors in the global forest ground biomass estimation model without the need for subjective human division of regions. Then, based on the first sample of each target sub-region, a corresponding local forest ground biomass estimation model is trained for each target sub-region. The trained local forest ground biomass estimation model is used to estimate the forest ground biomass of the pixels to be predicted in the target area, thereby improving the accuracy and reliability of forest ground biomass estimation. At the same time, a local model can be built without a large number of samples, which greatly reduces the modeling and computation costs and balances the efficiency of large-scale measurement and estimation accuracy.

[0075] Furthermore, in this embodiment, the forest aboveground biomass can be estimated for all pixels within the target area using a local forest aboveground biomass estimation model, and a spatial partitioning structure can be generated. Then, SHAP can be used to perform interpretability analysis on the local forest aboveground biomass estimation model for each partition sub-region, and the results of the interpretability analysis can be visualized to show the differences in the contribution of each input feature to the forest aboveground biomass estimation results, thereby revealing the potential driving factors of spatial heterogeneity of forest aboveground biomass.

[0076] Specifically, the forest aboveground biomass estimation method of this application mainly includes the following steps: Step 1: Select the GEDI L4A Version 2 aboveground biomass density product as the sample data of forest aboveground biomass; the aboveground biomass density value obtained by screening through quality indicators is used as the dependent variable in the subsequent construction of the baseline model.

[0077] Step 2: Using the geographic location of the GEDI footprint as the center, match its corresponding pixel location in the multi-source remote sensing data that has been resampled to 20m and spatially aligned; considering the scale of the GEDI footprint and spatial positioning error, extract its feature values ​​within a 60m range, and calculate its mean as the feature variable of the corresponding sample to construct the independent variable in the subsequent model. The feature variable includes one or more of the following: multispectral band features, multispectral band operational features, texture features, synthetic aperture radar dual polarization features, canopy height features, and diameter at breast height prediction features.

[0078] Step 3: Based on the dependent variable obtained in Step 1 and the independent variable obtained in Step 2, a global forest aboveground biomass estimation model (hereinafter referred to as the global model) is constructed using a machine learning regression algorithm, and the accuracy of the model is evaluated using the coefficient of determination, root mean square error, relative root mean square error, and bias.

[0079] Step 4: Based on the prediction residuals of the global model on the validation samples, construct a scoring function to identify spatial heterogeneity. The scoring function is based on the null hypothesis and alternative hypothesis, and is used to quantify the local offset of the residuals in the candidate sub-regions relative to the baseline distribution, and serves as the basis for subsequent recursive partitioning optimization.

[0080] Step 5: Overlay a regular grid onto the study area and aggregate the GEDI footprint samples and corresponding feature variables within each grid cell into a spatial cell; restrict the search for potential heterogeneous regions to combinations of predefined spatial cells, and perform bipartite optimization based on the scoring results using an efficient subset search strategy to identify the optimal candidate sub-regions.

[0081] Step 6: Consider the significance of spatial autocorrelation for the identified candidate sub-regions. Compare the difference in model prediction scores before and after partitioning based on the upper tail pairing T test, and correct the degrees of freedom according to spatial autocorrelation during the test. When the test result meets the preset significance condition, accept the current two-partition result; otherwise, stop the current region from splitting and use it as the final modeling sub-region.

[0082] Step 7: Continue to recursively execute steps 4 to 6 for the sub-regions that pass the significance test until the current region no longer meets the conditions for further splitting, thereby obtaining the final spatial heterogeneity partitioning results for forest aboveground biomass estimation.

[0083] Step 8: Based on the final spatial heterogeneity partitioning results obtained in Step 7, extract the GEDI footprint samples and their corresponding feature variables in each partition sub-region, and construct an independent local modeling sample set for each sub-region. The aboveground biomass density value retained in each sub-region is used as the dependent variable, and the feature variables extracted at the corresponding locations are used as independent variables.

[0084] Step 9: For each final sub-region, construct a local forest aboveground biomass estimation model (hereinafter referred to as the local model) to represent the local mapping relationship between forest aboveground biomass and characteristic variables in the corresponding spatial sub-region.

[0085] Step 10: Based on the final spatial heterogeneity partitioning results, all pixels to be predicted within the target area are automatically assigned to the corresponding partition sub-regions, and the local model corresponding to the sub-region is called to perform pixel-level prediction to obtain the forest aboveground biomass estimate for all pixels within the target area.

[0086] Step 11: Spatial stitching and output of the forest aboveground biomass estimate obtained in Step 10 to generate a spatial distribution map of forest aboveground biomass.

[0087] Step 12: Based on the final spatial partitioning results and the local models corresponding to each sub-region, select the local models to be interpreted and their corresponding sample data as the objects of interpretability analysis, and use the SHAP method to calculate the contribution value of each input feature to the prediction results of the local model.

[0088] Step 13: Based on the SHAP value, analyze the importance, contribution direction and contribution difference of the characteristic variables in each sub-region, and compare the characteristic contribution differences between different sub-regions to identify the key characteristics that dominate the estimation of forest aboveground biomass and reveal the potential driving mechanism of spatial heterogeneity.

[0089] Step 14: Visualize the analysis results obtained in Steps 12 and 13 to generate interpretable analysis results.

[0090] In summary, the main solutions in this application include: 1. A data-driven method for spatial heterogeneity identification and partitioning was developed. This method, centered on a residual-based scoring function, can explicitly identify spatial non-stationary relationships in forest aboveground biomass estimation, overcoming the limitations of traditional single global models in explicitly representing spatial heterogeneity. The scoring function, combined with efficient subset search and a recursive binary partitioning optimization strategy, can adaptively identify and partition potential heterogeneous regions, avoiding reliance on administrative divisions, vegetation types, or empirical rules. This improves the objectivity, applicability, and stability of the partitioning results. Furthermore, this method transforms spatial heterogeneity identification from an empirical judgment to a quantifiable statistical discrimination problem and reduces computational costs through efficient search, providing technical support for large-scale forest carbon storage assessment and regional carbon accounting.

[0091] 2. A partition significance assessment method considering spatial autocorrelation is designed. The upper-tail paired T-test is used to determine the significance of candidate partition results, and only the partition results that meet the requirements are retained, thereby reducing the risk of invalid splits and over-partitioning and improving the stability and reliability of spatial partition results. This method explicitly considers spatial autocorrelation factors in the partition determination process, which enhances the statistical robustness and spatial consistency of spatial heterogeneity identification results.

[0092] 3. A method for regional prediction and interpretability analysis of forest aboveground biomass considering spatial heterogeneity was designed. This method automatically matches the pixels of the sample or region to be predicted to the corresponding local model, realizing regional-scale prediction and continuous mapping of forest aboveground biomass. Simultaneously, the SHAP method is used to conduct interpretability analysis on multiple local models, quantifying the differences in the contribution and importance of various input features to the forest aboveground biomass estimation results in different sub-regions. This method not only achieves continuous mapping of forest aboveground biomass considering spatial heterogeneity but also reveals the key characteristics that dominate forest aboveground biomass estimation in different spatial sub-regions, enhancing the interpretability of the model results and providing richer analytical basis for large-scale forest carbon storage assessment and carbon accounting.

[0093] The following specific embodiment describes the forest aboveground biomass estimation method of this application. Figure 2 As shown, Figure 2 This is a flowchart of an explicit modeling method for spatial heterogeneity estimation in forest aboveground biomass, as described in this application.

[0094] Step 1: Acquire spaceborne lidar data, multispectral imagery, and synthetic aperture radar data for the study area. The spaceborne lidar data uses GEDI L4A products, the multispectral imagery uses Sentinel-2 data, and the radar data uses Sentinel-1 data.

[0095] Using the aboveground biomass density parameter from the GEDI L4A product as the dependent variable in the subsequent modeling sample set, quality control and sample screening of the GEDI L4A data were performed. Specifically, this included: removing poor-quality data by setting l4_quality_flag = 1 and degrade_flag = 0; retaining only full-power beam observation samples; setting the minimum waveform sensitivity threshold to 0.95; removing water samples; and retaining forest samples with a forest cover rate exceeding 20% ​​to obtain effective GEDI footprint samples that meet the modeling requirements.

[0096] Step 2: Extract canopy structure features based on high-resolution global canopy height maps.

[0097] Specifically, global canopy height data with a spatial resolution of 1m was acquired. This data consisted of canopy height estimates derived from the model inversion, rather than direct observations. The canopy height data was then aggregated to a spatial resolution of 20m to maintain consistency with other predictor variables. Based on the aggregated canopy height data, a fitting model between canopy height and diameter at breast height (DBH) was established using field plot observation data. The DBH predictor variable, DBH_pred, was estimated and used as an additional structural feature for subsequent forest aboveground biomass modeling.

[0098] Step 3: Resample the multispectral imagery, synthetic aperture radar data, and canopy structure feature data to a unified spatial resolution of 20m, and complete the spatial registration and alignment processing.

[0099] The acquired Sentinel-2 data underwent radiometric, geometric, and atmospheric corrections. Additionally, cloud removal was performed on all images before image synthesis based on the cloud mask band QA60. To ensure spatial resolution consistency, the original 10m and 20m bands were uniformly resampled to 20m. Multispectral reflectance characteristics of the target area were acquired, including one or more combinations of reflectance values ​​from bands B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12 in the Sentinel-2 images. For Sentinel-1 data, VV (Vertical-Vertical) and VH (Vertical-Horizontal) dual-polarization backscattering characteristics were acquired in the wide-swath acquisition interferometric mode.

[0100] Step 4: Construct feature variables based on band operations.

[0101] Based on the multispectral imagery and synthetic aperture radar data obtained in step 3, band-based operational feature variables for forest aboveground biomass estimation are constructed, as shown in Table 1. Table 1 is a feature variable table including vegetation index features, dual-polarization SAR features, and texture features. The vegetation index features are calculated from multispectral band reflectance, the dual-polarization SAR features are calculated from VV and VH backscattering coefficients, and the texture features are obtained from image grayscale statistical analysis. All feature variables in this embodiment are shown in Table 1.

[0102]

[0103] Step 5: Using the center position of each GEDI footprint as a reference, determine its corresponding center pixel position in each multi-source remote sensing image.

[0104] Considering that GEDI footprints have a certain spatial positioning error and their effective coverage area is greater than a single 20m pixel, this embodiment does not directly use the value of a single central pixel as the sample feature. Instead, based on the central pixel, it further extracts feature values ​​within a 60m range to obtain the statistical values ​​of the feature variables corresponding to the GEDI footprint sample, and constructs a sample set together with the corresponding aboveground biomass sample values.

[0105] Step 6: Further divide the obtained sample set into a training sample set, a validation sample set, and a test sample set. The training sample set is used for model training, the validation sample set is used for calculating model prediction residuals and subsequent spatial heterogeneity identification, and the test sample set is used for evaluating model accuracy.

[0106] Step 7: Based on the training sample set, a global baseline biomass estimation model that ignores spatial heterogeneity is constructed using machine learning regression algorithms. This model is used to characterize the global mapping relationship between forest aboveground biomass and multi-source remote sensing features within the study area. The machine learning regression algorithms include one or more of the following: random forest, gradient boosting tree, extreme gradient boosting, lightweight gradient boosting, class boosting, and multilayer perceptron.

[0107] Step 8: Based on the trained global model, make predictions on the validation and test sample sets to obtain the predicted values ​​and prediction residuals for the corresponding samples. The prediction residuals on the validation sample set are used for subsequent construction of the spatial heterogeneity scoring function and partition identification, while the prediction results on the test sample set are used for model accuracy evaluation.

[0108] Step 9: Use the coefficient of determination R 2 Root mean square error Relative root mean square error and deviation The accuracy of the global baseline model is evaluated using the following formula: ; ; ; ; in, This represents the total number of validation samples (i.e., the second sample). Indicates the first Reference values ​​for each validation sample. Indicates the first The predicted value of each validation sample. This represents the average of the reference values.

[0109] Step 10: Candidate region definition and spatial unit construction.

[0110] Based on the prediction residuals of the global model on the validation sample set, spatial heterogeneity identification analysis is carried out for forest aboveground biomass estimation.

[0111] To define candidate regions for spatial scanning, a regular grid is overlaid on the study area, and the GEDI footprint samples and corresponding remote sensing features within each grid cell are aggregated into spatial cells. The search for the optimal subset is restricted to combinations of predefined spatial cells. In this embodiment, the grid resolution is set to 0.2° to mitigate the sparse and uneven spatial distribution of the GEDI footprint and to avoid generating too many empty cells.

[0112] Step 11: Constructing the spatial heterogeneity hypothesis.

[0113] In this embodiment, record To verify the sample In spatial location Predicted residuals at location, spatial position This corresponds to the spatial unit obtained by the aforementioned regular grid division. For each spatial location... Record separately and Let $\mathbf{x}$ be the expected value and standard deviation of the baseline residual distribution at this location. Under the null hypothesis, it is assumed that the predicted residuals of each validation sample in the current region follow a Gaussian distribution centered at its baseline expected value. Under the alternative hypothesis, the candidate sub-regions... There is an additional mean offset relative to the baseline distribution. The offset This is used to characterize the degree of local deviation of the predicted residual from the baseline expectation within the candidate sub-region. Therefore, the null hypothesis is... With candidate sub-regions The alternative assumptions The definition is as follows: ; in, Let the model prediction residual be the i-th validation sample. and Spatial location The expected value and standard deviation of the baseline residual distribution. Set to 0, It follows a normal Gaussian distribution.

[0114] ; Step 12: Spatial heterogeneity scoring function construction module.

[0115] In this embodiment, candidate sub-regions The spatial heterogeneity scoring function is constructed based on the difference in residual distributions under the null and alternative hypotheses. By quantifying the local shift of the predicted residuals within a candidate sub-region relative to the baseline distribution, a score is obtained to evaluate the spatial heterogeneity of the candidate sub-region. The scoring function can be expressed as: ; in, , .

[0116] This represents the additional mean offset relative to the baseline distribution within the candidate sub-region. According to the maximum likelihood estimation principle, the estimated value of the additional mean offset q is: 。

[0117] Step 13: Spatial Heterogeneity Partition Optimization Module.

[0118] Based on the spatial heterogeneity scoring function obtained in step 12, candidate regions are searched for spatial units. Since the scoring function satisfies the additivity condition on the spatial units corresponding to the candidate sub-regions, this embodiment employs a linear-time subset scanning method for efficient searching of candidate sub-regions to avoid exhaustive calculation of all combinations of spatial units. Specifically, all spatial units are sorted, and candidate prefix subsets are formed sequentially according to the sorting results. The scoring value corresponding to each candidate subset is calculated, and the candidate sub-region S with the highest score is identified. Candidate sub-region S Together with its complement, they constitute the two-partition result of the current region, which serves as the input for subsequent partition validity checks.

[0119] Step 14: Spatial heterogeneity significance test module.

[0120] The candidate binary partitioning results obtained in step 13 are subjected to a significance test to determine whether the current partitioning can significantly improve the model's prediction performance. Specifically, the sample-level prediction scores in two scenarios are compared on the same validation sample set: one is the prediction score T of the model without partitioning. base Secondly, the prediction score T is obtained by training local models for each of the two sub-regions and making predictions using the current two-part partitioning results. part The current binary partitioning result is accepted only if the predicted performance improvement of the partitioned model relative to the baseline model before partitioning is significant.

[0121] In this embodiment, the paired T-test of the upper tail is used to compare the two cases, and the test statistic is expressed as: ; In this context, the predicted score for each sample is defined as the negative squared error, meaning that a higher score indicates better prediction performance. To test the statistic, This represents the error improvement sequence of the local-global model under the same sample. The sample mean of the difference sequence. is the sample standard deviation of the difference sequence.

[0122] To account for the spatial autocorrelation effect among validation samples, the validation samples are grouped according to the aforementioned spatial units, and a spatial correlation correction is applied to the significance test. This spatial autocorrelation correction reduces the significance inflation problem caused by spatially correlated samples, thereby improving the reliability of the partition validity determination. When the significance test result meets the preset significance condition, the current candidate binary partition result is accepted, and subsequent recursive partitioning is performed on the obtained sub-region; when the test result does not meet the preset significance condition, the current partition result is rejected, and the current region is regarded as the final modeling sub-region that will no longer be split.

[0123] Step 15: Continue recursively executing steps 11 to 14 until a stopping condition is triggered. The stopping condition includes at least one of the following: no substantial improvement in accuracy is achieved, i.e., the change in prediction accuracy does not meet the preset improvement condition, the significance test is not passed, the preset deepest stratification number is reached, or the number of samples within a sub-region is less than a preset sample number threshold. Through the above recursive partitioning process, the spatial heterogeneity partitioning results of the study area are finally obtained.

[0124] In this embodiment, forest region A is selected as the study area. The following explanation combines the accuracy comparison results of the specific global model and the model of the method in this application with the partitioning results. Figure 3 and Figure 4 As shown, Figure 3 This is a comparison chart of the accuracy of the global model and the model of the method in this application. Figure 4 The spatial heterogeneity partitioning result map generated by the method of this application.

[0125] Model results: such as Figure 3As shown, under various machine learning regression algorithms, including CatBoost (Categorical Boosting), GBDT (Gradient Boosting Decision Tree), LightGBM (Light Gradient Boosting Machine), MLP (Multi-Layer Perceptron), RF (Random Forest), and XGBoost (Extreme Gradient Boosting), compared with the global benchmark model that does not consider spatial heterogeneity, the model accuracy consistently improves after adopting the spatial heterogeneity identification, recursive partitioning, and local modeling methods of this application. Specifically, the determination coefficients of the method in this application are improved to varying degrees under different algorithms, while the relative root mean square error is reduced, indicating that this application can stably improve the performance of forest aboveground biomass estimation in different types of regression models. Considering that the GBDT algorithm shows good accuracy and significant gain in this embodiment, the results of the GBDT algorithm are used as representatives for subsequent spatial distribution, uncertainty analysis, and interpretability analysis.

[0126] Visualization results: such as Figure 4 As shown, the spatial heterogeneity partitioning results obtained by different machine learning algorithms differ in the number of partitions and the complexity of their boundaries. Specifically, GB and XGBoost produce relatively finer partitioning results, CatBoost and LightGBM show better spatial continuity, while RF and MLP produce relatively coarser partitions. Despite differences in local details, the partitioning results obtained by each algorithm are generally consistent in terms of large-scale spatial patterns, indicating that the method proposed in this application can stably identify spatial heterogeneous structures within the study area and has good algorithmic adaptability and spatial interpretation capabilities.

[0127] Step 16: Based on the final spatial heterogeneity partitioning results from Step 15, extract GEDI samples and their corresponding feature variables from each sub-region of the partition to construct an independent local modeling sample set for each sub-region. The aboveground biomass density value retained within each sub-region is used as the dependent variable, and the feature variables extracted at the corresponding locations are used as independent variables.

[0128] Step 17: For each final sub-region, construct a local forest aboveground biomass estimation model using the GBDT algorithm. During this process, hyperparameter optimization is performed on the algorithm to obtain the final global model and various local models based on GBDT.

[0129] Step 18: Based on the correspondence between the spatial location of each pixel to be predicted within the target area and the final zoning result, determine the sub-region to which it belongs, and establish an automatic matching relationship between the pixel and the local model. Call the local model corresponding to each sub-region to perform pixel-level prediction for the pixels within the region, and spatially stitch the prediction results of each sub-region to generate a spatial distribution map of forest aboveground biomass.

[0130] Step 19: Based on the field quadrat data, perform independent verification and uncertainty testing on the zonal mapping results that consider spatial heterogeneity and the global mapping results that do not consider spatial heterogeneity.

[0131] The following will combine Figure 5 and Figure 6 To describe the results of the independent validation and uncertainty test of the field plots in this application, such as Figure 5 As shown, Figure 5 This is a graph showing the independent validation results of field quadrats based on the GBDT algorithm. Figure 5 The results show that, compared with the global model that does not consider spatial heterogeneity, the accuracy of forest aboveground biomass estimation is improved by adopting the partitioned modeling method of this application. Under the GBDT algorithm, the correlation coefficient increased from 0.65 to 0.67, the coefficient of determination increased from 0.28 to 0.43, the root mean square error decreased from 65.31 Mg / ha to 58.23 Mg / ha, and the relative root mean square error decreased from 32.13% to 28.64%.

[0132] like Figure 6 As shown, Figure 6 The graph shows the results of the cross-validation uncertainty analysis based on the GBDT algorithm. Figure 6 As shown, the distribution of predicted standard deviations obtained based on five-fold cross-validation indicates that the uncertainty of the mapping results is significantly reduced after adopting the method of this application. Specifically, the mean of the predicted standard deviations decreased from 6.15 Mg / ha to 4.49 Mg / ha, and the 95th percentile decreased from 13.93 Mg / ha to 10.81 Mg / ha. These results demonstrate that the method of this application can improve the accuracy of forest aboveground biomass mapping while reducing prediction uncertainty.

[0133] Step 20: Select the local forest aboveground biomass estimation model and its input feature data corresponding to each sub-region as the interpretation object, and use the SHAP method to calculate the contribution value of each input feature to the model prediction result. Based on the SHAP value, analyze the importance, contribution direction and contribution difference of multi-source remote sensing features in each sub-region, and visualize the results to reveal the potential driving mechanism of spatial heterogeneity of forest aboveground biomass.

[0134] The following will combine Figure 7To describe the results of the interpretability analysis of this application, such as Figure 7 As shown, there are significant differences in feature importance and feature interaction relationships among different sub-regions and the global model, indicating significant spatial heterogeneity in the forest groundwater biomass estimation process. Among these, the canopy height feature (CH) exhibits high importance in all sub-regions and the global model, and is one of the key features dominating forest groundwater biomass estimation; features such as S2REP, B5, and DBHpred also contribute strongly in multiple sub-regions. The ranking and interaction strength of the dominant features are inconsistent across different sub-regions, indicating that the response mechanism of forest groundwater biomass to remote sensing features does indeed differ across different spatial sub-regions. These results demonstrate that the method proposed in this application can reveal the potential driving mechanism of spatial heterogeneity in forest groundwater biomass from the perspective of feature contribution and interaction.

[0135] In summary, the embodiments of this application can use GEDI spaceborne lidar data, Sentinel-2 multispectral imagery, Sentinel-1 synthetic aperture radar data, and canopy height data as inputs. By combining spatial heterogeneity identification, zoning, and verification methods with machine learning algorithms, it achieves large-scale forest aboveground biomass estimation and mapping that considers spatial heterogeneity. This application can explicitly characterize the spatial heterogeneity in forest aboveground biomass estimation in a data-driven manner, improve the model estimation accuracy and spatial expressive power, and obtain forest aboveground biomass estimation results with more reasonable spatial distribution.

[0136] According to the forest aboveground biomass estimation method proposed in this application, multiple forest aboveground biomass samples in a target area can be divided into a first sample and a second sample. A global forest aboveground biomass estimation model at the target area scale is trained using the first sample. The prediction residual of the second sample is determined based on the trained global forest aboveground biomass estimation model. The spatial heterogeneity score of the second sample is determined based on the prediction residual. The target area is spatially divided into multiple target sub-regions based on the spatial heterogeneity score, thereby accurately identifying and dividing areas with large estimation errors in the global forest aboveground biomass estimation model without the need for subjective human division of regions. Then, a corresponding local forest aboveground biomass estimation model is trained for each target sub-region based on the first sample of each target sub-region. The trained local forest aboveground biomass estimation model is used to estimate the forest aboveground biomass of the pixels to be predicted in the target area, thereby improving the accuracy and reliability of forest aboveground biomass estimation and obtaining a more spatially reasonable forest aboveground biomass estimation result. At the same time, a local model can be constructed without a large number of samples, significantly reducing the modeling and computation costs.

[0137] Next, the forest aboveground biomass estimation apparatus of this application is described with reference to the accompanying drawings.

[0138] Figure 8This is a schematic diagram of a forest aboveground biomass estimation device provided according to an embodiment of this application.

[0139] like Figure 8 As shown, the forest aboveground biomass estimation device 10 includes: an acquisition module 100, a first training module 200, a division module 300, a second training module 400, and an estimation module 500.

[0140] The acquisition module 100 is used to acquire multiple forest ground biomass samples of the target area, including a first sample and a second sample. The first sample is used for model training, and the second sample is used for spatial heterogeneity partitioning. The first training module 200 is used to train a global forest ground biomass estimation model at the target area scale using the first sample, and to determine the prediction residual of the second sample based on the trained global forest ground biomass estimation model. The partitioning module 300 is used to determine the spatial heterogeneity score of each second sample based on the prediction residual, and to spatially partition the target area to obtain multiple target sub-regions based on the spatial heterogeneity score of the second sample. The second training module 400 is used to extract the first sample of each target sub-region, and to train the corresponding local forest ground biomass estimation model based on the first sample of each target sub-region to obtain local forest ground biomass estimation models corresponding to multiple target sub-regions. The estimation module 500 is used to partition the pixels to be predicted in the target area to the corresponding target sub-regions, and to call the local forest ground biomass estimation model corresponding to the target sub-region to estimate the forest ground biomass of the pixels to be predicted.

[0141] It should be noted that the foregoing explanation of the forest aboveground biomass estimation method embodiment also applies to the forest aboveground biomass estimation device of this embodiment, and will not be repeated here.

[0142] According to the forest aboveground biomass estimation device proposed in this application, multiple forest aboveground biomass samples in a target area can be divided into a first sample and a second sample. A global forest aboveground biomass estimation model at the target area scale is trained using the first sample. The prediction residual of the second sample is determined based on the trained global forest aboveground biomass estimation model. The spatial heterogeneity score of the second sample is determined based on the prediction residual. The target area is spatially divided into multiple target sub-regions based on the spatial heterogeneity score. This accurately identifies and divides areas with large estimation errors in the global forest aboveground biomass estimation model, eliminating the need for subjective human division of regions. Then, a corresponding local forest aboveground biomass estimation model is trained for each target sub-region based on the first sample of each target sub-region. The trained local forest aboveground biomass estimation model is used to estimate the forest aboveground biomass of the pixels to be predicted in the target area, thereby improving the accuracy and reliability of forest aboveground biomass estimation. At the same time, a local model can be built without a large number of samples, significantly reducing the modeling and computation costs, and balancing the efficiency of large-scale measurement and estimation accuracy.

[0143] Figure 9 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 1101, the processor 1102, and the computer program stored on the memory 1101 and executable on the processor 1102.

[0144] When the processor 1102 executes the program, it implements the forest aboveground biomass estimation method provided in the above embodiments.

[0145] Furthermore, electronic devices also include: Communication interface 1103 is used for communication between memory 1101 and processor 1102.

[0146] The memory 1101 is used to store computer programs that can run on the processor 1102.

[0147] The memory 1101 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage.

[0148] If the memory 1101, processor 1102, and communication interface 1103 are implemented independently, then the communication interface 1103, memory 1101, and processor 1102 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0149] Optionally, in a specific implementation, if the memory 1101, processor 1102, and communication interface 1103 are integrated on a single chip, then the memory 1101, processor 1102, and communication interface 1103 can communicate with each other through an internal interface.

[0150] The processor 1102 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0151] This application also provides a computer-readable storage medium storing a computer program or instructions thereon, which, when executed by a processor, implements the above-described method for estimating forest aboveground biomass.

[0152] This application also provides a computer program product, including a computer program or instructions, which, when executed, implement the above-described method for estimating forest aboveground biomass.

[0153] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0154] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0155] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0156] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0157] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. A method for estimating forest aboveground biomass, characterized in that, Includes the following steps: Multiple forest aboveground biomass samples are obtained for the target area, wherein the forest aboveground biomass samples include a first sample and a second sample, the first sample is used for model training, and the second sample is used for spatial heterogeneity partitioning. The first sample is used to train a global forest aboveground biomass estimation model at the target area scale, and the prediction residual of the second sample is determined based on the trained global forest aboveground biomass estimation model. Based on the prediction residuals, a spatial heterogeneity score is determined for each second sample, and the target region is spatially partitioned into multiple target sub-regions based on the spatial heterogeneity scores of the second samples. The spatial partitioning of the target region into multiple target sub-regions based on the spatial heterogeneity scores of the second samples includes: dividing the target region into regular grids and matching the second samples to the corresponding regular grids to obtain multiple spatial units; calculating the spatial heterogeneity score of the corresponding spatial unit based on the prediction residuals of all second samples within the spatial unit; and using the target region as the region to be partitioned, partitioning the multiple spatial units within the region to be partitioned. The spatial heterogeneity scores of the units are sorted, and the spatial units are sequentially extracted based on the sorting results to divide the region to be partitioned into two candidate sub-regions. The two candidate sub-regions are used as binary partitioning results to be tested, and target tests are performed on these results, including accuracy improvement tests and significance tests. If the target tests pass, the binary partitioning results are retained, and the two candidate sub-regions are used as new regions to be partitioned. The binary partitioning and target tests are recursively performed until a preset stopping condition is met, resulting in multiple target sub-regions. The formula for calculating the spatial heterogeneity score of the candidate sub-regions is: ； in, The spatial heterogeneity score for candidate sub-region S is given. This represents the additional mean offset of the overall prediction residuals within the candidate sub-region relative to the baseline distribution. The sum of the inverse variance-weighted residuals of all second samples within the candidate sub-region S. , For the second sample In spatial location The predicted residual at the location, Spatial location The standard deviation of the baseline residual distribution. The sum of the inverse variance weights of all second samples within the candidate sub-region S. ; Extract the first sample of each target sub-region, and train the corresponding local forest aboveground biomass estimation model based on the first sample of each target sub-region to obtain local forest aboveground biomass estimation models corresponding to multiple target sub-regions; The pixels to be predicted within the target area are divided into corresponding target sub-regions, and the forest aboveground biomass of the pixels to be predicted is estimated by calling the local forest aboveground biomass estimation model corresponding to the target sub-region.

2. The method for estimating forest aboveground biomass according to claim 1, characterized in that, The target test on the binary partitioning result to be tested includes: Obtain the estimated value of the first forest aboveground biomass before the partitioning of the area to be partitioned, and the estimated value of the second forest aboveground biomass corresponding to the candidate sub-regions after the partitioning of the area to be partitioned. Based on the second forest aboveground biomass estimate, the first forest aboveground biomass estimate, and the corresponding forest aboveground biomass reference value, the change in prediction accuracy of the second sample of the candidate sub-region before and after partitioning is determined; If the change in the prediction accuracy does not meet the preset improvement condition, the binary partitioning result will not be retained, and the partitioning of the region to be partitioned will be stopped. If the change in the prediction accuracy meets the preset improvement condition, a significance test is performed on the change in the prediction accuracy. If the significance test passes, the binary partitioning result is retained; otherwise, the binary partitioning result is not retained, and the partitioning of the region to be partitioned is stopped.

3. The method for estimating forest aboveground biomass according to claim 1, characterized in that, The preset stop condition includes at least one of the following: The number of the first sample in the region to be partitioned is less than a preset sample number threshold; The number of spatial units in the area to be partitioned is less than a preset threshold for the number of spatial units. The change in prediction accuracy corresponding to the binary partitioning result does not meet the preset improvement condition. The binary partitioning results failed the significance test; The current recursive partitioning depth has reached the preset maximum partitioning depth.

4. The method for estimating forest aboveground biomass according to claim 1, characterized in that, Each forest aboveground biomass sample includes characteristic variables and forest aboveground biomass reference values. The acquisition of multiple forest aboveground biomass samples in the target area includes: Acquire multi-source remote sensing data of the target area; Feature variables of the corresponding forest ground biomass samples are extracted from the multi-source remote sensing data, and reference values ​​of forest ground biomass for multiple forest ground biomass samples in the target area are obtained.

5. The method for estimating forest aboveground biomass according to claim 4, characterized in that, The feature variables include a first feature variable and a second feature variable. The first feature variable includes one or more of the following: multispectral band features, multispectral band computational features, texture features, radar dual polarization features, and canopy height features. The second feature variable is a diameter at breast height (DBH) prediction feature variable. Feature variables corresponding to forest aboveground biomass samples are extracted from the multi-source remote sensing data, including: The multi-source remote sensing data is resampled to the target resolution; The resampled multi-source remote sensing data is spatially matched with multiple forest ground biomass samples to determine the sampling point corresponding to the forest ground biomass sample as the central pixel in the multi-source remote sensing data. Feature extraction is performed on the central pixel and the pixels within the target range of the central pixel to obtain the first feature variable; The second characteristic variable is determined based on the canopy height characteristic variable and the pre-built diameter at breast height prediction model.

6. The method for estimating forest aboveground biomass according to claim 5, characterized in that, The determination of the second characteristic variable based on the canopy height characteristic variable and the pre-built chest diameter prediction model includes: Obtain a third sample, wherein the third sample includes observations of canopy height and diameter at breast height; Based on the third sample, a fitting relationship between the coronal height and the thoracic diameter is constructed to obtain the thoracic diameter prediction model; The crown height feature variable is input into the diameter at breast height prediction model, and the diameter at breast height prediction model outputs the second feature variable.

7. A device for estimating forest aboveground biomass, characterized in that, The apparatus is used to implement the forest aboveground biomass estimation method as described in any one of claims 1-6, wherein the apparatus comprises: The acquisition module is used to acquire multiple forest aboveground biomass samples of the target area, wherein the forest aboveground biomass samples include a first sample and a second sample, the first sample is used for model training, and the second sample is used for spatial heterogeneity partitioning. The first training module is used to train a global forest aboveground biomass estimation model at the target area scale using the first sample, and to determine the prediction residual of the second sample based on the trained global forest aboveground biomass estimation model. The partitioning module is used to determine the spatial heterogeneity score of each second sample based on the prediction residual, and to spatially partition the target region based on the spatial heterogeneity score of the second sample to obtain multiple target sub-regions; The second training module is used to extract the first sample of each target sub-region and train the corresponding local forest aboveground biomass estimation model based on the first sample of each target sub-region, so as to obtain local forest aboveground biomass estimation models corresponding to multiple target sub-regions. The estimation module is used to divide the pixels to be predicted within the target area into corresponding target sub-regions, and call the local forest aboveground biomass estimation model corresponding to the target sub-region to estimate the forest aboveground biomass of the pixels to be predicted.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the forest aboveground biomass estimation method as described in any one of claims 1-6.

Images