Haloxylon ammodendron forest biomass estimation method based on unmanned aerial vehicle multi-source data collaborative inversion

Abstract

The invention discloses a haloxylon ammodendron forest biomass estimation method based on unmanned aerial vehicle multi-source data collaborative inversion, and relates to the technical field of environment remote sensing and information, and the method comprises the steps: obtaining a training sample set; obtaining unmanned aerial vehicle multispectral image data and unmanned aerial vehicle laser radar point cloud data of each training sample plot; extracting a first type of feature set comprising vegetation index features and texture features; extracting a second type of feature set comprising a height feature and a density feature; fusing the two types of feature sets to form a joint feature set; screening the joint feature set to construct a target feature subset; and carrying out model training by utilizing a machine learning algorithm to obtain an above-ground biomass estimation model. According to the method, the technical problem that the existing single remote sensing data source is low in precision in estimating the biomass on the desert sparse haloxylon ammodendron forest land is solved, high-precision and automatic estimation from the sample land to the regional scale is realized, and technical support is provided for accurate accounting of the desert carbon reserve in the arid region.

Description

Technical Field

[0001] This application relates to the fields of environmental remote sensing and information technology, and in particular to a method for estimating the biomass of Haloxylon ammodendron forests based on collaborative inversion of multi-source data from unmanned aerial vehicles. Background Technology

[0002] Desert shrublands are an important component of desert ecosystems, and accurate estimation of their aboveground biomass (AGB) is crucial for regional carbon sequestration and desertification monitoring. Currently, remote sensing technology has become the primary method for large-scale vegetation biomass estimation. Among these, unmanned aerial vehicle (UAV) remote sensing technology, with its advantages of high resolution and high flexibility, provides a new data source for desert areas where traditional ground surveys are difficult to conduct.

[0003] Existing research has attempted to apply UAV multispectral, visible light, or lidar data to individual shrub identification and biomass estimation, achieving some progress. However, current research generally focuses on the identification and biomass inversion of individual shrubs. While the potential of multi-source data has been validated at the individual shrub scale, a technical workflow for operational inversion of desert shrubland biomass at the regional scale is lacking.

[0004] When applied to desert sparse shrub forests such as Haloxylon ammodendron forests, the spectral signals of desert shrubs are weak and easily interfered with by complex soil backgrounds because the plants are short, have small crowns, and are sparsely distributed. When the single-plant scale model and method are extended to a region, they often face a series of complex technical problems such as plot representativeness, feature generalization ability, computational efficiency, and adaptation to spatial heterogeneity.

[0005] Therefore, there is a lack of effective biomass estimation methods in the existing technology for regional scales, such as desert sparse shrub forests like Haloxylon ammodendron. Summary of the Invention

[0006] The purpose of this application is to provide a method for estimating the biomass of Haloxylon ammodendron forests based on collaborative inversion of multi-source data from unmanned aerial vehicles. This method constructs complementary feature subsets by screening feature importance and uses machine learning models for training and estimation. It solves the technical problems of low accuracy and poor stability of existing single remote sensing data sources when estimating the biomass of sparse Haloxylon ammodendron forests in deserts. It achieves high-precision and automated estimation from plot to regional scale, providing technical support for accurate carbon storage accounting in arid desert areas.

[0007] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for estimating the biomass of Haloxylon ammodendron forests based on collaborative inversion of multi-source data from unmanned aerial vehicles (UAVs). The method is executed by a data processing device and includes: acquiring a training sample set, which includes ground biomass values ​​from multiple training plots; acquiring UAV multispectral image data and UAV lidar point cloud data for each training plot; extracting a first type of feature set from the UAV multispectral image data, the first type of feature set including vegetation index features calculated based on spectral information; extracting a second type of feature set from the UAV lidar point cloud data, the second type of feature set including height features statistically obtained based on point cloud height information; fusing the first type of feature set and the second type of feature set to form a joint feature set; and ranking features based on their importance using the training sample set. The method first filters the joint feature set to construct a target feature subset, wherein the target feature subset simultaneously contains at least one feature selected from the first type of feature set and at least one feature selected from the second type of feature set; using a machine learning algorithm, a model is trained based on the training sample set and the corresponding target feature subset to obtain a ground biomass estimation model; UAV multispectral image data and UAV lidar point cloud data of the area to be estimated are acquired; based on the target feature subset, the corresponding feature set is calculated from the UAV multispectral image data and UAV lidar point cloud data of the area to be estimated; the calculated corresponding feature set is input into the ground biomass estimation model to output the ground biomass estimation result of the Haloxylon ammodendron forest in the area to be estimated.

[0008] Optionally, the construction of the training sample set includes: obtaining the measured values ​​of aboveground biomass of predetermined measured sample plots as a first sample subset; for predetermined expanded sample plots, based on UAV lidar point cloud data covering the area, obtaining the height and crown area of ​​a single Haloxylon ammodendron tree through single-tree segmentation, and calculating the estimated value of aboveground biomass using a preset Haloxylon ammodendron allometric growth equation as a second sample subset; merging the first sample subset and the second sample subset to form the training sample set.

[0009] Optionally, the machine learning algorithm is an extreme gradient boosting algorithm.

[0010] Optionally, the first type of feature set also includes texture features extracted from the UAV multispectral image data based on the gray-level co-occurrence matrix.

[0011] Optionally, the vegetation index features include at least one of the following indices: normalized vegetation index, ratio vegetation index, and normalized greenness vegetation index.

[0012] Optionally, the second type of feature set also includes density features extracted from the UAV lidar point cloud data based on the point cloud density distribution.

[0013] Optionally, the height features include maximum height and percentile height.

[0014] Optionally, the feature importance ranking method is the random forest feature importance ranking method, which quantifies feature importance by calculating the percentage increase in mean squared error for each feature.

[0015] Optionally, the step of filtering the joint feature set to construct a target feature subset includes: selecting a predetermined number of features with the highest importance ranking based on the ranking result generated by the feature importance ranking method.

[0016] Secondly, this application provides a biomass estimation device for Haloxylon ammodendron forests based on collaborative inversion of multi-source data from unmanned aerial vehicles (UAVs), comprising: a sample set acquisition module configured to acquire a training sample set, the training sample set including ground biomass true values ​​of multiple training plots; a training data acquisition module configured to acquire UAV multispectral image data and UAV lidar point cloud data for each training plot; a first type of feature extraction module configured to extract a first type of feature set from the UAV multispectral image data, the first type of feature set including vegetation index features calculated based on spectral information; a second type of feature extraction module configured to extract a second type of feature set from the UAV lidar point cloud data, the second type of feature set including height features statistically obtained based on point cloud height information; a feature fusion module configured to fuse the first type of feature set and the second type of feature set to form a joint feature set; and a feature filtering module configured to select features based on the training sample set. A feature importance ranking method is used to filter the joint feature set to construct a target feature subset, wherein the target feature subset simultaneously contains at least one feature selected from the first type of feature set and at least one feature selected from the second type of feature set; a model training module is configured to perform model training using a machine learning algorithm based on the training sample set and the corresponding target feature subset to obtain a ground biomass estimation model; a data acquisition module for the area to be estimated is configured to acquire UAV multispectral image data and UAV lidar point cloud data of the area to be estimated; a feature calculation module for the area to be estimated is configured to calculate the corresponding feature set based on the target feature subset from the UAV multispectral image data and UAV lidar point cloud data of the area to be estimated; and a biomass estimation module is configured to input the calculated corresponding feature set into the ground biomass estimation model and output the ground biomass estimation result of the Haloxylon ammodendron forest in the area to be estimated.

[0017] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method for estimating the biomass of Haloxylon ammodendron forests based on collaborative inversion of multi-source data from UAVs. By acquiring and fusing vegetation index features from UAV multispectral imagery with height features from UAV lidar point clouds, a joint feature set is constructed. This overcomes the limitations of single optical data being susceptible to background interference or single structural data being insensitive to vegetation physiological states. It can comprehensively characterize the growth status of Haloxylon ammodendron forests from two complementary dimensions: the physiological state of the vegetation and its three-dimensional morphological structure. Furthermore, based on feature fusion, a feature importance ranking method is used to filter the joint feature set using a training sample set to construct target features. The subset effectively eliminates redundant features, reduces data dimensionality, and ensures that the model learns the most representative and complementary core feature combination of the two types of information, thereby improving its estimation accuracy in unknown areas. By training the aboveground biomass estimation model and determining the target feature subset, it can be quickly applied to any new area to be estimated, realizing the processing from raw data to the final biomass estimation result. Therefore, the UAV multi-source data collaborative inversion process constructed in this application realizes high-precision, non-destructive estimation of desert sparse shrub biomass from sample plots to the entire region, providing a reliable technical solution for accurate accounting of desert carbon storage. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an application environment diagram of a method for estimating the biomass of Haloxylon ammodendron forest based on collaborative inversion of multi-source data from unmanned aerial vehicles, according to one embodiment of this application. Figure 2 A schematic diagram of the model training phase provided in an embodiment of this application; Figure 3 This is a schematic diagram of the spatial distribution of four sample areas in one embodiment of this application; Figure 4 This is a schematic diagram showing the ranking of feature importance among three feature sets in one embodiment of this application; Figure 5 This is a statistical diagram illustrating the accuracy of different models under different data sources in one embodiment of this application. Figure 6 This is a scatter plot of AGB inversion values ​​and measured values ​​under different data sources and modeling methods in one embodiment of this application. Figure 7 This is a flowchart illustrating the construction method of the training sample set in one embodiment of this application; Figure 8This is a schematic diagram of linear regression analysis in one embodiment of this application; Figure 9 A flowchart illustrating the biomass estimation stage provided in an embodiment of this application; Figure 10 This is a high-resolution spatial distribution map of aboveground biomass in one embodiment of this application; Figure 11 A schematic diagram of the functional modules of a Haloxylon ammodendron forest biomass estimation device based on UAV multi-source data collaborative inversion is provided in an embodiment of this application; Figure 12 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] The method for estimating Haloxylon ammodendron forest biomass based on collaborative inversion of multi-source UAV data provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 11 communicates with server 12 via a network. A data storage system can store the data that server 12 needs to process. The data storage system can be set up independently, integrated into server 12, or placed in the cloud or on other servers. Terminal 11 can send UAV multispectral image data and lidar point cloud data to server 12. After receiving the data, server 12 executes the Haloxylon ammodendron biomass estimation method based on UAV multi-source data collaborative inversion proposed in this application. Server 12 can feed back the obtained Haloxylon ammodendron biomass estimation results or a visualized distribution map to terminal 11. Furthermore, in some embodiments, the Haloxylon ammodendron biomass estimation method based on UAV multi-source data collaborative inversion proposed in this application can also be implemented independently by server 12 or terminal 11.

[0023] The terminal 11 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, drones, drone ground stations, etc. The server 12 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0024] In one exemplary embodiment, a method for estimating the biomass of Haloxylon ammodendron forests based on collaborative inversion of multi-source data from unmanned aerial vehicles (UAVs) is provided. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is applied to… Figure 1 Taking server 12 as an example, this method comprises two stages: a model training stage and a biomass estimation stage. The model training stage is completed offline to generate a high-precision, dedicated estimation model; the biomass estimation stage can be applied online or offline, using the trained model to quickly estimate the biomass of a new region. These will be explained in detail below.

[0025] like Figure 2 As shown, the specific steps in the model training phase include: S101, Obtain the training sample set.

[0026] In practice, the training sample set includes the true aboveground biomass values ​​of multiple training plots, which include both measured and estimated aboveground biomass values.

[0027] In practice, a total of 91 training plots, each 30m × 30m in size, were established. These 91 plots were located in four sample areas within the study area (e.g., Figure 3 On the multispectral images (UAV-MSI) acquired by UAVs equipped with multispectral imagers in sample areas 1, 2, 3, and 4, some points were randomly placed to designate a total of 83 30m×30m rectangular areas as expanded sample plots, as well as 8 field measured sample plots. Specifically, 22, 20, 22, and 19 sample plot center points were set in the four sample areas, respectively, and a 30m×30m range was delineated with these points as the center. The size of these expanded sample plots (83) was consistent with that of the field measured sample plots (8), all being 30m×30m, to ensure data scale consistency.

[0028] S102, acquire UAV multispectral image data and UAV lidar point cloud data for each training sample site.

[0029] In practice, based on the location information of the training plots (such as corner coordinates), UAV multispectral imagery and UAV lidar point cloud data that perfectly match the 30m×30m training plot area are retrieved. The multispectral imagery data has been radiometrically calibrated to generate surface reflectance and has been stitched and registered (spatial resolution 7cm); the lidar point cloud data has undergone preprocessing such as denoising, ground point classification, DEM generation, and normalization.

[0030] S103, extract the first type of feature set from UAV multispectral image data.

[0031] In specific implementation, the first type of feature set includes vegetation index features calculated based on spectral information and texture features extracted from UAV multispectral image data based on gray-level co-occurrence matrix (GLCM). Specifically, for vegetation index features calculated based on spectral information, in this embodiment, based on the four bands of blue, green, red, and near-infrared, nine vegetation indices, including Normalized Difference Vegetation Index (NDVI), Ratio Vegetation Index (RVI), and Normalized Greenness Vegetation Index (GNDVI), are calculated for each sample plot, and the pixel mean within the sample plot is used as the feature value of that sample plot. Specifically, for texture features extracted from UAV multispectral image data based on gray-level co-occurrence matrix (GLCM), in this embodiment, GLCM calculation uses a 3×3 window, and eight texture features, including mean, variance, and contrast, are extracted from the four bands of blue, green, red, and near-infrared, respectively, resulting in a total of 32 texture indices, and the pixel mean of each index within the sample plot is calculated.

[0032] S104, extract the second type of feature set from the UAV lidar point cloud data.

[0033] In specific implementation, the second type of feature set includes height features obtained from point cloud height information statistics and density features extracted from UAV LiDAR point cloud data based on point cloud density distribution. Specifically, for the height features obtained from point cloud height information statistics, in this embodiment, the height features include maximum height (Hmax), average height (Hmean), and 99th percentile height (H99). Specifically, for the density features extracted from UAV LiDAR point cloud data based on point cloud density distribution, in this embodiment, the point cloud height is divided into 10 layers (d0-d9) from low to high, and the proportion of echoes in each layer is calculated as the density feature.

[0034] S105, merge the first type of feature set with the second type of feature set to form a joint feature set.

[0035] In practice, the first type of feature set and the second type of feature set are merged to form a record for each sample plot that includes all vegetation indices, textures, heights and densities, i.e., a joint feature set.

[0036] S106. Based on the training sample set, the joint feature set is filtered using a feature importance ranking method to construct a target feature subset.

[0037] In practice, the feature importance ranking method is the random forest feature importance ranking method. Feature importance is quantified by adding a percentage mean squared error (%IncMSE) metric; the higher the value, the greater the contribution of the corresponding feature. Feature importance analysis is implemented using the randomForest package in RStudio software, and the average importance score of each feature is calculated by iterating 100 times.

[0038] Furthermore, this embodiment analyzes the above three feature sets, and the analysis results are shown in the figure below. Figure 4 As shown, based on Figure 4 The results are as follows: When only the first type of feature set (MSI) is used, the most important feature is the normalized vegetation index (NDVI), which increases the mean squared error percentage (%IncMSE) by about 6%, indicating that spectral information makes a significant contribution to AGB inversion.

[0039] When only the second type of feature set (LiDAR) is used, the most important feature is the maximum height (Hmax), which increases the percentage mean squared error (%IncMSE) by about 12%, which is significantly higher than other structural features, indicating that the point cloud height feature has a stronger interpretive power in AGB inversion.

[0040] When using the fused joint feature set (MSI+LiDAR), the top 15 most important features selected include vegetation indices such as Normalized Difference Vegetation Index (NDVI), Ratio Vegetation Index (RVI), and Normalized Greenness Index (GNDVI) from the first feature set, as well as height features such as maximum height (Hmax) and 99th percentile height (H99) from the second feature set. Crucially, the importance scores of each feature are more evenly distributed in the joint feature set, with the percentage increase in mean squared error (%IncMSE) concentrated between 4% and 6%, demonstrating good complementarity. This indicates that fusing the two types of features can more comprehensively characterize the physiological and morphological information of Haloxylon ammodendron.

[0041] S107 utilizes machine learning algorithms to train a model based on a training sample set and the corresponding target feature subset, thereby obtaining a ground biomass estimation model.

[0042] In practice, instead of using a single algorithm, three different machine learning algorithms are initialized and trained for comparison in order to find the optimal solution, including: Random Forest (RF): An algorithm based on decision tree ensembles that is highly robust and resistant to overfitting; Support Vector Machine (SVM): An algorithm that uses the construction of an optimal hyperplane to handle small-sample, nonlinear regression problems; Extreme Gradient Boosting (XGBoost): A high-performance boosting tree model with significant advantages in feature interaction modeling and generalization capabilities.

[0043] In practice, the three machine learning algorithms mentioned above were implemented using the randomForest, e1071, and xgboost packages within the RStudio software environment. For each algorithm, the processor employed a grid search method to systematically tune its key parameters, thereby optimizing the potential performance and generalization ability of each model.

[0044] To evaluate model performance and mitigate overfitting issues when the sample size is small or the model complexity is high, this embodiment employs Leave-one-out cross-validation (LOOCV) for model training and validation. The specific process is as follows: First, for a training sample set containing 91 samples, one sample is selected as the test set each time, and the remaining 90 samples are used as the training set. This ensures that each sample is used to validate the model once, thereby obtaining more accurate and reliable evaluation results.

[0045] Then, in each round, the model is fitted using the training set and a prediction is made on the reserved test sample. This process is repeated 91 times, ensuring that each sample is used as the test set once.

[0046] Finally, the predicted values ​​of all 91 samples were collected and compared with the actual measured values.

[0047] In practical implementation, the validation results obtained from the above model training are used to determine the coefficient of determination (R²). 2 The performance of the three algorithms on the same target feature subset is systematically compared using two core metrics: R0 and root mean square error (RMSE). 2 The closer the value is to 1, the better the model fit and the stronger the explanatory power of the selected features for AGB; the smaller the RMSE, the closer the retrieved AGB is to the measured value.

[0048] The formula for calculating the coefficient of determination is: (Equation 1) The formula for calculating the root mean square error is: (Equation 2) In equations (1) and (2), Indicates the number of samples in the test set. Indicates the first The measured value of AGB for each sample. This represents the arithmetic mean of the measured AGB. Indicates the first The inversion value of AGB for each sample.

[0049] Model evaluation results are as follows Figure 5 and Figure 6 As shown, Figure 5 Accuracy statistics for different models built on different data sources. Figure 6 Scatter plots of AGB inversion values ​​and measured values ​​under different data sources and modeling conditions.

[0050] Depend on Figure 5 and Figure 6 The analysis results show that: For the performance ranking of a single algorithm: the extreme gradient boosting model performs best, with a coefficient of determination R. 2 The accuracy reached 0.89, and the root mean square error (RMSE) was 0.53 t / hm. 2 Support Vector Machine (SVM) models are the next best, followed by R... 2 The value was 0.88, and the RMSE was 0.54 t / hm. 2 R-squared of the random forest model 2 The value is 0.85, and the RMSE is 0.62 t / hm. 2 .

[0051] Advantages compared to single-data-source models: Compared to the XGBoost model that only uses MSI features (its R... 2 =0.84, RMSE=0.63 t / hm 2 R of the joint feature model 2 The efficiency improved by 0.05, while the RMSE decreased by 0.10 t / hm. 2 In contrast, the XGBoost model, which uses only LiDAR features (its R... 2 =0.71, RMSE=0.84t / hm 2 The advantages of the joint feature model are more significant.

[0052] Model bias analysis: such as Figure 6 As shown in the scatter plot, the XGBoost model based on the target feature subset has a more concentrated distribution of predicted and measured values ​​along the 1:1 line, which effectively reduces the systematic bias of underestimating high values ​​and overestimating low values, indicating that the model captures the range of biomass changes more accurately.

[0053] Based on the analysis of the above model evaluation results, the extreme gradient boosting model trained on the target feature subset performs best in terms of accuracy, stability and bias control. This extreme gradient boosting model (XGBoost) is determined as the final "aboveground biomass estimation model", and it and the corresponding target feature subset definition are saved together for use in subsequent biomass estimation stages.

[0054] By implementing steps S101 to S107, specifically in step S101, a training set with ecological representativeness was successfully constructed by combining a small number of field-measured plots with a large number of plots estimated based on UAV lidar point clouds through single-tree segmentation and allometric growth equations. This overcame the difficulties of conducting large-scale ground measurements in harsh desert environments, effectively expanded the modeling sample size, and laid a data foundation for constructing a high-precision regional model. Regarding steps S102 to S106, by extracting and fusing the first type of feature set reflecting the physiological state of vegetation with the second type of feature set reflecting the three-dimensional morphological structure, the deficiency of incomplete information from a single data source was overcome. Furthermore, a random forest feature importance ranking method was used for quantitative screening, revealing that spectral features and structural features have a clear complementarity in explaining the biomass of Haloxylon ammodendron forests. For step S107, by comparing the three algorithms of random forest, support vector machine and extreme gradient boosting, and using leave-one-out cross-validation for evaluation, the extreme gradient boosting algorithm was determined to be the optimal algorithm in this application scenario. This fully verified the effectiveness of multi-source data collaboration and feature selection. Moreover, the model effectively reduced the systematic bias of underestimating high values ​​and overestimating low values, and showed excellent generalization ability and stability.

[0055] For example, such as Figure 7 As shown, the construction method of the training sample set in step S101 above includes: S1011, Obtain the measured values ​​of aboveground biomass of the predetermined measured plots as the first sample subset; specifically, the predetermined measured plots can be the above-mentioned 8 field measured plots.

[0056] S1012, for the pre-determined expanded plots, based on the UAV lidar point cloud data covering the area, the height and crown area of ​​a single Haloxylon ammodendron are obtained by single-tree segmentation, and the estimated aboveground biomass is calculated using the pre-set Haloxylon ammodendron allometric growth equation, which serves as the second sample subset; specifically, the pre-determined expanded plots are the aforementioned 83 expanded plots.

[0057] S1013, merge the first sample subset and the second sample subset to form the training sample set.

[0058] In practical implementation, the calculation of the true aboveground biomass value (the measured aboveground biomass value in step S1011 and the estimated aboveground biomass value in step S1012) for the training sample set is as follows: The plant height (H) and crown area (C) of individual Haloxylon ammodendron plants in both the measured and expanded plots were substituted into the same Haloxylon ammodendron allometric growth equation to calculate the measured and estimated aboveground biomass values, respectively. The Haloxylon ammodendron allometric growth equation is shown in equation (3) below: (3) In formula (3), the plant height H is in meters (m); the crown area C is in square meters (m²). 2 The formula for calculating the crown area C is shown in equation (4) below: (4) In equation (4), L is the east-west crown width and W is the north-south crown width, both in meters.

[0059] In the specific implementation of step S1012, the expansion plot is obtained by dividing individual trees to obtain the height and crown area of ​​a single Haloxylon ammodendron. Since Haloxylon ammodendron is mostly distributed in shrubs and the crowns overlap significantly, the division of individual trees is prone to misclassification due to blurred boundaries. In order to reduce the division error, each 30m×30m expansion plot is divided into 9 10m×10m sub-units for processing. Thus, by reducing the processing unit, the accuracy of individual tree division can be improved.

[0060] In the specific implementation of step S1013, before merging the first sample subset and the second sample subset, it is necessary to ensure that the estimation method for calculating the AGB of the expanded sample area based on UAV lidar point cloud data is reliable, and the accuracy of the above-mentioned AGB estimation method needs to be verified.

[0061] The accuracy verification process is as follows: the 8 measured plots are also divided into sub-units, resulting in 72 sub-units. The measured values ​​of the 72 sub-units can be calculated using the measured plant height and crown area of ​​Haloxylon ammodendron and formula (1); then the estimated values ​​of the 72 sub-units generated in step S1012 are obtained from the 8 measured plots; and the 72 pairs of measured values ​​and estimated values ​​are subjected to linear regression analysis, such as... Figure 8 As shown, the results indicate that the scatter points are closely distributed on both sides of the 1:1 line (the solid black line in the figure), and the coefficient of determination (R²) of the fitted line (the dashed red line in the figure) is similar. 2 The value was 0.94, and the root mean square error (RMSE) was 0.49 t / hm. 2 This proves that the above point cloud-based estimation method has extremely high accuracy and reliable results.

[0062] like Figure 9 As shown, the specific steps in the biomass estimation stage include: S201, acquire UAV multispectral image data and UAV lidar point cloud data of the area to be estimated.

[0063] In practice, the acquired data needs to have undergone the same preprocessing steps as in the model training phase. For example, for UAV multispectral image data: radiometric calibration, band registration, and stitching have been performed to generate surface reflectance images; for UAV lidar point cloud data: point cloud denoising, ground point classification, DEM generation, and normalization have been completed.

[0064] S202, based on the target feature subset, calculates the corresponding feature set from the UAV multispectral image data and UAV lidar point cloud data of the area to be estimated.

[0065] In practice, the complete definition of the "target feature subset" determined and saved during the model training phase is read and loaded. This definition clearly specifies which specific features need to be calculated (e.g., including 15 features such as NDVI, RVI, Hmax, and H99) and their precise calculation methods. Subsequently, strictly following this definition, the exact same features are recalculated from the data of the region to be estimated obtained in step S201. For example, all vegetation indices and texture features (NDVI, RVI, etc.) included in the target feature subset are calculated from multispectral image data, and all height and density features (Hmax, H99, etc.) included in the target feature subset are calculated from lidar point cloud data. Finally, all the calculated feature values ​​are organized in the same order and format as in the model training phase to generate a feature set corresponding to the region.

[0066] S203, input the calculated feature set into the aboveground biomass estimation model, and output the aboveground biomass estimation results of the Haloxylon ammodendron forest in the area to be estimated.

[0067] In practice, the feature set calculated in step S202 is used as input and loaded into the optimal aboveground biomass estimation model (i.e., the XGBoost model trained based on the target feature subset) that was trained and saved during the model training phase. The model automatically runs forward propagation calculation and then obtains its output prediction value, which is the aboveground biomass estimate for the area to be estimated.

[0068] Furthermore, a high-resolution spatial distribution map of aboveground biomass in the area to be estimated can be generated based on the aboveground biomass estimate output in step S203.

[0069] As an application example of this method, the above-described biomass estimation process was applied to the four study areas involved in the training phase, yielding the following results: Figure 10 The image shows a high-resolution spatial distribution map of aboveground biomass. From... Figure 10 It can be seen that there is significant spatial differentiation in AGB levels among the sample areas: sample area 1 (oasis edge) has the highest average AGB (2.50 t / hm). 2 The average AGB of sample areas 2, 3, and 4 decreased with increasing oasis distance (0.90, 0.84, and 0.64 t / hm, respectively). 2The results show that the favorable hydrothermal conditions at the edge of the oasis are conducive to the growth of *Haloxylon ammodendron*, while in areas far from the oasis, the scarcity of environmental resources (such as water) restricts the growth and biomass accumulation of *Haloxylon ammodendron*. This spatial differentiation pattern reveals the heterogeneity of desert ecosystem resource distribution and its impact on vegetation growth. Therefore, this embodiment successfully achieved the scale transformation from discrete plot models to continuous regional mapping, demonstrating the application potential of the biomass estimation method at the regional scale and providing a solution for the accurate accounting of desert carbon storage.

[0070] In one exemplary embodiment, such as Figure 11 As shown, a biomass estimation device for Haloxylon ammodendron forests based on collaborative inversion of multi-source data from unmanned aerial vehicles is provided, comprising: The sample set acquisition module 301 is configured to acquire a training sample set, which includes ground biomass values ​​of multiple training plots.

[0071] The training data acquisition module 302 is configured to acquire UAV multispectral image data and UAV lidar point cloud data for each of the training sample sites.

[0072] The first feature extraction module 303 is configured to extract a first type of feature set from the UAV multispectral image data, the first type of feature set including vegetation index features calculated based on spectral information.

[0073] The second type of feature extraction module 304 is configured to extract a second type of feature set from the UAV lidar point cloud data. The second type of feature set includes height features obtained statistically based on point cloud height information.

[0074] The feature fusion module 305 is configured to perform the fusion of the first type of feature set and the second type of feature set to form a joint feature set.

[0075] The feature filtering module 306 is configured to perform filtering on the joint feature set based on the training sample set using a feature importance ranking method to construct a target feature subset, wherein the target feature subset simultaneously contains at least one feature filtered from the first type of feature set and at least one feature filtered from the second type of feature set.

[0076] The model training module 307 is configured to perform model training based on the training sample set and the corresponding target feature subset using a machine learning algorithm to obtain a ground biomass estimation model.

[0077] The data acquisition module 308 for the area to be estimated is configured to acquire UAV multispectral image data and UAV lidar point cloud data for the area to be estimated.

[0078] The feature calculation module 309 for the region to be estimated is configured to calculate the corresponding feature set based on the target feature subset from the UAV multispectral image data and UAV lidar point cloud data of the region to be estimated.

[0079] The biomass estimation module 310 is configured to input the calculated corresponding feature set into the aboveground biomass estimation model and output the aboveground biomass estimation result of the Haloxylon ammodendron forest in the area to be estimated.

[0080] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 11 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements the aforementioned method for estimating Haloxylon ammodendron biomass based on collaborative inversion of multi-source data from unmanned aerial vehicles (UAVs).

[0081] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0082] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0083] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0084] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0085] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Furthermore, any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory.

[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for estimating the biomass of Haloxylon ammodendron forests based on collaborative inversion of multi-source data from unmanned aerial vehicles (UAVs), wherein the method is executed by a data processing device, characterized in that, include: Obtain a training sample set, which includes ground biomass values ​​from multiple training plots; Acquire UAV multispectral image data and UAV lidar point cloud data for each of the training sample sites; A first type of feature set is extracted from the UAV multispectral image data, the first type of feature set including vegetation index features calculated based on spectral information; A second type of feature set is extracted from the UAV lidar point cloud data, the second type of feature set including height features obtained statistically based on point cloud height information; The first type of feature set is merged with the second type of feature set to form a joint feature set; Based on the training sample set, the joint feature set is filtered using a feature importance ranking method to construct a target feature subset, wherein the target feature subset simultaneously contains at least one feature selected from the first type of feature set and at least one feature selected from the second type of feature set. A ground biomass estimation model is obtained by using machine learning algorithms to train a model based on the training sample set and the corresponding target feature subset. Acquire UAV multispectral imagery data and UAV lidar point cloud data for the area to be estimated; Based on the target feature subset, the corresponding feature set is calculated from the UAV multispectral image data and UAV lidar point cloud data of the area to be estimated; The calculated feature set is input into the aboveground biomass estimation model, and the aboveground biomass estimation result of the Haloxylon ammodendron forest in the area to be estimated is output.

2. The method for estimating Haloxylon ammodendron forest biomass based on collaborative inversion of multi-source UAV data as described in claim 1, characterized in that, The construction of the training sample set includes: Obtain the measured values ​​of aboveground biomass from predetermined sample plots as the first sample subset; For the pre-determined expanded sample plots, based on the UAV lidar point cloud data covering the area, the height and crown area of ​​a single Haloxylon ammodendron are obtained by segmenting individual trees, and the estimated aboveground biomass is calculated using the pre-set Haloxylon ammodendron allometric growth equation, which serves as the second sample subset. The first sample subset and the second sample subset are merged to form the training sample set.

3. The method for estimating Haloxylon ammodendron forest biomass based on collaborative inversion of multi-source UAV data as described in claim 1, characterized in that, The machine learning algorithm mentioned is the extreme gradient boosting algorithm.

4. The method for estimating Haloxylon ammodendron forest biomass based on collaborative inversion of multi-source UAV data according to claim 1, characterized in that, The first type of feature set also includes texture features extracted from the UAV multispectral image data based on the gray-level co-occurrence matrix.

5. The method for estimating Haloxylon ammodendron forest biomass based on collaborative inversion of multi-source UAV data according to claim 1, characterized in that, The vegetation index features include at least one of the following indices: normalized vegetation index, ratio vegetation index, and normalized greenness vegetation index.

6. The method for estimating Haloxylon ammodendron forest biomass based on collaborative inversion of multi-source UAV data according to claim 1, characterized in that, The second type of feature set also includes density features extracted from the UAV lidar point cloud data based on the point cloud density distribution.

7. The method for estimating Haloxylon ammodendron forest biomass based on collaborative inversion of multi-source UAV data according to claim 1, characterized in that, The height characteristics include maximum height and percentile height.

8. The method for estimating Haloxylon ammodendron forest biomass based on collaborative inversion of multi-source UAV data according to claim 1, characterized in that, The feature importance ranking method is the random forest feature importance ranking method, which quantifies feature importance by calculating the percentage increase in mean squared error for each feature.

9. The method for estimating Haloxylon ammodendron forest biomass based on UAV multi-source data collaborative inversion according to claim 1, characterized in that, The step of filtering the joint feature set to construct a target feature subset includes: selecting a predetermined number of features with the highest importance ranking based on the ranking results generated by the feature importance ranking method.

10. A biomass estimation device for Haloxylon ammodendron forests based on collaborative inversion of multi-source data from unmanned aerial vehicles, characterized in that, include: The sample set acquisition module is configured to acquire a training sample set, which includes ground biomass true values ​​from multiple training plots. The training data acquisition module is configured to acquire UAV multispectral image data and UAV lidar point cloud data for each of the training sample sites; The first type of feature extraction module is configured to extract a first type of feature set from the UAV multispectral image data, the first type of feature set including vegetation index features calculated based on spectral information; The second type of feature extraction module is configured to extract a second type of feature set from the UAV lidar point cloud data, the second type of feature set including height features obtained statistically based on point cloud height information. The feature fusion module is configured to fuse the first type of feature set with the second type of feature set to form a joint feature set; The feature filtering module is configured to perform filtering on the joint feature set based on the training sample set using a feature importance ranking method to construct a target feature subset, wherein the target feature subset simultaneously contains at least one feature filtered from the first type of feature set and at least one feature filtered from the second type of feature set; The model training module is configured to perform model training based on the training sample set and the corresponding target feature subset using a machine learning algorithm to obtain a ground biomass estimation model. The data acquisition module for the area to be estimated is configured to acquire UAV multispectral image data and UAV lidar point cloud data for the area to be estimated. The feature calculation module for the region to be estimated is configured to calculate the corresponding feature set based on the target feature subset from the UAV multispectral image data and UAV lidar point cloud data of the region to be estimated. The biomass estimation module is configured to input the calculated corresponding feature set into the aboveground biomass estimation model and output the aboveground biomass estimation result of the Haloxylon ammodendron forest in the area to be estimated.

Images