Method for estimating forest aboveground biomass by integrating multi-source remote sensing data and an integrated model

Abstract

The present invention provides a method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model, relating to the field of forest management technology. The method comprises: obtaining airborne LiDAR data, high-resolution orthophotos, and multi-source remote sensing data of a predetermined sample plot; using the high-resolution orthophotos as an aid, obtaining the forest aboveground biomass of the predetermined sample plot based on the airborne LiDAR data; performing feature optimization on the multi-source remote sensing data to obtain an optimal combination of remote sensing feature variables; constructing and stacking multiple models based on the optimal combination of remote sensing feature variables and the forest aboveground biomass of the predetermined sample plot using an extreme gradient boosting algorithm, an adaptive boosting algorithm, a random forest algorithm, and a gradient boosting decision tree algorithm; determining the model with the highest accuracy among all the models and using it to estimate the forest aboveground biomass of the predetermined study area to obtain an estimated result. The present invention solves the problem of low accuracy in forest biomass estimation in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of forest management, and in particular to a method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model. Background Art

[0002] Accurate estimation of forest biomass is crucial for sustainable forest management and ecosystem monitoring. Traditional biomass estimation methods rely primarily on field surveys and plot measurements. However, these methods have significant limitations in large-scale forest areas with complex terrain. Field measurements are constrained by cost, time, and human resources, and they struggle to cover the entire forest area. The use of airborne LiDAR (LiDAR) makes the entire process more automated and efficient. It can cover large areas of forest in a relatively short period of time, acquire structural information, and provide accurate individual tree structural parameters through tree segmentation technology, providing reliable data support for forest resource management and ecosystem monitoring.

[0003] When estimating forest biomass over large areas, multispectral imagery (such as Sentinel-2) and synthetic aperture radar imagery (such as Sentinel-1) each have unique advantages and limitations. Multispectral imagery, with its high spatial resolution and multi-band information, can provide detailed information on surface structure and vegetation types. However, it is constrained by factors such as cloud cover and limited daytime and nighttime availability. Furthermore, multispectral imagery can face optical saturation in areas of dense vegetation, limiting the accuracy of its estimates. Synthetic aperture radar imagery offers all-weather observation capabilities, unrestricted by weather and cloud cover, ensuring stable and continuous data acquisition and exhibiting good sensitivity in areas of dense vegetation cover. However, its lower spatial resolution and relative weakness in vegetation type classification limit its ability to reveal detailed forest structure. By leveraging the complementary nature of active and passive remote sensing data, it is possible to more accurately obtain information on forest biomass and vegetation types.

[0004] In terms of machine learning models, currently widely used single models are insufficient in expressing complex nonlinear relationships. Because the relationship between forest biomass and remote sensing variables can be multi-layered, nonlinear, and interactive, traditional single machine learning models struggle to capture these complex relationships. Therefore, designing more powerful and expressive machine learning models to improve the accuracy and robustness of forest biomass estimation is a key research challenge. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide a method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model. The present invention solves the problem of low accuracy in forest biomass estimation in the existing technology.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model, including:

[0008] Dividing the study area to be tested to obtain a number of sample plots;

[0009] Obtain airborne lidar data, high-resolution orthophotos of the pre-set sample plots, and multispectral images, synthetic aperture radar images, and DEM images of the study area to be measured;

[0010] Obtain multi-source remote sensing data of the study area to be measured based on the multispectral image, synthetic aperture radar image and DEM image;

[0011] Using the high-resolution orthophoto as an aid, obtaining the forest aboveground biomass of the preset plot according to the airborne lidar data;

[0012] Performing feature optimization on spectral information, polarization characteristic factors, texture information, and terrain factors in the multi-source remote sensing data to obtain an optimal remote sensing characteristic variable combination;

[0013] Based on the extreme gradient boosting algorithm, the adaptive boosting algorithm, the random forest algorithm and the gradient boosting decision tree algorithm, according to the optimal remote sensing characteristic variable combination and the forest aboveground biomass of the preset sample plot, a first biomass prediction model, a second biomass prediction model, a third biomass prediction model and a fourth biomass prediction model are respectively constructed;

[0014] Based on an ensemble learning method, a stacked prediction model is constructed according to the output results of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model and a multilayer perceptron algorithm;

[0015] Determining a prediction model with the best accuracy among the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model, and the stacked prediction model;

[0016] Based on the predicted model with the best accuracy, the forest aboveground biomass in the study area is estimated to obtain the estimated results.

[0017] Preferably, obtaining the forest aboveground biomass of the preset sample plot based on the airborne laser radar data includes:

[0018] Preprocessing the airborne laser radar data and obtaining the predicted tree height of a single tree in the preset sample plot based on a single tree segmentation algorithm according to the preprocessed airborne laser radar data;

[0019] Construct a tree height and diameter at breast height prediction model;

[0020] Inputting the predicted tree height of a single tree into a tree height and diameter at breast height prediction model to obtain the predicted diameter at breast height of a single tree in a preset sample plot;

[0021] The forest aboveground biomass of the preset sample plot is obtained according to the single tree predicted diameter at breast height and the single tree predicted tree height.

[0022] Preferably, the method for constructing the tree height and diameter at breast height prediction model comprises:

[0023] Measure the DBH data and tree height data of each tree in the preset sample plot;

[0024] Based on the tree height curve basic model, a tree height and diameter at breast height prediction model was constructed according to the diameter at breast height data and tree height data of each individual tree.

[0025] Preferably, the calculation expression for the predicted DBH of a single tree is:

[0026] D=(2.453*(H-1.3) 0.3789 )*exp(0.03883*(H-1.3));

[0027] Among them, D is the predicted DBH of a single tree, and H is the predicted tree height of a single tree.

[0028] Preferably, the airborne laser radar data is preprocessed and, based on the preprocessed airborne laser radar data, the predicted tree height of a single tree in the preset plot is obtained based on a single tree segmentation algorithm, including:

[0029] A denoising algorithm based on spatial distance is used to remove noise points in the original point cloud of the airborne laser radar data to obtain first preprocessed data;

[0030] Using an improved progressive triangulated network filtering algorithm to classify ground points in the first lidar data to obtain second preprocessed data;

[0031] The second pre-processed data is normalized according to the ground points and the predicted tree height of a single tree is obtained by using a region growth combined with a threshold judgment algorithm.

[0032] Preferably, based on an ensemble learning method, a stacked prediction model is constructed according to the output results of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model and the fourth biomass prediction model and a multilayer perceptron algorithm, including:

[0033] Optimizing the parameters of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model by means of network search to obtain a first optimization model, a second optimization model, a third optimization model, and a fourth optimization model;

[0034] Based on the ensemble learning method, an initial model is obtained according to the output results of the first optimization model, the second optimization model, the third optimization model and the fourth optimization model;

[0035] The multilayer perceptron algorithm is used as a meta-model and the output results of the initial model are trained to construct a stacking prediction model.

[0036] Preferably, determining the prediction model with the best accuracy among the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model and the stacked prediction model comprises:

[0037] The coefficient of determination and root mean square error were calculated;

[0038] performing accuracy evaluation on the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model, and the stacking prediction model using the determination coefficient and the root mean square error to obtain an evaluation result;

[0039] The prediction model with the best accuracy is determined according to the evaluation results.

[0040] Preferably, it also includes:

[0041] Each tree in the pre-set sample plot was measured to obtain the actual diameter at breast height, actual height and actual coordinates of each tree;

[0042] Determine the measured forest aboveground biomass of the preset sample plot based on the measured diameter at breast height, measured tree height and measured coordinates of each individual tree;

[0043] The estimated results are verified based on the measured forest aboveground biomass.

[0044] The present invention discloses the following technical effects:

[0045] The present invention provides a method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model, comprising: dividing the study area to be measured to obtain a plurality of sample plots; obtaining airborne laser radar data, high-resolution orthophotos, and multispectral images, synthetic aperture radar images, and DEM images of the preset sample plots; obtaining multi-source remote sensing data of the study area to be measured based on the multispectral images, synthetic aperture radar images, and DEM images; using the high-resolution orthophotos as an aid, obtaining the forest aboveground biomass of the preset sample plot based on the airborne laser radar data; performing feature optimization on spectral information, polarization characteristic factors, texture information, and terrain factors in the multi-source remote sensing data to obtain an optimal remote sensing feature variable combination; and optimizing the remote sensing feature variable combination based on an extreme gradient boosting algorithm, an adaptive enhancement algorithm, and a random access algorithm. The machine forest algorithm and the gradient boosting decision tree algorithm are used to construct the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model respectively based on the optimal remote sensing characteristic variable combination and the forest aboveground biomass of the preset sample plot; based on the ensemble learning method, a stacked prediction model is constructed based on the output results of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model and the multi-layer perceptron algorithm; the prediction model with the best accuracy among the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model, and the stacked prediction model is determined; based on the determined prediction model with the best accuracy, the forest aboveground biomass of the study area to be measured is estimated to obtain an estimation result. In order to solve the problem that the collection of measured single tree biomass data is time-consuming and labor-intensive, the present invention is based on airborne lidar data, assisted by orthophoto high-resolution data, and uses single tree segmentation technology and establishes a tree height-diameter-at-breast height prediction model to accurately estimate forest biomass at the scale of single tree and sample plot. The present invention can provide a large number of reliable training samples for the construction of regional forest biomass models, and can also provide an effective means for forest resource inventory; the feature selection method adopted by the present invention can more effectively capture the nonlinear relationship between forest biomass and remote sensing characteristic factors compared to traditional Pearson correlation analysis. At the same time, by combining active and passive remote sensing factors and terrain factors, the richness of spectral information is improved, and the uncertainty of a single data source is reduced, thereby making the estimation results of forest biomass more accurate and reliable; the present invention significantly improves the generalization ability and robustness of the algorithm by stacking multiple independent machine learning models into an integrated model. This stacking and integration method gives full play to the unique advantages of each single model, and by integrating the predictive capabilities of different models, the final model is more adaptable and can better handle complex samples.This integration strategy not only helps improve the performance of the model on training data, but also enhances the practicality and reliability of the entire algorithm. Based on airborne lidar point cloud data, the present invention adopts single-tree segmentation technology to reduce the difficulty and cost of collecting single-tree biomass. It takes advantage of multiple data sources and uses stability selection for feature optimization to improve the stability of the model estimation of forest aboveground biomass. It also combines advanced machine learning algorithms to improve the estimation accuracy of forest aboveground biomass. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 A first flow chart of a method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model provided by an embodiment of the present invention;

[0048] Figure 2 A second flow chart of a method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model provided by an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of the accuracy of single tree parameter prediction based on airborne laser point cloud extraction provided by an embodiment of the present invention, wherein: Figure 3 (a) is a scatter plot of the fitting of the predicted tree height and the measured tree height of a single tree. Figure 3 (b) is a scatter plot of the predicted DBH of a single tree and the measured DBH. Figure 3 (c) is a scatter plot of the predicted biomass and measured biomass of individual trees.

[0050] Figure 4 A schematic diagram of estimating aboveground biomass of a large forest area using a stacking prediction model provided by an embodiment of the present invention;

[0051] Figure 5 A fitting scatter plot showing the accuracy of estimating forest aboveground biomass using a stacking prediction model using measured data as verified by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] like Figure 1 As shown, the present invention provides a method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model, comprising:

[0055] Step 100: Divide the study area to be tested to obtain a number of sample plots;

[0056] Step 200: Acquire airborne lidar data, high-resolution orthophotos of a preset sample site, and multispectral images, synthetic aperture radar images, and DEM images of the study area to be measured;

[0057] Step 300: obtaining multi-source remote sensing data of the study area to be measured based on the multispectral image, synthetic aperture radar image and DEM image;

[0058] Step 400: using the high-resolution orthophoto as an aid, obtaining the forest aboveground biomass of the preset sample plot according to the airborne lidar data;

[0059] Step 500: performing feature optimization on the spectral information, polarization characteristic factors, texture information, and terrain factors in the multi-source remote sensing data to obtain an optimal remote sensing characteristic variable combination;

[0060] Step 600: Based on the extreme gradient boosting algorithm, the adaptive boosting algorithm, the random forest algorithm, and the gradient boosting decision tree algorithm, a first biomass prediction model, a second biomass prediction model, a third biomass prediction model, and a fourth biomass prediction model are respectively constructed according to the optimal remote sensing characteristic variable combination and the forest aboveground biomass of the preset sample plot;

[0061] Step 700: Based on an ensemble learning method, a stacked prediction model is constructed according to the output results of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model and a multilayer perceptron algorithm;

[0062] Step 800: determining a prediction model with the best accuracy among the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model, and the stacked prediction model;

[0063] Step 900: Estimate the aboveground biomass of the forest in the study area to be measured based on the determined prediction model with the best accuracy to obtain an estimation result.

[0064] Specifically, such as Figure 2As shown in Figure 2, another way to express a method for estimating forest aboveground biomass by combining multi-source remote sensing data and integrated models is as follows:

[0065] Step 1: Collecting measured data. First, sample plots are laid out, including plots of different stand ages. Three plots of 20m x 20m are evenly distributed for each stand age to ensure that the measured data covers information from different stages of tree growth. Next, each tree with a DBH greater than or equal to 5cm in the sample plot is measured to obtain the DBH, height, and coordinates of each individual tree. The measured individual tree biomass is calculated based on existing allometric growth equations. Within the same time period and region, airborne LiDAR data and high-resolution orthophotos are collected, as well as multispectral imagery, synthetic aperture radar imagery, and DEM imagery.

[0066] Step 2: Using the measured DBH and tree height data of each individual tree obtained in step 1, a single tree DBH prediction model is established based on the tree height curve basic model;

[0067] Step 3: The airborne lidar data is preprocessed by flight adjustment, denoising, and ground point filtering. The individual tree height and coordinate information are then extracted using a single tree segmentation algorithm. The individual tree detection rate is evaluated using detection rate, accuracy, and F-score as indicators.

[0068] Step 4: Use the tree height of the correctly split tree extracted in step 3 as the independent variable in step 2 to calculate the predicted diameter at breast height (DBH); and substitute the predicted DBH and tree height into the allometric growth equation in step 1 to obtain the predicted single tree biomass.

[0069] Step 5: Based on the grid size generated by the multispectral imagery and supplemented by the high-resolution orthophotos from step 1, training samples are uniformly selected within the area of different forest ages. The split trees contained in the sample plot are cropped using a fixed grid size, and the predicted biomass of each split tree is obtained according to step 4, and the predicted biomass is summed up to obtain the biomass of the sample plot.

[0070] Step 6: Using the biomass of the sample plot obtained in step 5 as the dependent variable for constructing the regional model, the spectral information, texture information, and terrain factors obtained from the multi-source remote sensing images are subjected to feature optimization to obtain the optimal combination of remote sensing feature variables;

[0071] In step 7, the dependent variable and optimal feature variable from step 6 are used as the model input data. The biomass prediction model is constructed using eXtreme Gradient Boosting (XGBoost), Adaptive Boosting (Adaboost), Random Forest (RF), and Gradient Boosting Decision Trees (GBDT). Finally, an ensemble learning method is introduced to stack the prediction results of the above four machine learning algorithms into a single model. The multilayer perceptron algorithm model is used as the meta-model to generate the final prediction result. This combines multiple models into a more powerful model.

[0072] Step 8: Evaluate the estimation accuracy of different models in step 7, select the optimal model to estimate forest biomass in a large area, and verify the estimated results using the biomass of the sample plots measured in step 1.

[0073] Furthermore, obtaining the forest aboveground biomass of the preset sample plot based on the airborne laser radar data includes:

[0074] Preprocessing the airborne laser radar data and obtaining the predicted tree height of a single tree in the preset sample plot based on a single tree segmentation algorithm according to the preprocessed airborne laser radar data;

[0075] Construct a tree height and diameter at breast height prediction model;

[0076] Inputting the predicted tree height of a single tree into a tree height and diameter at breast height prediction model to obtain the predicted diameter at breast height of a single tree in a preset sample plot;

[0077] The forest aboveground biomass of the preset sample plot is obtained according to the single tree predicted diameter at breast height and the single tree predicted tree height.

[0078] Furthermore, the method for constructing the tree height and diameter at breast height prediction model includes:

[0079] Measure the DBH data and tree height data of each tree in the preset sample plot;

[0080] Based on the tree height curve basic model, a tree height and diameter at breast height prediction model was constructed according to the diameter at breast height data and tree height data of each individual tree.

[0081] Furthermore, the expression of the allometric growth equation is:

[0082] AGB=0.02407389*(D 2 H) 0.9768058 +0.00492556*(D 2 H) 0.8449044+0.0007088*(D 2 H) 0.935545 +0.0063525*(D 2 H) 0.8738162 ;

[0083] Where AGB is the biomass of a single tree, in kg; D is the predicted diameter at breast height of a single tree, in cm; and H is the height of the correctly segmented tree, in m.

[0084] Furthermore, the calculation expression for the predicted DBH of a single tree is:

[0085] D=(2.453*(H-1.3) 0.3789 )*exp(0.03883*(H-1.3)).

[0086] Furthermore, the airborne laser radar data is preprocessed and, based on the preprocessed airborne laser radar data, a predicted tree height of a single tree in the preset sample plot is obtained based on a single tree segmentation algorithm, including:

[0087] A denoising algorithm based on spatial distance is used to remove noise points in the original point cloud of the airborne laser radar data to obtain first preprocessed data;

[0088] Using an improved progressive triangulated network filtering algorithm to classify ground points in the first lidar data to obtain second preprocessed data;

[0089] The second pre-processed data is normalized according to the ground points and the predicted tree height of a single tree is obtained by using a region growth combined with a threshold judgment algorithm.

[0090] Specifically, the LiDAR360 software was used to process the airborne laser point cloud. First, a spatial distance-based denoising algorithm was used to remove noise points from the original point cloud. The number of neighborhood points was set to 10, and the standard deviation multiple was set to 10. Then, an improved progressive triangulation filtering algorithm was used to classify ground points. The scene used steep slope terrain, and the maximum terrain slope, iteration angle, and iteration distance were set to 88°, 13°, and 1.7m, respectively. The ground points were normalized and finally, a region growing combined with a threshold judgment algorithm was used to segment individual trees. The accuracy of the individual tree detection rate was evaluated using the accuracy rate (p), detection rate (r), and F score. The details are as follows:

[0091]

[0092]

[0093]

[0094] Where TP, FP, and FN represent the number of correctly segmented trees, the number of over-segmented trees, and the number of under-segmented trees, respectively. The F-score is the weighted harmonic mean obtained by considering both p and r.

[0095] Furthermore, based on an ensemble learning method, a stacked prediction model is constructed according to the output results of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model and a multilayer perceptron algorithm, including:

[0096] Optimizing the parameters of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model by means of network search to obtain a first optimization model, a second optimization model, a third optimization model, and a fourth optimization model;

[0097] Based on the ensemble learning method, an initial model is obtained according to the output results of the first optimization model, the second optimization model, the third optimization model and the fourth optimization model;

[0098] The multilayer perceptron algorithm is used as a meta-model and the output results of the initial model are trained to construct a stacking prediction model.

[0099] Furthermore, determining the prediction model with the best accuracy among the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model, and the stacked prediction model includes:

[0100] The coefficient of determination and root mean square error were calculated;

[0101] performing accuracy evaluation on the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model, and the stacking prediction model using the determination coefficient and the root mean square error to obtain an evaluation result;

[0102] The prediction model with the best accuracy is determined according to the evaluation results.

[0103] Specifically,

[0104]

[0105] in, Represents the model estimate, y i represents the sample measured value, represents the mean of the sample measured values, and n represents the number of samples.

[0106] Furthermore, each tree in the preset sample plot is measured to obtain the actual diameter at breast height, actual height and actual coordinates of each tree;

[0107] Determine the measured forest aboveground biomass of the preset sample plot based on the measured diameter at breast height, measured tree height and measured coordinates of each individual tree;

[0108] The estimated results are verified based on the measured forest aboveground biomass.

[0109] Furthermore, the training samples used to build the biomass prediction model all come from the segmented tree data extracted from the airborne lidar point cloud. First, the sample plots need to be selected. In QGIS software, the multi-source images are resampled to pixels consistent with the actual sample plot size (20m×20m), and a fishing net and feature-to-surface operations are created. Then, within the area containing the tree species and the airborne lidar data range, 165 pixels of different forest ages are evenly selected. Finally, the segmented trees contained in the 165 pixels are cropped out, and according to the predicted biomass of each segmented tree, the biomass of individual trees contained in different pixels is summarized to obtain the training samples of the model.

[0110] Texture, vegetation index, and single-band features were extracted from Sentinel-2 images, as well as texture and polarimetric features from Sentinel-1 images. Topographic features were extracted from DEM images using QGIS software, resulting in a total of 118 feature factors. Feature optimization was performed using a random forest algorithm in Python, based on the stability of the underlying model.

[0111] The stacked prediction model includes a base model and a meta-model. In the first stage, the parameters of each base model (XGBoost, Adaboost, RF, and GBDT) are optimized through network search. In the second stage, the multiple base models optimized in the first part are integrated into one model, and the meta-model (multi-layer perceptron algorithm) is trained using the prediction results of these base models to make the final prediction. This process aims to improve the generalization performance of a single learning model and achieve better performance by combining the advantages of different models; 80% of the forest aboveground biomass (the biomass of individual trees in each plot) is used for modeling, and 20% is used to evaluate the performance of the model. The evaluation index uses the determination coefficient R 2 The optimal model was evaluated to estimate the aboveground biomass of regional forests. The estimated biomass was then evaluated using measured data to verify the feasibility of the present invention.

[0112] The present invention also provides the following specific embodiments:

[0113] Acquisition of measured data:

[0114] The study area of this example is located in the transition zone from the southern subtropics to the mid-subtropics. A forest farm collected measured data from November to December 2022. Eucalyptus forests aged 2 to 7 years were selected, and 3 representative plots were evenly selected for each age of eucalyptus forest. After the plots were selected, 18 plots of 20×20m (400m 2 ) plots. The slope, aspect, altitude, stand density and other information of the plots were recorded. Each eucalyptus tree in each plot was measured, and there were a total of 930 measured trees. Each tree was measured three times using a laser altimeter and the average was taken as the final tree height value. The diameter at breast height greater than 5 cm was measured at a height of 1.3 meters. The single tree positioning in the plot was carried out using the Zhonghaida D8proRTK system. During the measurement process, it was ensured that the solution was fixed before recording. The diameter at breast height and tree height of the above 930 measured trees were substituted into the empirical allometric growth model to calculate the forest biomass of each eucalyptus tree as the measured biomass of the single tree, which was summarized to obtain the measured biomass of the plot.

[0115] Acquisition of airborne laser point cloud and orthophoto data:

[0116] During the same time period, a company's M300RTK quad-rotor drone equipped with a Zenmuse L1 laser scanning system was used to collect airborne laser point cloud data of eucalyptus trees. The Zenmuse L1 integrates a Livox lidar module, a high-precision IMU, and a 50-inch CMOS camera on a three-axis stabilized gimbal. Using the terrain-simulating flight mode, the flight altitude was set to 100m relative to the ground, the speed was set to 10m / s, the payload was set to triple echo mode, 140kHz laser pulse emission frequency and non-repetitive scanning mode, and the scanner was set to face the ground at 90°. A DJI M300RTK drone equipped with an H20T lens was used to collect orthophoto data. The altitude was set to 300m relative to the ground and the image resolution was 0.1m. Table 1 is a comparison table of the detection accuracy of single tree segmentation at different forest ages. Table 1 is as follows:

[0117] Table 1 Comparison of single tree segmentation detection accuracy at different forest ages

[0118]

[0119]

[0120] Acquisition of multispectral imagery, synthetic aperture radar imagery, and DEM imagery data:

[0121] The Sentinel-1 / 2 (Sentinel-1 / Sentinel-2) used in this embodiment are both from the Google Earth Engine (GEE) platform. The sensor carried by the Sentinel-1 satellite is a synthetic aperture radar (SAR) based on the C band. Through SAR technology, Sentinel-1 can capture microwave radiation from the surface of the earth at different times and weather conditions. The Sentinel-1 data provided by the GEE platform has undergone a series of key pre-processing steps, including orbit correction, thermal noise removal, radiation calibration, and terrain correction. The multispectral sensor carried by the Sentinel-2 satellite has 13 bands. This embodiment uses the Sentinel-2 2A-level product provided by the GEE platform. The 2A-level product has undergone key pre-processing steps such as atmospheric correction, geometric correction, and cloud and shadow removal. The DEM image comes from the polarimetric synthetic aperture radar (PolSAR) sensor carried on the ALOS satellite (Advanced Land Observing Satellite). This sensor can obtain polarimetric radar reflectivity data on the earth's surface, thereby providing high-resolution terrain information.

[0122] Extraction of single tree parameters and selection of model training samples based on airborne lidar point cloud:

[0123] LiDAR360 software was used to pre-process the airborne laser point cloud, including flight adjustment, denoising, and ground point filtering. The point cloud was then segmented into individual trees using a region growing algorithm combined with a threshold judgment algorithm. The individual tree segmentation results were evaluated based on both detection rate and structural parameter accuracy. The accuracy of the individual tree detection rate is shown in Table 1. The structural parameter accuracy can be calculated by calculating the R between the predicted tree height and the measured tree height of the matching trees. 2 and RMSE values, such as Figure 3 (a) shows that the predicted tree height after single tree segmentation extraction is substituted into the tree height and diameter at breast height prediction model to obtain the predicted diameter at breast height. The scatter plot of the predicted diameter at breast height and the measured diameter at breast height is shown in Figure 3 (b) As shown. Substituting the predicted tree height and predicted DBH into the empirical allometric equation, the predicted biomass of a single tree is obtained. The scatter plot of the predicted biomass and the measured biomass is shown in Figure 3 (c) shown.

[0124] Then, in QGIS software, resample the multi-source image to the measured plot size:

[0125] Consistent pixels (20m×20m) were used. Within the grid generated by multi-source imagery, 165 pixels of varying forest ages were evenly selected, supplemented by high-resolution orthophotos. Finally, the split trees contained within each of these 165 pixels were cropped, and the predicted biomass of each split tree was obtained according to step 4. The biomass of individual trees contained within different pixels was summed to obtain the biomass of the 165 training samples.

[0126] Selection and optimization of remote sensing characteristic factors:

[0127] like Figure 4 As shown, this embodiment selects 118 factors, including single-band features (bands B2-B8A in Sentinel-2 data), vegetation index, polarization features, texture features (window size is 9×9) and terrain features, as multi-source remote sensing feature variables (Table 2). Stability selection is used to optimize feature variables. The stability of the features is evaluated by randomly resampling the data, and the features are sorted according to the stability score of the features to determine the most important features. Through this method, the instability of the results of a single feature selection can be avoided, and features that have an important impact on the prediction performance of the model can be better identified. In this embodiment, the stability score of 10% is set as the threshold value, and the 12 factors with the highest stability scores are selected as the independent variables of the model. Table 2 is a collection table of remote sensing feature factors extracted from multi-source remote sensing data, and Table 2 is shown below:

[0128] Table 2 Collection of remote sensing feature factors extracted from multi-source remote sensing data

[0129]

[0130]

[0131] Table 3 is the stability score table of the optimal remote sensing feature variable combination, as shown below:

[0132] Table 3 Stability score table of the optimal remote sensing characteristic variable combination

[0133]

[0134] The _x after the different texture features in Table 3 refers to the different band information of Sentinel-1 / 2 data.

[0135] Construction and estimation of regional forest aboveground biomass model:

[0136] Taking 165 forest aboveground biomass training samples selected based on airborne point cloud data as dependent variables and the optimal remote sensing feature variable factor as the independent variable, XGBoost, Adaboost, RF and GBDT were used to construct regional forest aboveground biomass models respectively. Then, the ensemble learning method was introduced to stack the prediction results of the above four machine learning algorithms into one model, and the multi-layer perceptron algorithm model was used as the meta-model to generate the prediction results. By comparing Table 4, it can be seen that the estimation accuracy of the stacking prediction model provided by the embodiment of the present invention is better than XGBoost, Adaboost, RF and GBDT. Finally, the stacking prediction model with the best effect was selected to estimate the regional forest aboveground biomass, and the estimated results were verified using measured data, such as Figure 5 As shown, where R 2 The accuracy of different models in estimating regional forest aboveground biomass is compared in Table 4.

[0137] Table 4 Comparison of the accuracy of different models in estimating regional forest aboveground biomass

[0138]

[0139]

[0140] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0141] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model, characterized in that: include: The research area to be measured was divided into several sample plots; Obtain airborne lidar data, high-resolution orthophotos of the pre-set sample plots, and multispectral images, synthetic aperture radar images, and DEM images of the study area to be measured; Obtain multi-source remote sensing data of the study area to be measured based on the multispectral image, synthetic aperture radar image and DEM image; Using the high-resolution orthophoto as an aid, obtaining the forest aboveground biomass of the preset plot according to the airborne lidar data; Performing feature optimization on spectral information, polarization characteristic factors, texture information, and terrain factors in the multi-source remote sensing data to obtain an optimal remote sensing characteristic variable combination; Based on the extreme gradient boosting algorithm, the adaptive boosting algorithm, the random forest algorithm and the gradient boosting decision tree algorithm, according to the optimal remote sensing characteristic variable combination and the forest aboveground biomass of the preset sample plot, a first biomass prediction model, a second biomass prediction model, a third biomass prediction model and a fourth biomass prediction model are respectively constructed; Based on an ensemble learning method, a stacked prediction model is constructed according to the output results of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model and a multilayer perceptron algorithm; Determining a prediction model with the best accuracy among the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model, and the stacked prediction model; According to the prediction model with the best accuracy, the forest aboveground biomass in the research area to be measured is estimated to obtain the estimation results; Based on the ensemble learning method, a stacked prediction model is constructed according to the output results of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model and a multilayer perceptron algorithm, including: Optimizing the parameters of the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, and the fourth biomass prediction model by means of network search to obtain a first optimization model, a second optimization model, a third optimization model, and a fourth optimization model; Based on the ensemble learning method, an initial model is obtained according to the output results of the first optimization model, the second optimization model, the third optimization model and the fourth optimization model; Using the multilayer perceptron algorithm as a meta-model and training the output of the initial model to construct a stacked prediction model; Determining a prediction model with the best accuracy among the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model, and the stacked prediction model includes: The coefficient of determination and root mean square error were calculated; performing accuracy evaluation on the first biomass prediction model, the second biomass prediction model, the third biomass prediction model, the fourth biomass prediction model, and the stacking prediction model using the determination coefficient and the root mean square error to obtain an evaluation result; Determine the prediction model with the best accuracy according to the evaluation results; The obtaining of the forest aboveground biomass of the preset sample plot according to the airborne laser radar data includes: Preprocessing the airborne laser radar data and obtaining the predicted tree height of a single tree in the preset sample plot based on a single tree segmentation algorithm according to the preprocessed airborne laser radar data; Construct a tree height and diameter at breast height prediction model; Inputting the predicted tree height of a single tree into a tree height and diameter at breast height prediction model to obtain the predicted diameter at breast height of a single tree in a preset sample plot; Obtaining the forest aboveground biomass of the preset sample plot according to the single tree predicted diameter at breast height and the single tree predicted tree height; Preprocessing the airborne laser radar data and obtaining the predicted tree height of a single tree in the preset sample plot based on a single tree segmentation algorithm according to the preprocessed airborne laser radar data includes: A denoising algorithm based on spatial distance is used to remove noise points in the original point cloud of the airborne laser radar data to obtain first preprocessed data; Using an improved progressive triangulated network filtering algorithm to classify ground points in the first lidar data to obtain second preprocessed data; Normalizing the second pre-processed data according to the ground points and using a region growing combined with a threshold judgment algorithm to obtain a predicted tree height of a single tree; The method also includes: measuring each tree in the preset sample plot to obtain the actual diameter at breast height, actual height and actual coordinates of each tree; Determine the measured forest aboveground biomass of the preset sample plot based on the measured diameter at breast height, measured tree height and measured coordinates of each individual tree; The estimated results are verified based on the measured forest aboveground biomass.

2. The method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model according to claim 1, characterized in that: The method for constructing the tree height and diameter at breast height prediction model comprises: Measure the DBH data and tree height data of each tree in the preset sample plot; Based on the tree height curve basic model, a tree height and diameter at breast height prediction model was constructed according to the diameter at breast height data and tree height data of each individual tree.

3. The method for estimating forest aboveground biomass by combining multi-source remote sensing data and an integrated model according to claim 2, characterized in that: The calculation expression of the predicted DBH of a single tree is: D=(2.453*(H-1.3) 0.3789 )*exp(0.03883*(H-1.3)); Among them, D is the predicted DBH of a single tree, and H is the predicted tree height of a single tree.

Images