A method for estimating forest carbon storage based on multi-source remote sensing data and random forest method

Abstract

The present invention discloses a method for estimating forest carbon reserves based on multi-source remote sensing data and random forest method, establishes remote sensing estimation models of carbon reserves of different types of forests based on multi-source remote sensing data, ground sample plot survey data and random forest algorithm, and obtains the optimal remote sensing estimation model of forest carbon reserves after training and optimizing the model, so as to perform carbon reserve inversion mapping of different types of forests in the study area. In the present invention, the spectral characteristics of optical remote sensing images can well reflect the characteristic information of ground objects; laser radar, as an active remote sensing method, has the characteristics of strong penetration, not easily affected by bad weather, high-resolution information, etc.; random forest algorithm, as a machine learning algorithm, can better describe the nonlinear relationship between forest carbon reserves and remote sensing variables than traditional linear regression algorithms. These advantages make this method less susceptible to factors such as terrain, climate, time, etc. in the process of implementing forest carbon reserve estimation, and has higher estimation accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of forest carbon storage estimation, and relates to a method for estimating forest carbon storage based on multi-source remote sensing data and a random forest method. Background Art

[0002] As an important part of the terrestrial ecosystem, the forest ecosystem has the largest carbon pool on the land surface. Forest carbon sink is one of the most economical and effective ways to cope with climate change and achieve the "dual carbon goals". In this context, it is particularly important to accurately estimate forest carbon reserves.

[0003] Traditional carbon sink monitoring methods include plot inventory method and carbon flux positioning observation method. The plot inventory method usually relies on long-term and high-cost plot survey work, which is costly and its accuracy is affected by human factors; the carbon flux positioning observation method is affected by the distribution of the device, and the monitoring range is very limited, making it difficult to cover large areas of multi-type forest areas. In recent years, remote sensing and related technologies have developed by leaps and bounds. The spectral characteristics of optical remote sensing images can well reflect the characteristic information of ground objects. As an active remote sensing method, lidar has the characteristics of strong penetration, not easily affected by bad weather, all-weather operation, and high-resolution information, which provides new ideas for remote sensing estimation of forest carbon reserves.

[0004] Remote sensing-based model simulation methods are necessary for estimating large-scale forest carbon stocks. Using different models in remote sensing modeling will greatly affect the accuracy of carbon stock estimation. The random forest algorithm is a machine learning algorithm that can better describe the nonlinear relationship between forest carbon stocks and remote sensing variables than the traditional linear regression algorithm, and has better adaptability and higher prediction accuracy.

[0005] In summary, it is particularly necessary to develop a forest carbon stock estimation technology based on multi-source remote sensing data and random forest algorithm. Summary of the invention

[0006] The purpose of the present invention is to provide a method for estimating forest carbon reserves based on multi-source remote sensing data and the random forest method. Based on multi-source remote sensing data and ground sample plot survey data, high-precision estimation of carbon reserves of different types of forests is achieved, which solves the problems of insufficient accuracy and high labor costs in forest carbon reserve determination in existing methods, can greatly reduce labor costs, significantly improve work efficiency, and data is not easily restricted by factors such as terrain, climate, and time, and the estimation accuracy can meet the requirements of investigation and monitoring.

[0007] The technical solution adopted by the present invention is: a method for estimating forest carbon storage based on multi-source remote sensing data and random forest method, and the specific operation steps are as follows:

[0008] Step 1: Set up a ground survey plot in the study area, obtain survey data for each tree in the plot, and obtain optical remote sensing image data and lidar point cloud data of the area close to the survey time;

[0009] Step 2: Based on the tree survey data obtained in step 1, the forest carbon stock of each sample plot is calculated as a data set;

[0010] Step 3, pre-processing the optical remote sensing image and lidar point cloud data obtained in step 1 respectively, and finally unifying the coordinate system of the remote sensing image data and the point cloud data into CGCS2000;

[0011] Step 4: Based on the preprocessed optical remote sensing image and the coordinates of the center point of the sample plot, two types of remote sensing variables, namely, spectral information and texture information, are extracted; based on the preprocessed lidar point cloud data, three types of remote sensing variables, namely, height variable, density variable, and intensity variable, are extracted, and correlation and significance analysis is performed on the extracted remote sensing variables and the forest carbon stock calculated in step 2, and remote sensing variables with high correlation are selected to participate in the establishment of a remote sensing estimation model for forest carbon stock;

[0012] Step 5, using the remote sensing variables screened in step 4 as independent variables and the forest carbon storage data calculated in step 2 as dependent variables, selecting modeling samples, using the random forest algorithm to establish a remote sensing estimation model for forest carbon storage, and optimizing the model by adjusting model parameters;

[0013] Step 6: Based on the test samples, the accuracy of the remote sensing estimation model of forest carbon stock obtained in step 5 is verified and evaluated;

[0014] Step 7: Based on the optimal forest carbon storage remote sensing estimation model obtained in step 5, the forest carbon storage in the study area is inverted and mapped.

[0015] The present invention is also characterized in that:

[0016] In step 1, the ground sample survey content includes three categories:

[0017] 1) Stand factors, i.e. dominant tree species, origin, and canopy density;

[0018] 2) Tree factors, i.e. tree number, standing tree type, tree species, diameter at breast height, and tree height;

[0019] 3) Other information, i.e. plot number, location coordinates, investigator, and survey date.

[0020] In step 2, forest carbon stock calculations use the tree allometric equation and the percentage of carbon content in biomass;

[0021] The tree allometric growth equation is AGB=aD b Hc , where a, b, c are equation parameters, AGB is forest aboveground biomass, D is breast height, and H is tree height;

[0022] When the tree species is larch, a=0.06848, b=2.01549, c=0.59146; carbon content 47.42%;

[0023] When the tree species is red pine, a=0.027847, b=1.810004, c=0.905002; carbon content 48.09%;

[0024] When the tree species is oak, a=0.06149, b=2.14380, c=0.58390; carbon content 48.02%;

[0025] When the tree species is birch, a=0.06807, b=2.10850, c=0.52019; the carbon content is 45.56%.

[0026] Step 3: Preprocessing of remote sensing images includes radiation calibration, atmospheric correction, image mosaicking, and image cropping; preprocessing of point cloud data includes flight strip adjustment, data stitching, noise filtering, point cloud classification, projection conversion, and data cropping.

[0027] In step 4, the extracted remote sensing variables are subjected to Person correlation analysis and significance analysis with the forest carbon storage data calculated in step 2, and remote sensing variables with correlation coefficient r ≥ 0.3 and significance p ≤ 0.01 are selected as candidate variables for modeling.

[0028] In step 5, 70% of the forest carbon storage data calculated in step 2 is selected as the modeling sample, and the remaining 30% is selected as the test sample. The carbon storage remote sensing estimation model is trained using the random forest algorithm, and the model is optimized by adjusting the model parameters. The model parameters include the random seed number seed, the number of trees in the forest ntree, and the number of features used for each tree mtry.

[0029] In step 6, the coefficient of determination R 2 RMSE, MAE, and the determination coefficient R 2 The larger the value is, the smaller the root mean square error RMSE and relative root mean square error MAE are, indicating that the model accuracy is higher. The specific calculation methods of the three indicators are as follows:

[0030] (1)

[0031] (2)

[0032] (3)

[0033] Where: n is the number of sample plots; i Number the plots; x i For sample plot i Measured value of stand volume; For sample plot i The stand volume model estimate of It is the average value of stand volume measured in all sample plots.

[0034] The beneficial effects of the present invention are:

[0035] 1. Traditional forest carbon stock estimation relies on sample plot surveys, which is labor-intensive, inefficient, and has its accuracy greatly affected by human factors. This method extracts remote sensing variables based on multi-source remote sensing data and uses the random forest algorithm to establish a forest carbon stock remote sensing estimation model, which saves manual survey costs, improves efficiency, and can realize remote sensing estimation of carbon stocks in large areas of forests.

[0036] 2. The implementation of this method is not easily restricted by factors such as terrain, climate, and time. Multi-source remote sensing data has the characteristics of strong penetration, not susceptible to bad weather, all-weather operation, and high-resolution information, which makes this method less susceptible to external constraints during implementation.

[0037] 3. A higher estimation accuracy can be achieved. Based on this method, the forest carbon storage of Korean pine forest, larch forest, oak forest, birch forest and other forest stands was estimated, and the estimation accuracy can reach R 2 >0.7, which is better than all existing estimation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a technical flow chart of the method for estimating forest carbon storage based on multi-source remote sensing data and random forest method of the present invention.

[0039] Figure 2 is a schematic diagram of a ground sample plot in Example 3 of the present invention;

[0040] FIG3 (a) is a scatter plot of the estimated value and the measured value of the carbon storage of the Korean pine forest estimated by the method of the present invention;

[0041] FIG3 ( b ) is a scatter plot of the estimated and measured carbon storage values ​​of larch forests estimated by the method of the present invention;

[0042] FIG3 (c) is a scatter plot of the estimated and measured carbon storage values ​​of the oak forest estimated by the method of the present invention;

[0043] FIG3( d ) is a scatter plot of the estimated and measured carbon storage values ​​of birch forests estimated using the method of the present invention.

[0044] Among them, a is a red pine forest, b is a larch forest, c is an oak forest, and d is a birch forest. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0046] Example 1

[0047] like Figure 1 As shown in the figure, the method for estimating forest carbon storage based on multi-source remote sensing data and random forest method has the following specific steps:

[0048] Step 1: Set up a ground survey plot in the study area, obtain survey data for each tree in the plot, and obtain Sentinel-2A optical remote sensing image data and lidar point cloud data of the area close to the survey time;

[0049] Step 2: Based on the tree survey data obtained in step 1, the forest carbon stock of each plot is calculated as a data set using the tree allometric growth equation and the stand biomass carbon measurement model;

[0050] Step 3, pre-processing the optical remote sensing image and lidar point cloud data obtained in step 1 respectively, and finally unifying the coordinate system of the remote sensing image data and the point cloud data into CGCS2000;

[0051] Step 4: Based on the preprocessed optical remote sensing image and the coordinates of the center point of the sample plot, two types of remote sensing variables, namely, spectral information and texture information, are extracted; based on the preprocessed lidar point cloud data, three types of remote sensing variables, namely, height variable, density variable, and intensity variable, are extracted, and correlation analysis and significance analysis are performed on the extracted remote sensing variables and the forest carbon stock calculated in step 2, and remote sensing variables with correlation coefficient r ≥ 0.3 and significance p ≤ 0.01 are selected to participate in the establishment of the remote sensing estimation model of forest carbon stock;

[0052] Step 5, using the remote sensing variables screened in step 4 as independent variables and the forest carbon storage data calculated in step 2 as dependent variables, selecting modeling samples, using the random forest algorithm to establish a remote sensing estimation model for forest carbon storage, and optimizing the model by adjusting model parameters;

[0053] Step 6: Based on the test samples, the accuracy of the remote sensing estimation model of forest carbon stock obtained in step 5 is verified and evaluated;

[0054] Step 7: Based on the optimal forest carbon storage remote sensing estimation model obtained in step 5, the forest carbon storage in the study area is inverted and mapped.

[0055] Example 2

[0056] The difference from Example 1 is that

[0057] In step 1, the ground sample survey content includes three categories:

[0058] 1) Stand factors, i.e. dominant tree species, origin, and canopy density;

[0059] 2) Tree factors, i.e. tree number, standing tree type, tree species, diameter at breast height, and tree height;

[0060] 3) Other information, i.e. plot number, location coordinates, investigator, and survey date.

[0061] In step 2, forest carbon stock calculations use the tree allometric equation and the percentage of carbon content in biomass;

[0062] The tree allometric growth equation is AGB=aD b H c , where a, b, c are equation parameters, AGB is forest aboveground biomass, D is breast height, and H is tree height;

[0063] When the tree species is larch, a=0.06848, b=2.01549, c=0.59146; carbon content 47.42%;

[0064] When the tree species is red pine, a=0.027847, b=1.810004, c=0.905002; carbon content 48.09%;

[0065] When the tree species is oak, a=0.06149, b=2.14380, c=0.58390; carbon content 48.02%;

[0066] When the tree species is birch, a=0.06807, b=2.10850, c=0.52019; the carbon content is 45.56%.

[0067] Step 3: Preprocessing of remote sensing images includes radiation calibration, atmospheric correction, image mosaicking, and image cropping; preprocessing of point cloud data includes flight strip adjustment, data stitching, noise filtering, point cloud classification, projection conversion, and data cropping.

[0068] In step 4, the extracted remote sensing variables are subjected to Person correlation analysis and significance analysis with the forest carbon storage data calculated in step 2, and remote sensing variables with correlation coefficient r ≥ 0.3 and significance p ≤ 0.01 are selected as candidate variables for modeling.

[0069] In step 5, 70% of the forest carbon storage data calculated in step 2 is selected as the modeling sample, and the remaining 30% is selected as the test sample. The carbon storage remote sensing estimation model is trained using the random forest algorithm, and the model is optimized by adjusting the model parameters. The model parameters include the random seed number seed, the number of trees in the forest ntree, and the number of features used for each tree mtry.

[0070] In step 6, the coefficient of determination R 2 RMSE, MAE, and the determination coefficient R 2 The larger the value is, the smaller the root mean square error RMSE and relative root mean square error MAE are, indicating that the model accuracy is higher. The specific calculation methods of the three indicators are as follows:

[0071] (1)

[0072] (2)

[0073] (3)

[0074] Where: n is the number of sample plots; i Number the plots; x i For sample plot i Measured value of stand volume; For sample plot i The stand volume model estimate of It is the average value of stand volume measured in all sample plots.

[0075] Example 3

[0076] (1) Data acquisition: The ground sample plot data used in this example is located in the Northeast Forest Region, involving Jilin Province, Liaoning Province, and Inner Mongolia Autonomous Region. A circular ground sample plot with a radius of 13.82 m (such as Figure 2A total of 798 sample plots were collected (as shown in the table above) (including 201 Korean pine sample plots, 199 larch sample plots, 196 oak sample plots, and 202 birch sample plots). The sample plots were surveyed from September to November 2019. The survey content mainly included: 1) stand factors, namely dominant tree species (groups), origin, canopy density, etc.; 2) tree factors, namely sample tree number, standing tree type, tree species, breast diameter, tree height, etc.; 3) other information, namely sample plot number, location coordinates, investigator, survey date, etc. The forest carbon stock of each sample plot was calculated using the tree allometric growth equation and stand biomass carbon measurement model. The allometric growth equations and average carbon content of the four dominant tree species in the stand are shown in Table 1. The aerial photography time of LiDAR point cloud data is from September to October 2019; download Sentinel-2 Level-1C data products with good image quality that can cover the study area from August to November 2019.

[0077] Table 1 Allometric growth equation and average carbon content used in this method

[0078]

[0079] (2) Data preprocessing: Sentinel-2 Level-1C data products have undergone orthorectification and geometric correction. They also need to be atmospherically corrected using the sen2cor plug-in in the SNAP software to obtain Sentinel-2 Level-2A data, and then undergo image mosaicking, cropping and other processing procedures. The software used to preprocess LiDAR point cloud data is the TerraScan module in Micorstation and the domestic LiDAR360 software, which includes flight strip adjustment, data stitching, noise filtering, point cloud classification, projection conversion and data cropping.

[0080] (3) Remote sensing variable extraction: Based on multi-source remote sensing data, spectral information, texture information, point cloud height variables, density variables, and intensity variables are extracted, totaling 61 remote sensing variables. See Table 2 for details.

[0081] Table 2 Extracted information of various remote sensing variables

[0082]

[0083] (4) Variable screening: Person correlation analysis and significance analysis (t-test) were performed on the extracted remote sensing variables and forest carbon storage data, and remote sensing variables with correlation coefficient r ≥ 0.3 and significance p ≤ 0.01 were selected as candidate variables for modeling. The results of remote sensing variable screening for the four forest stands are shown in Table 3.

[0084] Table 3 Modeling variable screening results

[0085]

[0086] Note: P<0.01 ** Indicates significant difference at the 0.01 level (two-sided); P < 0.001 *** Indicates significant difference at the 0.001 level (two-sided)

[0087] (5) Model training: Based on the modeling samples, the random forest algorithm was used to train the carbon storage remote sensing estimation model. The remote sensing estimation model of forest carbon storage for the four forest stands was optimized by adjusting the model parameters. The optimization results are shown in Table 5.

[0088] Table 5 Model optimization results of remote sensing estimation models for different types of forest carbon stocks

[0089]

[0090] (6) Accuracy verification: based on the coefficient of determination R 2 The accuracy of the four trained remote sensing estimation models for forest carbon storage in different forest stands was verified and evaluated using three indicators: root mean square error (RMSE) and relative root mean square error (MAE). The evaluation results are shown in Table 6.

[0091] Table 6 Accuracy evaluation results of remote sensing estimation models for different types of forest carbon stocks

[0092]

[0093] (7) Forest carbon storage inversion mapping: The comparison results of the predicted values ​​and measured values ​​of carbon storage of different types of forests in the study area are shown in Figures 3(a) to 3(d). It can be seen from Figures 3(a) to 3(d) that the method of the present invention can accurately estimate the carbon storage of different types of forests and is applicable to a variety of different types of forests. It has good practicality and the carbon storage estimation accuracy (R 2 are all greater than 0.7) and are significantly better than the existing remote sensing model simulation methods.

Claims

1. A method for estimating forest carbon storage based on multi-source remote sensing data and random forest method, characterized in that: The specific steps are as follows: Step 1: Set up a ground survey plot in the study area, obtain survey data for each tree in the plot, and obtain optical remote sensing image data and lidar point cloud data of the area close to the survey time; In step 1, the ground sample survey content includes three categories: (1) Stand factors, i.e. dominant tree species, stand origin, and canopy density; (2) Tree factors, i.e. tree number, standing tree type, tree species, diameter at breast height, and tree height; (3) Other information, i.e., plot number, location coordinates, investigator, and survey date; Step 2: Based on the tree survey data obtained in step 1, the forest carbon stock of each sample plot is calculated as a data set; The forest carbon stock calculation uses the tree allometric growth equation and the percentage of carbon content in biomass; The tree allometric growth equation is AGB=aD b H c , where a, b, c are equation parameters, AGB is forest aboveground biomass, D is breast height, and H is tree height; When the tree species is larch, a=0.06848, b=2.01549, c=0.59146; carbon content 47.42%; When the tree species is red pine, a=0.027847, b=1.810004, c=0.905002; carbon content 48.09%; When the tree species is oak, a=0.06149, b=2.14380, c=0.58390; carbon content 48.02%; When the tree species is birch, a=0.06807, b=2.10850, c=0.52019; Carbon content 45.56%; Step 3, pre-processing the optical remote sensing image and lidar point cloud data obtained in step 1 respectively, and finally unifying the coordinate system of the remote sensing image data and the point cloud data into CGCS2000; The preprocessing of remote sensing images includes radiometric calibration, atmospheric correction, image mosaicking, and image cropping; the preprocessing of point cloud data includes flight strip adjustment, data splicing, noise filtering, point cloud classification, projection conversion, and data cropping; Step 4: Based on the preprocessed optical remote sensing image and the coordinates of the center point of the sample plot, two types of remote sensing variables, namely, spectral information and texture information, are extracted; based on the preprocessed lidar point cloud data, three types of remote sensing variables, namely, height variable, density variable, and intensity variable, are extracted, and correlation and significance analysis is performed on the extracted remote sensing variables and the forest carbon stock calculated in step 2, and remote sensing variables with high correlation are selected to participate in the establishment of a remote sensing estimation model for forest carbon stock; In step 4, the extracted remote sensing variables are subjected to Person correlation analysis and significance analysis with the forest carbon storage data calculated in step 2, and remote sensing variables with correlation coefficient r ≥ 0.3 and significance p ≤ 0.01 are selected as candidate variables for modeling; Step 5, using the remote sensing variables screened in step 4 as independent variables and the forest carbon storage data calculated in step 2 as dependent variables, selecting modeling samples, using the random forest algorithm to establish a remote sensing estimation model for forest carbon storage, and optimizing the model by adjusting model parameters; 70% of the forest carbon storage data calculated in step 2 are selected as modeling samples, and the remaining 30% are selected as test samples. The carbon storage remote sensing estimation model is trained using the random forest algorithm, and the model is optimized by adjusting the model parameters, wherein the model parameters include the random seed number seed, the number of trees in the forest ntree, and the number of features used for each tree mtry; Step 6: Based on the test samples, the accuracy of the remote sensing estimation model of forest carbon stock obtained in step 5 is verified and evaluated; Step 7: Based on the optimal forest carbon storage remote sensing estimation model obtained in step 5, the forest carbon storage in the study area is inverted and mapped.

2. The method for estimating forest carbon storage based on multi-source remote sensing data and random forest method according to claim 1, characterized in that: In step 6, the coefficient of determination R 2 RMSE, MAE, and the determination coefficient R 2 The larger the value is, the smaller the root mean square error RMSE and relative root mean square error MAE are, indicating that the model accuracy is higher. The specific calculation methods of the three indicators are as follows: （1） （2） （3） Where: n is the number of sample plots; i Number the plots; x i For sample plot i Measured value of stand volume; For sample plot i The stand volume model estimate of It is the average value of stand volume measured in all sample plots.

Images