Forest carbon reserve estimation method of stacked integrated fusion model

Abstract

The invention discloses a forest carbon reserve estimation method of a stacked integrated fusion model, and belongs to the technical field of forest resource carbon reserve estimation in high-altitude areas. The method comprises the following steps: collecting remote sensing data of a target area, executing preprocessing operation, and extracting remote sensing characteristic variables of the target area; performing multi-collinear processing on the target area remote sensing feature variables and the target prediction parameters to generate a preliminary input data set of a machine learning model; a recursive feature elimination method is used for the preliminary input data set to obtain an optimal input data set; and constructing a stacking model by using a k-fold cross validation method to obtain a forest carbon storage amount prediction value of the target area, and performing precision verification. According to the method, feature selection is carried out on remote sensing feature variables, redundant features are effectively removed, the model complexity is reduced, and meanwhile key features are reserved; and through stacking the integrated model, the estimation precision of the forest carbon reserve is remarkably improved, and an effective reference can be provided for dynamic monitoring of forest resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forest resource carbon storage estimation in high altitude areas, and particularly relates to a forest carbon storage estimation method of a stacked integrated fusion model. BACKGROUND Forests are an important part of the terrestrial ecosystem, and carbon storage has an important influence on the carbon sequestration capacity of forests. There are currently five main ways to estimate forest carbon storage: micrometeorology, plot inventory, box method, mathematical model method, and remote sensing technology. In recent years, remote sensing technology can quickly, accurately and non-destructively estimate aboveground biomass and carbon storage, and has become one of the important technical means for accurately estimating the carbon storage of high mountain pines in Shangri-La City. Traditional statistical regression methods often need to meet certain assumptions, but the continuous observation and hierarchy of forest growth data cannot meet the above assumptions. Compared with traditional plot survey methods, machine learning algorithms such as support vector machines, back propagation neural networks, extreme gradient boosting and random forests have significant advantages in processing large-scale, multi-source remote sensing data, and have been gradually used in the estimation of forest, grassland and crop biomass.

[0002] However, the prediction of forest carbon storage is affected by scale, method, forest type and condition changes, and there is a large amount of uncertainty. At present, a single machine learning algorithm still lacks effectiveness and rationality, and there are technical problems of underestimation of high values, uncertainty and low work efficiency. Therefore, how to provide a method for accurately estimating forest carbon storage with high precision and simple operation to improve the precision of forest carbon storage estimation is a technical problem that needs to be solved. SUMMARY

[0003] To solve the above technical problems, the present application aims to provide a forest carbon storage estimation method of a stacked integrated fusion model, which aims to solve the technical problems of complex operation, underestimation of high values, uncertainty and low work efficiency of the current forest carbon storage estimation method.

[0004] To achieve the above technical problems, the specific steps are as follows: S1, collecting remote sensing data of a target area, performing preprocessing operations on the collected remote sensing data of the target area, and extracting remote sensing feature variables of the target area; The remote sensing data includes 3 scenes of Landsat8OLI remote sensing images and environmental factors; The step of performing preprocessing operations on the collected remote sensing data of the target area and extracting remote sensing feature variables of the target area includes: S1.1, performing preprocessing operations on the collected remote sensing data of the target area through ENVI5.6 software, including: Radiometric calibration: converting the digital quantization value (Digital Number, DN value) in the target area remote sensing data to a physical value Convert to radiance value; Atmospheric correction: atmospheric correction is performed using the FLAASH model; Terrain correction: terrain correction is performed using the extended tool Topographic Correction in ENVI 5.6 software; S1.2, mosaic the pre-processed 3 Landsat8OLI remote sensing images The mosaic method is: after splicing 3 Landsat8OLI remote sensing images, the target region is obtained by cutting according to the administrative boundary; S1.3, extracting target region remote sensing feature variables from the 3 Landsat8OLI remote sensing images of the target region after mosaicking; The extraction method is: through the feature extraction operation of ArcGIS 10.8 software, the band combination variable characteristics, texture information variable characteristics and vegetation index variable characteristics of the 3 Landsat8OLI remote sensing images of the target region after mosaicking are obtained.

[0005] S2, multiple collinearity processing is performed on the obtained target region remote sensing feature variables and target prediction parameters to generate a preliminary input data set of the machine learning model; The target prediction parameter is: the carbon storage of the target species in the sample plot of the target region; wherein, the target prediction parameter is calculated according to the forest resource second class survey data, and the expression is as follows: In the formula, Indicates the carbon storage of the target species, and in the embodiment, the target species is Pinus densata; Indicates the biomass of the target forest sample plot; Indicates the carbon content coefficient; Indicates the stand volume; Indicates the wood density of the target species; Indicates the biomass expansion factor; The multiple collinearity processing method is: the Pearson correlation processing is adopted to obtain the correlation coefficient between each remote sensing feature variable and the target prediction parameter in the target region, and the remote sensing feature variable with a correlation coefficient not higher than a threshold value is removed; in the present application, the threshold value is set to 0.9.

[0006] S3, using recursive feature elimination (RFE) on the preliminary input data set to obtain the optimal input data set; Recursive feature elimination (RFE) is a model-based feature selection method, which iteratively trains the model and eliminates the least important features, and finally retains the optimal feature subset; in this embodiment, a random forest is used as the base model; The step of obtaining the optimal input data set comprises: Initializing the base model: constructing the feature matrix with the number of sample plots in the target prediction parameter and the preliminary input data set , , wherein, represents the number of sample plots in the target prediction parameter, represents the preliminary input data set; Setting the target variable: taking the real carbon storage value in the sample plot as the target variable ; Set the number of reserved features; in this embodiment, 9 remote sensing feature variables are reserved; Calculate the feature importance score by iteration of the base model and perform score sorting, eliminate unimportant features, and the iteration termination condition is that the remaining features and the set reserved feature number are the same; Output the optimal feature subset.

[0007] S4, using k-fold cross-validation method to construct a stacked model, obtaining the forest carbon storage prediction value of the target area, and completing the construction of the forest carbon storage estimation method; The stacked model includes: adaptive boosting algorithm (AdaBoost), gradient boosting decision tree (CatBoost), random forest (RF), light gradient boosting machine (LightGBM) and linear regression model (MLR), and the optimal hyperparameter combination of each model is searched using the hyperparameter tuning method network; The construction process comprises: Taking the optimal input data set and the real carbon storage value in the sample plot as the target variable Construct a training and validation data set; Using adaptive boosting algorithm (AdaBoost), gradient boosting decision tree (CatBoost), random forest (RF), light gradient boosting machine (LightGBM) to construct the first layer model; Divide the training and validation set (TrainValiData x ) into k folds, use k-1 folds as the training set (TrainData x ), and the rest as the validation set (ValiData x ); Train adaptive boosting algorithm (AdaBoost), gradient boosting decision tree (CatBoost), random forest (RF) and light gradient boosting machine (LightGBM) using the training set; verify adaptive boosting algorithm (AdaBoost), gradient boosting decision tree (CatBoost), random forest (RF) and light gradient boosting machine (LightGBM) using the verification set, obtain a one-dimensional prediction sequence for each iteration, repeat k times, and traverse all combinations; Stack the k one-dimensional prediction sequences into a meta-feature matrix; Construct a second layer model using the meta-feature matrix and a linear regression model (MLR), input the meta-feature matrix into the linear regression model (MLR), and obtain the final result through the linear regression model (MLR).

[0008] S5, verify the accuracy of the obtained forest carbon storage prediction value of the target area.

[0009] Advantages of the present application The present application effectively eliminates redundant features, reduces model complexity, and retains key features by selecting features from remote sensing characteristic variables. In addition, the construction of the super parameter optimization framework significantly reduces the model calculation efficiency and time cost.

[0010] The present application significantly improves the estimation accuracy of forest carbon storage by stacking the integrated model, solves the overfitting problem in the single machine learning regression estimation model, avoids the estimation failure caused by improper selection of a single model by stacking the integrated model, provides a more stable and reliable solution for forest carbon storage estimation, reduces the uncertainty in the model construction process, and provides an effective reference for forest resource dynamic monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 The step flowchart of the present application is shown in the figure. Figure 2 The stacked model schematic diagram of the present application is shown in the figure. Figure 3 The target area carbon storage of the present application embodiment is shown in the figure. DETAILED DESCRIPTION

[0012] The present application will be further described in detail below in combination with specific embodiments.

[0013] As shown in the figure, a stacked integrated fusion model forest carbon storage estimation method comprises the following steps: Figure 1 S1, collect remote sensing data of the target area, perform preprocessing operation on the collected remote sensing data of the target area, and extract remote sensing characteristic variables of the target area. S1, collect remote sensing data of the target area, perform preprocessing operation on the collected remote sensing data of the target area, and extract remote sensing characteristic variables of the target area. In this embodiment, the target area is Shangri-La City, Yunnan Province; Remote sensing data includes three Landsat 8 OLI remote sensing images and environmental factors. The three Landsat 8 OLI remote sensing images are from the Geospatial Data Cloud, covering Shangri-La City, Yunnan Province, from November to December, with a spatial resolution of 30 meters and a cloud cover of less than 8%. The steps of performing preprocessing operations on the collected remote sensing data of the target area and extracting remote sensing characteristic variables of the target area include: S1.1. Perform pre-processing operations on the collected remote sensing data of the target area using ENVI5.6 software, including: Radiometric calibration: converting the digital quantization value (DN value) in the remote sensing data of the target area into Convert to radiance value; Atmospheric correction: The FLAASH model used is used for atmospheric correction; Terrain correction: The extended tool Topographic Correction in ENVI 5.6 software was used for terrain correction; S1.2. Mosaic the three pre-processed Landsat 8OLI remote sensing images The mosaicking method is as follows: three Landsat8OLI remote sensing images are stitched together and then cropped according to the administrative boundary to obtain three Landsat8OLI remote sensing images of the target area; S1.3. Extract remote sensing characteristic variables of the target area from the three Landsat 8OLI remote sensing images of the mosaicked target area; The extraction method is: performing feature extraction operations using ArcGIS 10.8 software to obtain band combination variable features, texture information variable features, and vegetation index variable features of the three Landsat 8OLI remote sensing images of the target area after mosaicking. In this embodiment, there are 162 variable features in total.

[0014] S2. Perform multicollinearity processing on the obtained remote sensing characteristic variables of the target area and the target prediction parameters to generate a preliminary input data set for the machine learning model; The target prediction parameter is the carbon storage of the target species in the sample plots of the target area. The target prediction parameter is calculated based on the second-class forest resource survey data and is expressed as follows: Where, represents the carbon storage of the target species. In this embodiment, the target species is alpine pine; represents the biomass of the target forest plot; represents the carbon content coefficient, and 0.5 is taken in the embodiment; represents the stand volume; represents the wood density of the target species; represents the biomass expansion factor; The multiple collinearity processing mode is as follows: Pearson correlation processing is used to obtain the correlation coefficient between each remote sensing characteristic variable in the target area and the target prediction parameter, and the remote sensing characteristic variable with a correlation coefficient not higher than a threshold value is removed; in the present application, the threshold value is set to 0.9, and the preliminary input data set has a total of 58.

[0015] S3, using recursive feature elimination (RFE) on the preliminary input data set to obtain an optimal input data set; Recursive feature elimination (RFE) is a model-based feature selection method, which iteratively trains a model and removes the least important features to finally retain an optimal feature subset; in the present embodiment, a random forest is used as the base model; The step of obtaining the optimal input data set comprises: Initializing the base model: constructing a feature matrix with the number of sample plots in the target prediction parameter and the preliminary input data set , , wherein, represents the number of sample plots in the target prediction parameter, represents the preliminary input data set; Setting the target variable: taking the real carbon storage value in the sample plot as the target variable ; Setting the number of retained features; in the present application, 9 remote sensing characteristic variables are retained; Iteratively calculating the feature importance score by the base model and performing score sorting, removing unimportant features, and the iteration termination condition is that the remaining features and the set number of retained features are the same; Outputting the optimal feature subset; In the present embodiment, the 9 remote sensing characteristic variables obtained are: precipitation (pre); air temperature (Temp); first texture factor (R11B6ENT), wherein R11 represents a window size of 11, B6 represents the first infrared short wave of the target area, and ENT means entropy; B347 band combination, B347 is obtained by , wherein, represents the green light band, B4 represents the red light band, and B7 represents the second infrared short wave of the target area; second texture factor (R7B6ME), wherein R7 represents a window size of 7, and ME means mean; vegetation index (ND67); third texture factor (R9B5ME), wherein R5 represents a window size of 5, and B5 represents near-infrared light; perpendicular vegetation index (PVI); and fourth texture factor R11B5ME.

[0016] S4, such as Figure 2 As shown in the figure, the stacking model was constructed using the k-fold cross-validation method to obtain the predicted value of forest carbon storage in the target area, thus completing the construction of the forest carbon storage estimation method. In this embodiment, a 5-fold cross-validation method is used, and the stacked models include: adaptive boosting algorithm (AdaBoost), gradient boosting decision tree (CatBoost), random forest (RF), lightweight gradient boosting machine (LightGBM) and linear regression model (MLR), and the hyperparameter tuning method network search is used to determine the optimal hyperparameter combination for each model; The build process includes: The optimal input data set and the actual carbon storage value in the sample plot are used as the target variable Construct training and validation datasets; The first-layer model is constructed using the adaptive boosting algorithm (AdaBoost), gradient boosting decision tree (CatBoost), random forest (RF), and lightweight gradient boosting machine (LightGBM); The training validation set (TrainValiData x ) is divided into k-fold ( x =1, 2, 3, 4, 5), using k-1 folds as the training set (Train Data x , x =1, 2, 3, 4, 5), and the rest are used as validation sets (ValiData x , x =1,2,3,4,5); Use the training set to train the adaptive boosting algorithm (AdaBoost), gradient boosting decision tree (CatBoost), random forest (RF), and lightweight gradient boosting machine (LightGBM); use the validation set to verify the adaptive boosting algorithm (AdaBoost), gradient boosting decision tree (CatBoost), random forest (RF), and lightweight gradient boosting machine (LightGBM). Each iteration obtains a one-dimensional prediction sequence, and repeats the iteration k times to traverse all combinations; Stack k one-dimensional prediction sequences into a meta-feature matrix; The second-layer model is constructed with the meta-feature matrix and the linear regression model (MLR). The meta-feature matrix is ​​input into the linear regression model (MLR), and the final result is obtained through the linear regression model (MLR).

[0017] S5. Verify the accuracy of the predicted forest carbon storage value in the target area; Precision analysis indicators include: determination coefficient , root mean square error (RMSE), mean absolute error (MAE), relative root mean square error (RRMSE) and precision percentage (p%); the verification results are shown in Table 1; Table 1: Results of model performance evaluation in each algorithm From Table 1, it can be seen that the determination coefficient value of the present application is the highest, indicating that the fitting ability of the present application to the change of carbon storage is the strongest; the root mean square error is the smallest, and the accuracy is better than other single models; the mean absolute error is the lowest, which reflects that the stability of the model constructed by the present application is good; the relative root mean square error is the lowest, indicating that the proportion of the error of the present application to the mean value of the true value is the smallest; the precision percentage is the highest, indicating that the model constructed by the present application has the best effect compared with single model, which reflects that the present application has significant improvement in estimation accuracy.

[0018] Finally, the stacked ensemble fusion model is run in the Python 3.9 code program to carry out carbon storage inversion, and the carbon storage spatial distribution mapping of the research area is shown in Figure 3 .

[0019] In the present application, although the present application is described based on specific embodiments, it should be clear that these embodiments are only examples for embodying the principles and applications of the present application. Therefore, the exemplary embodiments can be modified, and other layouts can also be designed, provided that the spirit and scope of the present application defined by the appended claims are not violated. Different dependent claims and features described herein can be combined in a manner different from the original claim description. In addition, the features described in conjunction with a single embodiment can be applied to other described embodiments.

Claims

1. A forest carbon storage estimation method based on a stacked integrated fusion model, characterized in that: The following steps are involved: S1, collecting remote sensing data of the target area, performing preprocessing operations on the collected remote sensing data of the target area, and extracting remote sensing characteristic variables of the target area; S2. Perform multicollinearity processing on the obtained remote sensing characteristic variables of the target area and the target prediction parameters to generate a preliminary input data set for the machine learning model; The target prediction parameters are: carbon storage of target species in sample plots in the target area; S3, using recursive feature elimination on the preliminary input data set to obtain the optimal input data set; S4. Use the k-fold cross-validation method to construct a stacking model to obtain the predicted value of forest carbon storage in the target area, completing the construction of the forest carbon storage estimation method; The stacking models include: adaptive boosting algorithm, gradient boosting decision tree, random forest, lightweight gradient boosting machine and linear regression model; S5. Verify the accuracy of the predicted forest carbon storage value in the target area.

2. The method for estimating forest carbon stock using a stacked integrated fusion model according to claim 1, characterized in that: The steps of collecting remote sensing data of the target area, performing preprocessing operations on the collected remote sensing data of the target area, and extracting remote sensing characteristic variables of the target area include: S1.

1. Perform preprocessing operations on the collected remote sensing data of the target area using ENVI5.6 software, including: radiometric calibration, atmospheric correction, and terrain correction; S1.

2. Mosaicing the pre-processed remote sensing images The mosaicking method is: after stitching the remote sensing images, crop them according to the administrative boundaries to obtain the remote sensing images of the target area; S1.

3. Extract remote sensing characteristic variables of the target area from the mosaicked remote sensing image of the target area.

3. The method for estimating forest carbon stock using a stacked integrated fusion model according to claim 1, characterized in that: The obtained target area remote sensing characteristic variables and target prediction parameters are subjected to multicollinearity processing to generate a preliminary input data set for the machine learning model. The expression of the target prediction parameter is as follows: Where, represents the carbon storage of the target species. In this embodiment, the target species is alpine pine; represents the biomass of the target forest plot; Indicates carbon content coefficient; represents stand stock; represents the wood density of the target species; represents the biomass expansion factor; The method of multicollinearity processing is: using Pearson correlation processing to obtain the correlation coefficient between each remote sensing characteristic variable and the target prediction parameter in the target area, and eliminating the remote sensing characteristic variables with correlation coefficients not higher than the threshold.

4. The method for estimating forest carbon stock using a stacked integrated fusion model according to claim 1, wherein: The method of using recursive feature elimination to obtain the optimal input data set is as follows: Initialize the base model: construct a feature matrix based on the number of plots in the target prediction parameter and the initial input data set , ,in, represents the number of sample plots in the target prediction parameter, represents the preliminary input dataset; Set the target variable: take the actual carbon storage value in the sample plot as the target variable ; Setting the number of retained features; the number of retained features is a preset value; The feature importance scores are calculated and ranked by iterative calculation through the base model, and unimportant features are eliminated. The iteration termination condition is that the number of remaining features is the same as the number of retained features set; Output the optimal feature subset.

5. The method for estimating forest carbon stock using a stacked integrated fusion model according to claim 1, characterized in that: The method for constructing a stacking model using the k-fold cross-validation method to obtain the predicted value of forest carbon storage in the target area and complete the construction of the forest carbon storage estimation method is as follows: The optimal input data set and the actual carbon storage value in the sample plot are used as the target variable Construct training and validation datasets; The first layer model is constructed using the adaptive boosting algorithm, gradient boosting decision tree, random forest, and lightweight gradient boosting machine; Divide the training validation set into k folds, use k-1 folds as the training set, and the rest as the validation set; Use the training set to train the adaptive boosting algorithm, gradient boosting decision tree, random forest, and lightweight gradient boosting machine; Use the validation set to verify the adaptive boosting algorithm, gradient boosting decision tree, random forest, and lightweight gradient boosting machine. Each iteration obtains a one-dimensional prediction sequence. Repeat the iteration k times to traverse all combinations. Stack k one-dimensional prediction sequences into a meta-feature matrix; The second-layer model is constructed with the meta-feature matrix and the linear regression model. The meta-feature matrix is ​​input into the linear regression model, and the final result is obtained through the linear regression model.

Images