Tree maximum crown breadth height simulation prediction method based on ensemble learning
By constructing a coupled model of height at maximum crown width and spatial competition using an ensemble learning method, the problem of obtaining three-dimensional phenotypic parameters of trees is solved, and high-precision simulation and prediction of tree height at maximum crown width is achieved, meeting the needs of forestry applications.
Patent Information
- Application Number
- CN202411025068.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing phenotypic acquisition methods, such as lidar point cloud data, are insufficient to accurately obtain height parameters at the maximum crown width between tree canopies, especially in forests with high stand density, making it difficult to obtain three-dimensional phenotypic parameters of trees.
An ensemble learning-based approach was adopted, combining machine learning algorithms and ensemble learning methods to construct a coupled model of height competition between the maximum crown width and space. Through data collection, base model selection, optimization and fitting, ensemble model construction and result comparison, the optimal model was selected for simulation prediction.
It enables high-precision acquisition of height parameters at the maximum crown width in forests with high stand density, solving the problem of difficulty in obtaining three-dimensional phenotypic parameters of trees and meeting the needs of related research and forestry applications.
Smart Images

Figure CN118917445A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of three-dimensional phenotypic parameter fitting, and relates to a method for simulating and predicting the height of a tree at a maximum crown width based on ensemble learning. Background Art
[0002] The height at the maximum crown width is the height corresponding to the crown width position of the tree crown. For some coniferous trees, the crown can be divided into two parts, upper and lower. It is an important three-dimensional phenotypic parameter of trees. In the forest, due to the possible overlap between the crowns, some existing phenotypic collection methods, such as lidar point cloud data, cannot easily extract the phenotypic parameters of the crown such as the maximum crown width height parameter. Ensemble learning can integrate several base models and perform high-precision regression fitting on the data. As a height parameter located on the crown, there is a direct relationship between the maximum crown width height and the tree height and the spatial competition caused by the surrounding trees. Constructing a coupling model between the maximum crown width height and spatial competition is an important direction of related research. Summary of the invention
[0003] In view of the problems in the prior art, the present invention provides a method for simulating and predicting the height of a tree at its maximum crown width based on ensemble learning.
[0004] A method for simulating and predicting the height of a tree at its maximum crown width based on ensemble learning includes the following steps:
[0005] Step 1: Collect the data required for model construction, including the location (X, Y, Z), tree height, diameter at breast height, height of the maximum crown width in a single direction (east, west, south, north), and crown width in a single direction (east, west, south, north). Calculate the vertical spatial structure parameters and horizontal spatial structure parameters in a single direction (east, west, south, north).
[0006] Step 2: Base model selection: select 11 machine learning models from three categories as base models according to the commonly used regression fitting models.
[0007] (1) Tree-based models: decision tree, random forest, gradient boosting, and adaptive boosting.
[0008] (2) Gradient boosting tree variant models: extreme gradient boosting, lightweight gradient boosting machine, gradient boosting decision tree.
[0009] (3) Linear models: support vector machine, ridge regression, lasso algorithm, elastic net regression.
[0010] Step 3: Base model optimization and fitting. In order to improve the simulation accuracy, grid search cross-validation is introduced into the base model, hyperparameter tuning is performed, and data cleaning and feature scaling operations are performed to improve the model effect. Evaluation indicators are calculated.
[0011] Step 4: ensemble model construction and fitting. Select the five commonly used ensemble learning methods: bagging, boosting, voting, stacking, and mixing. Combine the 11 base models in a permutation and combination manner. Perform ensemble training on each combination for each ensemble learning method and calculate the evaluation index.
[0012] Step 5: Result comparison and optimal model selection: compare the fitting results of all base models and ensemble learning combination models, compare the evaluation indicators, and select the optimal fitting model.
[0013] The advantages of the present invention are: the implementation method is simple and effective, and the height of the maximum crown width in different directions of east, west, south and north is simulated by means of machine learning algorithms and integrated learning methods, which solves the problem of difficulty in obtaining three-dimensional phenotypic parameters of trees, and realizes the simulation and prediction of phenotypic parameters that are difficult to obtain by using high-precision phenotypic parameters such as tree height and crown width that are easily collected from trees, combined with environmental factors such as spatial competition as independent variables. Compared with lidar extraction, in forests with high stand density, the regression prediction method is easier to achieve the acquisition of the height parameters at the maximum crown width. And through the combination of different machine learning algorithms with different integrated learning methods, the model combination that is most suitable for simulating the height parameters at the maximum crown width is verified, and the accuracy of the simulation results can meet the needs of related research and forestry applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] 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 or the prior art descriptions. 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 creative work. As shown in the figure:
[0015] Figure 1 This is a flow chart of the tree maximum crown height simulation prediction technology based on ensemble learning of the present invention.
[0016] Figure 2 A schematic diagram of the phenotypic parameters and positions measured for each tree of the present invention (taking the east and west directions as an example).
[0017] Figure 3 (a) Random forest model training learning curve.
[0018] Figure 3(b) Learning curve of lightweight gradient boosting machine model training.
[0019] Figure 3(c) Support vector machine model training learning curve.
[0020] Figure 3(d) Lasso algorithm model training learning curve.
[0021] Fig. 3(e) Learning curve of elastic net regression model training.
[0022] Figure 3(f) Learning curve of adaptive enhancement model training.
[0023] Fig. 3 (g) Ridge regression model training learning curve.
[0024] Figure 3(h) Learning curve of extreme gradient boosting model training.
[0025] Figure 3 (i) Gradient boosting decision tree model training learning curve.
[0026] Figure 3(j) Gradient boosting model training learning curve.
[0027] Figure 3 (k) Decision tree model training learning curve.
[0028] Fig. 4 (a) Learning curve of combined model training with the first best fitting accuracy.
[0029] Fig. 4(b) Learning curve of training of the combined model with the second best fitting accuracy.
[0030] Fig. 4(c) is the training learning curve of the combined model with the third highest fitting accuracy.
[0031] Fig. 4(d) Learning curve of combined model training with the 4th highest fitting accuracy.
[0032] Fig. 4(e) Learning curve of combined model training with the fifth highest fitting accuracy.
[0033] Fig. 4(f) Learning curve of combined model training with the 6th highest fitting accuracy.
[0034] Fig. 4(g) Learning curve of combined model training with the 7th highest fitting accuracy.
[0035] Fig. 4(h) Learning curve of combined model training with the 8th highest fitting accuracy.
[0036] Fig. 4 (i) Learning curve of combined model training with the 9th highest fitting accuracy.
[0037] Fig. 4(j) Learning curve of combined model training with the 10th best fitting accuracy.
[0038] Figure 5 This is a distribution diagram of the true value predicted value of the random forest model of the present invention.
[0039] Figure 6 This is a distribution diagram of the true value predicted value of the combined model with the highest fitting accuracy of the present invention. DETAILED DESCRIPTION
[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.
[0041] Example 1: Figure 1 , Figure 2 , Figure 3(a), Figure 3(b), Figure 3(c), Figure 3(d), Figure 3(e), Figure 3(f), Figure 3(j), Figure 3(h), Figure 3(i), Figure 3(j), Figure 3(k), Figure 4(a), Figure 4(b), Figure 4(c), Figure 4(d), Figure 4(e), Figure 4(f), Figure 4(j), Figure 4(h), Figure 4(i), Figure 4(j), Figure 5 and Figure 6 As shown, a method for simulating and predicting the height at the maximum crown width of a tree based on ensemble learning mainly involves the simulation technology of the height at the maximum crown width of the tree in different directions.
[0042] By coupling the environmental and phenotypic parameters through an ensemble learning model, a simulation and prediction method for the tree maximum crown height based on ensemble learning was proposed.
[0043] A simulation and prediction method for the maximum crown height of trees based on ensemble learning constructs a coupling model of the maximum crown height and spatial competition, which makes up for the difficulty in obtaining the maximum crown height parameters and provides technical support for the three-dimensional phenotypic research of trees.
[0044] Including the construction of the coupling model of the maximum crown height and space competition, the overall technical process is as follows Figure 1 shown.
[0045] The specific steps are as follows:
[0046] Step 1: Data collection preparation.
[0047] Due to the different effects of spatial competition in different directions on the three-dimensional phenotype of trees, data collection is different from traditional forestry surveys. Four directions are divided: east, west, south, and north. For each direction, the height of the maximum crown width and the tree height are measured by a height meter. The crown width in each direction is measured by a tower ruler. The diameter at breast height of each tree is measured by a circumference ruler, and the position coordinates (X, Y, Z) of each tree are collected using RTK. The vertical and horizontal spatial structure parameters of each tree in the east, west, south, and north directions are calculated and processed through the collected data as the eigenvalues required by the model.
[0048] Step 2: Selection and optimization fitting of the height and space competition coupling basis model at the maximum crown width.
[0049] Many algorithms in machine learning are suitable for regression fitting of data. By screening the research literature on regression fitting using machine learning, 11 machine learning models were selected as base models according to the commonly used models for regression fitting. According to the model characteristics, they can be divided into three categories, including (1) tree-based models: decision tree, random forest, gradient boosting, and adaptive boosting. (2) Gradient boosting tree variant models: extreme gradient boosting, lightweight gradient boosting machine, gradient boosting decision tree. (3) Linear models: support vector machine, ridge regression, lasso algorithm, elastic network regression.
[0050] In order to improve the fitting accuracy, grid search cross-validation is introduced for hyperparameter tuning of all base models, and combined with data cleaning, feature scaling and other operations to improve the model effect.
[0051] Table 1 Parameter settings for hyperparameter tuning of base models
[0052]
[0053] Through operations such as hyperparameter optimization, data cleaning, and feature scaling, the noise of the data is reduced, the overfitting phenomenon in the model is reduced, and the fitting accuracy of the height at the maximum crown width is improved.
[0054] Step 3: Construct and fit the integrated model of height and spatial competition coupling at the maximum crown width.
[0055] In order to further improve the fitting accuracy of the model, an ensemble learning method is introduced to integrate multiple base models to obtain a strong learner and obtain a better fitting effect. Five commonly used ensemble learning methods are selected: bagging, boosting, voting, stacking, and mixing.
[0056] The 11 base models are combined in the form of permutations and combinations, starting from two-by-two combinations and gradually increasing the number until all 11 base models become ensemble learning combinations. Each ensemble learning method performs ensemble training on each combination.
[0057] The expression of the bagging model is:
[0058]
[0059] Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and Pr n (x) is the prediction result of the nth base model.
[0060] The expression of the lifting model is:
[0061]
[0062] Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and λ n is the weight of the nth model. n (x) is the prediction result of the nth base model.
[0063] The expression of the voting model is:
[0064]
[0065] Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and Pr n (x) is the prediction result of the nth base model.
[0066] Its expression is the same as that of the bagging model, but the difference is that the bagging model randomly generates multiple sample subsets in the training set during the training phase and trains them on the base models respectively. However, all base models of the voting model are trained on the entire training set.
[0067] The expression of the stacking model is:
[0068] Pr(x)=Meta(h1(x),h2(x),…,h N (x)) (4)
[0069] Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and h N (x) is the prediction result of the nth base model, and Meta is the meta-model, usually a linear regression or other regression model.
[0070] The expression of the mixed model is:
[0071] Pr(x)=Meta(h1(x),h2(x),…,h N (x)) (5)
[0072] Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and h N (x) is the prediction result of the nth base model, and Meta is the meta-model, usually a linear regression or other regression model.
[0073] The formulation is the same as for the stacked model, except that a fixed validation set is used when splitting the training data instead of using cross-validation.
[0074] The collected and processed parameters are provided as characteristic values to all models, and the height at the maximum crown width is used as the output value to train the coupling model between the height at the maximum crown width and spatial competition.
[0075] Step 4: Compare results and select the optimal model.
[0076] Through the previous steps 2 and 3, the fitting results of the single model and the combined model and the accuracy evaluation index of each model were obtained, and 6 indicators were selected: the coefficient of determination (R 2 ), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), explained variance score (EVS), and median absolute error (Med ian-AE) were used as evaluation criteria. 2 The RMSE and RMSE were used as the main evaluation indicators, and the others were used as auxiliary reference indicators to compare all the evaluation results. The model with the best accuracy was selected as the fitting model of the maximum crown height and applied to the simulation prediction of the maximum crown height.
[0077] This completes the entire process of simulating and predicting the height at the maximum crown width of trees based on ensemble learning.
[0078] Example 2: Figure 1 , Figure 2 , Figure 3(a), Figure 3(b), Figure 3(c), Figure 3(d), Figure 3(e), Figure 3(f), Figure 3(j), Figure 3(h), Figure 3(i), Figure 3(j), Figure 3(k), Figure 4(a), Figure 4(b), Figure 4(c), Figure 4(d), Figure 4(e), Figure 4(f), Figure 4(j), Figure 4(h), Figure 4(i), Figure 4(j), Figure 5 and Figure 6 As shown in the figure, a method for simulating and predicting the height at the maximum crown width of trees based on ensemble learning is used, and an example of simulating and predicting the height at the maximum crown width of Chinese fir trees in the Shanxia Forest Farm of Jiangxi Yalin Center is provided.
[0079] 1. Data Collection and Processing
[0080] We selected 27 plots of trees in the Yalin Center of Jiangxi Province, with a total of 580 Chinese fir trees, and measured the position coordinates (X, Y, Z) of each tree in the plot. We also measured the diameter at breast height, tree height, crown width in four directions, and the height at the maximum crown width in four directions, and obtained 2,320 sets of data. We also calculated the vertical and horizontal spatial structure parameters in each direction.
[0081] 2. Simulation prediction of the maximum crown height
[0082] Using 11 base models that have undergone hyperparameter optimization, data cleaning, feature scaling, etc., the data is divided into training set and test set in a ratio of 8:2 to train the model. The learning curves of the training process are shown in Figure 3(a), Figure 3(b), Figure 3(c), Figure 3(d), Figure 3(e), Figure 3(f), Figure 3(j), Figure 3(h), Figure 3(i), Figure 3(j) and Figure 3(k) (11 base model training learning curves).
[0083] The evaluation indicators of all models are shown in Table 2:
[0084] Table 2 Basic model evaluation indicators
[0085]
[0086] By combining models through ensemble learning, each ensemble learning method corresponds to 2046 combination models, and a total of 10180 combination models are constructed. The 10 models with the highest fitting accuracy are selected. The top 10 models are sorted from high to low in terms of accuracy, as shown in Table 3.
[0087] Table 3. The top 10 ensemble learning combination models with the best fitting accuracy
[0088]
[0089] The learning curves of the model are shown in Figures 4(a), 4(b), 4(c), 4(d), 4(e), 4(f), 4(j), 4(h), 4(i) and 4(j) (training learning curves of the top 10 combinations of fitting accuracy).
[0090] The evaluation indicators of these 10 models are shown in Table 4.
[0091] Table 4 Evaluation indicators of the top 10 combined models for fitting accuracy
[0092]
[0093] Comparing the fitting accuracy of 11 base models and 10 combined models, the random forest has the highest accuracy among the base models. 2 The results reached 0.8186, RMSE was 1.4708, MAE was 0.9043, MAPE was 0.0498, EVS was 0.8226, and MedAE was 0.4370. The bagging ensemble with the highest accuracy among the combined models - extreme gradient boosting + random forest + support vector machine + gradient boosting + ridge regression, achieved the highest accuracy in the test set R 2 The main evaluation index R 2 The RMSE and EVS evaluation are higher than those of the base model, and the MAE, MAPE and MedAE are slightly worse. Taking all factors into consideration, the model quality has been improved, and the simulation and prediction of the height at the maximum crown width can be more accurate.
[0094] The training results constructed a coupling model of height and spatial competition at the maximum crown width. The introduction of ensemble learning improved the accuracy of model fitting. The results can meet the technical and application needs of forest phenotypic parameter acquisition, three-dimensional phenotypic research, and phenotypic-environment interaction research.
[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for simulating and predicting the height of a tree at its maximum crown width based on ensemble learning, characterized in that: Contains the following steps: Step 1: Collect the data required for model construction, including the location (X, Y, Z), tree height, breast diameter, height of the maximum crown width in a single direction (east, west, south, north), and crown width in a single direction (east, west, south, north). Calculate the vertical spatial structure parameters and horizontal spatial structure parameters in a single direction (east, west, south, north). Step 2: Base model selection: select 11 machine learning models from 3 categories as base models according to the commonly used regression fitting models. (1) Tree-based models: decision tree, random forest, gradient boosting, adaptive boosting, (2) Gradient boosting tree variant models: extreme gradient boosting, lightweight gradient boosting machine, gradient boosting decision tree, (3) Linear models: support vector machine, ridge regression, lasso algorithm, elastic network regression, Step 3: Base model optimization and fitting. In order to improve the simulation accuracy, grid search cross-validation is introduced into the base model, hyperparameter tuning is performed, and data cleaning and feature scaling operations are performed to improve the model effect and calculate the evaluation indicators. Step 4: ensemble model construction and fitting. Select the five commonly used ensemble learning methods: bagging, boosting, voting, stacking and mixing. Combine the 11 base models in a permutation and combination manner. Perform ensemble training on each combination for each ensemble learning method and calculate the evaluation index. Step 5: Result comparison and optimal model selection: compare the fitting results of all base models and ensemble learning combination models, compare the evaluation indicators, and select the optimal fitting model.
2. The method for simulating and predicting the height of a tree at its maximum crown width based on ensemble learning according to claim 1, characterized in that: Step 1: Data collection, including the following steps: Considering the different effects of spatial competition in different directions on the three-dimensional phenotype of trees, data collection is different from traditional forestry surveys. Four directions are divided: east, west, south and north. For each direction, the height of the maximum crown width and the tree height in that direction are measured by a height meter. The crown width in each direction is measured by a tower ruler, and the breast diameter of each tree is measured by a circumference ruler. RTK is used to collect the position coordinates (X, Y, Z) of each tree. The vertical spatial structure parameters and horizontal spatial structure parameters of each tree in the east, west, south and north directions are calculated and processed through the collected data as the characteristic values required by the model.
3. The method for simulating and predicting the height of a tree at its maximum crown width based on ensemble learning according to claim 1, characterized in that: The machine learning model in step 2 contains the following steps: selection and optimization fitting of the base model of the coupling of the maximum crown width and spatial competition. Many algorithms in machine learning are suitable for regression fitting of data. By screening the research literature on regression fitting using machine learning, 11 machine learning models are selected as base models according to the commonly used models for regression fitting. They are divided into three categories according to the model characteristics, including (1) tree-based models: decision tree, random forest, gradient boosting, and adaptive boosting; (2) gradient boosting tree variant models: extreme gradient boosting, lightweight gradient boosting machine, and gradient boosting decision tree; (3) linear models: support vector machine, ridge regression, lasso algorithm, and elastic network regression. In order to improve the fitting accuracy, grid search cross-validation is introduced for all base models to perform hyperparameter tuning, and combined with data cleaning and feature scaling operations to improve the model effect. Through hyperparameter optimization, data cleaning, feature scaling and other operations, the noise of the data is reduced, the overfitting phenomenon in the model is reduced, and the fitting accuracy of the maximum crown width is improved.
4. The method for simulating and predicting the height of a tree at its maximum crown width based on ensemble learning according to claim 3, characterized in that: The construction and fitting of the integrated model of the coupling of the maximum crown height and spatial competition also includes the following steps: further improving the fitting accuracy of the model, introducing an integrated learning method, integrating multiple base models, obtaining a strong learner, and obtaining a better fitting effect, and selecting the five commonly used integrated learning methods: bagging, boosting, voting, stacking, and mixing. The 11 base models are combined in a permutation and combination manner, starting with two-by-two combinations and gradually increasing the number until all 11 base models become ensemble learning combinations. Each ensemble learning method performs ensemble training on each combination. The expression of the bagging model is: Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and Pr n (x) is the prediction result of the nth base model, The expression of the lifting model is: Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and λ n is the weight of the nth model, Pr n (x) is the prediction result of the nth base model, The expression of the voting model is: Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and Pr n (x) is the prediction result of the nth base model, Its expression is the same as that of the bagging model. The difference is that the bagging model randomly generates multiple sample subsets in the training set during the training phase and trains them on the base models respectively, while all base models of the voting model are trained on the entire training set. The expression of the stacking model is: Pr(x)=Meta(h1(x),h2(x),…,h N (x)) (4) Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and h N (x) is the prediction result of the nth base model, Meta is the meta-model, usually a linear regression or other regression model, The expression of the mixed model is: Pr(x)=Meta(h1(x),h2(x),…,h N (x)) (5) Among them, Pr(x) is the prediction result of the integrated model, N is the number of base models, and h N (x) is the prediction result of the nth base model, Meta is the meta-model, usually a linear regression or other regression model, The expression is the same as the stacked model, except that a fixed validation set is used when splitting the training data instead of cross-validation. The collected and processed parameters are provided as characteristic values to all models, and the height at the maximum crown width is used as the output value to train the coupling model between the height at the maximum crown width and spatial competition.
5. The method for simulating and predicting the height of a tree at its maximum crown width based on ensemble learning according to claim 1, characterized in that: The result comparison and optimal model selection in step 5 also include the following steps: through claims 3 and 4, the fitting results of the single model and the combined model and the accuracy evaluation index of each model are obtained, and 6 indicators are selected: the determination coefficient (R 2 ), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), explained variance score (EVS), and median absolute error (Median-AE) were used as evaluation criteria, with R 2 And RMSE were used as the main evaluation indicators, and the others were used as auxiliary reference indicators. All evaluation results were compared, and the model with the best accuracy was selected as the fitting model for the height at the maximum crown width, which was applied to the simulation prediction of the height at the maximum crown width.
Citation Information
Patent Citations
Rail transit vehicle-mounted data prediction method based on ensemble learning model
CN111612260A
Vibration prediction method for hot continuous rolling mill
CN116796646A
Office building rent prediction method and system based on Stacking ensemble learning algorithm
CN117635280A
Method for simulating heights of maximum crown breadths of tree three-dimensional model in different directions
CN118097009A
Porosity prediction method based on selective ensemble learning
US20230203925A1
Cited By
Machine learning and amino acid sequence information-based umami peptide recognition threshold prediction method
CN119479908A
