Needle-leaved arbor biomass estimation method and system based on PSO-BP neural network and SHAP interpreter

Abstract

The invention relates to the technical field of forest resource investigation and monitoring, in particular to a needlebush biomass estimation method and system based on a PSO-BP neural network and an SHAP interpreter. Comprising the following steps: step 1, collecting big data of the coniferous arbors, the big data comprising forest basic data and environment variable data; 2, preprocessing the big data to obtain significant correlation factors; 3, training the significant correlation factors through a PSO-BP neural network to construct a biomass model for estimating the biomass of the coniferous arbors; and 4, verifying the precision of the index evaluation model and analyzing the contribution degree of each factor based on a machine learning explanation tool SHAP interpreter. According to the method, the actually measured data and factor data of the single-tree biomass of the cunninghamia lanceolata in different growth areas are collected and sorted by utilizing published research literatures, the non-parametric model of the cunninghamia lanceolata is reconstructed under the condition of considering multiple factor variables, and the precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of forest resource investigation and monitoring, and particularly to a method and system for estimating the biomass of coniferous trees based on a PSO-BP neural network and a SHAP interpreter. Background Art

[0002] As the main body of the global carbon cycle, the carbon sequestration potential of forest ecosystems has attracted much attention. Forest biomass data is the prerequisite for understanding forest carbon sequestration. Coupled with the vast territory of our country, it is crucial to construct accurate biomass models under the action of different environmental factors.

[0003] Cunninghamia lanceolata is a tall tree of the genus Cunninghamia in the family Cupressaceae, with lanceolate or narrow leaves. As one of the most important artificial forest ecosystems in the subtropical region of China, Cunninghamia lanceolata plantations are widely distributed, accounting for 60% - 80% of the total area of subtropical artificial forests, and have important strategic significance for China's timber production and forestry economic development. Cunninghamia lanceolata plantations are characterized by fast growth and relatively high requirements for climate and soil environments. A large number of species investigations of Cunninghamia lanceolata plantations have found that the species diversity of their communities varies greatly with different forest ages, diameters at breast height, and densities, and the mechanism of this difference has rarely been explored, especially lacking large-scale research.

[0004] The growth of forest trees is a complex non-linear process, affected by multiple factors such as climate, site, competition, and disturbance and their interactive effects. Ignoring the spatial heterogeneity and temporal complexity of forest tree growth will lead to great uncertainty in the estimation of forest tree growth. A large amount of previous work on forest biomass estimation has shown that the influence of site environmental variables on forest biomass cannot be ignored. China has a vast territory and rich forest resources. Different site conditions directly act on various environmental variables, thus affecting forest growth and development, and further affecting the forest carbon sequestration function. Therefore, there are obvious differences in the growth of different tree species distributed in different regions of China, which are difficult to accurately observe.

[0005] Therefore, the present invention proposes a method and system for estimating the biomass of coniferous trees based on a PSO-BP neural network and a SHAP interpreter. Summary of the Invention

[0006] To solve at least one of the above-mentioned technical problems, the present invention proposes a method and system for estimating the biomass of coniferous trees based on a PSO-BP neural network and a SHAP interpreter.

[0007] The present invention is realized through the following technical solutions: A method for estimating the biomass of coniferous trees based on a PSO-BP neural network and a SHAP interpreter, comprising the following steps:

[0008] Step 1: Collect the big data of coniferous trees, where the big data includes forest tree basic data and environmental variable data;

[0009] Step 2: Preprocess the big data to obtain significantly correlated factors;

[0010] Step 3: Train the significantly correlated factors through a PSO-BP neural network to construct a biomass model for estimating the biomass of coniferous trees;

[0011] Step 4: Evaluate the model accuracy using verification metrics and analyze the contribution degree of each factor based on the machine learning interpretation tool SHAP interpreter.

[0012] Furthermore, the relevant data includes stand variable data, measured biomass data, and location variable data; the environmental variable data includes climate variable data, terrain variable data, and soil variable data.

[0013] Furthermore, the preprocessing includes:

[0014] Eliminate abnormal data with tree height ≤ 1.3m;

[0015] Use the Min-max method to normalize the input variables to eliminate the dimension difference;

[0016] Screen significantly correlated factors through correlation analysis.

[0017] Furthermore, in Step 3, the PSO-BP neural network model establishes a mapping function from microscopic parameters to macroscopic parameters through a BP neural network, and then embeds the mapping function into the fitness function of the PSO algorithm to search for the optimal microparameters.

[0018] Furthermore, the applications of the SHAP interpreter include:

[0019] Calculate the SHAP values of each input variable to quantify its contribution to biomass prediction;

[0020] Show the interaction between variables through a swarm SHAP plot to identify key driver factor combinations.

[0021] Furthermore, the key driver factors include the synergistic effect of stand variables and climate factors, and their interaction is analyzed through the non-linear superposition of SHAP values.

[0022] Furthermore, in Step 4, the verification metrics include the coefficient of determination, adjusted coefficient of determination, relative standard error, root mean square error, and Akaike information criterion, where the prediction accuracy of the PSO-BP model for the total biomass of Chinese fir satisfies the coefficient of determination ≥ 0.98 and the root mean square error ≤ 10.0.

[0023] The present invention also provides a system for implementing the method according to any one of claims 1-7, characterized in that it comprises:

[0024] A data input module: used to import forest tree basic data and environmental variable data;

[0025] A preprocessing module: performing data cleaning, normalization, and feature screening;

[0026] A PSO-BP model training module: realizing neural network structure optimization and parameter training;

[0027] A SHAP interpreter module: generating variable contribution degree maps and interaction effect analysis reports.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] This application uses publicly published research literature to collect and organize the measured data of individual Chinese fir biomass and factor data in different growth regions. Considering multiple factor variables, a non-parametric model of Chinese fir is reconstructed, and the accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A data graph for variable factors.

[0031] Figure 2 A correlation graph between the biomass of the whole plant and each organ and each factor.

[0032] Figure 3 A response graph of the total Chinese fir biomass to significantly correlated factors.

[0033] Figure 4 A comparison graph between the predicted values and the true values of three neural networks for the whole Chinese fir plant and different organs.

[0034] Figure 5 A bee colony SHAP graph of the optimal biomass non-parametric model for the whole Chinese fir plant and each organ. In each graph, the biomass is respectively (a) stem, (b) branch, (c) leaf, (d) root, and (e) total. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0036] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this invention belongs. The terms used in the description of the present invention herein are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "or / and" used herein includes any and all combinations of one or more of the related listed items.

[0037] Please refer to Figures 1 - 5 , this embodiment provides a method for estimating the biomass of coniferous trees based on the PSO-BP neural network and the SHAP interpreter, specifically including:

[0038] I. Data acquisition:

[0039] 1.1. Acquisition of data materials:

[0040] This application mainly uses four parts of data materials: published forestry basic data, including Cunninghamia lanceolata (148 articles, 460 biomass data, including stand variables, measured biomass data, and location variables), climate variable data extracted based on ClimateAP, terrain variable data extracted based on DEM (Digital Elevation Model, DEM) data, and soil variable data extracted based on HWSD (Harmonized World Soil Database, HWSD) data.

[0041] 1.2. Forestry basic data:

[0042] Through domestic and foreign academic databases such as CNKI Database, VIP Journal Database Network, China Forestry Digital Library, Sciencedirect, Wanfang Data Knowledge Service Platform, and Web of Science, search for relevant literature and books on "carbon storage", "biomass model", "biomass", "biomass", and "biomass model", etc. Based on the journal literature, conference literature, dissertations, forestry books, and annual reports published or carried in China on the measured biomass of Cunninghamia lanceolata from 1970 to 2022, establish a measured dataset of forest biomass, and summarize data including the location of forest plots, longitude and latitude, annual average temperature, annual rainfall, slope, aspect, forest age, diameter at breast height, tree height, stand density, biomass of the whole plant and each organ, collection time, etc. Among them, longitude and latitude are used as location variables, and diameter at breast height, tree height, forest age, and stand density are stand variables.

[0043] Screening requirements: ① The research object is a pure forest with stable growth status, that is, forest age meets more than 3 years, and there is no severe natural disaster (such as insect pests, fire, etc.) or frequent human disturbance (such as thinning, environmental pollution, etc.) in the recent period for the forest data. ② Since the diameter at breast height is usually measured at 1.3 m above the ground, data with H less than or equal to 1.3 m are deleted. ③ The research object is the (dry weight) biomass data obtained based on the direct harvest method and sampled strictly according to the standardized biomass survey method during actual measurement. It does not include estimated data and data cited from other literatures, that is, biomass data calculated based on previous biomass models and data not actually measured separately for each organ (branches, trunks, leaves, roots). ④ For underground biomass data, only data that include all roots (root stumps and roots) and the excavation area and depth meet the specifications during actual operation are considered. Finally, 460 pieces of data related to the effective biomass of Chinese fir were sorted out, involving 148 papers.

[0044] 1.3. Environmental variable data:

[0045] (1) Climate variables: Using the ClimateAP software developed by Wang et al., climate variables were extracted according to the longitude, latitude coordinates and elevation data of the sample plots. The growth time span of all the analytical trees sorted out in this application is from 1970 to 2022. Therefore, six annual climate variables (annual average temperature, annual average rainfall, annual average evaporation, annual wet and heat index, cold and hot temperature difference, frost-free period) for each year from 1970 to 2022 and seven data (highest temperature, lowest temperature, average temperature, highest rainfall, lowest rainfall, average rainfall, average humidity) in the growing season months (the growing season of Chinese fir is from April to November) were extracted, and the average value was calculated as the multi-year average climate variable of each sample plot ( Figure 1 ).

[0046] (2) Topographic variables: In this application, topographic variables were extracted based on SRTM DEM data. The absolute elevation accuracy of SRTM data is ±16 m, and the absolute plane accuracy is ±20 m (https: / / data.tpdc.ac.cn / / ). Based on the SRTM DEM data, surface analysis was carried out in GIS software to calculate the slope and aspect data of the sample plots with missing data.

[0047] (3) Soil variables: The soil variable data in this application is HWSD, which is jointly provided by the Food and Agriculture Organization, the International Institute for Applied Systems Analysis (IIASA), the World Soil Information (ISRIC), the Institute of Soil Science of the Chinese Academy of Sciences and the Joint Research Centre of the European Commission

[36] , and is considered to be one of the most comprehensive global soil databases. The HWSD v1.2 data for the Chinese region is based on the 1:1 million soil data of the second national land survey provided by the Nanjing Institute of Soil Science.

[0048] II. Data analysis:

[0049] 2.1. Data preprocessing: Before constructing the forest biomass model, the factors are uniformly normalized based on the Min-max method to avoid the excessive influence of factors with large dimensions during training, which helps improve the convergence speed of the model. Through correlation analysis, factors with significant correlations are screened and directly introduced into the model to prepare for the construction of the biomass non-parametric model.

[0050] 2.2. Model construction: In this application, a non-parametric model (machine learning) is used for model construction. The non-parametric models include three types: BP (Backpropagation Neural Network, BP) neural network model, GA-BP (Genetic Algorithm-Backpropagation Neural Network, GA-BP) neural network model, and PSO-BP (Particle Swarm Optimization-Backpropagation Neural Network, PSO-BP) neural network model. The contribution degree of each factor is analyzed and measured using the SHAP method.

[0051] Non-parametric model: The BP neural network model consists of an input layer, n hidden layers, and an output layer. Neurons in the same layer are independent of each other. This model does not need to master the relationship between input and output and has a strong non-linear mapping ability. The GA-BP neural network can solve the problem of difficult determination of the global optimal solution in the traditional BP neural network. The core of its algorithm is to randomly conduct global search based on simulating natural selection and genetics, so as to optimize the initial weights and thresholds of the network and improve the performance of the neural network. The main idea of PSO-BP is to establish a mapping function from microscopic parameters to macroscopic parameters through the BP neural network, and then embed the mapping function into the fitness function of the PSO algorithm to search for the optimal micro-parameters.

[0052] 2.3. Model verification and interpretation: Five indicators are selected to evaluate the accuracy of the biomass model, namely the coefficient of determination (R-squared, R 2 ), adjusted coefficient of determination (Adjusted R-squared, Adj-R 2)、Relative Standard Error (RSE), Root Mean Square Error (RMSE), and Akaike Information Criterion (AIC) are used to reflect the overall deviation of the model and evaluate the accuracy of the biomass model. In this application, the SHAP interpretation tool is introduced to evaluate the feature weights of the input factors of the neural network model. The combination of machine learning and machine learning interpretation tools has higher predictability and interpretability.

[0053] III. Results:

[0054] 3.1. Screening and processing of model feature independent variables:

[0055] 3.1.1. Correlation analysis of factors affecting Chinese fir biomass:

[0056] Correlation analysis of Chinese fir biomass in different organs and the whole plant with each influencing factor is as Figure 2 , Note: ** indicates significant correlation at the 0.01 level (two-tailed), and * indicates significant correlation at the 0.05 level (two-tailed). It is found that the correlation coefficients of the four stand variables, namely diameter at breast height, tree height, stand age, and stand density, are all very significant and are among the factors with higher influence. Among them, diameter at breast height and tree height, as the most commonly used independent variable factors in biomass models, have the highest influence degree among stand variables, and the correlation coefficient ranges from 0.616 to 0.877. Except for stand variables, the correlation coefficients of the remaining factors are all within ±0.3, and there are certain differences in the responses of different organs to the factors. Among the location variables, longitude is higher in terms of both influence degree and significance compared to latitude; among the climate variables, except that the highest temperature in the growing season has a significant impact on all organs and the whole of Chinese fir, the annual average temperature, average temperature in the growing season, lowest temperature in the growing season, and frost-free period only have relatively large influences on individual organs; among the terrain variables, only the elevation factor rarely shows significant influence; and there is no significant correlation in the soil variables and the overall influence degree is low. It shows that different variables have different effects on the biomass of different organs of Chinese fir, but the stand variables have the greatest influence, are significantly affected by temperature in the climate variables, and the soil variables have the least influence.

[0057] The method of the upper envelope is used to analyze the response of Chinese fir total biomass data to the changes of each significantly influencing single factor. For the longitude factor in the location variables, the response of Chinese fir annual average total biomass to it shows a fifth-degree polynomial. The annual average total biomass gradually increases as the longitude increases eastward between 106° and 108°, and the maximum value is located at 107.96° east longitude. Subsequently, as the longitude continues to increase eastward, the biomass slowly decreases, and finally there is a gentle slope near 118.17° east longitude, but it is significantly lower than that in the western region. The optimal longitude range is between 107° and 110°.

[0058] For climate variables such as Figure 3 (b), (c), and (d) show that the overall trend of the annual average total biomass and the three climate factors is to increase first and then decrease. Among them, the effects of the frost-free period and the highest temperature in the growing season on the annual average total biomass of Chinese fir have two peaks and valleys, and the first peak and valley are both the minimum values, while the annual average temperature has only one peak and valley. The optimal frost-free period range is 290 d to 310 d, the optimal highest temperature in the growing season is 29 °C to 31 °C, and the optimal annual average temperature range is 19 °C to 21 °C.

[0059] Regarding the stand variable, the growth process of the stand age factor and the total biomass conforms to the Richard model, that is, the change trend of "first slower - then accelerating - then slow", while the response of the stand density factor and the annual average total biomass of Chinese fir shows a power function and is significantly negatively correlated.

[0060] 3.2 Research on the non-parametric model of neural network biomass

[0061] 3.2.1 Construction of the non-parametric model of Chinese fir biomass

[0062] The factors that have a significant impact on the whole Chinese fir plant and its organs and the biomass data are used as samples for model training. By optimizing the parameters of the above three neural networks and comparing the model results of the total biomass and its components constructed under different parameter values, the optimal network parameters of the whole Chinese fir plant and its organs are determined (Table 1).

[0063] Table 1 Optimal parameters of the three neural network models for the biomass of the whole Chinese fir plant and its organs:

[0064]

[0065] For the three models of Chinese fir leaves, compared with the BP neural network, the fitting degree of the GA-BP model slightly decreases but the error precision improves. And it can be seen from the comparison chart that the GA-BP model better improves the abnormal situation of negative values, but the underestimation situation is still relatively serious. The PSO-BP neural network can not only significantly improve the accuracy of the Chinese fir leaf biomass model but also reduce the occurrence of abnormal situations (Table 2). Comparing the models of Chinese fir roots, it can be seen that the estimation results of the BP neural network are on the high side when the biomass value is between 20 and 40, and there are more underestimation phenomena after 45. The model accuracies of the GA-BP and PSO-BP neural networks are relatively consistent and both are improved compared with the BP neural network. Among them, the overestimation situation of the PSO-BP model is significantly reduced. The accuracies of the three models of the total Chinese fir biomass are higher than those of each organ. Among them, the fitting degree and error precision of the PSO-BP model are the best ( Figure 4 ).

[0066] Table 2 Prediction accuracy of the neural network model for the biomass of the whole Chinese fir plant and its organs

[0067]

[0068] 3.2.2 Analysis of the contribution degree of influencing factors of the non-parametric model of Chinese fir biomass:

[0069] The SHAP diagram of the bee colony describes the contribution ranking of factors in the optimal non-parametric model of the whole Chinese fir plant and each organ and the influence degree of factors of a single sample. It can be seen that the four stand variables of diameter at breast height (DBH), tree height (H), stand age (Age), and stand density (Den) and their interaction have relatively high contribution degrees in the whole plant and each organ ( Figure 5 , Note: Among them, DBH is the diameter at breast height, H is the tree height, Den is the stand density, Age is the stand age, Lon is the longitude, Lat is the latitude, Tmax_M is the highest temperature in the growing season, Tmin_M is the lowest temperature in the growing season, Tave_M is the average temperature in the growing season, Tmax_M is the average temperature, Tave_Y is the annual average temperature, NFFD is the frost-free period, and Ele is the altitude.). In addition to the stand variables, the contribution of longitude in the optimal non-parametric model of Chinese fir trunk is the highest and shows a negative correlation; while the contribution of the frost-free period and longitude to the optimal non-parametric model of branches and the whole plant is relatively large and both show negative correlations. It can be seen from the SHAP value that the influence of the frost-free period is greater than that of longitude, and although the interaction of the two factors in the optimal non-parametric model of branches is not more obvious than their single effects, it cannot be ignored; in the optimal non-parametric model of leaves, longitude, NFFD, and TMIN_M all show relatively large negative correlations; the Chinese fir roots are the most sensitive to longitude and TAVE_Y. Among them, it can be seen that there are cases where the SHAP values are positive and negative at the same time for the points with low eigenvalue (showing blue), indicating that the influence of longitude on root biomass is not a simple positive (negative) correlation. Generally speaking, the contribution degree of stand variables is the highest, and there are obvious differences in the contribution degrees of the remaining factors to the whole Chinese fir plant and different organs.

[0070] The above-described embodiments only represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A coniferous tree biomass estimation method based on PSO-BP neural network and SHAP interpreter, characterized in that: The following steps are involved: Step 1: Collecting big data of coniferous trees, including basic forest data and environmental variable data; Step 2: preprocess the big data to obtain significant correlation factors; Step 3: Use PSO-BP neural network to train significant correlation factors and build a biomass model for estimating coniferous tree biomass; Step 4: Verify the accuracy of the model evaluation indicators and analyze the contribution of each factor based on the machine learning interpretation tool SHAP interpreter.

2. The coniferous tree biomass estimation method based on PSO-BP neural network and SHAP interpreter according to claim 1 is characterized in that: The basic forest data include stand variable data, biomass measured data and location variable data; the environmental variable data include climate variable data, terrain variable data and soil variable data.

3. The coniferous tree biomass estimation method based on PSO-BP neural network and SHAP interpreter according to claim 2 is characterized in that: The pre-processing comprises: Abnormal data with tree height ≤1.3 m were eliminated; The Min-max method is used to normalize the input variables to eliminate dimensional differences; The significant correlation factors were screened through correlation analysis.

4. The coniferous tree biomass estimation method based on PSO-BP neural network and SHAP interpreter according to claim 3 is characterized in that: In the step three, the PSO-BP neural network model establishes a mapping function from micro parameters to macro parameters through the BP neural network, and then embeds the mapping function into the fitness function of the PSO algorithm to search for the optimal micro parameters.

5. The coniferous tree biomass estimation method based on PSO-BP neural network and SHAP interpreter according to claim 3 is characterized in that: Applications of the SHAP interpreter include: The SHAP value of each input variable was calculated to quantify its contribution to biomass prediction; The bee swarm SHAP diagram is used to display the interaction between variables and identify the key driving factor combination.

6. The coniferous tree biomass estimation method based on PSO-BP neural network and SHAP interpreter according to claim 5 is characterized in that: The key driving factors include the synergistic effects of forest stand variables and climate factors, and their interactions are analyzed through nonlinear superposition of SHAP values.

7. The coniferous tree biomass estimation method based on PSO-BP neural network and SHAP interpreter according to claim 1 is characterized in that: In the step 4, the verification indicators include the coefficient of determination, the corrected coefficient of determination, the relative standard error, the root mean square error and the Akaike information criterion, wherein the prediction accuracy of the PSO-BP model for the total biomass of Chinese fir satisfies the coefficient of determination ≥ 0.98 and the root mean square error ≤ 10.

0.

8. A system for implementing the method according to any one of claims 1 to 7, characterized in that: include: Data input module: used to import forest basic data and environmental variable data; Preprocessing module: performs data cleaning, normalization and feature screening; PSO-BP model training module: realizes neural network structure optimization and parameter training; SHAP Interpreter Module: Generates variable contribution graphs and interaction effect analysis reports.

Images