Chinese fir large-diameter log growth dynamic prediction method and system based on deep learning-sem fusion

Abstract

This invention discloses a method and system for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion, relating to the fields of forestry resource management and large-diameter timber cultivation. Through a collaborative process of deep learning feature extraction, SEM dynamic modeling, and iterative simulation, it integrates the advantages of wide coverage of remote sensing data, the accuracy of measured data, and the driving force of environmental data. It eliminates the need for frequent manual surveys, significantly reducing data acquisition costs and time consumption. Furthermore, the advantages of multi-model fusion greatly improve the consistency between prediction results and actual growth patterns, ultimately achieving efficient and accurate prediction of the growth dynamics of large-diameter Chinese fir timber. This provides a scientific basis for adjusting tending measures and planning rotation cycles, promoting the transformation of Chinese fir cultivation from experience-driven to data-driven approaches.

Description

Technical Field

[0001] This invention relates to the field of forestry resource management and large-diameter timber cultivation technology, and in particular to a method and system for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion. Background Technology

[0002] The precise cultivation of large-diameter Chinese fir timber is of great significance for ensuring timber supply and improving forestry economic benefits. The key to achieving this goal lies in the accurate prediction of Chinese fir growth dynamics, so as to formulate scientific tending and management strategies accordingly. For a long time, the prediction of Chinese fir growth dynamics has been one of the core focuses of forestry scientific research and production practice.

[0003] However, in the actual cultivation of large-diameter Chinese fir, traditional growth prediction methods have many limitations: on the one hand, relying on manual field surveys to obtain growth data is not only inefficient and costly, but also has limited spatiotemporal scale for data acquisition, making it difficult to comprehensively and timely reflect the growth dynamics of Chinese fir stands; on the other hand, growth models based on experience are insufficient in characterizing the complex nonlinear relationships between environmental factors and the stand's own growth feedback during the growth process of Chinese fir, and cannot accurately capture the dynamic changes at different growth stages, resulting in a large deviation between the prediction results and the actual growth conditions, making it difficult to meet the requirements for growth prediction accuracy in the cultivation of large-diameter timber.

[0004] Specifically, existing technologies have significant shortcomings in predicting the growth dynamics of large-diameter Chinese fir timber: First, the data acquisition methods limit the timeliness and comprehensiveness of the predictions, as manual surveys are insufficient for large-scale, high-frequency growth monitoring; second, the models lack the ability to integrate multiple influencing factors and simulate the cumulative effects of growth dynamics, making the prediction results unreliable for providing reliable support for the refined management of large-diameter timber cultivation. These problems seriously hinder the intelligent and precise development of large-diameter Chinese fir timber cultivation. Summary of the Invention

[0005] This invention provides a method and system for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion, which solves the technical problem of how to achieve efficient and accurate prediction of the growth dynamics of large-diameter Chinese fir timber.

[0006] The first aspect of this invention provides a method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion, comprising:

[0007] Acquire remote sensing image data, measured Chinese fir growth data, and environmental data for the target area;

[0008] Based on a pre-set hybrid feature extraction model, feature extraction is performed using the remote sensing image data, the measured Chinese fir growth data, and the environmental data to obtain a comprehensive feature vector.

[0009] Based on the pre-set SEM dynamic path model, the growth dynamics are predicted using the comprehensive feature vector and the measured Chinese fir growth data, and the annual growth dynamic simulation results of large-diameter Chinese fir in the target area are obtained.

[0010] Optionally, the pre-set hybrid feature extraction model includes a CNN canopy structure feature extraction module and an LSTM temporal growth feature extraction module. Based on the pre-set hybrid feature extraction model, feature extraction is performed using the remote sensing image data, the measured Chinese fir growth data, and the environmental data to obtain a comprehensive feature vector, including:

[0011] The remote sensing image data is input into the CNN canopy structure feature extraction module for feature extraction to obtain deep canopy features;

[0012] Based on the deep canopy features and the measured Chinese fir growth data, a growth-related feature sequence was constructed.

[0013] The growth-related feature sequence is input into the LSTM temporal growth feature extraction module for feature extraction to obtain temporal growth features;

[0014] Environmental factors are extracted from the environmental data, and the canopy depth features, the temporal growth features, and the environmental factors are concatenated to obtain a comprehensive feature vector.

[0015] Optionally, the step of using the remote sensing image data to input the CNN canopy structure feature extraction module for feature extraction to obtain deep canopy features includes:

[0016] Multi-layer convolution operations are performed on the remote sensing image data to obtain convolutional features;

[0017] The convolutional features are subjected to multiple pooling operations to obtain pooled features;

[0018] Flattening and fully connecting the pooled features yields deep canopy features.

[0019] Optionally, the step of using the growth-related feature sequence as input to the LSTM temporal growth feature extraction module for feature extraction to obtain temporal growth features includes:

[0020] Perform LSTM gating operation on the growth-related feature sequence to obtain the hidden state sequence;

[0021] A fully connected operation is performed on the hidden state to obtain the temporal growth features.

[0022] Optionally, the step of constructing a growth-related feature sequence based on the deep canopy features and the measured Chinese fir growth data includes:

[0023] Based on the growth mechanism of Chinese fir, multiple candidate indicator features were selected from the measured Chinese fir growth data and the deep canopy features;

[0024] Calculate the first Pearson correlation coefficient among the features of each candidate indicator;

[0025] Based on each of the first Pearson correlation coefficients, a preset number of candidate index features exceeding a preset first correlation coefficient threshold are selected as growth-related features;

[0026] A growth-related feature sequence is constructed using multiple of the aforementioned growth-related features.

[0027] Optionally, the step of predicting growth dynamics based on a pre-set SEM dynamic path model, using the comprehensive feature vector and the measured Chinese fir growth data, to obtain the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area includes:

[0028] Calculate the second Pearson correlation coefficient of each comprehensive feature within the comprehensive feature vector;

[0029] Based on each of the second Pearson correlation coefficients, a preset number of the comprehensive features exceeding the preset second correlation coefficient threshold are selected as exogenous variables;

[0030] The annual increase in diameter at breast height, annual increase in tree height, and cumulative growth were extracted from the measured Chinese fir growth data as endogenous variables.

[0031] Multiple causal paths are constructed using the exogenous and endogenous variables;

[0032] The path coefficients of each causal path are obtained by using the pre-set SEM dynamic path model with the input of the exogenous variables.

[0033] Based on the path coefficients, combined with the current values ​​of the exogenous variables and the historical values ​​of the endogenous variables, an iterative growth dynamic simulation is performed to obtain the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area.

[0034] Optionally, the iterative growth dynamic simulation based on the path coefficient, combined with the current value of the exogenous variable and the historical value of the endogenous variable, to obtain the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area includes:

[0035] Initialize the current value of the exogenous variable and the historical value of the endogenous variable;

[0036] For each forecast year, the predicted value of the endogenous variable for the current forecast year is calculated based on the path coefficient, the current value of the exogenous variable, and the historical value of the endogenous variable.

[0037] The predicted endogenous variable value for the current forecast year is used as the historical endogenous variable value for the next year. Then, the process jumps to the step of calculating the predicted endogenous variable value for the current forecast year based on the path coefficient, the current value of the exogenous variable, and the historical value of the endogenous variable for each forecast year, until the growth dynamic simulation of all forecast years is obtained, thus obtaining the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area.

[0038] Optionally, the predicted values ​​of the endogenous variables include the predicted values ​​of annual diameter at breast height (DBH), annual tree height, and cumulative growth.

[0039] Optionally, the predicted values ​​of the endogenous variables for the current forecast year are specifically calculated as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] In the formula, Indicates the first The predicted annual increase in chest diameter for the year; Indicates the first The predicted annual tree height increment for the year; Indicates the first The predicted cumulative growth for the year; , All indicate the first The year corresponds to the current value of the exogenous variable, where, , Index for exogenous variables; Indicates the first Measured or predicted annual increase in chest diameter; Indicates the first The measured or predicted annual tree height increment for the year; Indicates the first Measured or predicted cumulative annual growth; Indicates the first The path coefficients of each exogenous variable on the annual increase in chest diameter; Indicates the first The path coefficients of each exogenous variable on the annual tree height increment; This represents the path coefficient of the increase in chest diameter in the previous year to the increase in chest diameter in the current year; This represents the path coefficient of the increase in tree height in the previous year to the increase in tree height in the current year. This represents the path coefficient of annual increase in diameter at breast height to cumulative growth. , Indicates the error term; This indicates the number of exogenous variables affecting the annual increase in chest diameter; This indicates the number of exogenous variables that affect the annual increase in tree height.

[0044] The second aspect of this invention provides a system for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion, comprising:

[0045] The acquisition module is used to acquire remote sensing image data, measured Chinese fir growth data, and environmental data of the target area.

[0046] The extraction module is used to extract features from the remote sensing image data, the measured Chinese fir growth data, and the environmental data based on a pre-set hybrid feature extraction model, to obtain a comprehensive feature vector.

[0047] The prediction module is used to predict the growth dynamics of large-diameter Chinese fir trees in the target area based on a pre-set SEM dynamic path model, using the comprehensive feature vector and the measured Chinese fir growth data, and to obtain the annual growth dynamic simulation results of large-diameter Chinese fir trees in the target area.

[0048] As can be seen from the above technical solutions, the present invention has the following advantages:

[0049] (1) A hybrid feature extraction model consisting of CNN and LSTM was used to achieve deep fusion and accurate representation of multi-source data. The CNN module's multi-layer convolution and pooling processing of remote sensing images can effectively mine deep features of the canopy structure, breaking through the limitations of traditional manual surveys in capturing canopy information incompletely; the LSTM module's gating operation on growth-related feature sequences can accurately capture the temporal dynamics of the growth process, solving the problem that traditional models are unable to characterize the cumulative effect of growth. The combination of the two makes feature extraction more in line with the multi-dimensional characteristics of Chinese fir growth, providing high-quality input for the prediction of large-diameter Chinese fir growth, and improving prediction accuracy from the data level.

[0050] (2) The combination of the SEM dynamic path model and feature selection mechanism realizes the quantification and dynamic simulation of growth influence relationships. By screening exogenous variables through Pearson correlation coefficient, it is ensured that the factors included in the model are highly correlated with growth, reducing redundant information interference; the causal path and path coefficients constructed by the SEM model accurately quantify the dynamic relationship between exogenous variables (such as environmental factors, canopy features) and endogenous variables (diameter at breast height increment, tree height increment, etc.), overcoming the shortcomings of traditional empirical models in describing complex influence relationships vaguely. Iterative simulation based on path coefficients can output growth dynamic results year by year, realizing the leap from single-point prediction to full-cycle dynamic simulation, greatly improving the timeliness and continuity of prediction.

[0051] (3) Through the collaborative process of “deep learning feature extraction - SEM dynamic modeling - iterative simulation”, the advantages of the wide coverage of remote sensing data, the accuracy of measured data and the driving advantage of environmental data are integrated. It does not require high-frequency manual surveys, significantly reducing the cost and time consumption of data acquisition. At the same time, through the advantages of multi-model fusion, the consistency between the prediction results and the actual growth pattern is greatly improved. Finally, it achieves efficient and accurate prediction of the growth dynamics of large-diameter Chinese fir, providing a scientific basis for the adjustment of tending measures and the planning of rotation cycle, and promoting the transformation of Chinese fir cultivation from experience-driven to data-driven. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart illustrating the steps of a method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion, provided in an embodiment of the present invention;

[0054] Figure 2 This is a structural block diagram of a deep learning-SEM fusion-based system for predicting the growth dynamics of large-diameter Chinese fir timber, provided in an embodiment of the present invention. Detailed Implementation

[0055] This invention provides a method and system for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion. By integrating multiple types of data and utilizing advanced analysis methods, it achieves efficient and accurate prediction of the growth dynamics of large-diameter Chinese fir timber, thereby effectively overcoming the shortcomings of existing technologies and providing strong technical support for the scientific cultivation of large-diameter Chinese fir timber.

[0056] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0057] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion, as provided in an embodiment of the present invention.

[0058] This invention provides a method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion, comprising:

[0059] Step 101: Obtain remote sensing image data, measured Chinese fir growth data, and environmental data for the target area.

[0060] In this embodiment of the invention, the target area refers to the geographical area of ​​a specific Chinese fir stand where dynamic prediction of the growth of large-diameter Chinese fir timber is required. It can be defined according to the actual cultivation range or research needs, and its boundaries can be determined by latitude and longitude coordinates or vector boundaries in a geographic information system. Remote sensing image data refers to surface image data of the target area obtained through remote sensing platforms such as satellites and drones, containing multispectral or hyperspectral information, reflecting characteristics such as the canopy structure and vegetation coverage of the Chinese fir stand. Specifically, satellite imagery or drone aerial imagery with a resolution of not less than 3 meters can be used, and preprocessing such as radiometric correction and geometric correction is required to ensure data accuracy. Measured Chinese fir growth data refers to raw data reflecting the growth status of Chinese fir obtained through field surveys, including but not limited to the annual increase in diameter at breast height (the annual increase in diameter at 1.3 meters) and annual increase in tree height (the annual increase in diameter at breast height of a single Chinese fir tree). The data includes annual growth of tree height, cumulative diameter at breast height (DBH, cumulative growth of trunk diameter at 1.3 meters), and cumulative tree height (cumulative growth of tree height). The survey period must match the prediction time scale, usually annual observation data over several consecutive years. Environmental data refers to data on external environmental factors affecting Chinese fir growth, including topographic environmental data (such as altitude, slope, and aspect, extracted through digital elevation models), climate environmental data (such as average annual temperature, annual precipitation, and sunshine hours, obtained from meteorological stations or meteorological databases), soil environmental data (such as soil organic matter content, nitrogen, phosphorus, and potassium content, obtained through soil sampling analysis), and vegetation structure environmental data (such as stand density and canopy closure, obtained through field surveys or remote sensing inversion). The acquisition of the above data must cover the time period corresponding to the measured Chinese fir growth data to ensure the temporal consistency between feature extraction and model training.

[0061] Step 102: Based on the pre-set hybrid feature extraction model, feature extraction is performed using remote sensing image data, measured Chinese fir growth data, and environmental data to obtain a comprehensive feature vector.

[0062] Furthermore, the pre-built hybrid feature extraction model includes a CNN canopy structure feature extraction module and an LSTM temporal growth feature extraction module. Step 102 may include the following sub-steps:

[0063] S11. Use remote sensing image data to input the CNN canopy structure feature extraction module to extract features and obtain deep canopy features.

[0064] Furthermore, S11 may include the following sub-steps:

[0065] S111. Perform multi-layer convolution operations on the remote sensing image data to obtain convolutional features.

[0066] In this embodiment of the invention, the remote sensing image data is multispectral or hyperspectral data that has undergone radiometric and geometric correction preprocessing, which can truly reflect the canopy appearance and spectral characteristics of the Chinese fir forest in the target area. Multi-layer convolution operation refers to the process of extracting features from the input data through multiple consecutive convolutional layers. The core operation unit is the convolution kernel, which is a two-dimensional matrix with learnable parameters used to capture local spatial correlation features in the data. The convolution feature is a multi-channel feature map containing local feature information of the Chinese fir canopy, which is output after the convolution operation.

[0067] In practice, the preprocessed remote sensing image data is first converted into a tensor format acceptable to the CNN canopy structure feature extraction module, with dimensions of [batch size, image height, image width, number of bands]. This tensor data is then input into the first convolutional layer, which is configured with 16 3×3 convolutional kernels and a stride of 1. A zero-padding strategy is employed to ensure the output feature map size matches the input image. By sliding the convolutional kernels across the image tensor, dot product summation is performed on each local pixel region to obtain the initial feature response value. This is then transformed non-linearly using the ReLU activation function to enhance the model's non-linear representation of features, resulting in the output of the first convolutional layer, such as canopy edges and simple textures. The first convolutional layer extracts low-level features. Then, the output features are fed into the second convolutional layer, where the number of kernels is increased to 32. The kernel size remains 3×3, with a stride of 1 and zero padding. Convolution and activation operations are repeated to further aggregate low-level features into more discernible mid-level features such as canopy texture combinations and local structures. The number of kernels is then gradually increased through 3-4 convolutional layers (e.g., 64, 128). Each layer uses the same kernel parameters and activation function, allowing feature extraction to progress from basic spatial information to complex features strongly correlated with the canopy structure of the Chinese fir. The resulting multi-channel feature map, output after multiple convolutional operations, is the convolutional feature, which fully preserves the spatial structural correlation of the Chinese fir canopy.

[0068] S112. Perform multi-level pooling operations on the convolutional features to obtain pooled features.

[0069] In this embodiment of the invention, multi-layer pooling refers to the process of downsampling convolutional features through multiple consecutive pooling layers. The core objective is to reduce data dimensionality, decrease model computation, and enhance the translation invariance of features while preserving key feature information. The pooled features are condensed key feature maps output after pooling operations, which can better focus on core information strongly correlated with the canopy structure of Chinese fir. This embodiment preferably employs max pooling, a strategy that accurately preserves significant structural information of the Chinese fir canopy, such as the outline of dense canopy regions and key texture peaks, while effectively filtering out secondary redundant information.

[0070] In practice, the convolutional features output by S111 (dimensions of [batch size, feature map height, feature map width, number of channels]) are first input into the first pooling layer. This layer is configured with a 2×2 pooling kernel and a stride of 2. The pooling kernel slides uniformly across the convolutional feature map at the set stride, filtering out the maximum value of all feature values ​​within each 2×2 local window and using it as the output of that window. This compresses the height and width of the feature map to half of their original size, while keeping the number of channels unchanged, resulting in the intermediate core features after the first pooling layer. Then, this intermediate core feature is passed to the second pooling layer, using the same 2×2 pooling kernel and stride of 2. By using long parameter settings and repeating max pooling operations, the data dimensionality is further compressed and more critical canopy structure features are focused, such as the overall morphological association of the canopy and the spatial distribution of core functional areas. The number of pooling layers is adapted to the number of convolutional layers in S111. If there are 4 convolutional layers, only 2 pooling layers are needed to ensure that the feature dimensionality matches the subsequent processing requirements. The feature map output after multiple pooling operations is the pooled feature, which not only fully preserves the key identification information of the Chinese fir canopy structure, but also greatly reduces data redundancy and subsequent computational pressure, providing efficient and accurate feature input for flattening and fully connected operations.

[0071] S113. Flatten and fully connect the pooling features to obtain the deep features of the canopy.

[0072] In this embodiment of the invention, the flattening operation refers to the process of converting a multidimensional pooled feature map into a one-dimensional feature vector. The core purpose is to adapt to the input format of the fully connected layer and ensure the orderly transmission of feature information. The fully connected operation is to use a fully connected layer containing learnable weight parameters to fully connect all feature nodes of the previous layer with neurons of the current layer, thereby realizing feature fusion and nonlinear transformation. The deep canopy features are high-dimensional feature vectors that can accurately characterize the core structure and growth correlation characteristics of the Chinese fir canopy after flattening and fully connected processing. They are an important foundation for subsequent temporal feature extraction.

[0073] In practice, the input pooling features are first flattened. Following a fixed "row-first" or "column-first" order, all pixel values ​​from each channel in the pooling feature map are sequentially arranged and integrated, converting the original four-dimensional tensor data into a one-dimensional feature vector. The converted data dimension is [batch size, pooled feature map height × pooled feature map width × number of channels], ensuring that the key canopy information is not lost or disordered during pooling. This one-dimensional feature vector is then input into a pre-configured fully connected layer network. The first fully connected layer has 256 neurons. A matrix multiplication operation is performed between the one-dimensional feature vector and a preset initial weight matrix, followed by a non-linear transformation using the ReLU activation function. This effectively fuses canopy features from different dimensions, filters out invalid information, and strengthens the expression of core features. Finally, the output features from the first fully connected layer are transmitted... The second fully connected layer, with 128 neurons, is then used. Repeated weight matrix operations and ReLU activation further refine the deep correlations between features, enhancing their discriminative power and relevance. The weight parameters of the fully connected layer are continuously iteratively optimized during model training via backpropagation to ensure the output features are highly correlated with the canopy structure of the Chinese fir (such as canopy thickness, density, and spatial distribution patterns). Finally, after feature integration via flattening and feature deepening processing through two fully connected layers, the output high-dimensional feature vector represents the deep canopy features. This feature integrates key spatial information from pooling features and uncovers potential correlations between features through nonlinear transformations. It provides high-quality static feature input to the LSTM temporal growth feature extraction module, laying a solid foundation for improving the overall accuracy of the dynamic growth prediction model.

[0074] S12. Based on the deep features of the canopy and measured growth data of Chinese fir, a growth-related feature sequence was constructed.

[0075] Furthermore, S12 may include the following sub-steps:

[0076] S121. Based on the growth mechanism of Chinese fir, multiple candidate indicator features were selected from the measured growth data of Chinese fir and the deep characteristics of the canopy.

[0077] In this embodiment of the invention, the growth mechanism of Chinese fir refers to the inherent law of Chinese fir under the synergistic effect of its own physiological characteristics, canopy structure and environmental factors during the growth process. Specifically, it includes: the canopy structure indirectly regulates the annual growth increment of diameter at breast height and tree height by affecting the light interception efficiency and nutrient competition intensity; the tree's own growth has a cumulative effect, and the growth status of the previous year will affect the growth potential of the current year; candidate index features refer to the set of features that may be related to the growth dynamics of Chinese fir, which are initially screened and used for further screening of core features that are strongly correlated with growth. In practice, based on the "canopy-growth" correlation in the growth mechanism of Chinese fir, features related to the photosynthetically effective radiation absorption capacity and resource allocation efficiency within the canopy are selected from deep canopy features. For example, features related to canopy closure (reflecting light competition) and canopy vertical structure (affecting nutrient transport) are extracted from the dimensions with higher weights in the feature vector. At the same time, based on the time-series cumulative mechanism of Chinese fir growth, indicators that directly reflect the growth status are selected from measured Chinese fir growth data, including annual diameter at breast height (DBH) increment, annual tree height increment, cumulative DBH (reflecting growth foundation), cumulative tree height (reflecting growth potential), and the growth rate ratio derived from DBH and tree height (the ratio of annual DBH increment to annual tree height increment, reflecting the tendency of growth resource allocation). The selection process must ensure that the candidate indicator features cover the key correlation dimensions of "canopy structure features - growth status indicators", and that each candidate feature corresponds to a specific influence path in the Chinese fir growth mechanism (e.g., canopy structure features correspond to resource acquisition capacity, and growth status indicators correspond to growth results), providing a comprehensive and growth-compliant basic dataset for subsequent selection of core features through correlation analysis.

[0078] It is worth mentioning that the purpose of this invention in screening candidate indicator features based on the growth mechanism of Chinese fir is to break through the limitations of traditional "experience-based screening" or "pure data-driven screening" and construct a feature system deeply bound to the essential laws of Chinese fir growth. Its unique value is reflected in three aspects:

[0079] Firstly, by accurately anchoring the growth mechanism (such as photosynthetic drive, resource competition, temporal accumulation, and stage specificity), the selected candidate indicator features all correspond to the core regulatory path of Chinese fir growth. This eliminates redundant and noisy features that are not substantially related to growth from the source, avoiding the problems of "redundant feature dimensions leading to model overfitting" and "key features being masked by irrelevant information" in traditional screening. This makes the subsequent feature sequence construction more focused on the core contradictions, laying the foundation for improving the generalization ability and stability of the prediction model.

[0080] Secondly, the mechanism-based screening achieves interpretability of the "feature-growth" relationship. The selection of each candidate indicator feature is supported by clear biological logic (such as leaf area index being associated with total photosynthetic products and cumulative diameter at breast height being associated with nutrient transport efficiency). This overcomes the shortcomings of pure data-driven screening, where "features and growth only have statistical correlation and lack practical significance". This allows subsequent model training to not only capture the surface patterns of the data, but also to explore the underlying mechanisms of growth, significantly improving the reliability and credibility of the prediction results.

[0081] Third, the selection method of candidate indicator features based on the growth mechanism of Chinese fir can adapt to the core needs of different growth stages of Chinese fir. Based on the stage-specific mechanism of "prioritizing tree height in the young stage and accelerating diameter at breast height in the middle stage", indicator features that match the growth focus of the corresponding stage are selected (such as focusing on canopy porosity in the young stage and focusing on cumulative diameter at breast height in the middle stage). This solves the pain point that the traditional fixed feature set is difficult to adapt to the dynamic changes of the entire growth cycle of Chinese fir, making the feature sequence more targeted and providing high-quality input that fits the actual growth law for annual growth dynamic simulation. Ultimately, it realizes the leap from "data association prediction" to "mechanism-driven prediction", which greatly improves the creativity and practicality of the technical solution.

[0082] S122. Calculate the first Pearson correlation coefficient between the characteristics of each candidate indicator.

[0083] In this embodiment of the invention, the candidate indicator features are a set of features selected in S121 based on the growth mechanism of Chinese fir, covering canopy structure and growth status, including but not limited to canopy leaf area correlation features and canopy porosity correlation features in the deep canopy features, as well as annual diameter at breast height (DBH) increment, annual tree height increment, cumulative DBH, and cumulative tree height in the measured Chinese fir growth data; the first Pearson correlation coefficient is a statistic used to quantify the degree of linear correlation between two candidate indicator features, and its value ranges from [-1, 1], where the closer the absolute value is to 1, the stronger the linear correlation between the two features (positive value is positive correlation, negative value is negative correlation), and the closer the absolute value is to 0, the weaker the linear correlation. This coefficient is calculated by the formula:

[0084] ;

[0085] In the formula, This represents the first Pearson correlation coefficient. Indicates the number of samples. , They represent the first The observed values ​​of two candidate indicator features in a sample. Indicates the characteristics of candidate indicators The sample mean, Indicates the characteristics of candidate indicators The sample mean.

[0086] In practice, the candidate indicator features are first organized into a standardized two-dimensional data matrix (rows represent samples, such as observation data from different years or locations; columns represent candidate features), ensuring that all feature data are of the same magnitude (dimension differences can be eliminated through Z-score standardization); then, based on this matrix, the first Pearson correlation coefficient between each pair of candidate features is calculated, generating a symmetric correlation coefficient matrix, where the first Pearson correlation coefficient is the first Pearson correlation coefficient between each pair of candidate features. Line number The value of the column is the number of... The candidate feature and the first The correlation coefficients of candidate features are calculated. The calculation process needs to cover all candidate feature pairs to fully reflect the degree of linear correlation between features. For example, the correlation between "canopy leaf area correlation feature" and "annual diameter at breast height (DBH) increment" (reflecting the correlation between photosynthetic productivity and radial growth) and "cumulative DBH" and "annual tree height increment" (reflecting the impact of cumulative growth on longitudinal growth) need to be calculated. This step can accurately identify feature pairs with strong linear redundancy among candidate features (such as two canopy structure features that may be highly correlated because they both reflect canopy closure). This provides a quantitative basis for selecting core features based on thresholds, avoids the reduction of model learning efficiency or overfitting due to feature redundancy, and ensures that the retained features can cover the key mechanism path of Chinese fir growth and enhance the scientificity and conciseness of the feature set through statistical correlation verification.

[0087] S123. Based on each first Pearson correlation coefficient, a preset number of candidate index features exceeding the preset first correlation coefficient threshold are selected as growth-related features.

[0088] In this embodiment of the invention, the preset first correlation coefficient threshold is a critical value set based on experience and pre-experiment results in the field of Chinese fir growth. In this embodiment, it is preferably 0.6, and the value range can be adjusted between 0.5 and 0.7. It is used to define the feature standard of "strong correlation with growth". The preset number is the number of features to be retained determined according to the input dimension requirements and feature coverage of the subsequent LSTM time-series growth feature extraction module. In this embodiment, it is set to 8-12 to ensure that the model complexity is not too high due to excessive feature dimension, while covering the core driving path of Chinese fir growth. Growth-related features refer to the core feature set that, after screening, has a significant linear correlation with the growth dynamics of Chinese fir (such as annual diameter at breast height increase and tree height increase) and has low redundancy among features. It is the basis for constructing the growth-related feature sequence.

[0089] In practice, based on the correlation coefficient matrix generated by S122, the first Pearson correlation coefficient between each candidate indicator feature and the core growth indicators of Chinese fir (such as annual diameter at breast height increase and annual tree height increase) is extracted. Candidate features with an absolute value of the coefficient exceeding the preset first correlation coefficient threshold (0.6) are initially retained to form a preliminary screening set. If the number of features in the preliminary screening set exceeds the preset number (such as 12), the first Pearson correlation coefficient between the features in the set is further calculated. For feature pairs with an absolute value of the correlation coefficient exceeding 0.8 (considered as highly redundant features), the feature with higher correlation is retained by comparing the correlation strength between the two and the core growth indicators until the number of features is reduced to the preset number. If the number of features in the initial screening set is less than the preset number (e.g., less than 8), the threshold is appropriately lowered (e.g., adjusted to 0.5) and the screening is repeated. At the same time, key features that conform to the growth mechanism of Chinese fir are given priority (such as features related to canopy photosynthetic efficiency and historical incremental features reflecting the cumulative effect of growth). This ensures that the selected growth-related features are not only statistically strongly correlated, but also cover the complete mechanism chain of "canopy structure-resource acquisition-growth increment". This avoids the omission of key mechanism features due to relying solely on statistical indicators. The final set of growth-related features will provide a concise and efficient input for constructing time series sequences, improving the accuracy of the LSTM module in capturing the dynamic laws of growth.

[0090] S124. Construct a growth-related feature sequence using multiple growth-related features.

[0091] In this embodiment of the invention, the growth-related features are a set of core features obtained by S123 through screening with the first Pearson correlation coefficient. These features are significantly correlated with the growth dynamics of Chinese fir and have low redundancy. Specifically, they include canopy structure features (such as canopy leaf area correlation features and canopy porosity correlation features) and growth status features (such as annual diameter at breast height (DBH) increment, annual tree height increment, cumulative DBH, and cumulative tree height). Each feature corresponds to a key regulatory path in the growth mechanism of Chinese fir. The growth-related feature sequence refers to a two-dimensional time-series data structure formed by arranging multiple growth-related features in an orderly manner according to the time dimension. Its dimension is [number of time steps, number of growth-related features]. The time step corresponds to the continuous growth year of Chinese fir (e.g., if the growth observation is continuous for 5 years, the number of time steps is 5). Each time step contains the observed values ​​of all growth-related features for that year. Its core purpose is to adapt to the input format of the LSTM time-series growth feature extraction module, so that the model can capture the dynamic correlation law of feature changes over time.

[0092] In practice, the growth-related features selected by S123 are first time-series aligned to ensure that the observation time periods of all features are completely consistent (all on an annual scale, covering the same consecutive growth years). For example, the canopy leaf area correlation features and annual diameter at breast height (DBH) increment features for "2018-2022" are matched year by year. Then, using time steps as units, all growth-related feature observations for each year are integrated into a one-dimensional feature vector (the dimension is [number of growth-related features]). Then, the one-dimensional feature vectors of each year are concatenated sequentially according to time order (e.g., from early to late growth years) to form a complete sequence of growth-related features. If some years have individual features... For missing observations of the features, linear interpolation or predicted values ​​based on the growth mechanism of Chinese fir are used to fill in the missing values ​​(e.g., filling in missing values ​​for intermediate years based on the changing trends of canopy features in the previous and following years) to ensure the continuity and integrity of the sequence. The final constructed growth-related feature sequence not only retains the individual attributes of each growth-related feature, but also presents the dynamic relationship between features through temporal arrangement (e.g., the temporal response relationship between annual changes in canopy structure features and growth increment). It can completely convey the temporal dynamic information of Chinese fir growth, providing structurally standardized and informationally complete input data for the subsequent LSTM temporal growth feature extraction module to mine the long-term dependencies of the growth process and capture the temporal growth pattern.

[0093] S13. Use the growth-related feature sequence as input to the LSTM time-series growth feature extraction module to extract features and obtain time-series growth features.

[0094] Furthermore, S13 may include the following sub-steps:

[0095] S131. Perform LSTM gating operation on the growth-related feature sequence to obtain the hidden state sequence.

[0096] In this embodiment of the invention, the growth-related feature sequence is a set of core features of Chinese fir growth constructed by S124 and arranged in an ordered manner according to time steps, with dimensions of [number of time steps, number of growth-related features]. Each time step corresponds to the canopy structure and growth status features of Chinese fir in a growing year (such as canopy leaf area correlation features, annual diameter at breast height, etc.). LSTM gating operation refers to the operation process of Long Short-Term Memory (LSTM) selectively memorizing and updating temporal input information through three gate control units: forget gate, input gate, and output gate. Its core is to capture long-term dependencies through the sigmoid activation function (outputting a value between 0 and 1 to control the information retention ratio) and the tanh activation function (generating candidate update information). The hidden state sequence is a high-dimensional feature vector sequence output by LSTM at each time step, with dimensions of [number of time steps, number of hidden layer neurons]. Each vector integrates the key information of the current time step input and the historical time steps, reflecting the dynamic correlation law of Chinese fir growth features changing over time.

[0097] In practice, the growth-related feature sequences are first converted into a three-dimensional tensor that the LSTM module can accept (dimensions are [batch size, number of time steps, number of growth-related features], in this embodiment the batch size is set to 1, i.e., a single time-series data input); then the weight matrices and bias vectors of the LSTM's forget gate, input gate, and output gate are initialized. The forget gate is used to decide which historical information in the cell state is irrelevant to the current growth stage (e.g., discarding canopy vertical features that are strongly correlated with rapid tree height growth but have no effect on radial growth in the middle age after the juvenile stage), the input gate is used to determine which new information to include in the cell state (e.g., the contribution of the current year's canopy leaf area correlation features to the annual diameter at breast height increase), and the output gate is used to control the output content of the hidden state; for each time step, the input... The input feature vector is first passed through a forgetting gate; then an input gate is performed to fuse historical effective information with current new information (such as retaining the continuous impact of historical cumulative diameter at breast height on nutrient transport, while incorporating the new contribution of canopy porosity to photosynthetic efficiency in the current year); finally, an output gate is obtained through an output gate operation, which is then combined with cell state to generate the hidden state of the current time step; after performing the above gating operation on all time steps in sequence, the final sequence is the hidden state sequence. This sequence fully records the dynamic correlation of Chinese fir growth characteristics at different time steps (such as how changes in canopy structure affect growth increment across years), providing a temporal dimension of deep feature representation for fully connected layers to extract growth dynamics. This is a conventional LSTM gating operation, which will not be elaborated further here.

[0098] S132. Perform a fully connected operation on the hidden state to obtain the temporal growth features.

[0099] In this embodiment of the invention, the hidden state integrates the growth-related features of the corresponding time step with the key information of the historical time step, reflecting the dynamic correlation of Chinese fir growth at that time point (such as the coupling relationship between the current year's canopy structure changes and the growth increments of the previous two years). The fully connected operation refers to the process of linearly transforming and nonlinearly fusing the hidden state vector through a fully connected layer containing learnable weight parameters, thereby compressing the feature dimension and extracting the core information. Each neuron of the fully connected layer is fully connected to all dimensions of the hidden state, ensuring that the multi-dimensional correlations in the hidden state are fully captured. The temporal growth feature is a low-dimensional feature vector output after the fully connected operation, which can accurately characterize the temporal dynamic law of Chinese fir growth (such as the periodic fluctuation of the annual growth increment, the lag effect of the canopy structure on growth, etc.), and is the core input of the growth prediction model.

[0100] In practice, considering that the hidden state at the last time step of the LSTM has integrated the temporal information of the entire growth-related feature sequence (including all growth dynamic correlations from the initial observation year to the latest year), the hidden state at the last time step of the hidden state sequence is selected as the input of the fully connected operation (the dimension is [number of hidden layer neurons], and in this embodiment, the number of hidden layer neurons is set to 64). Then, this hidden state vector is input into a pre-set fully connected layer with 32 neurons. A matrix multiplication operation is performed between the weight matrix (dimension [32, 64]) and the hidden state vector. After superimposing the bias vector (dimension

[32] ), a nonlinear transformation is performed using the ReLU activation function to filter redundant information and strengthen features strongly correlated with the temporal growth pattern of Chinese fir (such as the effect of cross-year canopy leaf area changes on the increase in diameter at breast height). The cumulative impact, the differences in characteristics between peak and off-peak growing seasons, etc.); the weight parameters of the fully connected layer are continuously optimized through backpropagation during model training to ensure that the output features focus on the core law of "temporal correlation-growth response" (e.g., a certain weight parameter strengthens the influence of the third-year canopy porosity feature on the tree height increment in the fifth year, which fits the lag effect mechanism of Chinese fir growth); finally, the 32-dimensional feature vector output by the fully connected operation is the temporal growth feature. This feature not only retains the long-term temporal dependencies captured by LSTM, but also improves the computational efficiency of the prediction model through dimensional compression. At the same time, each dimension corresponds to a specific temporal dynamic pattern in the growth process of Chinese fir (e.g., the transition feature from rapid growth in the juvenile stage to stable growth in the middle-aged stage), providing deep feature support of the temporal dimension for accurate prediction of the future growth status of Chinese fir.

[0101] S14. Extract environmental factors from environmental data and concatenate them using canopy depth features, temporal growth features, and environmental factors to obtain a comprehensive feature vector.

[0102] In this embodiment of the invention, environmental data refers to the set of external environmental parameters that affect the growth of Chinese fir, including meteorological data (such as annual precipitation, average annual temperature, and annual sunshine hours), soil data (such as soil pH, soil organic matter content, and available nitrogen, phosphorus, and potassium concentrations), and topographic data (such as slope, aspect, and altitude). These data are obtained through long-term monitoring station records or remote sensing inversion, and directly or indirectly regulate the photosynthetic efficiency, nutrient absorption, and growth rhythm of Chinese fir. Environmental factors are key environmental indicators obtained from environmental data after screening and standardization, and must be directly related to the growth mechanism of Chinese fir (such as annual precipitation related to water supply, and available nitrogen content related to protein synthesis). In this embodiment, 8-12 core factors are selected (such as annual precipitation, average annual temperature, soil organic matter content, and slope). The comprehensive feature vector is a fusion feature vector formed by splicing canopy deep features, temporal growth features, and environmental factors in dimensional order. Its dimension is the sum of the dimensions of the three. Its core function is to integrate the three key information of Chinese fir growth: "internal structural features - dynamic growth law - external environmental driving force", and provide comprehensive input for growth prediction.

[0103] In practice, the environmental data is first preprocessed: outliers are removed, and environmental parameters of different dimensions are converted into standardized data with a mean of 0 and a standard deviation of 1 through Z-score standardization to ensure consistency with the numerical scale of canopy depth features and temporal growth features. Then, environmental factors are extracted from the preprocessed environmental data, with the selection criteria based on the growth mechanism of Chinese fir (e.g., the average annual temperature must be within the suitable growth range of 15-23℃ for Chinese fir; samples with temperatures below or above this range need to be labeled and associated with growth inhibition features). Next, features are concatenated in a fixed order: "canopy depth features → temporal growth features → environmental factors." Specifically, the 128-dimensional vector of canopy depth features is retained first, and then the 32-dimensional vector of temporal growth features and environmental factors are appended sequentially to the end. The 10-dimensional vector of the sub-samples is used to form a continuous 170-dimensional comprehensive feature vector. The splicing process must ensure that each feature dimension is aligned (all features are from the same batch of samples) and that there is no dimensional duplication or information conflict. For example, although the "annual sunshine hours" in the environmental factors and the "light interception features" in the canopy features are related, they belong to external supply and internal utilization, respectively, and must be retained at the same time to reflect the supply and demand relationship. The final comprehensive feature vector includes the static structural basis of the Chinese fir itself (deep canopy features), the dynamic change law of growth (temporal growth features), and the driving and limiting factors of the external environment (environmental factors), which fully covers the growth regulation closed loop of "structure-dynamic-environment", providing comprehensive and synergistic feature support for the prediction model to accurately capture the complex response mechanism of Chinese fir growth.

[0104] Step 103: Based on the pre-set SEM dynamic path model, growth dynamics are predicted using comprehensive feature vectors and measured Chinese fir growth data to obtain the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area.

[0105] Furthermore, step 103 may include the following sub-steps:

[0106] S21. Calculate the second Pearson correlation coefficient of each comprehensive feature within the constructed comprehensive feature vector.

[0107] In this embodiment of the invention, the comprehensive feature vector is a fused feature vector formed by splicing "canopy deep features → temporal growth features → environmental factors" in S14. Its dimension is the sum of the dimensions of the three (e.g., 170 dimensions). The comprehensive features it contains cover the internal structural features of Chinese fir growth (each dimension of canopy deep features), dynamic growth law features (each dimension of temporal growth features), and external environmental driving features (each dimension of environmental factors). Each comprehensive feature corresponds to a specific link in the Chinese fir growth regulation chain. The second Pearson correlation coefficient is a statistic used to quantify the degree of linear correlation between any two comprehensive features in the comprehensive feature vector. It has the same definition as the first Pearson correlation coefficient in S122, but the application scenario is different. The calculation process will not be described again.

[0108] In practice, the comprehensive feature vector is first preprocessed. Since the canopy depth features and time-series growth features have been standardized in previous steps, only the environmental factor features (if any) that have not been standardized need to be Z-score standardized to ensure that all comprehensive features are of the same numerical magnitude and to eliminate the influence of dimensional differences on correlation calculation. Then, the preprocessed comprehensive feature vector is organized into a two-dimensional data matrix (rows represent samples, and columns represent each comprehensive feature dimension). Based on this matrix, the second Pearson correlation coefficient between any two comprehensive features is calculated, generating a symmetric comprehensive feature correlation coefficient matrix. The value in the p-th row and q-th column of the matrix is ​​the correlation coefficient between the p-th and q-th comprehensive features. The calculation process must comprehensively cover all comprehensive feature pairs, including the correlation within the same type of feature (such as the "canopy leaf area correlation feature" and "canopy pore" in the canopy depth features). The correlation between features (such as the correlation between "gap correlation features") and between different types of features (such as the correlation between "cross-year growth correlation features" and "annual precipitation" in environmental factors) is also included. This also includes potential redundant correlations across dimensions (such as the correlation between "light interception features" in canopy depth features and "annual sunshine hours" in environmental factors). This step can accurately identify feature pairs with strong linear redundancy in the comprehensive feature vector (such as a certain dimension of a time-series growth feature being highly correlated with a certain dimension of the canopy depth feature because they both reflect the cumulative effect of growth). This provides a quantitative basis for subsequent threshold-based selection of core comprehensive features, avoiding the decrease in model training efficiency and weakened generalization ability caused by hidden redundancy resulting from the splicing of different types of features. Simultaneously, it ensures that the retained features comprehensively and without repetition cover the complete growth regulation path of "structure-dynamics-environment," laying the foundation for improving the accuracy of growth prediction.

[0109] S22. Select the preset number of comprehensive features that exceed the preset second correlation coefficient threshold as exogenous variables based on each second Pearson correlation coefficient.

[0110] In this embodiment of the invention, the preset second correlation coefficient threshold is a critical value set in combination with the cross-category characteristics of comprehensive features (the correlation strength between different types of features is usually lower than that between similar features). In this embodiment, it is preferably 0.55, and the value range can be adjusted between 0.5 and 0.65. It is lower than the first correlation coefficient threshold in S123, which avoids omitting key cross-category correlation features and effectively eliminates weak correlation redundancy. The preset number is a value determined according to the upper limit of the exogenous variable input dimension and the completeness of feature information of the growth prediction model (such as ARIMA, LSTM fusion model). In this embodiment, it is set to 20-30 to ensure the balance between model operation efficiency and feature coverage. Comprehensive features are the features in the comprehensive feature vector that cover the canopy depth features, time-series growth features, and environmental factors. Each feature corresponds to a specific link in the regulation of Chinese fir growth. Exogenous variables are the core comprehensive feature set that has a significant linear correlation with the growth target of Chinese fir (such as the increase in diameter at breast height and the increase in tree height in future years) and has low redundancy between features after screening. They are used as external input variables of the prediction model to quantify the driving and constraining effects of external factors on the growth of Chinese fir.

[0111] In practice, based on the comprehensive feature correlation coefficient matrix generated by S21, the second Pearson correlation coefficient between each comprehensive feature and the Chinese fir growth target (such as the predicted growth increment in the next 1-3 years) is extracted. Comprehensive features with an absolute value exceeding the preset second correlation coefficient threshold (0.55) are initially retained to form a preliminary cross-category screening set. If the number of features in the preliminary screening set exceeds the preset number (such as 30), the second Pearson correlation coefficient between features within the set is further calculated. For highly redundant feature pairs with an absolute value exceeding 0.85, cross-category related features (such as "cross-year cumulative increment in time-series growth features" and "average annual temperature in environmental factors") or key features that fit the Chinese fir growth mechanism (such as "canopy leaf area related features" and "soil organic matter content", corresponding to the photosynthetic capacity-nutrient supply synergistic mechanism) are preferentially retained until the number of features is reduced to a minimum. The threshold is set to a preset number. If the number of features in the initial screening set is less than the preset number (e.g., less than 20), the threshold is lowered to 0.5 and the screening is repeated. At the same time, the cross-category core features of "structure-environment" and "dynamic-environment" (such as the correlation between canopy porosity and annual sunshine hours, and the correlation between temporal growth lag features and annual precipitation) are forcibly retained to avoid the loss of cross-category regulation paths due to reliance on statistical indicators alone. The exogenous variables obtained by the final screening include both the core structure and dynamic features of Chinese fir itself (such as the light interception correlation dimension in the deep canopy features and the stage growth pattern dimension in the temporal growth features) and key external environmental driving factors (such as soil available nitrogen content and average annual precipitation). There is no significant redundancy between the features, which can provide the core input of "internal foundation-external driving" synergy for the prediction model, ensuring that the model accurately captures the complex response mechanism of Chinese fir growth and improves the scientificity and reliability of growth prediction.

[0112] S23. Extract the annual increase in diameter at breast height, annual increase in tree height, and cumulative growth from the measured Chinese fir growth data as endogenous variables.

[0113] In this embodiment of the invention, the annual diameter at breast height (DBH) increment refers to the increase in the diameter of the Chinese fir trunk at 1.3 meters within a certain growth year. It is calculated by subtracting the measured DBH value of the previous year from the measured DBH value of the current year, directly reflecting the radial growth intensity of the Chinese fir. The annual tree height increment refers to the increase in the tree height within a certain growth year. It is calculated by subtracting the measured tree height value of the previous year from the measured tree height value of the current year, reflecting the longitudinal growth rate of the Chinese fir. The cumulative growth refers to the cumulative growth value from the year of planting to the observation year, including the cumulative DBH (the sum of the DBH increments of each year) and the cumulative tree height (the sum of the tree height increments of each year), reflecting the long-term cumulative effect of Chinese fir growth and the current growth foundation. The endogenous variable refers to the core growth indicator in the growth prediction model that is driven by the exogenous variable (the core comprehensive feature selected by S22) and has a time-series dependency relationship. It is the prediction target and internal feedback variable of the model. Its core role is to quantify the dynamic changes in Chinese fir growth and to take over the regulatory effect of the exogenous variable.

[0114] In practice, the measured Chinese fir growth data are first aligned to ensure that the annual diameter at breast height (DBH) increment and annual tree height increment for each observation sample correspond to a complete growth year (e.g., from January 1st to December 31st of the calendar year), avoiding calculation errors due to differences in observation time points. Then, the increment data is calculated year by year using the formulas "Annual DBH increment = Measured DBH of the current year - Measured DBH of the previous year" and "Annual tree height increment = Measured tree height of the current year - Measured tree height of the previous year". Outliers in the calculation results (e.g., negative increments due to measurement errors) are corrected or removed based on the Chinese fir growth mechanism (e.g., the growth rate slows down in the near-maturity period but negative growth is rare), and replaced with linear interpolation results based on increments from adjacent years. Finally, the annual increments are calculated by summing the results. Cumulative diameter at breast height (DBH) and cumulative tree height are used to form complete cumulative growth data. Finally, the annual DBH increment, annual tree height increment, cumulative DBH, and cumulative tree height are integrated into an endogenous variable set. Z-score standardization is performed on all endogenous variables to ensure that their numerical scale is consistent with that of exogenous variables, avoiding the impact of dimensional differences on model parameter optimization. The resulting endogenous variables include both annual increment indicators reflecting short-term growth dynamics and cumulative indicators reflecting long-term growth foundations. They can comprehensively characterize the core results of Chinese fir growth and form a "driving factor-growth result" correspondence with exogenous variables. This provides a complete and synergistic core input for the subsequent construction of a "exogenous variable driven + endogenous variable time-series feedback" predictive model, ensuring that the model can accurately capture the complex mapping relationship between growth dynamics and driving factors.

[0115] S24. Construct multiple causal paths using exogenous and endogenous variables.

[0116] In this embodiment of the invention, the causal path refers to the logical association path of "driving factor-growth result" or "preceding growth-subsequent growth" formed by clarifying the direction of regulation of exogenous variables on endogenous variables and the temporal dependence between endogenous variables based on the growth mechanism of Chinese fir. Each path corresponds to a clear biological mechanism, rather than a simple statistical correlation.

[0117] In practical implementation, the construction logic is based on "mechanism anchoring + variable matching": First, for environmental factors among exogenous variables, a direct causal path is constructed in conjunction with the "environment-driven growth" mechanism, such as "annual precipitation (environmental factor) → annual increase in diameter at breast height (endogenous variable)" (path logic: precipitation determines soil moisture supply, affects root water absorption and nutrient transport efficiency, and thus regulates radial growth intensity), and "average annual temperature (environmental factor) → annual increase in tree height (endogenous variable)" (path logic: temperature affects photosynthetic enzyme activity, determines the rate of photosynthetic product accumulation, and adapts to the longitudinal growth of Chinese fir). Firstly, regarding the temperature requirements for growth; secondly, for the exogenous variables of canopy depth characteristics, pathways are constructed based on the mechanism of "canopy structure affecting resource acquisition," such as "canopy leaf area correlation characteristics (canopy depth characteristics) → annual diameter at breast height increase (endogenous variable)" (path logic: leaf area determines light interception, and the accumulation of photosynthetic products directly supports radial growth), and "canopy porosity correlation characteristics (canopy depth characteristics) → annual tree height increase (endogenous variable)" (path logic: porosity affects understory light distribution, reduces light competition among individuals, and promotes longitudinal growth under apical dominance); furthermore... To address the correlation between exogenous variables and endogenous variables related to temporal growth characteristics, a "dynamic law feedback" path is constructed, such as "cross-year growth correlation characteristics (temporal growth characteristics) → next year's tree height increase (endogenous variable)" (path logic: the growth rhythm captured by temporal characteristics (such as peak season growth rate) directly predicts subsequent growth trends). Simultaneously, based on the "cumulative effect" mechanism of Chinese fir growth, a temporal causal path between endogenous variables is constructed, such as "cumulative diameter at breast height (DBH) (endogenous variable) → current year's DBH increase (endogenous variable)" (path logic: cumulative DBH determines the scale of the trunk's vascular tissue). The higher the transport efficiency, the better it can support the annual radial growth increment. The path logic is: "Annual tree height increment (endogenous variable) → Cumulative tree height (endogenous variable)" (the continuous accumulation of short-term longitudinal growth increment forms the basis for long-term growth). In addition, considering the synergy of growth regulation, a causal path driven by multiple exogenous variables is constructed, such as "soil organic matter content (environmental factor) + canopy light interception characteristics (deep canopy characteristics) → annual diameter at breast height increment (endogenous variable)" (the path logic is: soil organic matter provides nutrients, canopy characteristics enhance photosynthetic capacity, and the two synergistically support radial growth).

[0118] After all causal pathways are constructed, they need to be screened through "mechanism consistency verification" (to ensure that the pathways conform to the physiological characteristics of Chinese fir) and "second Pearson correlation coefficient verification" (to ensure that there is a significant association between the pathway variables and no spurious causality). Pathways that conflict with the growth mechanism or have extremely weak associations are eliminated, and finally 8-12 core causal pathways are retained. Each pathway clearly defines the direction of action and mechanistic basis between variables, and comprehensively covers the complete growth regulation chain of "environment-driven - structural support - dynamic feedback - cumulative effect". This provides a clear logical framework for building an accurate growth prediction model, enabling the model to make predictions based on causal relationships rather than simple statistical associations, and significantly improving the interpretability and reliability of the prediction results.

[0119] S25. Using exogenous variables as input, a pre-set SEM dynamic path model is used for path estimation to obtain the path coefficients of each causal path.

[0120] In this embodiment of the invention, the pre-set SEM dynamic path model refers to a parameterized model pre-set based on the Structural Equation Model (SEM) framework and combined with the dynamic characteristics of the growth time series of Chinese fir. Its core is to quantify the synergistic regulatory effect of multiple factors on growth by analyzing the correlation between latent and observable variables and the path relationships between variables. The "dynamic" aspect is reflected in the inclusion of a time series dimension in the model, allowing path coefficients to be dynamically adjusted with each time step (growth year) to adapt to the differences in regulatory patterns at different growth stages of Chinese fir. Path estimation refers to solving the model parameters using a statistical algorithm (maximum likelihood estimation in this embodiment) to enable the model to predict... The process of minimizing the deviation between the value and the actual observed value (the measured value of the endogenous variable) and thus determining the influence strength between variables in each causal path; the causal path is a "drive-response" logical association path constructed by S24 based on the growth mechanism; the path coefficient is the core result of path estimation, which is a standardized coefficient that quantifies the influence strength of the independent variable on the dependent variable in the causal path. The value range is usually [-1,1]. Positive values ​​indicate positive driving force (such as increased precipitation promoting growth), negative values ​​indicate negative constraint (such as extreme high temperature inhibiting growth), and the larger the absolute value, the higher the influence strength. Due to the dynamic characteristics of the model, the same causal path may correspond to different path coefficients at different growth stages.

[0121] In practice, the time-series data of exogenous and endogenous variables are first integrated into the model input dataset. Exogenous variables serve as "exogenous latent variables" or "observable exogenous variables," while endogenous variables serve as "endogenous latent variables" or "observable endogenous variables." The structural equations of the model are pre-defined according to the causal path constructed in S24, including measurement equations and structural equations. The measurement equations describe the association between latent and observable variables, while the structural equations describe the causal paths between variables. Subsequently, the model parameters are iteratively solved using maximum likelihood estimation. During the iteration process, the model continuously adjusts the coefficients of each causal path to achieve the optimal goodness of fit between the predicted and measured values ​​of the endogenous variables based on the current coefficients. For example, for the path "canopy leaf area association characteristics → annual diameter at breast height increase," the path estimation might yield a coefficient of 0.42. This indicates that for every 1 standardized unit increase in canopy characteristics, the annual diameter at breast height (DBH) increases by an average of 0.42 standardized units, and this coefficient may rise to 0.51 in the middle-aged stage (the stage of vigorous radial growth), which is consistent with the mechanism that photosynthetic products in the middle-aged stage tend to be distributed radially. For the path of "average annual temperature → annual tree height increase", if the temperature exceeds the upper limit suitable for Chinese fir (e.g., 25℃), a negative coefficient of -0.28 may be obtained, reflecting the inhibitory effect of high temperature on longitudinal growth. Finally, the path coefficients of each causal path obtained by path estimation not only quantify the influence strength of exogenous variables on endogenous variables and among endogenous variables, but also reflect the specificity of growth stages through dynamic adjustment. This provides quantitative mechanistic support for growth prediction based on causal relationships, enabling predictions to not only rely on data correlation, but also to achieve accurate extrapolation based on clear influence strength.

[0122] S26. Based on the path coefficient, combined with the current value of the exogenous variable and the historical value of the endogenous variable, iterative growth dynamic simulation is performed to obtain the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area.

[0123] Furthermore, S26 may include the following sub-steps:

[0124] S261. Initialize the current value of the exogenous variable and the historical value of the endogenous variable.

[0125] In this embodiment of the invention, the current value of the exogenous variable refers to the latest observed value of the exogenous variable corresponding to the prediction start time, reflecting the external driving state of the prediction start point; the historical value of the endogenous variable refers to the observed value of the endogenous variable for several consecutive growth years before the prediction start time (in this embodiment, it is set to the previous 5 years, i.e., 2017-2021), including the increase in diameter at breast height / tree height and the cumulative growth in each year, which is used to capture the temporal dependence of the endogenous variable.

[0126] In practice, the prediction time window is first defined, and the latest observed value at the starting moment is extracted from the time series database of exogenous variables as the current value. For example, if the latest observation of the canopy leaf area correlation feature is data from 2022, then its current value is the standardized feature value of that year. For exogenous variables of environmental factors (such as precipitation in 2022), the standardized value of the measured record is directly used as the current value to ensure consistency with the standardization scale during model training. For the historical values ​​of endogenous variables, complete records for five consecutive years prior to the prediction start time are extracted from the measured Chinese fir growth data and arranged in chronological order to form a historical sequence. For example, the annual diameter at breast height (DBH) increments from 2017 to 2021 are [1.2cm, 1.5cm, 1.3cm, 1.6cm, 1.4cm]. If there are missing endogenous variables in a certain year, the interpolation is performed based on the "temporal dependence between endogenous variables" mechanism in the causal path constructed in S24 (such as the correlation between annual tree height increment and cumulative tree height), combined with data from adjacent years (such as using the mean of tree height increments in 2018 and 2020 to complete the 2019 data) to ensure the continuity of the historical value sequence. During initialization, it is necessary to verify the time alignment between the current values ​​of exogenous variables and the historical values ​​of endogenous variables to avoid cross-sample data confusion. At the same time, the initialized current and historical values ​​are converted into tensor formats that the model can recognize. The current values ​​of exogenous variables are one-dimensional vectors, and the historical values ​​of endogenous variables are two-dimensional time series matrices. This provides a standardized starting point input for iterative prediction based on the SEM dynamic path model, ensuring that the prediction process can start from the real growth state and environmental conditions, accurately inherit the historical growth patterns and current driving factors, and lay a data foundation for improving prediction accuracy.

[0127] S262. For each forecast year, calculate the forecast value of the endogenous variables for the current forecast year based on the path coefficient, the current value of the exogenous variables, and the historical value of the endogenous variables.

[0128] Furthermore, the predicted values ​​of endogenous variables include the predicted values ​​of annual diameter at breast height (DBH), annual tree height, and cumulative growth.

[0129] It should be noted that the selection of the annual diameter at breast height (DBH) increase, annual tree height increase, and cumulative growth as the predicted values ​​of endogenous variables is based on a comprehensive consideration of the biological characteristics of Chinese fir growth, the needs of forestry practice, and the prediction logic of the model. The three values ​​form a complete characterization of the growth status of Chinese fir from different dimensions, and there is a clear mechanistic relationship and functional complementarity.

[0130] From the perspective of growth mechanism, the annual increase in diameter at breast height (DBH) is a direct reflection of the radial growth of Chinese fir, closely related to the development of trunk vascular tissue and the radial distribution of photosynthetic products. It directly reflects the annual increase in timber yield and is a core indicator for assessing the quality and economic value of forest timber. The annual increase in tree height corresponds to longitudinal growth and is regulated by factors such as apical dominance and canopy light competition, reflecting the ability of forests to compete for space resources. Its dynamic changes can reflect the long-term impact of stand density, light conditions, etc. The cumulative growth (cumulative DBH and cumulative tree height) is the cumulative result of the annual increases. It represents both the current growth base of the forest (e.g., cumulative DBH determines the efficiency of trunk nutrient transport, directly affecting radial growth potential) and the transformation of growth stages (e.g., after the cumulative tree height reaches a threshold, the increase in tree height will slow down due to the weakening of apical dominance). It is a key link connecting short-term dynamics and long-term trends.

[0131] From the perspective of prediction logic, the three form a closed loop of "short-term increment - long-term accumulation - stage feedback": the prediction of annual increment needs to be based on the current foundation represented by the cumulative growth (e.g., for individuals with large cumulative diameter at breast height, the annual increase in diameter at breast height may slow down due to intensified resource competition), while the prediction of cumulative growth depends on the continuous superposition of annual increment. This intrinsic connection enables the model to capture the cumulative effect of growth and stage specificity (e.g., the increase in tree height in the young stage dominates the cumulative growth, while the contribution of the increase in diameter at breast height in the middle-aged stage increases).

[0132] From a forestry practice perspective, the annual increase in diameter at breast height (DBH) is directly related to timber production forecasts, the annual increase in tree height is used to assess the stand's spatial expansion capacity, and the cumulative growth serves decisions such as harvesting cycle planning and stand maturity assessment. Together, these three constitute the core indicator system of "growth dynamics - resource assessment - management" in forestry production, ensuring that the forecast results can directly support actual production needs.

[0133] Therefore, selecting these three types of predicted values ​​as endogenous variables can not only comprehensively cover the short-term dynamics and long-term trends of Chinese fir growth, but also strengthen the predictive logic of the model through their inherent mechanism association, while meeting the quantitative needs of forestry practice for key growth indicators, so that the prediction results have both biological rationality and application value.

[0134] The specific calculation of the predicted values ​​of endogenous variables for the current forecast year is as follows:

[0135] ;

[0136] ;

[0137] ;

[0138] In the formula, Indicates the first The annual diameter at breast height (DBH) increase is an endogenous variable reflecting the radial growth efficiency of Chinese fir. Indicates the first The annual tree height increase forecast is an endogenous variable reflecting the longitudinal growth potential of Chinese fir. Indicates the first The annual cumulative growth forecast is an endogenous variable for judging the effectiveness of large-diameter timber cultivation. , All indicate the first The year corresponds to the current value of the exogenous variable, where, , This serves as an index for exogenous variables, used to distinguish different exogenous variables, which are driving factors affecting the growth of Chinese fir. Indicates the first The measured or predicted annual increase in diameter at breast height is a historical value of an endogenous variable that reflects the growth status of the previous year. Indicates the first The measured or predicted annual tree height increment is a historical value of an endogenous variable reflecting the growth status of the previous year. Indicates the first The measured or predicted cumulative growth for a year is a historical value of an endogenous variable that reflects the growth performance of the previous year. Indicates the first The path coefficients of exogenous variables on the annual increase in chest diameter are used to quantify the influence of exogenous variables on endogenous variables. Indicates the first The path coefficients of exogenous variables on the annual tree height increment are used to quantify the influence of exogenous variables on endogenous variables. It represents the path coefficient of the increase in chest diameter in the previous year to the increase in chest diameter in the current year, and quantifies the strength of the dynamic association between endogenous variables; This represents the path coefficient of the increase in tree height in the previous year to the increase in tree height in the current year, quantifying the strength of the dynamic association between endogenous variables. It represents the path coefficient of annual diameter at breast height (DBH) increase on cumulative growth, usually 1, since cumulative growth = historical cumulative value + current year increase, quantifying the strength of the cumulative association between endogenous variables; , This represents the error term, which reflects the growth variation that the model does not fully explain. It is usually taken as the mean of the model training residuals or 0. This indicates the number of exogenous variables that affect the annual increase in chest diameter, such as the number of canopy porosity and annual mean temperature. This indicates the number of exogenous variables that affect the annual increase in tree height, such as the number of slopes and annual precipitation.

[0139] In this embodiment of the invention, the predicted annual diameter at breast height (DBH) increase, the predicted annual tree height increase, and the predicted cumulative growth can be calculated using the above formulas.

[0140] S263. Use the predicted value of the endogenous variable for the current forecast year as the historical value of the endogenous variable for the next year, and jump to execute the step of calculating the predicted value of the endogenous variable for the current forecast year based on the path coefficient, the current value of the exogenous variable and the historical value of the endogenous variable for each forecast year, until the growth dynamic simulation of all forecast years is obtained, and the annual growth dynamic simulation results of large-diameter Chinese fir in the target area are obtained.

[0141] In this embodiment of the invention, for ease of understanding, after obtaining the predicted values ​​of endogenous variables for the current prediction year (e.g., 2023) (predicted annual diameter at breast height (DBH) increment 1.87 cm, predicted annual tree height increment 1.18 m, and predicted cumulative growth 14.37 m), these values ​​are directly updated to the historical values ​​of endogenous variables required for calculation in the next year (2024). That is, the historical values ​​of endogenous variables in 2024 include the above-mentioned predicted values ​​for 2023 (which, together with the historical measured values ​​from 2018 to 2022, constitute a continuous time series). At the same time, the current values ​​of exogenous variables for 2023 are obtained (if they are measured data, they are used directly; if they are prediction years, they are standardized values ​​based on the 2023 canopy leaf area correlation characteristics and annual precipitation obtained from the environmental trend model), keeping the path coefficients (e.g., the path coefficient of the canopy leaf area correlation characteristics to the annual DBH increment 0.35, the lag term coefficient of the annual DBH increment 0.40, etc.) stable within the prediction period (because the short-term growth mechanism does not change significantly). Subsequently, following the same logic as calculating the 2023 forecast, based on the current values ​​of exogenous variables in 2024, the updated historical values ​​of endogenous variables, and path coefficients, the predicted values ​​of endogenous variables for 2024 are calculated (e.g., the annual diameter at breast height (DBH) increase may be slightly adjusted due to the increased cumulative growth in 2023, resulting in 1.79 cm; the annual tree height increase is affected by the rainfall trend in 2023, resulting in 1.12 m; and the cumulative growth is updated to 14.37 + 1.79 = 16.16). Similarly, the predicted values ​​for 2024 are used as historical values ​​for 2025, and the current values ​​of exogenous variables in 2024 are updated simultaneously as input for the 2025 calculation. The steps of "current value of exogenous variable + historical value of endogenous variable + path coefficient → predicted value for the current year" are repeated to obtain the predicted values ​​of endogenous variables for 2025, 2026, and 2027, until all preset forecast years (e.g., the next 5 years) are covered. Finally, the predicted values ​​of annual diameter at breast height (DBH), annual tree height, and cumulative growth for each year from 2023 to 2027 were integrated in chronological order to form the dynamic simulation results of the growth of large-diameter Chinese fir in the target area from the current starting year to each future year. The results fully present the annual variation trends of radial and longitudinal growth as well as the long-term evolution of cumulative growth. They can reflect the specific values ​​of growth in each year and also reflect the dynamic relationship that "early growth foundation affects later growth potential". This provides a continuous and quantitative time-series basis for assessing the growth trend, planning the management cycle, and formulating tending measures for large-diameter Chinese fir in the target area.

[0142] Please see Figure 2 , Figure 2 This is a structural block diagram of a deep learning-SEM fusion-based system for predicting the growth dynamics of large-diameter Chinese fir timber, provided in an embodiment of the present invention.

[0143] This invention provides a deep learning-SEM fusion-based system for predicting the growth dynamics of large-diameter Chinese fir timber, comprising:

[0144] The acquisition module 201 is used to acquire remote sensing image data, measured Chinese fir growth data and environmental data of the target area;

[0145] Extraction module 202 is used to extract features based on a pre-set hybrid feature extraction model, using remote sensing image data, measured Chinese fir growth data and environmental data, to obtain a comprehensive feature vector;

[0146] The prediction module 203 is used to predict the growth dynamics of large-diameter Chinese fir trees in the target area by using a pre-set SEM dynamic path model, comprehensive feature vectors and measured Chinese fir growth data, and obtain the annual growth dynamic simulation results of large-diameter Chinese fir trees in the target area.

[0147] Furthermore, the pre-built hybrid feature extraction model includes a CNN canopy structure feature extraction module and an LSTM temporal growth feature extraction module. The extraction module 202 includes:

[0148] The first extraction submodule is used to extract features from the CNN canopy structure feature extraction module by inputting remote sensing image data into the canopy structure feature extraction module, and obtain deep canopy features.

[0149] The processing submodule is used to construct growth-related feature sequences based on deep canopy features and measured Chinese fir growth data;

[0150] The second extraction submodule is used to extract features by inputting the growth-related feature sequence into the LSTM temporal growth feature extraction module to obtain temporal growth features;

[0151] The splicing submodule is used to extract environmental factors from environmental data and splice them using deep canopy features, temporal growth features, and environmental factors to obtain a comprehensive feature vector.

[0152] Furthermore, the first extraction submodule includes:

[0153] Convolutional feature units are used to perform multi-layer convolution operations on remote sensing image data to obtain convolutional features.

[0154] Pooling feature unit is used to perform multiple pooling operations on convolutional features to obtain pooled features;

[0155] The canopy deep feature unit is used to flatten and fully connect pooled features to obtain canopy deep features.

[0156] Furthermore, the second extraction submodule includes:

[0157] The hidden state sequence unit is used to perform LSTM-gated operations on the growth-related feature sequences to obtain the hidden state sequence.

[0158] The temporal growth feature unit is used to perform a fully connected operation on the hidden state to obtain temporal growth features.

[0159] Furthermore, the processing submodule includes:

[0160] The candidate indicator feature unit is used to screen multiple candidate indicator features from measured Chinese fir growth data and deep canopy features based on the Chinese fir growth mechanism.

[0161] The first Pearson correlation coefficient unit is used to calculate the first Pearson correlation coefficient between the features of each candidate indicator.

[0162] The growth-related feature unit is used to select a preset number of candidate index features that exceed the preset first correlation coefficient threshold as growth-related features based on each first Pearson correlation coefficient.

[0163] Growth-related feature sequence unit, used to construct a growth-related feature sequence using multiple growth-related features.

[0164] Furthermore, the prediction module 203 includes:

[0165] The second Pearson correlation coefficient submodule is used to calculate the second Pearson correlation coefficient of each comprehensive feature within the constructed comprehensive feature vector;

[0166] The exogenous variable submodule is used to select a preset number of comprehensive features that exceed the preset second correlation coefficient threshold as exogenous variables based on each second Pearson correlation coefficient.

[0167] The endogenous variable submodule is used to extract annual diameter at breast height (DBH) increment, annual tree height increment, and cumulative growth as endogenous variables from measured Chinese fir growth data.

[0168] The causal path submodule is used to construct multiple causal paths using exogenous and endogenous variables;

[0169] The path coefficient submodule is used to perform path estimation by inputting exogenous variables into a pre-set SEM dynamic path model, and to obtain the path coefficients of each causal path.

[0170] The iterative growth dynamic simulation submodule is used to perform iterative growth dynamic simulation based on path coefficients, combined with the current values ​​of exogenous variables and the historical values ​​of endogenous variables, to obtain the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area.

[0171] Furthermore, the iterative growth dynamic simulation submodule includes:

[0172] An initialization unit is used to initialize the current values ​​of exogenous variables and the historical values ​​of endogenous variables;

[0173] The endogenous variable forecast unit is used to calculate the endogenous variable forecast value for each forecast year based on the path coefficient, the current value of the exogenous variable, and the historical value of the endogenous variable.

[0174] The jump unit is used to take the predicted value of the endogenous variable for the current forecast year as the historical value of the endogenous variable for the next year, and jump to execute the step of calculating the predicted value of the endogenous variable for the current forecast year based on the path coefficient, the current value of the exogenous variable and the historical value of the endogenous variable for each forecast year, until the growth dynamic simulation of all forecast years is obtained, and the annual growth dynamic simulation results of large-diameter Chinese fir in the target area are obtained.

[0175] Furthermore, the predicted values ​​of endogenous variables include the predicted values ​​of annual diameter at breast height (DBH), annual tree height, and cumulative growth.

[0176] Furthermore, the specific calculation of the predicted values ​​of endogenous variables for the current forecast year is as follows:

[0177] ;

[0178] ;

[0179] ;

[0180] In the formula, Indicates the first Predicted annual increase in chest diameter for the year; Indicates the first Predicted annual tree height increase for the year; Indicates the first Forecast of cumulative annual growth; , All indicate the first The year corresponds to the current value of the exogenous variable, where, , Index for exogenous variables; Indicates the first Measured or predicted annual increase in chest diameter; Indicates the first The measured or predicted annual tree height increment for the year; Indicates the first Measured or predicted cumulative annual growth; Indicates the first The path coefficients of each exogenous variable on the annual increase in chest diameter; Indicates the first The path coefficients of each exogenous variable on the annual tree height increment; This represents the path coefficient of the increase in chest diameter in the previous year to the increase in chest diameter in the current year; This represents the path coefficient of the increase in tree height in the previous year to the increase in tree height in the current year. This represents the path coefficient of annual increase in diameter at breast height to cumulative growth. , Indicates the error term; This indicates the number of exogenous variables affecting the annual increase in chest diameter; This indicates the number of exogenous variables that affect the annual increase in tree height.

[0181] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0182] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0183] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0184] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0185] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0186] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion, characterized in that, include: Acquire remote sensing image data, measured Chinese fir growth data, and environmental data for the target area; Based on a pre-set hybrid feature extraction model, feature extraction is performed using the remote sensing image data, the measured Chinese fir growth data, and the environmental data to obtain a comprehensive feature vector. The pre-set hybrid feature extraction model includes a CNN canopy structure feature extraction module and an LSTM temporal growth feature extraction module. Based on the pre-set hybrid feature extraction model, feature extraction is performed using the remote sensing image data, the measured Chinese fir growth data, and the environmental data to obtain a comprehensive feature vector, including: The remote sensing image data is input into the CNN canopy structure feature extraction module for feature extraction to obtain deep canopy features; Based on the deep canopy features and the measured Chinese fir growth data, a growth-related feature sequence was constructed. The growth-related feature sequence is input into the LSTM temporal growth feature extraction module for feature extraction to obtain temporal growth features; Environmental factors are extracted from the environmental data, and the canopy depth features, the temporal growth features, and the environmental factors are concatenated to obtain a comprehensive feature vector; Based on the pre-set SEM dynamic path model, the growth dynamics of Chinese fir are predicted by using the comprehensive feature vector and the measured Chinese fir growth data, and the annual growth dynamic simulation results of large-diameter Chinese fir in the target area are obtained. The method, based on a pre-set SEM dynamic path model, uses the comprehensive feature vector and the measured Chinese fir growth data to predict growth dynamics, obtaining the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area, including: Calculate the second Pearson correlation coefficient of each comprehensive feature within the comprehensive feature vector; Based on each of the second Pearson correlation coefficients, a preset number of the comprehensive features exceeding the preset second correlation coefficient threshold are selected as exogenous variables; The annual increase in diameter at breast height, annual increase in tree height, and cumulative growth were extracted from the measured Chinese fir growth data as endogenous variables. Multiple causal paths are constructed using the exogenous and endogenous variables; The path coefficients of each causal path are obtained by using the pre-set SEM dynamic path model with the input of the exogenous variables. Based on the path coefficients, iterative growth dynamics are performed by combining the current value of the exogenous variable and the historical value of the endogenous variable.

2. The method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion according to claim 1, characterized in that, The process of inputting the remote sensing image data into the CNN canopy structure feature extraction module for feature extraction to obtain deep canopy features includes: Multi-layer convolution operations are performed on the remote sensing image data to obtain convolutional features; The convolutional features are subjected to multiple pooling operations to obtain pooled features; Flattening and fully connecting the pooled features yields deep canopy features.

3. The method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion according to claim 1, characterized in that, The growth-related feature sequence is input into the LSTM temporal growth feature extraction module for feature extraction to obtain temporal growth features, including: Perform LSTM gating operation on the growth-related feature sequence to obtain the hidden state sequence; A fully connected operation is performed on the hidden state to obtain the temporal growth features.

4. The method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion according to claim 1, characterized in that, The construction of a growth-related feature sequence based on the deep canopy features and the measured Chinese fir growth data includes: Based on the growth mechanism of Chinese fir, multiple candidate index features were selected from the measured Chinese fir growth data and the deep canopy features; Calculate the first Pearson correlation coefficient among the features of each candidate indicator; Based on each of the first Pearson correlation coefficients, a preset number of candidate index features exceeding a preset first correlation coefficient threshold are selected as growth-related features; A growth-related feature sequence is constructed using multiple of the aforementioned growth-related features.

5. The method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion according to claim 1, characterized in that, The iterative growth dynamic simulation based on the path coefficient, combined with the current value of the exogenous variable and the historical value of the endogenous variable, yields the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area, including: Initialize the current value of the exogenous variable and the historical value of the endogenous variable; For each forecast year, the predicted value of the endogenous variable for the current forecast year is calculated based on the path coefficient, the current value of the exogenous variable, and the historical value of the endogenous variable. The predicted endogenous variable value for the current forecast year is used as the historical endogenous variable value for the next year. Then, the process jumps to the step of calculating the predicted endogenous variable value for the current forecast year based on the path coefficient, the current value of the exogenous variable, and the historical value of the endogenous variable for each forecast year, until the growth dynamic simulation of all forecast years is obtained, thus obtaining the annual growth dynamic simulation results of large-diameter Chinese fir timber in the target area.

6. The method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion according to claim 5, characterized in that, The predicted values ​​of the endogenous variables include the predicted values ​​of annual diameter at breast height (DBH), annual tree height, and cumulative growth.

7. The method for predicting the growth dynamics of large-diameter Chinese fir timber based on deep learning-SEM fusion according to claim 6, characterized in that, The predicted values ​​of the endogenous variables for the current forecast year are specifically calculated as follows: ； In the formula, Indicates the first The predicted annual increase in diaphragm diameter for the year; Indicates the first The predicted annual tree height increment for the year; Indicates the first The predicted cumulative growth for the year; , All indicate the first The year corresponds to the current value of the exogenous variable, where, , Index for exogenous variables; Indicates the first Measured or predicted cumulative annual growth; Indicates the first The measured or predicted annual tree height increment for the year; Indicates the first Measured or predicted cumulative annual growth; Indicates the first The path coefficients of each exogenous variable on the annual increase in chest diameter; Indicates the first The path coefficients of each exogenous variable on the annual tree height increment; This represents the path coefficient of the increase in chest diameter in the previous year to the increase in chest diameter in the current year; This represents the path coefficient of the increase in tree height in the previous year to the increase in tree height in the current year; This represents the path coefficient of annual increase in diameter at breast height to cumulative growth. , Indicates the error term; This indicates the number of exogenous variables affecting the annual increase in chest diameter; This indicates the number of exogenous variables that affect the annual increase in tree height.

8. A growth dynamic prediction system for large-diameter Chinese fir timber based on deep learning-SEM fusion, characterized in that, The deep learning-SEM fusion-based dynamic prediction system for the growth of large-diameter Chinese fir timber is used to implement the deep learning-SEM fusion-based dynamic prediction method for the growth of large-diameter Chinese fir timber as described in any one of claims 1-7. The deep learning-SEM fusion-based dynamic prediction system for the growth of large-diameter Chinese fir timber includes: The acquisition module is used to acquire remote sensing image data, measured Chinese fir growth data, and environmental data of the target area. The extraction module is used to extract features from the remote sensing image data, the measured Chinese fir growth data and the environmental data based on a pre-set hybrid feature extraction model, and obtain a comprehensive feature vector. The prediction module is used to predict the growth dynamics of large-diameter Chinese fir trees in the target area based on a pre-set SEM dynamic path model, using the comprehensive feature vector and the measured Chinese fir growth data, and to obtain the annual growth dynamic simulation results of large-diameter Chinese fir trees in the target area.

Images