Palaeoecological plant species estimation method and system based on graph computation and feature fusion

Abstract

The method and system for calculating the paleoecological plant species based on graph calculation and feature fusion are disclosed. The method obtains the multi-dimensional original morphological features of a fossil sample, constructs a graph structure model representing the correlation of the features and extracts graph structure scalar features. The graph scalar features and the original features are spliced and fused, and then input into a pre-trained target classification model to output a prediction probability. At the same time, the range similarity of the sample is calculated based on the botanical reference specification. Finally, the final similarity is obtained by weighting the prediction probability and the range similarity to determine the target species. The target classification model is obtained by joint training based on a mixed data set containing a double-engine mechanism to simulate the sample and real confirmed data. Through decoupled graph feature fusion, double-track hierarchical discrimination and virtual-real combined training paradigm, the present application effectively overcomes the bias caused by the data scarcity and fossil defects in the field of paleontology, greatly improving the robustness of the inference and the interpretability of the identification results.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and paleontology, and in particular to a method and system for estimating paleoecological plant species based on graph computation and feature fusion. Background Technology

[0002] Typical paleontological assemblages serve as crucial windows for studying the evolution of ecosystems during specific geological periods. The abundant plant fossils they preserve are of immense scientific value for reconstructing paleoclimate and plant evolution. Taking a typical tropical rainforest fossil site as an example, the unearthed biodiversity is extremely rich, encompassing hundreds of species. Dominant species such as the Dipterocarpaceae family, with their tall canopies forming the upper forest layer, are vital for maintaining the vertical hierarchy of the ecosystem. Due to the high evolutionary stability of these iconic species, their fossil morphology and chemical characteristics are highly consistent with extant species, providing key evidence for accurately indicating the paleotropical environment. However, due to geological changes, unearthed fossils often lack complete population structures, especially the fractures or incompleteness of large tree structures, severely hindering the accurate reconstruction of the vertical ecological structure of ancient forests.

[0003] Against the backdrop of a paradigm shift in archaeological research, while artificial intelligence (AI) has made progress in fossil fragment identification and structural reconstruction through high-precision scanning and deep learning algorithms, technological gaps remain in the field of paleoecological species distribution reconstruction and prediction. Traditional paleobotanical identification relies heavily on expert morphological comparison experience, which suffers from limitations such as high subjectivity, low standardization, and limited efficiency. Existing machine learning methods, when inferring species, mostly treat fossil samples as independent individuals, failing to fully consider the intrinsic correlations between various morphological characteristics of plants and ignoring the symbiotic relationships and niche constraints among species in fossil communities. This isolated data processing model leads to insufficient accuracy in species prediction, making it difficult to objectively reconstruct the spatial distribution and vertical structure of complex paleoecological systems. Therefore, a computational technique capable of integrating multidimensional features and characterizing complex topological relationships among species is urgently needed. Summary of the Invention

[0004] To address the aforementioned technical problems in the existing technology, this invention proposes a method and system for estimating paleoecological plant species based on graph computation and feature fusion.

[0005] According to a first aspect of the present invention, a method for estimating paleoecological plant species based on graph computation and feature fusion is proposed, comprising:

[0006] S1, Obtain the original morphological feature data of the plant fossil sample to be tested. The original morphological feature data includes at least numerical features, categorical features and probabilistic features. S2, Based on the original morphological feature data, construct a graph structure model to characterize the correlation between plant morphological features and generate the corresponding adjacency matrix; S3, map the measured values ​​of the plant fossil sample under test in each feature dimension to the node signal vector on the graph structure model, extract the graph structure scalar features through graph neighborhood signal aggregation operation; and splice and fuse the graph structure scalar features with the original morphological feature data to generate an extended feature matrix; S4, based on the pre-trained target classification model, the extended feature matrix is ​​inferred and processed to output the predicted probability that the tested plant fossil sample belongs to the target species; wherein, the target classification model is a machine learning model built based on gradient boosting tree; S5. Based on the preset botanical reference standard, calculate the range similarity of the plant fossil sample to be tested on the original morphological feature data; calculate the final similarity by weighting and fusing the predicted probability and the range similarity, and determine whether the plant fossil sample to be tested belongs to the target species based on the final similarity.

[0007] The above technical solution overcomes the shortcomings of traditional paleontological classification methods, which rely excessively on isolated physical features and ignore the coupling relationships between features. This method innovatively introduces a graph computing architecture, transforming static morphological features into dynamic network topology signals. Simultaneously, by coupling machine learning prediction probabilities with range similarity based on expert prior norms, a dual-track judgment mechanism of "data-driven + rule-verified" is constructed. This mechanism significantly improves the model's fault tolerance and inference robustness when dealing with paleoecological fossil samples (especially those with severe feature deficiencies or extreme variations), effectively preventing absurd misjudgments that might result from a single black-box model.

[0008] In some specific embodiments, the graph structure model in step S2 uses a two-stage merging strategy to generate the adjacency matrix, specifically including: Phase 1: Invoke the large language model, output the correlation relationship and initial weight between features based on the feature names and their descriptions; perform node validity verification on the correlation relationship output by the large language model, and extract weights for repeatedly generated feature edges based on a preset optimization strategy to generate the first adjacency matrix. The second stage: For feature pairs that did not have edges established in the first stage, backoff connections are established based on prior knowledge of plant taxonomy; specifically, this includes: establishing connections between feature nodes belonging to the same business or attribute category, assigning preset default weights, and generating a second adjacency matrix; Merging strategy: Based on a preset matrix fusion rule, the first adjacency matrix and the second adjacency matrix are merged by corresponding elements to generate the final adjacency matrix. The preset matrix fusion rule includes taking the maximum value of the corresponding position weight.

[0009] In the above technical solution, a two-stage merging strategy is adopted. First, a large language model is used to automatically mine the implicit semantic relationships between massive features. Then, prior knowledge of plant taxonomy is used as a "safety net" to establish backlinks for features with the same attribute. This strategy not only ensures the semantic richness and objectivity of the graph structure, but also eliminates the feature aggregation discontinuity caused by isolated feature nodes, ensuring the integrity and biological rationality of the global graph topology.

[0010] In some specific embodiments, step S3 includes: S31, Computation Node Safety level ,in For nodes The degree, For the corresponding adjacency matrix elements, It is a preset minimum positive number; S32, Computation Node Neighbor weighted mean ,in, For nodes The value of on the signal vector; S33, Based on the security degree and the weighted mean of the neighbors, extract the scalar features of the graph structure; S34, the graph structure scalar features are concatenated with the original morphological feature data to generate an extended feature matrix.

[0011] In the above technical solution, by introducing the concept of safety degree and neighbor-weighted aggregation logic, the comprehensive performance of each morphological feature in its local ecological / biological context can be effectively captured while ensuring the absolute stability of the algorithm's operation. Furthermore, the complex high-dimensional information of the graph structure is losslessly reduced in dimensionality and refined into high-quality scalar features that are easily ingested by downstream classification models, greatly enhancing the expressive power of the input feature space.

[0012] In some specific embodiments, the graph structure scalar features include graph smoothing energy, the global average of the neighbor aggregation mean, the dispersion of the neighbor aggregation mean, degree-weighted features, and local variability, wherein the graph smoothing energy is: The global average of the neighbor aggregation mean: The dispersion of the aggregated mean of the neighbors: The degree-weighted features are as follows: The local variability: In the formula, For nodes The value taken on the signal vector, The total number of the numerical feature nodes and the probabilistic feature nodes participating in the graph construction.

[0013] The aforementioned technical solution abandons the redundant calculation of directly inputting a massive graph matrix. Instead, it quantifies the coupling perturbations between features from multiple structural physical dimensions, such as the overall smoothness of the graph, central tendency, data dispersion, node importance weighting, and the intensity of local variation. This provides downstream advanced gradient boosting tree models with extremely rich "high-order derived biological information," enabling the model not only to identify "what leaves look like" but also to perceive "the synergistic changing trends between leaf features and climate features," thereby significantly improving the fitting upper limit and classification accuracy of species inference.

[0014] In some specific embodiments, the target classification model is an XGBoost classifier. Before inputting the extended feature matrix into the pre-trained XGBoost classifier, the following feature preprocessing process is also included: A feature preprocessing pipeline is constructed to perform one-hot encoding on the categorical features in the extended feature matrix and configure an unknown category ignoring strategy; standardization is performed on the numerical features in the extended feature matrix to make the mean of each feature component 0 and the variance 1; the preprocessed feature components are concatenated to generate a high-dimensional feature vector that is input to the XGBoost classifier.

[0015] In the above technical solution, an end-to-end feature preprocessing pipeline is constructed to map unstructured discrete classification features (such as leaf edge shape) and continuous numerical features (such as precipitation) to a unified low-dimensional continuous feature space. In particular, the introduction of the unknown category ignoring strategy greatly improves the system robustness of the model when dealing with "unseen features" that often occur in paleoecological fossil excavation, and avoids model prediction collapse.

[0016] In some specific embodiments, the model prediction probability output by the XGBoost classifier in step S4 is obtained by mapping the original score of the model output to the [0,1] interval using a logistic regression function, and its calculation formula is as follows:

[0017] In the formula, The model prediction probability output by the XGBoost classifier; This is a high-dimensional feature vector that has been transformed through the feature preprocessing pipeline and then input into the classifier; In the classifier The weighted sum of the regression trees is calculated using the following formula: ,in, For the number of iteration rounds, For the specific first A tree of return, For the first Regression tree for feature vectors The given output score, This is the preset learning rate.

[0018] In the above technical solution, by introducing a logistic regression mapping function, the weighted scores of the unbounded tree model are strictly nonlinearly compressed into a standardized probability space. This transformation not only gives the classification results an intuitive "statistical confidence level," but more importantly, it enables the model output to be within the same dimensional range as the "range similarity" subsequently determined by human experts, laying the necessary mathematical foundation for the subsequent dual-track weighted fusion.

[0019] In some specific embodiments, step S5 specifically includes: S51: Extract effective features from the original morphological feature data of the plant fossil sample to be tested, and exclude the graph structure scalar features; based on the preset botanical reference standard, calculate the numerical similarity component of the numerical features, the categorical similarity component of the categorical features, and the probabilistic similarity component of the probabilistic features in the effective features respectively. S52: Calculate the arithmetic mean of the similarity components of all the effective features obtained, and obtain the total range similarity of the plant fossil sample to be tested; S53: The predicted probability and the total range similarity are weighted and summed according to a preset fusion coefficient to calculate the final similarity; S54: Compare the final similarity with a preset comprehensive similarity threshold. If the final similarity is greater than or equal to the comprehensive similarity threshold, then determine that the tested plant fossil sample belongs to the target species; otherwise, determine that it does not belong to the target species.

[0020] In the above technical solution, by stripping away the structural features of the graph and regressing to the most basic physical space to calculate the "total range similarity," it is equivalent to adding an "expert common sense safety valve." The final weighted fusion and threshold comparison mechanism ensures that the inference results have both the sensitivity of AI in mining nonlinear features and absolutely do not violate the macroscopic physical laws of paleoecology.

[0021] In some specific embodiments, step S51 includes: For the aforementioned numerical characteristics, let the positive class reference interval given in the botanical reference standard be... Set the interval width ,in This is the preset minimum interval tolerance; if the measured value of the feature is... Its numerical similarity components The calculation formula is: ; For the aforementioned categorical features, let the set of preference values ​​specified in the botanical reference standard be . If the feature observation value is Its categorical similarity components The calculation formula is: ; Regarding the probabilistic feature, let the positive reference probability given in the botanical reference standard be... If the characteristic observation value is Its probabilistic similarity components The calculation formula is: ; The botanical reference standard's characteristic attribute system is divided into a basic data space including environmental characteristics, plant morphological characteristics, ecological characteristics, chemical composition characteristics, and geological period characteristics. When constructing the botanical reference standard, a hierarchical parameter nesting construction logic is adopted. In the multidimensional feature space, a broad background benchmark interval for general plants is set, as well as a specific core suitable interval for target groups that is precisely nested within the background benchmark interval.

[0022] In the above technical solution, three major computational operators were designed to address the heterogeneous characteristics of different data types: adaptive tolerance continuous distance penalty (numerical), strict hit matching (categorical), and absolute error decay (probabilistic). In particular, the design of the "hierarchical parameter nesting construction logic" can effectively distinguish between the "universal survival boundary" of the plant kingdom and the "core reproductive boundary" of the target species, thereby achieving an extremely detailed biological characterization of the target group at the algorithm level.

[0023] In some specific embodiments, the pre-trained target classification model is obtained by performing joint training based on a hybrid training dataset that combines virtual and real data; The mixed training dataset consists of positive and negative samples. The ratio of positive to negative samples is dynamically adjusted and controlled within a preset equilibrium range to suppress long-tail class bias during model fitting. The data sources for the hybrid training dataset include large-scale simulated samples generated based on a dual-engine mechanism, as well as real plant fossil confirmation data directly injected through an external structured configuration table. The dual-engine mechanism includes a basic sampling engine and an adversarial perturbation engine. The large-scale simulated samples are specifically composed of a hybrid of baseline simulated samples and adversarial simulated samples. The baseline simulated samples are generated by the basic sampling engine based on the parameter distribution patterns of a preset botanical reference standard, using a random sampling strategy including the Monte Carlo method. The adversarial simulated samples are generated by the adversarial perturbation engine by introducing an adversarial learning mechanism and applying feature offset perturbations to the classification decision boundary.

[0024] The above technical solution constructs a closed-loop hybrid data-driven paradigm: on the one hand, it utilizes a dual-engine mechanism (Monte Carlo to ensure complete distribution and adversarial learning to uncover weaknesses in the decision boundary) to construct an extremely robust "virtual feature ocean" out of thin air; on the other hand, it precisely injects valuable "real-world anchor points" through a structured configuration table. This joint training method combining virtual and real elements enables the target classification model to possess both extremely broad generalization capabilities and extremely accurate entity verification capabilities.

[0025] According to a second aspect of the invention, a computer-readable storage medium is provided on which one or more computer programs are stored, which, when executed by a computer processor, implement the method described above.

[0026] According to a third aspect of the present invention, a paleoecological plant species estimation system based on graph computation and feature fusion is proposed, comprising: The morphological feature acquisition module is configured to acquire the original morphological feature data of the plant fossil sample to be tested. The original morphological feature data includes at least numerical features, categorical features and probabilistic features. The graph topology construction module is configured to construct a graph structure model representing the correlation between plant morphological features based on the original morphological feature data, and generate the corresponding adjacency matrix. The feature extraction and fusion module is configured to map the measured values ​​of the plant fossil sample under test in each feature dimension to the node signal vectors on the graph structure model, extract the graph structure scalar features through graph neighborhood signal aggregation operation, and splice and fuse the graph structure scalar features with the original morphological feature data to generate an extended feature matrix.

[0027] The classification probability prediction module is configured to perform inference processing on the extended feature matrix based on a pre-trained target classification model, and output the predicted probability that the tested plant fossil sample belongs to the target species; wherein, the target classification model is a machine learning model built based on gradient boosting tree; The dual-track determination output module is configured to calculate the range similarity of the tested plant fossil sample on the original morphological feature data based on a preset botanical reference standard; to perform a weighted fusion calculation of the predicted probability and the range similarity to obtain the final similarity; and to determine whether the tested plant fossil sample belongs to the target species based on the final similarity.

[0028] This invention proposes a method and system for estimating paleoecological plant species based on graph computation and feature fusion, which has the following technical effects: 1. This invention models the implicit relationships between morphological features of paleoecological plants as a weighted undirected graph. By introducing core concepts such as safety degree and local variability to aggregate graph neighborhood signals, it successfully transforms complex network topology signals into high-dimensional graph scalar features. This feature, along with the original morphological features, is input into the XGBoost classifier. This approach not only completely overcomes the limitation of traditional machine learning methods that over-rely on isolated physical features while ignoring the coupled symbiotic relationships between features, but also cleverly avoids the overfitting curse that traditional graph neural networks (GNNs) are prone to in small-sample paleontological scenarios. This mechanism can accurately identify feature loss or extreme value anomalies caused by incomplete fossil preservation, achieving robust inference for incomplete specimens.

[0029] 2. This invention innovatively introduces "range similarity" based on botanical reference standards as a physical safety valve, and weights it in fusion with the "model prediction probability" output by the classifier. This dual-track parallel design ensures that the final species diagnosis result not only possesses the sensitivity of underlying AI in uncovering nonlinear correlations, but is also strictly constrained by macroscopic paleobotanical prior rules. The output simultaneously presents model probability and expert consistency scores, effectively intercepting absurd misjudgments with high confidence levels induced by extreme anomalies. This aligns perfectly with the thinking habits and verification needs of paleontologists, significantly enhancing the academic credibility of the identification results.

[0030] 3. Addressing the challenge of the "sparse quantity and extreme imbalance" of fossil samples in a long-tailed distribution, this invention introduces a dual-engine data augmentation strategy during the training phase. This strategy combines Monte Carlo random sampling (constructing the basic distribution framework) with an adversarial learning mechanism (mining weaknesses in the decision boundary), creating an extremely robust and noise-resistant virtual feature space. Simultaneously, it supports the precise injection of small-sample, real-world verification data from expert testing via an external structured configuration table. This joint training paradigm, which dynamically balances the ratio of positive to negative samples, enables the target inference model to possess both extremely broad generalization adaptability and highly accurate physical entity verification capabilities.

[0031] 4. This invention transforms discrete paleontological identification experience into a rigorously executed computer pipeline, connecting the entire process from the standardized configuration of plant hierarchical features, graph topology construction jointly led by natural language and expert priors, feature preprocessing and pipeline cleaning, to joint model fitting and output of the judgment. This fully automated and standardized algorithm architecture completely eliminates the predicament of traditional species inference relying heavily on scholars' personal experience and intuition, greatly improving the analytical efficiency, objectivity, accuracy, and experimental reproducibility of identifying large batches of complex paleontological fossil specimens. Attached Figure Description

[0032] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Many anticipated advantages of the embodiments and other embodiments of the invention will be readily recognized as they become better understood through reference to the following detailed description. Other features, objects, and advantages of this application will become more apparent from reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart of a paleoecological plant species estimation method based on graph computation and feature fusion, according to an embodiment of this application. Figure 2 This is a schematic diagram of the hierarchical parameter setting structure of some features in a botanical reference specification according to a specific embodiment of this application; Figure 3 This is a framework diagram of a paleoecological plant species estimation system based on graph computing and feature fusion, which is a specific embodiment of this application. Figure 4 This is a data generation visualization interface, which is a specific embodiment of this application. Figure 5 This is a training parameter setting interface of a specific embodiment of this application; Figure 6 This is a calculation example diagram of the similarity determination interface in a specific embodiment of this application; Figure 7 This is a schematic diagram of the interactive interface of a positive and accurate prediction system based on target group, which is a specific embodiment of this application. Figure 8 This is a schematic diagram of the interactive interface of an exclusive interception verification system for near-field interference groups, which is a specific embodiment of this application. Detailed Implementation

[0033] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0034] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0035] Figure 1 A flowchart illustrating a paleoecological plant species estimation method based on graph computation and feature fusion according to an embodiment of this application is shown. Figure 1 As shown, the method includes the following steps: S1: Obtain the original morphological feature data of the plant fossil sample to be tested. The original morphological feature data includes at least numerical features, categorical features and probabilistic features.

[0036] In some specific embodiments, before obtaining the original morphological characteristic data of the plant fossil sample to be tested in step S1, a botanical reference standard is pre-constructed to support species inference. In order to comprehensively describe the target tree group (such as Dipterocarpaceae plants and their related groups in the Zhangpu Paleontological Group), in the absence of complete plant genetic data, this embodiment divides the characteristic attribute system of the botanical reference standard into five basic data spaces based on multidimensional extant and fossil data: environmental characteristics, plant morphological characteristics, ecological characteristics, chemical composition characteristics, and geological period characteristics. Specifically, environmental characteristics include paleoclimate and geomorphological parameters that characterize habitats, such as annual mean temperature range, annual precipitation, average humidity, soil pH, and altitude; plant morphological characteristics include observable physical morphological parameters, such as leaf length, leaf width, leaf margin morphology, leaf vein structure, tree height, and root type; ecological characteristics include inferred biological characteristics, such as growth rate, drought resistance, and pollination and dispersal methods; chemical composition characteristics include extractable or inferred macromolecular marker information, such as resin content and terpene content; and geological period characteristics include spatiotemporal background parameters such as geological age, paleotemperature, and paleocarbon dioxide levels.

[0037] Based on the establishment of the aforementioned multidimensional feature attribute space, reasonable reference base parameters are configured for different feature attributes. The benchmark for these parameters is extracted from the statistical average of historical fossils and extant phylogenetic data of target plants (such as Dipterocarpaceae), and compiled into a basic database for artificial intelligence recognition and learning. Taking tree height as an example, it is known that the average tree height of general large tropical trees is between 30 and 50 meters, while typical Dipterocarpaceae trees are usually between 40 and 80 meters tall, also influenced by accompanying shorter varieties (such as Flame Tree, with an average tree height between 10 and 20 meters). Considering the ecological diversity range and extreme variation of the above groups, the feature value boundary for the tree height of the target group is reasonably set to 20 to 90 meters in the reference specification. This process is repeated for all types of data, completing the parameter range and preference attribute settings. Further, the reference... Figure 2 , Figure 2 A schematic diagram illustrating the hierarchical parameter setting structure of certain features in a botanical reference specification according to a specific embodiment of this application is shown. For example... Figure 2 As shown, taking the parameter construction of environmental conditions as an example, the specification not only sets a broad background benchmark range for general plants (e.g., annual average temperature set at 18-32℃, annual precipitation set at 800-5000mm), but also precisely nests the specific core suitability ranges for target groups within the specification (e.g., for Dipterocarpaceae, the core suitable annual average temperature is locked at 24-28℃, and the core suitable annual precipitation is locked at 2000-4000mm). This hierarchical and refined parameter construction mechanism completes the setting of boundary and preference attributes for all types of data, forming a highly structured prior knowledge base.

[0038] Based on the pre-constructed botanical reference standard, step S1 obtains the actual observational data of the current plant fossil sample to be tested. Due to the objective physical state of preservation of paleontological specimens (such as impression fossils, imprint fossils, etc.), the actual raw morphological feature data obtained from the plant fossil sample to be tested is mainly mapped to the plant morphological feature dimension in the reference standard. Specifically, the raw morphological feature data includes at least numerical features (e.g., continuous numerical parameters such as leaf length and width obtained through physical measurements) and categorical features (e.g., discrete state parameters such as smooth or serrated leaf margin types defined through morphological observation); in a preferred embodiment, the raw morphological feature data also includes probabilistic features (e.g., binary state markers characterizing the presence or absence of a specific fruiting organ). The above-mentioned multi-dimensional raw measured data collected from fossil entities, combined with the five-dimensional botanical reference standard, constitute the physical data foundation for subsequent feature similarity comparison and high-dimensional scoring inference using artificial intelligence.

[0039] S2: Based on the original morphological feature data, construct a graph structure model to characterize the correlation between plant morphological features and generate the corresponding adjacency matrix.

[0040] In some specific embodiments, to infer species from plant fossil samples, a graph structure model is constructed based on the original morphological feature data to characterize the relationships between plant morphological features. Specifically, quantifiable plant morphological features involved in graph signal propagation are used as graph nodes; if two morphological features have an intrinsic correlation in plant taxonomy or physiological function (e.g., "leaf shape index" and "leaf vein density" of a specific family and genus have a developmental correlation), then an edge connection is established between the corresponding feature nodes. The edge connection is an undirected weighted edge, and its weight... The preset value range is 0 to 1, used to characterize the correlation strength between two plant morphological features. Furthermore, to ensure the integrity and scientific validity of the graph structure, the source and weight assignment of the edges are constructed using a two-stage merging strategy, thus forming a weighted undirected graph. Specifically, the graph nodes include: numerical feature nodes, used to characterize the continuously quantifiable morphological parameters of the plant fossil, such as leaf length (cm), leaf width (cm), leaf shape index (length-to-width ratio), vein density (veins / cm²), epidermal cell morphological parameters, stomatal length (μm), stomatal density (numbers / mm²), or pollen grain diameter (μm); and probabilistic feature nodes, used to characterize the binary attribute features of the plant fossil, with values ​​of 0 or 1, such as the presence of a petiole, serrated leaf margin, reticulate venation, or stomatal depression. In addition, categorical features (such as leaf margin type encoding, stomatal type encoding, midrib orientation encoding, or leaf tip morphology encoding) do not participate in the construction of the graph structure, but are only input as independent feature dimensions to the subsequent classifier for processing.

[0041] Furthermore, based on the above definitions of nodes and edges, the generated weighted undirected graph is visualized as a [missing information] in the underlying data structure. adjacency matrix Specifically, suppose that the numerical feature nodes and probabilistic feature nodes participating in the graph construction are arranged in a fixed order, and the total number is... Adjacency matrix elements in Represents feature nodes With feature nodes The strength of the association between them (i.e., edge weights) , Since it is an undirected graph, it satisfies symmetry. If there is no direct edge connecting two feature nodes, the corresponding element value is 0. This adjacency matrix... As the core foundation of graph computation in this invention, it will be used to aggregate neighbor feature information on the original feature vectors of each sample and calculate graph structure scalar features such as graph smoothing energy accordingly.

[0042] In some specific embodiments, the graph structure model employs a two-stage merging strategy to generate the adjacency matrix, specifically including: The first stage is the large model construction: The large language model is invoked, and based on the feature names and their botanical descriptions, the relationships and initial weights between features are output. Specifically, after the edge set output by the large language model undergoes node validity verification, the maximum weight of the repeated feature edges is taken and written into the first adjacency matrix. For example, "leaf length" and "leaf width" are strongly correlated because they belong to the same leaf morphology dimension (e.g., a weight of 0.8); "vein density" and "stomatal density" are moderately correlated because they are related in plant water transport function (e.g., a weight of 0.6).

[0043] The second stage involves constructing a backpropagation graph: for feature pairs for which edges were not established in the first stage, backpropagation connections are established based on prior knowledge of plant taxonomy to supplement them. Specifically, numerical feature nodes within the same feature category (such as "leaf morphology", "vein features", or "epidermal structure") are paired with edges and assigned a constant weight (e.g., 0.5); probabilistic feature nodes are grouped into the same category and connected with each other, also assigned a constant weight (e.g., 0.5), thereby generating a second adjacency matrix. .

[0044] Merging Strategy: To ensure all feature nodes have a reasonable association structure, the maximum value of corresponding elements in the first and second adjacency matrices is taken, and the matrices are merged to generate the final adjacency matrix. Mathematically, let's assume... The feature nodes are arranged in a fixed order, resulting in the adjacency matrix. The merging rules are expressed as follows: Furthermore, the final adjacency matrix satisfies the symmetry constraint, meaning that the elements in the matrix... , representing nodes With nodes The association weight; if there is no direct edge connection, then The final adjacency matrix and graph structure features generated by the merger can be persistently saved (e.g., as GraphArtifact graph artifact data), which contains a list of numerical feature names, an adjacency matrix, and edge data to ensure the consistency of the underlying graph structure during subsequent graph signal aggregation and scalar feature extraction for each fossil sample.

[0045] In a specific embodiment, the first stage is large model construction: A large language model is invoked, and based on the names and descriptions of each feature, JSON data containing the relationships between feature edges is output. The data structure format is edges:[{source, target, weight}]. The output of the large language model is parsed and validated, retaining only valid node pairs. For repeatedly generated edges, the maximum weight is extracted and written into the first adjacency matrix. The second stage involves constructing the backplane graph: based on the business attributes of the features, pairs of numerical feature nodes with the same category field are connected with edges and assigned a weight constant of 0.5; simultaneously, probabilistic features are separately categorized into the probability category and their internal nodes are interconnected, thereby generating the second adjacency matrix. Merging rule: The adjacency matrices generated in the above two stages are subjected to an element-wise maximum operation to generate the final adjacency matrix. The merging rule is expressed as follows: Subsequently, the final adjacency matrix and graph structure features generated by the merger will be persistently saved as a GraphArtifact object. This object fully encapsulates core data, including a list of numeric feature names, an adjacency matrix, and edge data, for unified use in subsequent model training and prediction phases.

[0046] S3: Map the measured values ​​of the plant fossil sample to be tested in each feature dimension to the node signal vector on the graph structure model, and extract the graph structure scalar features through graph neighborhood signal aggregation operation; splice and fuse the graph structure scalar features with the original morphological feature data to generate an extended feature matrix.

[0047] In some specific embodiments, step S3 includes: S31, for each plant fossil sample to be tested, the quantized values ​​(i.e., original morphological measurements) of its feature dimensions involved in the graph composition are mapped to node signal vectors on the graph structure model. (in (total number of nodes), the first of the signal vectors The component corresponds to the fossil sample in the first... The values ​​at each feature node are taken. Specifically, for numerical features, the standardized measured values ​​are taken, and missing values ​​are treated as 0; for probabilistic features, real numbers of 0 or 1 are directly used in the calculation.

[0048] S32, to perform graph neighborhood information aggregation, record the nodes. The safety factor is ,in These are the elements of the adjacency matrix. To avoid division by zero errors during calculation, a very small positive number is introduced. (For example (Order level), defining the node security degree. (If calculated) If , then 1.0 is used instead. Based on this, the computing node Neighbor weighted mean ,in For nodes The signal value, the weighted mean of the neighbors reflects the node's value. The average level of its neighborhood features after comprehensively considering the topological associations of features.

[0049] S33 extracts the following five fixed-dimensional graph structure scalar features as a bridge between graph topological information and subsequent classifier fusion: graph smoothing energy : This feature measures the degree of difference in values ​​for associated morphological features among fossil samples. The lower the energy, the more consistent the values ​​of the associated features, and the more the fossil morphology conforms to botanical laws; the higher the energy, the more incomplete the fossil preservation or the existence of contradictions between features. Global average of neighbor aggregation means : This feature reflects the average level of all morphological features of the fossil sample after the graph structure is smoothed, and can be used as a comprehensive representation of the overall morphology. Dispersion of neighbor aggregation mean : (when When the dispersion is defined as 0, this feature measures the degree of difference between feature nodes after mixing with their neighbors. High dispersion indicates significant differences in neighborhood information between different features, reflecting the heterogeneity of fossil features. Degree-weighted features and : This feature makes the contribution of high-number nodes (i.e., morphological indicators with dense correlation features) to the overall signal greater, reflecting the dominant role of core morphological features in fossil identification. Local variability : This feature measures the average deviation of each feature value from the average value of its neighborhood, reflecting the degree of anomalousness of the fossil sample in the local feature association structure, and is used to identify specimens that are incompletely preserved or morphologically specialized.

[0050] The five extracted graph structure scalar features are concatenated with the original morphological feature data to form an extended feature matrix. This feature transformation mechanism no longer retains the explicit multi-layer graph neural network structure, but instead adopts a single-layer neighborhood aggregation plus scalar statistics processing method, which facilitates the efficient combination of graph structure features with classic table learners such as gradient boosting trees.

[0051] In a specific implementation, firstly, for each fossil sample in the dataset to be tested (i.e., a single row of data in the data table), a numerical vector is constructed that is completely consistent with the order of the graph nodes. Among them, the first of the numerical vectors Each component corresponds to the fossil sample in the feature list. The actual values ​​are taken on the matrix; during the data mapping process, if missing values ​​or non-numerical types are detected, they are uniformly treated as 0 to ensure the integrity of the matrix dimensions; while for probabilistic features, real numbers of 0 or 1 are directly extracted for subsequent calculations. Next, based on the core data structure generated in the graph construction phase—the final adjacency matrix... Calculate the node degree vector , its first Each component is represented as To ensure computational robustness when dealing with isolated feature nodes (i.e., to avoid fatal exceptions caused by division by zero in the underlying code), a very small constant is introduced. (In actual code execution) (magnitude), defining the degree of security. When it is determined that a node is an isolated node without any connected edges (i.e. When ), force the denominator to be replaced with 1.0. Then, calculate the neighbor weighted mean vector node by node, where the ... The weighted average of neighbors of each feature node The calculation formula is defined as follows: This value mathematically precisely reflects the node After comprehensively considering all topological feature associations, the average baseline level of its neighborhood features is determined. Based on this numerical calculation, the five graph structure scalar features extracted above (i.e., to The feature matrix is ​​concatenated with the original table feature columns to form an extended feature matrix. This feature fusion mechanism no longer retains the explicit multi-layer graph neural network (GNN) message passing structure, but instead adopts a processing method that combines single-layer neighborhood aggregation and scalar statistics, which greatly facilitates the seamless and efficient combination of graph topological prior information and classic table learners.

[0052] S4: Based on the pre-trained target classification model, perform inference processing on the extended feature matrix and output the predicted probability that the tested plant fossil sample belongs to the target species; wherein, the target classification model is a machine learning model built based on gradient boosting tree.

[0053] In some specific embodiments, the target classification model is an XGBoost classifier. Before inputting the extended feature matrix into the pre-trained XGBoost classifier, the following feature preprocessing process is also included: A feature preprocessing pipeline is constructed to perform one-hot encoding on the categorical features in the extended feature matrix and configure an unknown category ignoring strategy; standardization is performed on the numerical features in the extended feature matrix to make the mean of each feature component 0 and the variance 1; the preprocessed feature components are concatenated to generate a high-dimensional feature vector that is input to the XGBoost classifier.

[0054] In some specific embodiments, deep integration and inference of the classifier are performed based on the extracted extended feature matrix, and its end-to-end workflow specifically includes a model training stage and a model prediction stage.

[0055] Specifically, during the model training phase, a plant fossil training dataset (df) with species classification labels (e.g., the binary classification target field is_dipterocarpaceae_like, i.e., "whether it is similar to Dipterocarpaceae") is first acquired. Then, the graph structure model generated based on the aforementioned steps is loaded (this model is persisted as a graph artifact object, i.e., GraphArtifact, containing a list of numerical feature names, an adjacency matrix, and edge metadata). For each row of fossil samples in the training dataset, signal vectors are extracted strictly according to node order, and the adjacency matrix is ​​used to calculate... to Five graph structure scalar features, which are uniformly named gf_ Add the series columns (e.g., gf_smoothness, gf_nb_mean_avg, etc.) to the data frame.

[0056] Furthermore, a feature preprocessing pipeline (ColumnTransformer) is constructed for different data types. For categorical features (such as leaf edge type, leaf tip morphology, etc.), a one-hot encoder is called for processing, and the parameter handle_unknown='ignore' is explicitly set to automatically ignore unknown categories; for features containing original morphological measurements and the aforementioned graph structure scalar features (i.e., the aforementioned gf_ The numerical features of the series are normalized using a StandardScaler to ensure that the mean of each feature is 0 and the variance is 1. Then, a data processing pipeline is constructed, sequentially connecting the preprocessing pipeline with an XGBoost classifier. In this architecture, the XGBoost classifier fits an additive tree model only to the high-dimensional real-valued feature vectors generated after preprocessing; the topological prior information inherent in the graph structure model has already been processed... to These five scalar features are indirectly injected into the real number vector, thereby achieving a decoupled and efficient fusion of graph structure prior information and discriminative model, without having to repeatedly build the graph network inside the classifier.

[0057] During the model prediction phase, the previously trained and saved data processing pipeline and the same set of graph artifacts used during training are loaded. For the input plant fossil sample to be tested, only its original morphological feature columns are obtained, and the extracted features are recalculated strictly according to the same node order and adjacency matrix used during training. to Five graph-structured scalar features. Subsequently, the spliced ​​extended features are fed into the data processing pipeline for unified feature transformation and inference, and finally output the classification prediction probability that the fossil sample belongs to the positive class (i.e., is determined to be a species similar to Dipterocarpaceae).

[0058] Furthermore, in the above classification process, the classification prediction probability value output by the XGBoost classification model is obtained by smoothly mapping the original score output by the model to the [0,1] interval using a logistic regression function. The mathematical formula for calculating the classification prediction probability value is as follows: In the formula, The model prediction probability output by the XGBoost classifier, i.e., given the input feature vector. In this case, it represents the probability that the sample belongs to the positive class (e.g., belonging to the target paleoecological plant species). This value is mapped to the interval [0,1]. This is a high-dimensional feature vector that is transformed by the feature processing pipeline and then input to the classifier. for The weighted sum of the regression trees, i.e., the constructed additive model, is calculated using the following formula: In the formula, This indicates the number of basic learners set (i.e., the number of iteration rounds / the number of trees); Indicates the first CART regression trees for the feature vector The output score is given (i.e., the score of the sample falling into the leaf node). This represents the preset shrinkage step size (i.e., learning rate), which serves as a scaling factor for the contribution of each tree to prevent the model from overfitting.

[0059] Furthermore, to efficiently fit and classify the aforementioned extended feature matrix, specific configuration optimizations were performed on the core hyperparameters of the XGBoost classification model. Specifically, for binary classification tasks, the model's objective function was set by default to a binary loss based on logistic regression. The number of boosting epochs (n_estimators) was set to 300 to ensure sufficient fitting capability; the maximum depth of a single regression tree (max_depth) was set between 3 and 10 (specifically 6 in some preferred embodiments) to reasonably control the order of plant morphological feature interactions and the overall model complexity; the learning rate (i.e., the eta parameter) was set between 0.01 and 0.3 (specifically 0.1 in some preferred embodiments) to limit the contribution ratio of each tree to the cumulative output result, thereby effectively preventing overfitting. To enhance the model's generalization ability and effectively reduce the correlation between multiple trees, the system employs a random sampling strategy. The sampling ratio (subsample) of randomly used sample rows in each training iteration is set to 0.6 to 0.9 (specifically 0.8 in some preferred embodiments), and the sampling ratio (colsample_bytree) of randomly used feature columns in each tree is also set to 0.6 to 0.9 (specifically 0.8 in some preferred embodiments). Furthermore, regarding model training monitoring and underlying computation mechanisms, the model uses log loss as an evaluation metric to monitor the calibration quality of predicted probabilities in real time. A histogram approximation splitting algorithm (tree_method set to hist) is used as the method for constructing the underlying trees, significantly accelerating the training computation process for large-scale paleontological fossil data. Simultaneously, a global random seed (e.g., random_state=42) is set to ensure the reproducibility and stability of the entire training and splitting process. In the underlying fitting process of the XGBoost classification model, each regression tree employs a greedy splitting algorithm during training. The optimal splitting node is found by minimizing the second-order Taylor expansion approximation objective function, which includes the first and second derivatives (g and h) of the loss function. Simultaneously, an L2 regularization term for the leaf node weights is introduced into the objective function. This L2 regularization mechanism, along with previously set hyperparameters such as the maximum depth (e.g., max_depth=6), works synergistically to control the overall model capacity of the underlying model. This underlying mathematical optimization mechanism effectively suppresses overfitting, which is easily induced by the sparse and noisy nature of paleontological fossil sample data, while ensuring the fitting accuracy of high-dimensional morphological features.

[0060] In some optional embodiments, considering the extreme imbalance between positive and negative sample classes that may exist in actual paleontological fossil sample collection (i.e., the number of target species fossils is far less than that of non-target species), the XGBoost classification model can also be configured with a positive sample weight scaling parameter (scale_pos_weight). This parameter is dynamically adjusted according to the distribution ratio of positive and negative samples to increase the penalty for misclassification of rare positive samples during gradient updates, thereby effectively mitigating the prediction bias problem caused by class imbalance and further improving the model's generalization and recognition accuracy in extreme fossil excavation scenarios.

[0061] S5: Based on the preset botanical reference standard, calculate the range similarity of the plant fossil sample to be tested on the original morphological feature data; calculate the final similarity by weighting and fusing the predicted probability and the range similarity, and determine whether the plant fossil sample to be tested belongs to the target species based on the final similarity.

[0062] In some specific embodiments, step S5 specifically includes: S51, calculate the range similarity independent of the model classifier to quantitatively measure the consistency between the tested fossil sample and the prior knowledge of botanical experts (i.e., the positive class reference range). It should be noted that the aforementioned extracted graph structure scalar features are not included in this range similarity calculation to ensure that this indicator purely reflects the degree of consistency between the sample and the objective botanical knowledge base. The range similarity is calculated component-wise for the three types of features in the original morphological feature data of the tested plant fossil sample: For numerical features, let the positive class reference interval given by the expert or the training set (i.e., let the positive class reference interval given in the botanical reference specification be) be... Set the interval width ,in For the preset minimum interval tolerance, if the measured value of the feature is Its numerical similarity components The calculation formula is: In the formula, The boundary of the allowed value range given by similar samples or experts; This represents the measured value of the current sample to be judged on this feature; The normalization scale. When the measured value Falling into the reference range specified by the experts At the specified point, the model assigns a 100% similarity score (value 1); once the boundary is exceeded, the similarity score is adjusted based on the distance from the boundary. relative interval width The ratio decreases linearly and smoothly towards 0. This mechanism ensures that when the measured value deviates from the reference range by a full interval width... When the similarity decreases smoothly to 0, the similarity decreases to 0.

[0063] For the categorical features, let the set of preferred values ​​specified by experts (i.e., the set of preferred values ​​specified in the botanical reference standard) be . (For example, the preference set for leaf margin type is "serrated margin" or "entire margin"), if the feature observation value is Its categorical similarity components Calculated based on the exact matching principle: .

[0064] For probabilistic features (with values ​​of 0 or 1), let the positive class reference probability be... The characteristic observation value is Its probabilistic similarity components The formula for the absolute deviation attenuation is: .

[0065] S52, after calculating the similarity components of each morphological feature, let the total number of features actually extracted and validly existing on the fossil sample to be tested be... The total range similarity is obtained by taking the arithmetic mean of all the aforementioned feature components. ,but In the formula, The overall range similarity measures the degree of consistency between the current plant fossil sample being tested and the prior reference range of the target species set by experts across all extracted valid features. This indicates the total number of features that actually exist. Since paleontological fossils are often incomplete during preservation (e.g., only the leaves are preserved but the petioles are lost), this parameter specifically refers to the number of morphological features that were actually collected from the current sample to be tested and participated in the similarity calculation. The traversal index number represents the feature and is used to refer to the first feature in the cumulative summation operation. Each feature has a value ranging from 1 to... . Indicates the first The similarity component of each feature. Depending on the data type of the feature, it corresponds to the numerical similarity component calculated above. Classification similarity components or probabilistic similarity components The specific calculation results are shown in the text. When the sample is extremely incomplete... At that time, the Defined as 0.

[0066] S53, to balance the discriminative power of the machine learning model with the stability of expert rules, integrates the model output with range similarity to calculate the final similarity of the test samples. The calculation formula is as follows: In the formula, The model prediction probability that the tested plant fossil sample belongs to the target plant group is output by the target classification model; The fusion coefficient is a preset fusion coefficient between model prediction and expert knowledge. The value is set to 0.7, which assigns 70% decision weight to the model's prediction and 30% decision weight to the consensus of expert rules; this coefficient can also be adjusted based on the degree of preservation of the fossil being tested. Make dynamic adjustments.

[0067] S54, based on the established binary discrimination rule, completes the final species classification inference. Given the minimum comprehensive similarity threshold for class determination... (This application preferably sets it to 0.6-0.7), if it meets the requirements If the tested plant fossil sample belongs to the target species (same type); If the threshold is not met, it is determined that the species does not belong to the target species (not of the same kind). It can be flexibly and adaptively configured according to the business needs regarding the trade-off between model interpretability and false positives and false negatives.

[0068] Furthermore, regarding the calculation logic for the range similarity, this embodiment employs a decoupled dual-track verification mechanism. Specifically, the graph structure scalar features (i.e., those calculated via the graph structure model) mentioned above... to Abstract topological feature components are explicitly excluded from the feature traversal calculation of the range similarity. This design ensures that the range similarity only reflects the consistency between the original physical morphological features of the tested plant fossil sample and the expert business specification table (i.e., the objective botanical taxonomy knowledge base). Through this decoupling process, while utilizing a high-dimensional extended matrix containing graph topological information to ensure the high-dimensional fitting accuracy of the classifier (such as XGBoost), the expert rule verification of purely physical features also guarantees the human interpretability and biological rationality of the final judgment result.

[0069] Continue to refer to Figure 3 As an implementation of the above method, this application provides an embodiment of a framework diagram 300 for a paleoecological plant species estimation system based on graph computation and feature fusion. This system embodiment is similar to... Figure 1 Corresponding to the illustrated method embodiment, this system can be specifically applied to various electronic devices. The system 300 includes a morphological feature acquisition module 301, a graph topology construction module 302, a feature extraction and fusion module 303, a classification probability prediction module 304, and a dual-track decision output module 305, all interconnected. The morphological feature acquisition module 301 is configured to acquire the original morphological feature data of the plant fossil sample to be tested. The original morphological feature data includes at least numerical features, categorical features and probabilistic features. Graph topology construction module 302 is configured to construct a graph structure model representing the correlation between plant morphological features based on the original morphological feature data, and generate a corresponding adjacency matrix. The feature extraction and fusion module 303 is configured to map the measured values ​​of the plant fossil sample to be tested in each feature dimension to the node signal vector on the graph structure model, extract the graph structure scalar features through graph neighborhood signal aggregation operation, and splice and fuse the graph structure scalar features with the original morphological feature data to generate an extended feature matrix.

[0070] The classification probability prediction module 304 is configured to perform inference processing on the extended feature matrix based on a pre-trained target classification model, and output the predicted probability that the tested plant fossil sample belongs to the target species; wherein, the target classification model is a machine learning model constructed based on a gradient boosting tree; The dual-track determination output module 305 is configured to calculate the range similarity of the tested plant fossil sample on the original morphological feature data based on a preset botanical reference standard; to perform a weighted fusion calculation of the predicted probability and the range similarity to obtain the final similarity; and to determine whether the tested plant fossil sample belongs to the target species based on the final similarity.

[0071] Example The paleoecological plant species estimation method based on graph computation and feature fusion provided by this invention is modularly divided into a model training stage and a model prediction stage at the system's underlying engineering implementation level. Its complete end-to-end workflow is as follows: First, the model training phase (corresponding to the graph feature construction and classifier fitting module). Step 1: Loading Preset Specifications. First, obtain and load the user-configured or externally imported botanical reference specifications (marked as "spec" in the underlying configuration dictionary). These specifications pre-define the numerical reference ranges for each morphological feature, the set of preference options for categorical attributes, and the standard thresholds for probabilistic features, among other prior expert knowledge.

[0072] Step 2: Heterogeneous Graph Topology Construction. If the graph feature computation branch is detected as enabled, an initial base backtracking graph is generated based on the aforementioned botanical reference specification (spec). Further, if the system is configured with an external Large Language Model (LLM) interface, the edge set information returned by the LLM-assisted inference is parsed, and it is merged with the adjacency matrix of the base backtracking graph using a max-pooling / element-wise max operation. Subsequently, the fused complete graph network structure is serialized into a graph artifact object containing a node list and adjacency matrix information and persistently saved.

[0073] Step 3: Data Augmentation and Balanced Training Set Construction Based on a Dual-Engine Mechanism. Addressing the technical challenge of scarce real-world paleontological fossil sampling data, a data generation visualization interface (such as...) is provided on the front end. Figure 4 As shown in the figure, deep data augmentation is performed at the underlying level. When constructing the training set, the macroscopic sample generation scale and class distribution target are first set. For example, a large-scale (e.g., 5000) simulated samples are automatically generated, and a dynamic adjustment mechanism is used to control the positive class samples (e.g., Dipterocarpaceae) and negative class samples (other associated trees) within a preset equilibrium range (e.g., a 1:1 equilibrium ratio or a 1:3 bias ratio) to prevent long-tailed class bias from occurring in subsequent model fitting.

[0074] To achieve the aforementioned high-quality and highly balanced data expansion, the underlying engine employs a dual-engine strategy that integrates Monte Carlo random sampling and Adversarial Learning. Specifically, the basic sampling module is first activated. Based on the parameter distribution patterns in the prior knowledge base, and combined with oversampling and downsampling mechanisms, a large number of "benchmark simulated samples" conforming to conventional statistical laws are generated to construct the main framework of the data distribution.

[0075] Subsequently, the adversarial perturbation module is activated. Based on the user-configured "adversarial strength" (e.g., set to 0.10), an adversarial generative network (GAN) or gradient-directed perturbation algorithm is introduced to generate "adversarial simulated samples" (i.e., difficult samples) specifically targeting decision boundaries where the classifier is prone to confusion. Taking specific environmental and morphological features as an example: if the base module generates a very typical positive class baseline sample with an average annual temperature of 26℃ and a leaf width of 10cm, the adversarial module will apply a small but deceptive feature shift (e.g., slightly perturb the temperature to the marginal value of 23.9℃ while retaining the typical leaf width), thereby deliberately generating "difficult negative samples" or "marginal positive samples" that approach the classification threshold.

[0076] Finally, the two types of samples are mixed and spliced ​​according to a macroscopically preset balanced ratio to automatically generate a structured training data table with clear species classification labels. This mechanism not only ensures sufficient balance in the global distribution and category ratio of the training data, but also forces the subsequent high-dimensional classification model (XGBoost) to learn a more sensitive and noise-resistant physical decision boundary, greatly improving the model's generalization robustness when dealing with real and incomplete paleontological fossil specimens.

[0077] Furthermore, the system supports a hybrid data-driven paradigm that combines virtual and real data in the construction of the training data stream. As shown in Table 1, the underlying layer maintains a highly structured feature parameter table for target groups (such as Dipterocarpaceae), clearly defining the feature categories, feature names, variable types (continuous or categorical), value boundaries (minimum and maximum values) of multidimensional features, and candidate categorical value sets for categorical variables. Based on the aforementioned generation of large-scale random simulated samples using this parameter table, the system also supports the direct injection of real data. Specifically, users or experts can modify the system's preset structured configuration table to include real Dipterocarpaceae fossil data that has been discovered and clearly possesses many morphological features. The large-scale simulated samples and small-sample real-confirmed data will be merged, and the same underlying feature transformation and joint training logic will be executed. By comprehensively utilizing the large-scale generalization exploration of random data and the precise anchoring and fine-tuning of real data, the barrier of scarcity of real data in the paleontology field can be significantly overcome, significantly enhancing the robustness and inference accuracy of the final model on complex samples.

[0078] Table 1: Example of Characteristic Parameter Configuration in Reference Specifications for Dipterocarpaceae Botany

[0079] Furthermore, as shown in Table 2, when domain experts obtain highly complete real plant characteristics (including the target group Dipterocarpaceae and its associated comparative species), they can batch inject measured values ​​through this multidimensional structured feature matrix format. The system will automatically parse multiple sets of real confirmed data in Table 2, strictly align their feature dimensions, and stitch them into a large-scale virtual sample set generated by the rules in Table 1, and send them together into the preprocessing pipeline. It is worth noting that non-target associated groups such as Apocynaceae and Arecaceae are deliberately introduced in Table 2 as real negative class samples. Some of their physical characteristics (such as leaf size up to 500cm, genus names not registered in Table 1, etc.) will far exceed the positive class target reference specifications set in Table 1. When this part of the data is input into the system, it will perfectly trigger the system's set range similarity penalty mechanism and unknown category ignoring strategy. This is the core underlying logic of this system to achieve highly robust exclusion determination of target species in extremely similar ancient tropical rainforest associated communities.

[0080] Table 2: Structured Interactive Import Table for Receiving Real Fossil Confirmation Data

[0081] Step 4: Graph Feature Extraction and Expansion. For each row of samples in the training data table, the original node signal vectors are extracted strictly according to the graph node arrangement order. Then, the adjacency matrix from the previously generated graph artifact object (GraphArtifact) is used to perform neighborhood signal aggregation calculations to extract the relevant features. to Five scalar features of the graph structure. These five scalar features are then used as entirely new extended columns (e.g., gf_). The series of feature columns are added to the training data table.

[0082] Step 5: Feature Derivation and Preprocessing Pipeline Construction. To further enhance the dimensionality of the input feature space and the upper limit of model fitting, the derived feature construction logic is executed first before performing feature separation on the expanded training data table. Specifically, higher-order biological and ecological indices are calculated based on the acquired basic original features, including but not limited to: constructing "leaf length-to-width ratio" and "leaf area estimation" based on the mathematical relationship between leaf length and leaf width; estimating "trunk volume" and "height-to-diameter ratio" based on trunk diameter at breast height and tree height; constructing a "climate stress index" characterizing habitat stress based on annual temperature difference and annual precipitation; constructing a "resin quality index" based on the product feature of resin content and terpene content; and constructing a comprehensive "ecological adaptability index" by calculating the arithmetic mean based on tolerance features such as drought resistance, flood resistance, and fire resistance. After completing the above feature derivation, a separation operation is performed on the training data table containing basic features, derived features, and graph structure scalar features to extract the expanded feature matrix and classification labels. Before entering the feature preprocessing pipeline, the training parameter setting module (e.g., ...) is used to... Figure 5 (As shown) Automated data cleaning and feature matrix expansion are performed on the separated data. An automatic removal strategy is implemented for data rows containing severely missing values. Furthermore, the Interquartile Range (IQR) algorithm is used for outlier detection and removal. By calculating the first quartile (Q1) and the third quartile (Q3), outliers are automatically filtered out. Extremely variable noise data outside physical boundaries is used to ensure the purity of the data distribution fed into the fitting pipeline. Subsequently, a feature preprocessing pipeline (ColumnTransformer) is constructed for the extended feature matrix, performing standardization and normalization on the numerical features, one-hot encoding on the categorical features, and configuring an unknown category ignoring strategy.

[0083] Step Six: Hierarchical Cross-Validation and Grid Search Hyperparameter Tuning. Instantiate the target classification model (such as an XGBoost classifier with preset core hyperparameters) and combine it with the aforementioned feature preprocessing pipeline to form a complete data processing pipeline. During model training, to more accurately evaluate the model's generalization performance and avoid the risk of overfitting, the system employs hierarchical cross-validation. Cross-validation mode (e.g., configuration) That is, the configuration parameter skf=StratifiedKFold(n_splits=cv,...)). Alternate between these parameters. The system performs pipeline fitting on a subset of data, and then uses the remaining subset for performance evaluation (calculating metrics such as 'accuracy', 'precision', 'recall', 'f1', and 'roc_auc') to comprehensively assess the model's generalization ability. Based on this, grid search cross-validation (configured by `grid_search=GridSearchCV(...,cv=3,scoring='roc_auc',...)`) is used for global hyperparameter tuning. After hyperparameter tuning, the optimal parameter combination is used to perform end-to-end joint fitting of the data processing pipeline on the extended feature matrix and classification labels. Simultaneously, at the model's classification output, the default classification threshold is dynamically set to 0.6 (i.e., when the predicted probability output by the model is greater than or equal to 0.6, it is classified as a positive target species). Furthermore, the system supports importing newly added real plant confirmation data through external configuration files, enabling continuous incremental updates to the basic training set and model iteration.

[0084] Step 7: Model artifact serialization. After training is complete, the fitted data processing pipeline, botanical reference specification, graph artifact, and their associated metadata are packaged, serialized, and saved as a complete model artifact for use by the online prediction service.

[0085] Second, the model prediction stage (corresponding to the feature reasoning and expert rule fusion module). Step 1: Model Artifact Restoration. Load the saved model artifacts and deserialize and restore the pre-trained data processing pipeline, botanical reference specification, and graph artifact in runtime memory.

[0086] Step 2: Input of the test sample. Receive the original morphological feature columns of the plant fossil sample to be tested, either input by the user or imported in batches from external sources. At this point, the feature dimensions are strictly aligned with the reference specification used during training and do not yet include graph structure features.

[0087] Step 3: Real-time Graph Feature Calculation. If a recovered graph artifact is detected, the adjacency matrix and node order rules, identical to those used in the training phase, are used to extract the graph structure scalar features (gf_) for the test sample in real time. (Series of features), and then horizontally merge them into the original feature table to generate an extended feature matrix.

[0088] Step 4: High-dimensional classifier inference. The data processing pipeline is invoked (executing the `Pipeline.predict_proba` method), feeding the extended feature matrix into the pipeline for unified normalization and one-hot processing. Finally, the underlying classification model outputs the model's predicted probability that the test sample belongs to the positive target species. .

[0089] Step 5: Expert Rule Fusion Calculation. Using the restored botanical reference specification (spec), the range similarity calculated independently of the graph is calculated for each sample. Subsequently, the decoupling and fusion logic is executed to calculate the model's predicted probability and range similarity. The final similarity is obtained by combining the convex components. (For example, execute) (Weighted operation).

[0090] Step Six: Binary Classification Determination. The calculated final similarity score... The system compares the results with the user-preset or system-default comprehensive similarity threshold and outputs the final judgment result (e.g., "same type" or "different type").

[0091] To more intuitively illustrate the calculation logic of the dual-track fusion and threshold determination, combined with Figure 6 The similarity determination interface shown in the diagram illustrates the following operation: Assume the system receives input data from a fossil sample to be tested. After preprocessing and graph feature expansion, the data is fed into a pre-trained XGBoost classifier, which outputs the model's predicted probability of belonging to the target class. Simultaneously, the system calculates the range similarity of its original physical characteristics based on the aforementioned botanical reference specifications. If the system sets the weighted fusion coefficient to (70% model weighting and 30% range weighting), then the final similarity will be... The system extracts the dynamic judgment thresholds set in the human-computer interaction interface. ,because The system ultimately outputs the conclusion that "it is impossible for (different classes) to exist." This example fully demonstrates that the decoupling mechanism can effectively combine expert rules to provide a comprehensive and reasonable judgment output when the model prediction has extremely low confidence.

[0092] refer to Figure 7 and Figure 8 , Figure 7 and Figure 8 The diagrams show the interactive interfaces of the verification system for positive and accurate prediction based on the target species group and exclusive interception of its closely related interfering species, according to embodiments of this application. This application embodiment fully demonstrates the system's dual-track discrimination logic when facing "target species hit" and "closely related species interception" through a set of highly representative comparative experiments. First, as... Figure 7 As shown, the system performed positive validation on the real "[Dipterocarpaceae] Tokyo Dipterocarpium". The input parameters for this sample, such as tree height of 45.0 and average annual temperature of 23.0, highly matched the target specifications. Calculations showed that the model's predicted probability reached a high confidence level of 1.000, the physical range similarity was 0.886, and the final fusion similarity was as high as 0.966, with the system accurately outputting a "same class" judgment. This result fully demonstrates the extremely high sensitivity of this invention in identifying target groups. Subsequently, as... Figure 8 As shown, the system performed an exclusive interception test on the closely related species "[Dipterocarpaceae] Yunnan Poluo Shuang". Due to the high taxonomic similarity between this species and the target species, their physical range similarity reached 0.829. Using traditional single-track rules for judgment would easily lead to false positives. However, thanks to the high-dimensional topological feature recognition capability introduced in this invention, the model's predicted probability was only 0.017. Under the dual-track fusion mechanism, the final similarity was corrected to 0.261, far below the judgment threshold, and the system accurately gave a "dissimilar" judgment. Therefore, this invention not only achieves high-precision positive identification but also effectively intercepts "difficult negative samples" with highly similar physical characteristics by relying on the model's ability to capture fine-grained features, greatly enhancing the anti-interference ability and judgment robustness of paleoecological species estimation in complex symbiotic communities.

[0093] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for estimating paleoecological plant species based on graph computation and feature fusion, characterized in that, include: S1, Obtain the original morphological feature data of the plant fossil sample to be tested. The original morphological feature data includes at least numerical features, categorical features and probabilistic features. S2, Based on the original morphological feature data, construct a graph structure model to characterize the correlation between plant morphological features and generate the corresponding adjacency matrix; S3, map the measured values ​​of the plant fossil sample to be tested in each feature dimension to the node signal vector on the graph structure model, and extract the graph structure scalar features through graph neighborhood signal aggregation operation; The graph structure scalar features are concatenated and fused with the original morphological feature data to generate an extended feature matrix; S4, based on the pre-trained target classification model, the extended feature matrix is ​​inferred and processed to output the predicted probability that the tested plant fossil sample belongs to the target species; wherein, the target classification model is a machine learning model built based on gradient boosting tree; S5. Based on the preset botanical reference standard, calculate the range similarity of the plant fossil sample to be tested on the original morphological feature data; calculate the final similarity by weighting and fusing the predicted probability and the range similarity, and determine whether the plant fossil sample to be tested belongs to the target species based on the final similarity.

2. The paleoecological plant species estimation method based on graph computation and feature fusion according to claim 1, characterized in that, In step S2, the graph structure model uses a two-stage merging strategy to generate the adjacency matrix, specifically including: Phase 1: Invoke the large language model, output the correlation relationship and initial weight between features based on the feature names and their descriptions; perform node validity verification on the correlation relationship output by the large language model, and extract weights for repeatedly generated feature edges based on a preset optimization strategy to generate the first adjacency matrix. The second stage: For feature pairs that did not have edges established in the first stage, backoff connections are established based on prior knowledge of plant taxonomy; specifically, this includes: establishing connections between feature nodes belonging to the same business or attribute category, assigning preset default weights, and generating a second adjacency matrix; Merging strategy: Based on a preset matrix fusion rule, the first adjacency matrix and the second adjacency matrix are merged by corresponding elements to generate the final adjacency matrix. The preset matrix fusion rule includes taking the maximum value of the corresponding position weight.

3. The paleoecological plant species estimation method based on graph computation and feature fusion according to claim 1, characterized in that, Step S3 includes: S31, Computation Node Safety level ,in For nodes The degree, For the corresponding adjacency matrix elements, It is a preset minimum positive number; S32, Computation Node Neighbor weighted mean ,in, For nodes The value of on the signal vector; S33, Based on the security degree and the weighted mean of the neighbors, extract the scalar features of the graph structure; S34, the graph structure scalar features are concatenated with the original morphological feature data to generate an extended feature matrix.

4. The paleoecological plant species estimation method based on graph computation and feature fusion according to claim 3, characterized in that, The graph structure scalar features include graph smoothing energy, global average of neighbor aggregation mean, dispersion of neighbor aggregation mean, degree-weighted feature, and local variability, wherein the graph smoothing energy is: The global average of the neighbor aggregation mean: The dispersion of the aggregated mean of the neighbors: The degree-weighted features are as follows: The local variability: In the formula, For nodes The value taken on the signal vector, The total number of the numerical feature nodes and the probabilistic feature nodes participating in the graph construction.

5. The paleoecological plant species estimation method based on graph computation and feature fusion according to claim 1, characterized in that, The target classification model is an XGBoost classifier. Before inputting the extended feature matrix into the pre-trained XGBoost classifier, the following feature preprocessing process is also included: A feature preprocessing pipeline is constructed to perform one-hot encoding on the categorical features in the extended feature matrix and configure an unknown category ignoring strategy; standardization is performed on the numerical features in the extended feature matrix to make the mean of each feature component 0 and the variance 1; the preprocessed feature components are concatenated to generate a high-dimensional feature vector that is input to the XGBoost classifier.

6. The paleoecological plant species estimation method based on graph computation and feature fusion according to claim 5, characterized in that, In step S4, the model prediction probability output by the XGBoost classifier is obtained by mapping the original score of the model output to the [0,1] interval using a logistic regression function. The calculation formula is as follows: In the formula, The model prediction probability output by the XGBoost classifier; This is a high-dimensional feature vector that has been transformed through the feature preprocessing pipeline and then input into the classifier; In the classifier The weighted sum of the regression trees is calculated using the following formula: ,in, For the number of iteration rounds, For the specific first A tree of return, For the first Regression tree for feature vectors The given output score, This is the preset learning rate.

7. The paleoecological plant species estimation method based on graph computation and feature fusion according to claim 1, characterized in that, The S5 step specifically includes: S51: Extract effective features from the original morphological feature data of the plant fossil sample to be tested, and exclude the graph structure scalar features; based on the preset botanical reference standard, calculate the numerical similarity component of the numerical features, the categorical similarity component of the categorical features, and the probabilistic similarity component of the probabilistic features in the effective features respectively. S52: Calculate the arithmetic mean of the similarity components of all the effective features obtained, and obtain the total range similarity of the plant fossil sample to be tested; S53: The predicted probability and the total range similarity are weighted and summed according to a preset fusion coefficient to calculate the final similarity; S54: Compare the final similarity with a preset comprehensive similarity threshold. If the final similarity is greater than or equal to the comprehensive similarity threshold, then determine that the tested plant fossil sample belongs to the target species; otherwise, determine that it does not belong to the target species.

8. The paleoecological plant species estimation method based on graph computation and feature fusion according to claim 1, characterized in that, Step S51 includes: For the aforementioned numerical characteristics, let the positive class reference interval given in the botanical reference standard be... Set the interval width ,in This is the preset minimum interval tolerance; if the measured value of the feature is... Its numerical similarity components The calculation formula is: ; For the aforementioned categorical features, let the set of preference values ​​specified in the botanical reference standard be . If the feature observation value is Its categorical similarity components The calculation formula is: ; Regarding the probabilistic feature, let the positive reference probability given in the botanical reference standard be... If the characteristic observation value is Its probabilistic similarity components The calculation formula is: ; The botanical reference standard's characteristic attribute system is divided into a basic data space including environmental characteristics, plant morphological characteristics, ecological characteristics, chemical composition characteristics, and geological period characteristics. When constructing the botanical reference standard, a hierarchical parameter nesting construction logic is adopted. In the multidimensional feature space, a broad background benchmark interval for general plants is set, as well as a specific core suitable interval for target groups that is precisely nested within the background benchmark interval.

9. The paleoecological plant species estimation method based on graph computation and feature fusion according to claim 1, characterized in that, The pre-trained target classification model is obtained by joint training based on a hybrid training dataset that combines virtual and real data. The hybrid training dataset includes large-scale simulated samples generated based on a dual-engine mechanism, as well as real plant fossil confirmation data directly injected through an external structured configuration table. The large-scale simulation samples are composed of a mix of baseline simulation samples and adversarial simulation samples. The baseline simulation samples are generated by random sampling based on the parameter distribution patterns of a preset botanical reference standard. The adversarial simulation samples are generated by introducing an adversarial learning mechanism and applying feature offset perturbations to the classification decision boundary.

10. A paleoecological plant species estimation system based on graph computation and feature fusion, characterized in that, include: The morphological feature acquisition module is configured to acquire the original morphological feature data of the plant fossil sample to be tested. The original morphological feature data includes at least numerical features, categorical features and probabilistic features. The graph topology construction module is configured to construct a graph structure model representing the correlation between plant morphological features based on the original morphological feature data, and generate the corresponding adjacency matrix. The feature extraction and fusion module is configured to map the measured values ​​of the plant fossil sample under test in each feature dimension to the node signal vector on the graph structure model, and extract the graph structure scalar features through graph neighborhood signal aggregation operation; The graph structure scalar features are concatenated and fused with the original morphological feature data to generate an extended feature matrix. The classification probability prediction module is configured to perform inference processing on the extended feature matrix based on a pre-trained target classification model, and output the predicted probability that the tested plant fossil sample belongs to the target species; wherein, the target classification model is a machine learning model built based on gradient boosting tree; The dual-track determination output module is configured to calculate the range similarity of the tested plant fossil sample on the original morphological feature data based on a preset botanical reference standard; to perform a weighted fusion calculation of the predicted probability and the range similarity to obtain the final similarity; and to determine whether the tested plant fossil sample belongs to the target species based on the final similarity.

Images