Species distribution and abundance prediction method based on enhanced integrated machine learning

Abstract

The invention belongs to the field of artificial intelligence, and particularly relates to a species distribution and abundance prediction method based on enhanced integrated machine learning. Aiming at the problems in the prior art, the invention provides a double-model integration framework, and classification and regression tasks are combined, so that the prediction accuracy and adaptability are remarkably improved. The invention specifically relates to a method for predicting species distribution and abundance by using an ensemble learning and transfer learning algorithm. The method is suitable for the fields of biodiversity monitoring, biological invasion risk management and the like. By processing the relationship between the high-dimensional environment characteristic data and the species generation distribution and abundance, the quantitative prediction of the species distribution and abundance is realized. The technology has high universality, can be expanded to various ecological scenes, supports flexible application from large-scale data sets to small sample adaptation, and supports ecological protection, risk prediction and decision support.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method for predicting species distribution and abundance based on augmented ensemble machine learning. Background Technology

[0002] In current technologies, species distribution and abundance prediction mainly rely on traditional ecological models, such as generalized linear models (GLM), generalized additive models (GAM), maximum entropy models (MaxEnt), or simple random forest algorithms. These methods perform well when dealing with complex, nonlinear environmental variables, but they have significant limitations: first, they cannot effectively capture the interactions between features and multimodal distributions, leading to prediction biases in high-dimensional data; second, they are sensitive to imbalanced data or small sample scenarios, exhibiting insufficient generalization ability; and third, they neglect model transferability, failing to adapt from large-scale datasets to smaller regional datasets for fine-tuning. With increasing environmental change and data availability, the need for species distribution prediction is becoming increasingly urgent, but existing technologies struggle to provide a comprehensive and accurate tool that improves both prediction accuracy and model generality. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting species distribution and abundance based on enhanced ensemble machine learning.

[0004] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, the method includes the following steps:

[0005] Data processing: Input high-dimensional environmental characteristic data and species observation data, and perform standardization processing on the input data;

[0006] A species distribution prediction model and a species abundance prediction model are constructed. The species distribution prediction model uses the predicted probabilities of the base learner as meta-features, and the final distribution is output through the meta-learner.

[0007] The abundance prediction model regression model uses the predicted values ​​of the base learner as meta-features and outputs the abundance prediction values ​​through the meta-learner.

[0008] The distribution prediction model and the species abundance prediction model are pre-trained and fine-tuned, wherein,

[0009] In the pre-training phase: the base learners and meta-learners of the distribution prediction model and the species abundance prediction model are simultaneously trained on a large-scale dataset to learn the general patterns and relationships between environmental features and species distribution, and the pre-training results are saved.

[0010] During the fine-tuning phase: Based on the results of the pre-training, all base learner parameters are frozen, and the meta-learner parameters of the distribution prediction model and the species abundance prediction model are updated only using regional small sample data;

[0011] Species distribution prediction and species abundance prediction: Input preprocessed environmental feature data of the area to be predicted, output species distribution through the pre-trained and fine-tuned distribution prediction model, output abundance prediction value through the pre-trained and fine-tuned species abundance prediction model, and generate joint prediction results including probability and abundance.

[0012] Accuracy assessment: Outputs multi-dimensional visualizations and metrics.

[0013] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, a heterogeneous fusion base learner formed by combining AdaBoost and ExtraTrees is constructed in the species distribution prediction model.

[0014] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, the species distribution prediction model uses only the probability output of the base classifier as meta-feature, and the meta-learner of the species distribution prediction model uses Logistic regression for discrimination; the species abundance prediction model uses only the continuous predicted values ​​of the base regressor as meta-feature, and the meta-learner of the species abundance prediction model uses Ridge regression with L2 regularization for fitting.

[0015] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the species distribution prediction model, an adaptive threshold selection mechanism based on ROC curves is used to determine the optimal threshold for classification, and this mechanism is consistently applied in the pre-training and fine-tuning of the distribution prediction model and the species abundance prediction model, as well as in the species distribution prediction and species abundance prediction steps.

[0016] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the training phase, a log(x+1) transformation is applied to the abundance data to alleviate the data skewness problem, and in the species distribution prediction and species abundance prediction steps, the results are restored to the original scale by an inverse exponential transformation.

[0017] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the accuracy evaluation step, a permutation-based importance measurement method is adopted, and the ranking of key environmental factors is output through multiple random permutations and cross-validation.

[0018] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the preprocessing step, input data containing high-dimensional environmental features and species observation data is collected.

[0019] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the species distribution prediction and species abundance prediction steps, the input data containing environmental features of the area to be predicted is consistent with the feature dimensions of the large-scale dataset training data used in the pre-training stage.

[0020] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the species distribution prediction and species abundance prediction steps, the species distribution is output by the pre-trained and fine-tuned distribution prediction model, which is a value between 0 and 1, and the closer it is to 1, the higher the probability of its existence.

[0021] Beneficial technical effects;

[0022] This invention addresses the problems existing in the prior art by proposing a dual-model ensemble framework that combines classification and regression tasks, significantly improving the accuracy and adaptability of predictions.

[0023] Specifically, this invention relates to a method for predicting species distribution and abundance using ensemble learning and transfer learning algorithms. This method is applicable to fields such as biodiversity monitoring and biological invasion risk management. By processing the relationship between high-dimensional environmental characteristic data (such as climate variables, soil type, topographic features, and human activity indicators) and species occurrence, distribution, and abundance, this invention achieves quantitative prediction of species distribution and abundance. This technology is highly versatile, scalable to various ecological scenarios, and supports flexible applications from large-scale datasets to small-sample adaptations, supporting ecological protection, risk prediction, and decision support.

[0024] According to the technical solution of this application, firstly, an innovative dual-stacked ensemble framework integrating classification and regression is constructed. By combining classification base learners and regression base learners with targeted meta-learners (LogisticRegression and Ridge), collaborative prediction of species distribution and abundance is achieved. Compared with traditional models, this scheme captures the complex interactions of high-dimensional environmental features through the stacking of multiple base learners, reducing prediction bias. Secondly, a "pre-training, fine-tuning" transfer learning mechanism is introduced. This mechanism utilizes large-scale data to train general feature extraction capabilities. By freezing base learner parameters and updating only the meta-learners to adapt to small sample scenarios, the problem of insufficient generalization ability of traditional models with limited data is solved, achieving flexibility in cross-scale prediction. Thirdly, a standardized preprocessing and multi-dimensional evaluation system are integrated. This optimizes data distribution through feature normalization and target transformation, and enhances model interpretability through visualization tools such as ROC curves, residual analysis, and feature importance ranking. This provides a versatile and targeted technical tool for scenarios such as biodiversity monitoring and invasive species risk prediction. Attached Figure Description

[0025] Figure 1This is a schematic diagram of the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention; Figure 2 This is a structural diagram of a species distribution and abundance prediction system based on augmented ensemble machine learning. Detailed Implementation

[0026] According to the technical solution of this application, it is achieved by constructing a dual-model framework (classifier and regressor).

[0027] In this invention, species distribution refers to the probability of a species appearing under specific environmental conditions (a value between 0 and 1). Species abundance refers to the number of individuals or coverage of a species per unit area, and is a continuous value.

[0028] like Figure 1 As shown, the species distribution and abundance prediction method based on augmented ensemble machine learning according to the present invention includes the following steps:

[0029] Data processing: Input high-dimensional environmental feature data and species observation data, and perform standardization processing on the input data, including: ① Feature standardization: normalize the environmental feature data; ② Target variable transformation: retain 0 / 1 labels for classification tasks, and perform logarithmic transformation on abundance values ​​for regression tasks. If there are zero values, log(x+1) is used to improve the normality of the data.

[0030] Constructing species distribution prediction models and species abundance prediction models, specifically including:

[0031] Construct a species distribution prediction model to predict species distribution: use the predicted probabilities of the base learners as meta-features, and output the final distribution through the meta-learner.

[0032] Construct a species abundance prediction model: use the prediction values ​​of the base learner as meta-features, and output abundance prediction values ​​through the meta-learner;

[0033] The distribution prediction model and species abundance prediction model are pre-trained and fine-tuned, wherein,

[0034] In the pre-training phase: base learners and meta-learners for both types of models are trained simultaneously on a large-scale dataset to learn the general patterns and relationships between environmental features and species distribution. The pre-trained models are then saved.

[0035] In the fine-tuning stage: load the pre-trained model, freeze all base learner parameters, update the meta-learner parameters of the two models using only regional small sample data, generate the fine-tuned model, and achieve local adaptive adjustment;

[0036] Species distribution prediction and species abundance prediction: Input environmental characteristic data of the area to be evaluated. After preprocessing, the species distribution (0-1) is output through the pre-trained and fine-tuned species distribution prediction model, and the abundance prediction value is output through the pre-trained and fine-tuned species abundance prediction model (regression model). At the same time, a joint prediction result including probability and abundance is generated.

[0037] Accuracy Assessment: Outputs multi-dimensional visualizations and metrics: ① Classification Model: ROC curve (AUC value), confusion matrix, Youden's J optimal threshold; ② Regression Model: Residual analysis (residual distribution histogram, QQ plot), R... 2 , MSE, MAE; ③ Environmental characteristics: Ranking of the top 20 environmental characteristics (such as annual average temperature, soil pH) based on their importance.

[0038] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, a lightweight optimization strategy is proposed in the transfer learning stage. This strategy involves keeping all base learner parameters frozen during small-sample fine-tuning in a region, generating meta-features only using data from the new region, and then retraining the meta-learners. Regularization constraints are introduced during retraining to prevent overfitting in small samples. This approach effectively reduces the computational cost of cross-regional transfer while maintaining the generalization ability of the base learners acquired during large-scale pre-training, thus achieving a dual improvement in prediction accuracy and stability under small-sample conditions. This strategy has not yet been publicly applied in species distribution and abundance prediction, and for the first time solves the challenges of cross-regional transfer and insufficient sample size in ecological prediction tasks.

[0039] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, a heterogeneous fusion base learner composed of AdaBoost and ExtraTrees is constructed in the species distribution prediction model to achieve complementary effects of bias reduction and variance suppression. Unlike existing single ensemble models, this combination ensures the diversity of base learners within the stacked ensemble system, avoids redundancy caused by high correlation between homogeneous models, and improves the ability of the meta-learner to discriminate complex distribution patterns. Experiments show that this fusion structure can significantly improve the robustness and accuracy of classification prediction, and its application in the field of species distribution and abundance prediction has not yet been publicly reported.

[0040] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, a separate meta-feature representation mechanism is designed to address the differences between the species distribution and abundance prediction model tasks: the species distribution prediction model uses only the probability output of the base classifier as meta-features, and the meta-learner uses Logistic regression for discrimination; the species abundance prediction model uses only the continuous predicted values ​​of the base regressor as meta-features, and the meta-learner uses Ridge regression with L2 regularization for fitting. This approach avoids the interference of inconsistencies in the units and distribution characteristics of probability and numerical values ​​on the meta-learner, thereby improving the fitting ability and generalization performance of the meta-learner on their respective tasks.

[0041] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, an adaptive threshold selection mechanism based on ROC curves is proposed in the species distribution prediction model. Specifically, the optimal classification threshold is determined by calculating and identifying the maximum value of Youden's J index (sensitivity + specificity – 1), and this threshold is consistently applied throughout the training, evaluation, and prediction processes. This optimization avoids the instability of traditional fixed or empirical threshold methods, making the classification results more scientific and reproducible. Simultaneously, it achieves a balance between sensitivity and specificity in the species distribution prediction task, improving the reliability of the model's predictions.

[0042] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, this invention proposes for the first time a bidirectional consistent transformation-inverse transformation mechanism for training and prediction in the species abundance prediction model: during the training phase, a log(x+1) transformation is applied to the abundance data to alleviate data skewness; during the prediction phase, an exponential inverse transformation is used to restore the results to the original scale, and the results before and after the transformation are simultaneously output in the joint results table. This mechanism ensures both the numerical stability of the model training and the physical interpretability of the prediction output, effectively avoiding the problem of predicted values ​​deviating from the actual abundance level in traditional methods.

[0043] According to the species distribution and abundance prediction method based on enhanced ensemble machine learning of the present invention, in the accuracy evaluation step, a permutation-based importance metric method is adopted, which outputs a robust ranking of the top twenty key environmental factors through multiple random permutations and cross-validation. Compared with the traditional MaxEnt contribution rate method, this method can reduce the bias caused by differences in feature distribution or model structure, making the feature ranking results more robust and objective. This optimization not only improves the interpretability of the model, but also provides a scientifically reliable environmental factor reference for ecological research and management practice.

[0044] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the preprocessing step, input data containing high-dimensional environmental features and species observation data is collected. The data format must meet the structured requirements, that is, the first three columns are metadata, such as sample ID, longitude, and latitude, used to identify the sample source and spatial location; the fourth column is explicitly the target variable, where the classification task (species distribution) is a binary label, 1 indicates presence and 0 indicates absence, and the regression task (species abundance) is a continuous value, such as the number of individuals per unit area; the fifth column and thereafter are environmental feature columns, covering climate variables (such as annual mean temperature and annual precipitation), topographic features (such as altitude, slope, and aspect), soil properties (such as pH value and organic matter content), human activity indicators (such as land use type and road density), etc. The number of features can be expanded according to the actual scenario, but it must be ensured that the feature dimensions are consistent with those of the training data to ensure the consistency of the model input.

[0045] Standardization and transformation operations are performed on the input data to eliminate dimensional differences and optimize data distribution. For environmental features, StandardScaler is used for standardization, calculated as x' = (x - μ) / σ, where μ is the feature mean and σ is the feature standard deviation. All features are scaled to a distribution with a mean of 0 and a standard deviation of 1 to avoid interference from extreme values ​​in model training. For the target variable, the classification task directly retains the 0 / 1 binary label, while the regression task performs a logarithmic transformation on the abundance values. If zero or negative values ​​exist, a transformation of log(x + 1) is used to make the potentially skewed distribution closer to a normal distribution, reducing the impact of uneven data distribution on the prediction accuracy of the regression model. At the same time, missing value imputation and outlier removal further purify the data to provide high-quality input for model training. For missing value imputation, the mean is used for numerical features and the mode is used for classification features. Outliers are removed using the Z-score method, with samples |Z| > 3 considered outliers.

[0046] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the steps of constructing the species distribution prediction model (classification model) and the species abundance prediction model (regression model),

[0047] For the species distribution prediction model (classification model), multi-level learning of features is achieved. The classification model (StackingClassifier) ​​uses a base learner with StandardScaler to process the standardized environmental features in parallel. The output prediction probability is used as a meta-feature input to the LogisticRegression meta-learner. The final species distribution is generated through the weighted combination of the meta-learners. The base learner with StandardScaler includes AdaBoost classifier, XGBoost classifier, GBM classifier, HGBM classifier, Bagging classifier, RandomForest classifier, ExtraTrees classifier, and a fusion classifier of AdaBoost and ExtraTrees.

[0048] For the species abundance prediction model (regression model), a base learner with StandardScaler is used as the bottom layer. The predicted values ​​output by the base learner are used as meta-features input to the Ridge meta-learner (linear regression with L2 regularization). Regularization optimization reduces the risk of overfitting and outputs the final species abundance prediction value. The two models run in parallel and share the environmental feature preprocessing results to ensure the consistency of feature processing. The base learners with StandardScaler include AdaBoost regressors, XGBoost regressors, GBM regressors, HGBM regressors, Bagging regressors, RandomForest regressors, and ExtraTrees regressors.

[0049] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, the distribution prediction model and the species abundance prediction model are pre-trained and fine-tuned, employing a two-stage training strategy of "pre-training-fine-tuning" to achieve knowledge transfer and local adaptation.

[0050] During the pre-training phase, a large-scale dataset, such as species survey data covering the entire country, is used. Through cross-validation (the default validation set ratio is 0.2), all base learners and meta-learners of the constructed species distribution prediction model and species abundance prediction model are trained simultaneously. This enables the constructed species distribution prediction model and species abundance prediction model to learn the general rules between environmental characteristics and species distribution, such as the correlation between climate zones and species distribution. After training, the model parameters are saved as pre-trained classification model and pre-trained regression model, respectively.

[0051] During the fine-tuning phase, the pre-trained model is loaded, the parameters of all base learners are fixed, the learned general feature extraction capabilities are frozen, and the meta-learners are retrained using only regional small sample data, such as monitoring data from a protected area. The weight parameters of the meta-learners are adjusted to adapt to local ecological features, such as the impact of regional unique terrain on species. After fine-tuning, a fine-tuned classification model and a fine-tuned regression model are generated, which not only retains the general knowledge learned from large-scale data, but also improves the adaptability to small sample scenarios.

[0052] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the species distribution prediction and species abundance prediction steps, the new area data is predicted based on the trained model to generate structured results. The input environmental feature data of the new area must be consistent with the feature dimensions of the training data. After standardization by the preprocessing layer, it is input into the fine-tuned classification model and the fine-tuned regression model respectively. If local data is lacking, the pre-trained model can be used directly for prediction. The classification model outputs the species distribution in the area, a value between 0 and 1, with a higher probability of existence as the value is closer to 1. The regression model outputs continuous predicted values ​​of species abundance. For example, the converted logarithmic abundance needs to be restored to the actual abundance through exponential transformation. The prediction results are integrated into a joint data table containing sample ID, distribution, and abundance prediction values. It also supports outputting binarized occurrence predictions according to user needs, and uses Youden's J optimal threshold to distinguish between presence and absence, providing directly usable quantitative results for subsequent ecological analysis.

[0053] According to the species distribution and abundance prediction method based on augmented ensemble machine learning of the present invention, in the accuracy evaluation step, the accuracy of the model is evaluated using multi-dimensional indicators and visualization tools, wherein...

[0054] The evaluation of the species distribution prediction model includes: plotting the ROC curve and calculating the AUC value to assess the discriminative power of the probability prediction; the closer the AUC is to 1, the better the performance. A confusion matrix is ​​generated to show the number of true positives, false positives, true negatives, and false negatives, quantifying the types of classification errors. Youden's J statistic (J = sensitivity + specificity - 1) is calculated to determine the optimal classification threshold. The evaluation of the species abundance prediction model includes: residual analysis and commonality analysis. The residual analysis includes: plotting residual distribution histograms and QQ plots to verify whether the prediction error conforms to a normal distribution and calculating ROC. 2 ( Evaluate the model's explanatory power, MSE, and mean squared error. MAE, Mean Absolute Error Quantifying prediction bias; the commonality analysis includes ranking the influence of features based on their importance, randomly shuffling a feature value and observing the decline in model performance to determine the top 20 environmental features that have the greatest impact on the prediction results, such as annual average temperature and soil moisture. All evaluation results are output in the form of charts to intuitively show the model performance and ecological influencing factors, providing a scientific basis for decision-making on biodiversity conservation, invasive species control, and other issues.

[0055] This invention aims to address the shortcomings of existing ecological prediction methods, such as the lack of a unified interface and standardized results, insufficient cross-regional adaptability, and low automation levels, in order to support the practical application of large-scale ecological assessment and management decisions. The method constructs a unified framework through parameterized and scripted interfaces, providing four operating modes: training, fine-tuning, evaluation, and prediction. Input data uses a fixed column format (the first three columns are metadata, the fourth column is the target variable, and the rest are feature columns). Output results and image files use a unified naming convention and directory structure to ensure reproducibility and batch inference capabilities. Furthermore, it integrates a direct prediction mode based on a pre-trained model, enabling rapid prediction and adaptation without retraining when facing data from different regions or new species. This achieves end-to-end processing from data preprocessing, model training, regional migration, prediction inference to result output.

[0056] Example 1: Prediction of the distribution and abundance of the invasive plant *Eupatorium adenophorum* in my country

[0057] First, a large-scale national ecological dataset was collected, containing approximately all high-dimensional environmental features of the country, such as annual mean temperature, precipitation, altitude, soil pH, and land use type, as well as species occurrence labels and abundance values. Features were standardized using StandardScaler and the abundance values ​​were transformed using a log(x+1) preprocessing unit. Then, during the pre-training phase, StackingClassifier (8 base learners, including AdaBoost, XGBoost, etc., with LogisticRegression as the meta-learner) and StackingRegressor (7... The base learner (and the meta learner, Ridge) is trained on this dataset, and the pre-trained model is saved. During the fine-tuning stage, the model is loaded, the base learner is frozen, and only the meta learner parameters are updated using approximately 1,000 small samples of data from the province. When making predictions, the environmental characteristics of the new area are input, and the model outputs a distribution (e.g., 0.85 indicates a high risk) and an abundance prediction value (e.g., 5 individuals per square meter after restoration), and generates evaluation charts such as ROC curves (AUC = 0.95) and Top 20 feature importance maps (with annual average temperature ranking first). This example demonstrates the efficient application of the solution in intrusion risk management, improving prediction accuracy by approximately 15%.

[0058] Other alternative technical solutions of this invention include: in the selection of base learners, some models can be replaced (e.g., LightGBM can replace XGBoost, and CatBoost can replace GBM), or the number of base learners can be adjusted (6-8 for classification models and 5-7 for regression models) while still maintaining the ensemble effect; in the selection of meta-learners, classification tasks can be replaced with SVM or neural networks with regularization, and regression tasks can be replaced with Lasso or ElasticNet, suitable for different regularization requirements; in the training mechanism, the top-level parameters of some base learners can be unfrozen (rather than completely frozen) during the fine-tuning stage, and local adaptability can be improved through a small number of iterations, suitable for medium-sized regional data scenarios. These alternative solutions all maintain the core architecture of "dual model stacking + pre-training - fine-tuning", only adjusting the specific implementation details, and are suitable for different ecological data characteristics and application needs.

[0059] The above embodiments are used to understand the technical solutions of this application and do not limit the scope of protection of this application.

Claims

1. A method for predicting species distribution and abundance based on augmented ensemble machine learning, characterized in that, The method includes the following steps: Data processing: Input high-dimensional environmental characteristic data and species observation data, and perform standardization processing on the input data; A species distribution prediction model and a species abundance prediction model are constructed. The species distribution prediction model uses the predicted probabilities of the base learner as meta-features, and the final distribution is output through the meta-learner. The abundance prediction model regression model uses the predicted values ​​of the base learner as meta-features and outputs the abundance prediction values ​​through the meta-learner. The distribution prediction model and the species abundance prediction model are pre-trained and fine-tuned, wherein, In the pre-training phase: the base learners and meta-learners of the distribution prediction model and the species abundance prediction model are simultaneously trained on a large-scale dataset to learn the general patterns and relationships between environmental features and species distribution, and the pre-training results are saved. During the fine-tuning phase: Based on the results of the pre-training, all base learner parameters are frozen, and the meta-learner parameters of the distribution prediction model and the species abundance prediction model are updated only using regional small sample data; Species distribution prediction and species abundance prediction: Input preprocessed environmental feature data of the area to be predicted, output species distribution through the pre-trained and fine-tuned distribution prediction model, output abundance prediction value through the pre-trained and fine-tuned species abundance prediction model, and generate joint prediction results including probability and abundance. Accuracy assessment: Outputs multi-dimensional visualizations and metrics.

2. The species distribution and abundance prediction method based on augmented ensemble machine learning according to claim 1, characterized in that, A heterogeneous fusion base learner, consisting of a combination of AdaBoost and ExtraTrees, was constructed in the species distribution prediction model.

3. The species distribution and abundance prediction method based on augmented ensemble machine learning according to claim 1, characterized in that, The species distribution prediction model uses only the probability output of the base classifier as meta-feature, and the meta-learner of the species distribution prediction model uses Logistic regression for discrimination; the species abundance prediction model uses only the continuous predicted values ​​of the base regressor as meta-feature, and the meta-learner of the species abundance prediction model uses Ridge regression with L2 regularization for fitting.

4. The species distribution and abundance prediction method based on augmented ensemble machine learning according to claim 1, characterized in that, In the species distribution prediction model, an adaptive threshold selection mechanism based on ROC curves determines the optimal threshold for classification and is consistently applied in the pre-training and fine-tuning of the distribution prediction model and the species abundance prediction model, as well as in the species distribution prediction and species abundance prediction steps.

5. The species distribution and abundance prediction method based on augmented ensemble machine learning according to claim 1, characterized in that, During the training phase, a log(x+1) transformation is applied to the abundance data to alleviate the data skewness problem. In the species distribution prediction and species abundance prediction steps, the results are restored to the original scale through an inverse exponential transformation.

6. The method for predicting species distribution and abundance based on augmented ensemble machine learning according to claim 1, characterized in that, In the accuracy assessment step, a permutation-based importance measurement method is used, which outputs the ranking of key environmental factors through multiple random permutations and cross-validation.

7. The method for predicting species distribution and abundance based on augmented ensemble machine learning according to claim 1, characterized in that, In the preprocessing step, input data containing high-dimensional environmental features and species observation data is collected.

8. The method for predicting species distribution and abundance based on augmented ensemble machine learning according to claim 1, characterized in that, In the species distribution prediction and species abundance prediction steps, the input data containing environmental features of the area to be predicted has the same feature dimensions as the large-scale dataset used in the pre-training phase.

9. The method for predicting species distribution and abundance based on augmented ensemble machine learning according to claim 1, characterized in that, In the species distribution prediction and species abundance prediction steps, the species distribution is output by the pre-trained and fine-tuned distribution prediction model, which is a value between 0 and 1. The closer it is to 1, the higher the probability of its existence.

Images