Ecological system health assessment method based on machine learning

Abstract

The invention relates to the technical field of ecological system monitoring and evaluation, and discloses an ecological system health evaluation method based on machine learning. The method comprises the following steps: collecting multi-source continuous monitoring data of an ecological system, and carrying out integrity check and noise filtering to generate a standardized data set; performing multi-scale feature mining on the data set, and extracting ecological feature sequences of a time domain and a frequency domain; performing community discovery on the feature sequence by adopting a spectral clustering algorithm, and outputting a clustering result of the ecological health state; constructing a multi-class classification model according to the clustering result, and adjusting the weight of the model through iterative learning; and inputting the real-time ecological data flow into the trained multi-class classification model, performing probability prediction of the health state, and automatically generating a state evaluation chart. According to the method, the health state can be adaptively defined from the data, and objective and dynamic evaluation of ecological system health is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ecosystem monitoring and evaluation, in particular to an ecosystem health evaluation method based on machine learning. BACKGROUND

[0002] Currently, the evaluation of ecosystem health generally relies on threshold-based index evaluation methods or expert experience models. These methods usually require domain experts to pre-set a series of health thresholds for key ecological indicators, and then determine the health grade by comparing and weightedly integrating the monitoring data with these fixed thresholds. Another common approach is to directly apply supervised machine learning models, but this requires a large amount of historical data with accurately labeled health states as training samples.

[0003] The effectiveness of threshold-based methods highly depends on the accuracy and universality of the set thresholds. Ecosystems have high complexity and dynamics, and there are complex nonlinear interactions between their internal elements. Artificially set static thresholds are difficult to fully capture this dynamic correlation and may not accurately reflect the critical state of the system from quantitative change to qualitative change, resulting in overly rigid evaluation results that are not sensitive to early degradation risks. Directly using supervised learning faces the labeling bottleneck. It is costly and extremely difficult to obtain a large amount of high-quality and consistently labeled historical health state data, especially for emerging or data-poor ecological regions. Subjective differences in expert labeling can directly affect the learning effect and generalization ability of the model.

[0004] The defect of the prior art solution lies in the definition method of the health state classification standard. Whether it is a static threshold or relies on prior labeling, the learning process of the model is limited to artificially pre-set rules or labels that may not be perfect. This makes it difficult for the evaluation model to autonomously discover potential and unrecognized health state patterns from the data itself, and it cannot adapt to the long-term evolution of the ecosystem over time. How to objectively and adaptively define health state categories from continuous monitoring data and establish a framework that can make real-time and dynamic predictions is a key problem that needs to be solved by the prior art. SUMMARY

[0005] The purpose of the present application is to provide an ecosystem health evaluation method based on machine learning to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides an ecosystem health evaluation method based on machine learning, which comprises:

[0007] Collecting multi-source continuous monitoring data of the ecosystem, and performing integrity verification and noise filtering on the data to generate a standardized data set;

[0008] Multi-scale feature mining is performed on the normalized data set to extract time domain and frequency domain ecological feature sequences.

[0009] A spectral clustering algorithm is used to perform community discovery on the time domain and frequency domain ecological feature sequences to output ecological health state clustering results.

[0010] According to the ecological health state clustering results, a multi-class classification model is constructed, and the model weights are adjusted through iterative learning.

[0011] Real-time ecological data stream is input into the multi-class classification model for health state probability prediction, and state evaluation charts are generated.

[0012] Preferably, the multi-source continuous monitoring data of the ecosystem is collected, and the data is subjected to integrity check and noise filtering to generate a normalized data set, which includes:

[0013] Multi-source continuous monitoring data, including meteorological data, soil data and hydrological data, are obtained from an ecological monitoring network, and missing value interpolation and outlier detection are performed.

[0014] A sliding window filtering algorithm is applied to smooth the multi-source continuous monitoring data, remove high-frequency noise, and obtain denoised data sequences.

[0015] The denoised data sequences are subjected to standardization transformation to convert parameters of different dimensions to a unified scale to generate a normalized data set.

[0016] The normalized data set is input into a feature mining module, and a wavelet analysis algorithm is used to decompose the data sequences and extract multi-resolution feature components.

[0017] Preferably, the multi-scale feature mining of the normalized data set to extract time domain and frequency domain ecological feature sequences includes:

[0018] Time domain feature calculation is performed on the normalized data set, including mean, variance and autocorrelation function, to obtain a time domain feature vector.

[0019] Frequency domain transformation is performed on the normalized data set, and a fast Fourier algorithm is used to calculate the power spectral density to extract a frequency domain feature vector.

[0020] The time domain feature vector and the frequency domain feature vector are combined, and a feature selection algorithm is used to remove redundant features to obtain dimension-reduced time domain and frequency domain ecological feature sequences.

[0021] The time domain and frequency domain ecological feature sequences are input into a clustering analysis module for community discovery.

[0022] Preferably, the spectral clustering algorithm is used to perform community discovery on the time domain and frequency domain ecological feature sequences to output ecological health state clustering results, which includes:

[0023] Constructing the similarity matrix of time domain and frequency domain ecological feature sequence, using Gaussian kernel function to calculate the similarity between feature points;

[0024] Performing Laplace transform on the similarity matrix to obtain the Laplace matrix, and calculating the eigenvalues and eigenvectors of the Laplace matrix;

[0025] Based on the eigenvectors, k-means clustering is performed to divide the ecological feature sequence into multiple communities, each community representing a health status category;

[0026] Outputting the ecological health status clustering results, including community labels and central feature points.

[0027] Preferably, the method of constructing a multi-class classification model based on the ecological health status clustering results and adjusting the model weights through iterative learning includes:

[0028] Using the ecological health status clustering results as training labels and extracting the time domain and frequency domain ecological feature sequence as training features;

[0029] Initializing the parameters of the multi-class classification model, applying the gradient descent algorithm to optimize the loss function, and gradually adjusting the model weights;

[0030] Using the cross-validation method to evaluate the model performance to prevent overfitting, and obtaining the optimized multi-class classification model;

[0031] Using the optimized multi-class classification model for real-time health status prediction.

[0032] Preferably, the method of inputting real-time ecological data stream into the multi-class classification model for health status probability prediction and generating state evaluation charts includes:

[0033] Performing the same preprocessing and feature extraction on the real-time ecological data stream as the training data to obtain real-time feature sequence;

[0034] Inputting the real-time feature sequence into the multi-class classification model to calculate the probability value of each health status category;

[0035] Generating a health status distribution map according to the probability value to display the confidence level of different states;

[0036] Combining geographic information system to spatially visualize the health status distribution map and outputting the state evaluation chart.

[0037] Preferably, the method of applying a sliding window filtering algorithm to smooth the multi-source continuous monitoring data includes:

[0038] Defining the size and step length of the sliding window to segment the denoised data sequence;

[0039] The median value of data in each window is calculated to replace the center point value of the window to realize median filtering;

[0040] The window parameters are adjusted to adapt to the data change rate to ensure the consistency of the smoothing effect.

[0041] Preferably, the frequency domain transformation of the normalized data set and the application of the fast Fourier algorithm to calculate the power spectral density include:

[0042] The normalized data set is resampled to ensure an equal interval time sequence.

[0043] The fast Fourier algorithm is applied to convert the time domain data into a frequency domain representation, calculate the square of the frequency domain amplitude, and obtain the power spectral density.

[0044] Preferably, the k-means clustering based on the feature vector includes:

[0045] The initial cluster center is selected, the data points are assigned to the nearest cluster according to the feature vector distance, the cluster center is recalculated, and the iteration is performed until the cluster assignment is stable, and the final clustering result and the sum of squares of errors in the cluster are output.

[0046] Preferably, the spatial visualization of the health state distribution map combined with the geographic information system includes:

[0047] The health state probability value is mapped to the geographic coordinate point to generate point distribution data, the point data is converted into a continuous surface using a spatial interpolation method, the basic map layer is superimposed to generate a color-rendered state evaluation chart.

[0048] Compared with the prior art, the beneficial effects of the present application are:

[0049] The spectral clustering algorithm is used to perform community discovery on the ecological feature sequences extracted from the time domain and the frequency domain, which can process complex nonlinear relationships between high-dimensional features. This method does not rely on the health state threshold set by humans, but automatically divides according to the similarity of the distribution of data points in the feature space. This unsupervised learning method based on graph theory can identify the internal structure of data that cannot be captured by traditional linear methods, thereby objectively revealing different health state patterns determined by the operating rules of the ecological system. This avoids the bias or omission that may be caused by subjective threshold setting, so that the state classification result is completely data-driven, more scientific and objective, and can identify the transition state or sub-health state that cannot be defined based on simple rules.

[0050] On the basis of obtaining the health state classification through unsupervised clustering, the results are used as labels for training a supervised multi-class classification model, and a two-stage learning framework is constructed. The state discovery and state prediction are decoupled and seamlessly connected. The classification model learns the mapping relationship between the ecological characteristics and the clustering labels, and obtains the ability to predict the state attribution of the new real-time data stream. Through iterative learning to adjust the model weight, the classifier can continuously optimize its discrimination boundary. This design overcomes the limitations of simple unsupervised learning methods that cannot make real-time judgments in new areas. At the same time, the probability prediction output by the classification model can provide more abundant and hierarchical state reliability information than simple binary judgment or hard classification, providing more detailed basis for decision support. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The working principle diagram of the machine learning-based ecosystem health assessment method described in the application;

[0052] Figure 2 The flowchart of data acquisition and standardized processing;

[0053] Figure 3 The flowchart of multi-scale feature mining and feature selection;

[0054] Figure 4 The scatter plot of the ecosystem health state spectrum clustering result;

[0055] Figure 5 The interpolation diagram of the spatial distribution of the ecosystem health state. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0057] Please refer to Figure 1 The application provides a machine learning-based ecosystem health assessment method, which comprises the following steps: collecting multi-source continuous monitoring data of an ecosystem and performing data preprocessing to generate a standardized data set, then performing multi-scale feature mining on the standardized data set to extract time-domain and frequency-domain ecological feature sequences, performing community discovery on the feature sequences by using a spectral clustering algorithm and outputting an ecological health state clustering result, constructing a multi-class classification model based on the clustering result and adjusting the model weight through iterative learning, finally inputting real-time ecological data stream into the model for health state probability prediction and generating a state assessment chart, and realizing dynamic health monitoring and assessment of the ecosystem.

[0058] Example 1: See Figure 2 In practical implementation, multi-source continuous monitoring data is acquired from the ecological monitoring network. This multi-source continuous monitoring data includes meteorological data such as temperature and humidity readings, soil data such as pH and nutrient content, and hydrological data such as flow velocity and turbidity. Missing value imputation is performed to fill data gaps using linear interpolation methods, and a statistical outlier detection algorithm is applied to identify anomalies. A sliding window filtering algorithm is then applied to smooth the multi-source continuous monitoring data. The sliding window filtering algorithm segments the data sequence by defining the window size and step size, calculates the median of the data within each window, and replaces the value of the window's center point to achieve median filtering and remove high-frequency noise, resulting in a denoised data sequence. The denoised data sequence is then standardized using the Z-score standardization method to convert parameters of different dimensions to a unified scale, generating a normalized dataset. This normalized dataset is input into the feature mining module, where wavelet analysis algorithms such as discrete wavelet transform are used to decompose the data sequence and extract multi-resolution feature components, including approximation coefficients and detail coefficients. In some embodiments, the sliding window filtering algorithm sets the window size to 10 data points and the step size to 1 data point to adapt to the rate of data change and ensure consistent smoothing results. It is understandable that missing value imputation uses a linear interpolation formula:

[0059] ;

[0060] in: Indicates the location of the missing point. and Indicates the positions of adjacent known points. and Represents the values ​​of adjacent known points. This represents the interpolated value. Optionally, outlier detection uses a three-standard-deviation method to identify data points that exceed three standard deviations of the mean. In some embodiments, the wavelet analysis algorithm employs the Daubechies wavelet basis function to perform multi-level decomposition on the normalized dataset and extract time-frequency domain features. It can be understood that the normalization transformation converts the original data into a distribution with a mean of 0 and a standard deviation of 1 by calculating the mean and standard deviation of each parameter. Optionally, the sliding window filtering algorithm dynamically adjusts the window size when data changes drastically, for example, reducing the window size from 10 points to 5 points to maintain consistent smoothing effects.

[0061] Example 2: See Figure 3In specific implementations, the normalized data set is subjected to time domain feature calculation, which includes mean, variance and autocorrelation function. The mean is calculated by arithmetic average method to obtain the average value of the data sequence. The variance is calculated by square difference method to obtain the dispersion degree of the data points. The autocorrelation function is calculated by lag time parameter to obtain the correlation coefficient of the time sequence itself. A time domain feature vector is obtained. The normalized data set is subjected to frequency domain transformation. Fast Fourier algorithm is applied to calculate the power spectral density. The fast Fourier algorithm converts the time domain data sequence into frequency domain representation. A frequency domain feature vector is extracted. The time domain feature vector and the frequency domain feature vector are combined. Redundant features are removed by a feature selection algorithm. The feature selection algorithm uses principal component analysis method to reduce the feature dimension. A reduced dimension time domain and frequency domain ecological feature sequence is obtained. The time domain and frequency domain ecological feature sequence is input into a clustering analysis module for community discovery. In some embodiments, the frequency domain transformation process resamples the normalized data set. Linear interpolation method is used to adjust the time interval to ensure equal interval time sequence. Fast Fourier algorithm is applied to convert the time domain data into frequency domain representation. The square of the frequency domain amplitude is calculated to obtain the power spectral density. It can be understood that the power spectral density formula is represented as:

[0062] ;

[0063] wherein: represents the power spectral density at frequency f, represents the complex output after fast Fourier transform at frequency , represents the complex modulus operation. Optionally, in the time domain feature calculation, the variance is obtained by calculating the average value of the square difference between each data point and the mean. In some embodiments, the feature selection algorithm uses principal component analysis method to calculate the contribution rate of the feature value. The principal components with high contribution rate are retained to realize dimension reduction. It can be understood that the resampling process converts the irregular time sequence into a sequence with fixed time interval by interpolation method. Optionally, in the autocorrelation function calculation, the maximum lag time parameter is used to ensure that the autocorrelation value is calculated within the effective range.

[0064] In some embodiments, the similarity matrix is calculated using a Gaussian kernel function, and the Laplacian matrix is calculated by performing a Laplacian transform on the similarity matrix. In some embodiments, the initial cluster centers are selected using a random initialization method, and the data points are assigned to the nearest cluster based on the distance of the feature vectors. The cluster centers are recalculated, and the iteration is performed until the cluster assignment is stable. The final clustering result and the within-cluster sum of squares are output. It is understood that the similarity matrix is calculated using the Gaussian kernel function as follows:

[0065] ;

[0066] wherein: represents the similarity value between feature point and feature point , represents the Euclidean distance, represents the Gaussian kernel width parameter. Optionally, the Laplacian transform process uses a normalized Laplacian matrix form, and the feature decomposition is calculated after diagonalizing the similarity matrix by the degree matrix. In some embodiments, the initial cluster centers of the k-means clustering are selected using a random initialization method, and k points are randomly selected from the feature vector data points as the initial centers. It is understood that the within-cluster sum of squares is calculated by accumulating the sum of squares of the distances from each data point in the cluster to the cluster center. Optionally, the community label assignment is determined based on the cluster membership of each data point in the clustering result, and the output includes the identifier of each community and the representative feature point coordinates.

[0067] Referring to Figure 4 , this figure is a visualization result of community discovery of ecosystem health status based on the spectral clustering algorithm. In this figure, feature 1 and feature 2 are used as coordinate axes, and the distribution pattern of multi-source ecological monitoring data after multi-scale feature mining and dimensionality reduction processing in the two-dimensional feature space is presented. Different colored points correspond to three health states: blue represents health state 1, orange represents health state 2, and green represents health state 3. This figure directly verifies the ability of the spectral clustering algorithm to objectively divide the ecological health status without manual preset thresholds in an unsupervised scenario, only through the inherent characteristics of the data. This provides a basis for subsequent construction of multi-class classification models and realization of dynamic prediction of health status, and highlights the scientific nature of the algorithm in identifying complex health patterns of ecological systems.

[0068] In some embodiments, the gradient descent algorithm is applied to optimize the loss function of the multi-class classification model, where the loss function uses cross-entropy loss function, and the model parameters include weight matrix and bias vector. The gradient descent algorithm adjusts the model weights step by step through an iterative process. In some implementations, the algorithm calculates the gradient of the loss function with respect to the model parameters, updates the values of the weight matrix and bias vector in the opposite direction of the gradient, re-evaluates the loss function value after each update to ensure that the optimization direction is towards the minimum loss, and continues the adjustment process until the model weight changes tend to be stable or reach the preset number of iterations, thereby completing the optimization of the model parameters. In some embodiments, the gradient descent algorithm uses a stochastic gradient descent variant that updates the model parameters with small batches of data, and the loss function optimization process minimizes the difference between the predicted probabilities and the true labels. It can be understood that the model weight update formula is represented as:

[0069] ;

[0070] wherein: represents the weight matrix at the t-th iteration, represents the learning rate parameter, represents the gradient of the loss function at the weight matrix , represents the updated weight matrix. Optionally, the k-fold cross-validation method randomly divides the training features and labels into k mutually exclusive subsets, and uses k-1 subsets to train the model and 1 subset to validate the model performance. In some embodiments, the initialization parameters of the multi-class classification model are randomly generated using a uniform distribution, and the elements of the weight matrix are set in the range of [-0.1, 0.1]. Optionally, referring to Table 1, the output layer of the softmax regression model generates a probability distribution for each health state category, and the predicted class is determined by the maximum probability value.

[0071] Table 1: Training feature and label data table

[0072] In some embodiments, the real-time ecological data stream is input into a multi-class classification model for health status probability prediction, the real-time ecological data stream is pre-processed and feature extracted in the same way as the training data, the pre-processing includes data cleaning and standardization operations, the feature extraction includes time domain feature calculation and frequency domain feature calculation, the real-time feature sequence is obtained, the real-time feature sequence is input into the multi-class classification model, the multi-class classification model uses a softmax function to calculate the probability value of each health status category, a health status distribution map is generated according to the probability value, the health status distribution map shows the confidence level of different states, the health status distribution map is spatially visualized by combining a geographic information system, and a state assessment chart is output. In some embodiments, the spatial visualization process maps the health status probability value to a geographic coordinate point to generate point distribution data, uses a spatial interpolation method to convert the point data into a continuous surface, superimposes a basic map layer to generate a color-rendered state assessment chart. In some embodiments, the output of the multi-class classification model is a probability distribution vector of each health status category, and the dominant health status category is determined by taking the maximum value. It can be understood that the spatial interpolation method uses an inverse distance weighted interpolation algorithm, and the interpolation formula is represented as:

[0073] ;

[0074] wherein: represents the estimated value of the point to be interpolated, represents the health status probability value of the jth known point, represents the Euclidean distance from the point to be interpolated to the jth known point, represents the distance decay parameter, represents the number of known points. Optionally, the health status distribution map uses a gradient color system to represent the probability value range, with low probability value areas represented in cool tones and high probability value areas represented in warm tones. In some embodiments, a digital elevation model is loaded as a basic map layer during the geographic information system integration process to enhance the spatial visualization effect. It can be understood that the extraction process of the real-time feature sequence is completely consistent with the training stage, including time domain mean calculation and frequency domain power spectral density calculation. Optionally, the state assessment chart is output in raster image format, supporting geographic coordinate reference system positioning.

[0075] Referring to Figure 5The figure is the core result of the health status space visualization link. It is based on the health status probability prediction results of real-time ecological data flow, combined with geographic information system to realize spatial interpolation and visualization. In the figure, longitude and latitude are used as coordinate axes, and the ecological health status probability distribution of different regions is presented through color gradient: yellow area represents high health status probability, and purple area represents low probability. First, the health status probability of real-time monitoring points is mapped to geographic coordinates to form point data; then the point data is converted into a continuous probability surface through a spatial interpolation algorithm; finally, the basic geographic layer is superimposed and presented in color rendering. It intuitively presents the spatial heterogeneity of the health status of the ecological system, helps decision-makers identify hotspots and degradation risk areas, and provides a visual basis for regional management strategies such as ecological restoration and resource allocation, which embodies the technical advantages of real-time prediction and spatial visualization in the project and solves the lack of spatial dimension analysis in traditional evaluation methods.

[0076] It should be noted that, in this document, the terms such as first and second are used merely to distinguish one entity or operation from another, and are not necessarily required to be taken in an actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0077] Although the embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A machine learning based ecosystem health assessment method, characterized in that, The method comprises: Collecting multi-source continuous monitoring data of the ecosystem, and performing integrity verification and noise filtering on the data to generate a standardized data set; Performing multi-scale feature mining on the standardized data set to extract time-domain and frequency-domain ecological feature sequences; Using a spectral clustering algorithm to perform community discovery on the time-domain and frequency-domain ecological feature sequences, and outputting ecological health state clustering results; According to the ecological health state clustering results, a multi-class classification model is constructed, and the model weight is adjusted through iterative learning; Inputting real-time ecological data stream into the multi-class classification model for health state probability prediction, and generating state evaluation charts.

2. The machine learning based ecosystem health assessment method of claim 1, wherein, The method comprises: Obtaining multi-source continuous monitoring data from an ecological monitoring network, including meteorological data, soil data and hydrological data, performing missing value interpolation and anomaly point detection; Applying a sliding window filtering algorithm to smooth the multi-source continuous monitoring data, removing high-frequency noise, and obtaining a denoised data sequence; Performing standardization transformation on the denoised data sequence to convert parameters of different dimensions to a unified scale, and generating a standardized data set; Inputting the standardized data set into a feature mining module, and using a wavelet analysis algorithm to decompose the data sequence and extract multi-resolution feature components.

3. The machine learning based ecosystem health assessment method of claim 2, wherein, The method comprises: Performing time-domain feature calculation on the standardized data set, including mean, variance and autocorrelation function, to obtain a time-domain feature vector; Performing frequency-domain transformation on the standardized data set, and applying a fast Fourier algorithm to calculate power spectral density to extract a frequency-domain feature vector; Combining the time-domain feature vector and the frequency-domain feature vector, removing redundant features through a feature selection algorithm to obtain dimension-reduced time-domain and frequency-domain ecological feature sequences; Inputting the time-domain and frequency-domain ecological feature sequences into a clustering analysis module for community discovery.

4. The machine learning based ecosystem health assessment method of claim 3, wherein, The method comprises: Constructing a similarity matrix of the time-domain and frequency-domain ecological feature sequences, and using a Gaussian kernel function to calculate the similarity between feature points; Performing Laplace transformation on the similarity matrix to obtain a Laplace matrix, and calculating the eigenvalues and eigenvectors of the Laplace matrix; Based on the eigenvectors, performing k-means clustering to divide the ecological feature sequences into multiple communities, and each community represents a health state category; Outputting the ecological health state clustering results, including community labels and center feature points.

5. The machine learning based ecosystem health assessment method of claim 4, wherein, The method comprises: Using the ecological health state clustering results as training labels, and extracting the time-domain and frequency-domain ecological feature sequences as training features; Initializing the parameters of the multi-class classification model, applying a gradient descent algorithm to optimize the loss function, and gradually adjusting the model weight; Using a cross-validation method to evaluate the model performance to prevent overfitting, and obtaining an optimized multi-class classification model; Using the optimized multi-class classification model for real-time health state prediction.

6. The machine learning based ecosystem health assessment method of claim 5, wherein, The real-time ecological data stream is input into the multi-class classification model, the health state probability is predicted, and a state evaluation chart is generated, which includes: The real-time ecological data stream is preprocessed and feature extracted in the same way as the training data to obtain a real-time feature sequence; The real-time feature sequence is input into the multi-class classification model to calculate the probability value of each health state category; A health state distribution chart is generated according to the probability value to display the confidence level of different states; The health state distribution chart is spatially visualized in combination with a geographic information system to output a state evaluation chart.

7. The machine learning based ecosystem health assessment method of claim 2, wherein, The application of the sliding window filtering algorithm to the smoothing processing of multi-source continuous monitoring data includes: Defining the size and step length of the sliding window, segmenting the denoising data sequence; Calculate the median of the data in each window, replace the center point value of the window, and realize median filtering; Adjust the window parameters to adapt to the data change rate to ensure the consistency of the smoothing effect.

8. The machine learning based ecosystem health assessment method of claim 3, wherein, The frequency domain transformation of the normalized data set and the calculation of the power spectral density by the fast Fourier algorithm include: Resample the normalized data set to ensure equal interval time series; Apply the fast Fourier algorithm to convert time domain data to frequency domain representation, calculate the square of the frequency domain amplitude, and obtain the power spectral density.

9. The machine learning based ecosystem health assessment method of claim 4, wherein, The k-means clustering based on the feature vector includes: Select the initial cluster center, assign the data points to the nearest cluster according to the feature vector distance, recalculate the cluster center, and iterate until the cluster assignment is stable, output the final clustering result and the sum of squared errors within the cluster.

10. The machine learning based ecosystem health assessment method of claim 6, wherein, The spatial visualization of the health state distribution chart in combination with the geographic information system includes: Map the health state probability value to the geographic coordinate point to generate point distribution data, convert the point data to a continuous surface using spatial interpolation methods, superimpose the basic map layer to generate a color-rendered state evaluation chart.

Images