Plot quality evaluation and abnormality identification method and device based on multi-source forestry data fusion

Abstract

The application discloses a plot quality evaluation and abnormality identification method and device based on multi-source forestry data fusion, relates to the technical field of forestry plot quality evaluation, and comprises the following steps: collecting multi-source forestry data and performing pretreatment, generating a standardized multi-source data set and performing dimension reduction processing, orthogonally decomposing the dimension-reduced data set to extract each principal component direction; projecting the standardized multi-source data set to each principal component direction to obtain each principal component score, performing denoising correction on the principal component score, and performing polarity correction on the corrected score; determining the weight of each principal component based on the polar-corrected principal component score and weighted calculating the comprehensive quality score of each plot; constructing the conditional probability distribution of the signal strength of each principal component and calculating the significance measure; determining whether each plot is abnormal according to the significance measure; and outputting diagnosis prompt information in combination with the comprehensive quality score and the abnormality determination result of each plot. Thus, the objectivity of plot quality comprehensive evaluation and the statistical credibility of abnormality determination are realized.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of forestry quality assessment technology, and in particular to a method and apparatus for plot quality assessment and anomaly identification based on multi-source forestry data fusion. Background Technology

[0002] Forest resource surveys and ecological monitoring are crucial foundations for sustainable forestry management and ecosystem health management. With the rapid development of remote sensing, IoT sensing, and high-throughput sequencing technologies, plot monitoring has evolved from simple ground-based manual surveys to a multi-source data acquisition system encompassing multiple scales, types, and frequencies. However, faced with increasingly abundant and complex multi-source forestry monitoring data, existing plot quality assessment and anomaly identification methods still suffer from the following significant shortcomings:

[0003] Insufficient depth of data fusion: Existing methods mostly rely on a single data source or simple weighted fusion, making it difficult to fully explore the inherent correlations and complementary information between multiple data sources.

[0004] The determination of evaluation index weights is highly subjective: traditional methods often use the analytic hierarchy process (AHP) or expert scoring to determine weights, which are greatly influenced by subjective factors and lack an objective weighting mechanism based on the statistical characteristics of the data itself.

[0005] Anomaly identification lacks statistical inference safeguards: Existing anomaly detection methods are mostly based on thresholds or empirical rules, which cannot provide statistical significance tests and make it difficult to distinguish between genuine ecological anomalies and random data fluctuations.

[0006] The evaluation results lack uncertainty quantification: Existing methods usually provide point estimates and cannot provide uncertainty measures such as confidence intervals, which affects the reliability of the evaluation results and their reference value for decision-making.

[0007] Therefore, there is an urgent need for a method for plot quality assessment and anomaly identification that can deeply integrate multi-source heterogeneous forestry data, adaptively assign objective weights, and provide statistical inference assurance. Summary of the Invention

[0008] This invention addresses the problems of insufficient depth of multi-source forestry data fusion, strong subjectivity of evaluation index weights, lack of statistical inference guarantee for anomaly identification, and lack of uncertainty quantification in evaluation results. It provides a method and device for plot quality assessment and anomaly identification based on multi-source forestry data fusion. By introducing block heteroscedasticity noise modeling and conditional principal component analysis, it achieves deep fusion of multi-source forestry data, realizes objectivity of comprehensive plot quality assessment, statistical reliability of anomaly judgment, and quantification of evaluation results.

[0009] In a first aspect, embodiments of the present invention provide a method for plot quality assessment and anomaly identification based on multi-source forestry data fusion, including:

[0010] Collect multi-source forestry data from the target forest area and perform data preprocessing to form a standardized multi-source dataset;

[0011] The standardized multi-source dataset is subjected to dimensionality reduction, and the dimensionality-reduced multi-source dataset is subjected to orthogonal decomposition to extract the directions of each principal component;

[0012] The standardized multi-source dataset is projected onto the directions of each principal component to obtain the scores of each principal component. The scores of each principal component are then denoised and corrected based on the noise estimation results of the dimensionality-reduced multi-source dataset. Finally, the corrected scores are polarized according to the index attributes of each principal component direction.

[0013] The weights of each principal component are determined based on the scores of each principal component after polarity correction, and the overall quality score of each sample plot is calculated by weighting the principal components according to their weights.

[0014] Based on the orthogonal decomposition results, a conditional probability distribution of the signal intensity of each principal component is constructed. A significance measure is calculated based on the conditional probability distribution, and the significance measure is used to determine whether there are anomalies in each principal component direction for each sample plot.

[0015] The system classifies plots identified as abnormal and outputs graded early warning signals. It also outputs diagnostic prompts based on the comprehensive quality score of each plot and the abnormality determination results.

[0016] In a preferred embodiment, the multi-source forestry data is preprocessed to form a standardized multi-source dataset, including:

[0017] Data preprocessing is performed on the collected multi-source forestry data to obtain a dataset to be transformed with a unified spatial reference and dimensional standard.

[0018] Based on the data type distribution characteristics of each data source in the dataset to be transformed, corresponding transformation processes are performed to transform each data source into a comparable data format.

[0019] The data sources that have been transformed into comparable data formats are associated according to their sample site identifiers to form a standardized multi-source dataset.

[0020] In a preferred embodiment, the standardized multi-source dataset is subjected to dimensionality reduction processing, and the dimensionality-reduced multi-source dataset is then orthogonally decomposed to extract the directions of each principal component, including:

[0021] When the feature dimension of the standardized multi-source dataset is greater than the number of sample plots, pre-compression is performed on the standardized multi-source dataset, and the orthogonal decomposition is performed on the pre-compressed dataset to extract the directions of each principal component.

[0022] When the feature dimension of the standardized multi-source dataset is not greater than the number of sample plots, the orthogonal decomposition is directly performed on the standardized multi-source dataset to extract the directions of each principal component.

[0023] In a preferred embodiment, the standardized multi-source dataset is projected onto each principal component direction to obtain the principal component score. The principal component scores are then denoised and corrected based on the noise estimation results of the dimensionality-reduced multi-source dataset. Finally, the corrected scores are polarity-corrected according to the index attributes of each principal component direction, including:

[0024] The standardized multi-source dataset is projected onto each principal component direction to obtain the raw scores in each principal component direction in various ways;

[0025] Based on the noise estimation results of the multi-source dataset after dimensionality reduction, the noise energy component is separated from the total observation energy in the principal component direction, and the original score is shrunk and corrected according to the remaining signal energy to obtain the denoised principal component score.

[0026] Based on the index attributes of each principal component direction, determine whether the principal component direction is consistent with the positive trend of sample plot quality. If not, perform polarity reversal correction on the denoised principal component scores of each sample plot in the principal component direction.

[0027] In a preferred embodiment, the noise estimation result is obtained in the following manner:

[0028] The features of the dimensionality-reduced multi-source dataset are divided into multiple data blocks according to their original data sources;

[0029] Based on the singular value spectrum tail information of each data block, the noise variance within each data block is determined using random matrix theory.

[0030] The noise variance within each data block is combined to obtain the noise estimation result.

[0031] As a preferred implementation, determining whether the principal component direction is consistent with the positive trend of plot quality based on the index attributes of each principal component direction includes:

[0032] Obtain the degree of correlation between each principal component direction and each original index;

[0033] For any principal component direction, several original indicators are selected according to the degree of correlation.

[0034] Determine whether the number of negative indicators among the aforementioned raw indicators reaches a preset proportion or a preset number;

[0035] If so, then the direction of the principal component is determined to be inconsistent with the positive trend of the plot quality.

[0036] In a preferred embodiment, the weight of each principal component is determined based on the polarity-corrected principal component scores, and the overall quality score of each sample plot is calculated by weighting the principal component scores according to their respective weights, including:

[0037] The scores of each plot in each principal component direction are converted into the proportion of the total scores of all plots in that direction, which is used as the probability distribution of the scores of each principal component.

[0038] The information entropy of each principal component among all sample plots is calculated based on the probability distribution of its score, and the weight of each principal component is determined based on the information entropy of its score.

[0039] The scores of each plot in each principal component direction are weighted and summed according to their corresponding weights to obtain the comprehensive quality score of each plot.

[0040] In a preferred embodiment, a conditional probability distribution of the signal intensity of each principal component is constructed based on the orthogonal decomposition results. A significance metric is calculated based on the conditional probability distribution, and the presence of anomalies in each principal component direction is determined based on the significance metric. This includes:

[0041] Based on the orthogonal decomposition results, the conditional probability distribution of the signal intensity of each principal component under the preset principal component screening conditions is constructed;

[0042] Based on the conditional probability distribution, calculate the significance measure of signal intensity in each principal component direction;

[0043] The significance measure is compared with a preset significance level threshold;

[0044] If the significance measure is less than or equal to the preset significance level threshold, then the sample plot is determined to have an anomaly in the corresponding principal component direction.

[0045] As a preferred implementation, based on the orthogonal decomposition results, a conditional probability distribution of the signal intensity of each principal component under preset principal component screening conditions is constructed, including:

[0046] Based on the singular values ​​and singular vectors in the orthogonal decomposition results, the signal intensity parameters of each principal component direction are determined.

[0047] The data-driven screening process using the number of principal components retained is used as a conditional event, and the screening process includes determining whether each principal component should be retained.

[0048] Given the occurrence of the aforementioned conditional event, construct the conditional probability distribution of the signal intensity of each principal component.

[0049] In a preferred embodiment, the sample plots determined to be abnormal are classified and graded early warning signals are output, including:

[0050] Based on preset anomaly type determination criteria, sample plot anomalies are classified into data layer anomalies, environmental layer anomalies, and ecological layer anomalies.

[0051] Based on the value range of the significance metric and the duration of the anomaly, the corresponding warning level is output according to the preset multi-level correspondence rules.

[0052] In a preferred embodiment, the method further includes:

[0053] When the significance metric meets the value range condition corresponding to a certain warning level, but the duration of the abnormality does not reach the duration condition corresponding to that level, the output is reduced by one level according to that warning level.

[0054] Secondly, embodiments of the present invention also provide a plot quality assessment and anomaly identification device based on multi-source forestry data fusion, used to execute the plot quality assessment and anomaly identification method based on multi-source forestry data fusion described in any of the above embodiments, the device comprising:

[0055] The data acquisition and preprocessing module is used to collect multi-source forestry data from the target forest area and perform data preprocessing to form a standardized multi-source dataset.

[0056] The dimensionality reduction decomposition module is used to perform dimensionality reduction processing on the standardized multi-source dataset and to perform orthogonal decomposition on the dimensionality-reduced multi-source dataset to extract the directions of each principal component.

[0057] The projection correction module is used to project the standardized multi-source dataset onto the directions of each principal component to obtain the scores of each principal component. Based on the noise estimation results of the dimensionality-reduced multi-source dataset, the scores of each principal component are denoised and corrected. The corrected scores are then polarized according to the index attributes of each principal component direction.

[0058] The comprehensive scoring module is used to determine the weight of each principal component based on the scores of each principal component after polarity correction, and to calculate the comprehensive quality score of each sample plot by weighting the principal components according to their weights.

[0059] An anomaly identification module is used to construct a conditional probability distribution of the signal intensity of each principal component based on the orthogonal decomposition results, calculate a significance measure based on the conditional probability distribution, and determine whether there are anomalies in each sample plot in each principal component direction based on the significance measure.

[0060] The early warning and diagnosis output module is used to classify sample plots that are judged to be abnormal and output graded early warning signals, and output diagnostic prompts based on the comprehensive quality score of each sample plot and the abnormality judgment results.

[0061] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising:

[0062] One or more processors;

[0063] Storage device for storing one or more programs;

[0064] When the one or more programs are executed by the one or more processors, the one or more processors implement the plot quality assessment and anomaly identification method based on multi-source forestry data fusion as described in any embodiment of the present invention.

[0065] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for plot quality assessment and anomaly identification based on multi-source forestry data fusion as described in any embodiment of the present invention.

[0066] In summary, the present invention achieves the following beneficial effects:

[0067] This invention achieves deep fusion and differentiated noise reduction of multi-source forestry data by constructing a block-based heteroscedastic noise model and integrating conditional principal component analysis, effectively overcoming the information distortion problem of traditional equal-variance fusion. Through an information entropy-driven objective weighting mechanism, it avoids subjective weighting bias, enhancing the objectivity and repeatability of the assessment results. Furthermore, the introduction of a selective inference framework to construct a conditional probability distribution provides rigorous statistical significance testing for anomaly identification, effectively controlling the risk of misjudgment. In addition, this invention ensures consistency between the comprehensive score and the positive trend of ecological quality through an automatic load polarity correction mechanism, and enables the model to adaptively evolve with the expansion of the monitoring system through an incremental dynamic update mechanism, ensuring the temporal consistency of long-term assessment results. Attached Figure Description

[0068] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0069] Figure 1 This is a flowchart of a method for plot quality assessment and anomaly identification based on multi-source forestry data fusion provided in an embodiment of the present invention;

[0070] Figure 2 This is an overall flowchart of the method for plot quality assessment and anomaly identification based on multi-source forestry data fusion provided in this embodiment of the invention;

[0071] Figure 3 This is a flowchart of the three-level anomaly tracing and four-color graded early warning system provided in the embodiments of the present invention, consisting of a data layer, an environment layer, and an ecological layer.

[0072] Figure 4 This is a schematic diagram of the structure of the sample plot quality assessment and anomaly identification device based on multi-source forestry data fusion provided in an embodiment of the present invention;

[0073] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0074] The present invention 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, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0075] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations (or steps) may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0076] Example 1

[0077] like Figure 1 As shown, Embodiment 1 of the present invention provides a flowchart of a method 100 for plot quality assessment and anomaly identification based on multi-source forestry data fusion. The method 100 specifically includes the following steps:

[0078] Step S110: Collect multi-source forestry data of the target forest area and perform data preprocessing to form a standardized multi-source dataset.

[0079] Specifically, fixed monitoring plots were established in the target forest area to collect the following four types of multi-source data:

[0080] (1) Remote sensing data: Multispectral / hyperspectral image data were acquired by using UAVs equipped with multispectral or hyperspectral sensors. After radiometric correction, geometric correction and atmospheric correction, vegetation indices and canopy structure parameters were extracted. Among them, vegetation indices include normalized vegetation index (NDVI), ratio vegetation index (RVI), enhanced vegetation index (EVI), etc., and canopy structure parameters include canopy height model parameters, which are used to characterize the spatial distribution and photosynthetic activity of the forest canopy.

[0081] (2) Ground survey data of sample plots: Tree measurements and community surveys were conducted within fixed sample plots to obtain the following characteristics:

[0082] Population characteristics: including but not limited to plant height, diameter at ground level, crown width, root width, biomass, etc.; flowering characteristics: including but not limited to flowering period, number of flowers, inflorescence size; leaf functional traits: including but not limited to leaf length, width, thickness, dry matter content, specific leaf area, etc.; pest and disease factors: including but not limited to insect population density, disease incidence rate, etc.; interspecific relationships: including but not limited to symbiotic fungi, epiphytes, etc.

[0083] (3) Soil microenvironment data: Soil profile samples were collected from the plots to determine soil physical properties (soil texture, bulk density, porosity, water holding capacity, soil moisture content, etc.), soil chemical properties (pH value, organic matter content, total nitrogen, available phosphorus, available potassium, etc.) and hydrological characteristics (evaporation intensity, soil infiltration characteristics, etc.) to characterize the site's nutrient supply capacity and water transport conditions.

[0084] (4) Litter decomposition and microbial data: The decomposition rate of litter was determined by the decomposition bag method. At the same time, litter layer and humus layer samples were collected. High-throughput sequencing technology was used to obtain microbial community structure data, and functional microbial groups and their relative abundance were extracted to characterize the material cycling function of the ecosystem.

[0085] As a preferred embodiment, the following preprocessing operations are performed sequentially on the collected multi-source data:

[0086] (1) Format unification and anomaly screening: Convert all data sources into a structured numerical table format and perform preliminary screening for outliers for each indicator to remove obvious outliers caused by recording errors or sensor malfunctions.

[0087] (2) Spatiotemporal registration: Match the spatial resolution of remote sensing images with the spatial scale of the sample plot ground survey. For example, resample the images to a scale consistent with the sample plot area, while ensuring that all data correspond to the same monitoring time window, such as aligning with "year-month" as the time unit.

[0088] (3) Scale unification: Convert indicators with different dimensions into dimensionless standardized dimensions so that all indicators are at the same reference benchmark in terms of numerical magnitude.

[0089] (4) Transformation processing for data type distribution: The original forestry data includes count data, proportional data and continuous data, and their distribution patterns are different. Directly performing multivariate statistical analysis would violate the assumptions of the statistical model regarding data distribution. Therefore, this invention implements transformation processing according to different data types:

[0090] For count-type data, such as high-throughput microbial sequencing reads and insect population density counts, variance stabilization transformations are used to convert the count distribution to an approximately normal distribution, including logarithmic transformations. Or inverse hyperbolic sine transformation To compress the long-tailed distribution and stabilize the variance,

[0091] For proportional data, such as litter decomposition residue rate, species cover, and pest and disease incidence, a logarithmic transformation after boundary truncation is used. Specifically, let the original proportional data... To avoid the logarithmic transform diverging at the 0 and 1 boundaries, a truncation parameter is used. Let the truncated intermediate variable Then transform into ,in, This is the original count data. The transformed value is... These are intermediate variables after truncation.

[0092] For continuous data, such as pH value, organic matter content, plant height, and diameter at breast height, the original numerical form is preserved without the need for distribution transformation.

[0093] After the above transformation, the distribution of indicators from each data source is approximately continuous, satisfying the model assumptions for subsequent multivariate statistical analysis.

[0094] All data sources after the above preprocessing and transformation are associated according to the sample plot identifier (such as sample plot number) to form standardized multi-source data. ,in, Number of sample plots; Original feature dimension.

[0095] Step S120: Perform dimensionality reduction on the standardized multi-source dataset, and perform orthogonal decomposition on the dimensionality-reduced multi-source dataset to extract the directions of each principal component.

[0096] To address the problem of high-dimensionality, low-sample data where the original feature dimension may exceed the number of sample plots, and to fully consider the differences in noise levels among multiple data sources, this invention employs a data-driven dimensionality reduction strategy, combined with... Figure 2 As shown, the process includes three sub-steps: pre-compression, block heteroscedasticity noise modeling, and conditional principal component analysis.

[0097] In a preferred embodiment, before performing the pre-compression step, the relationship between the feature dimension of the standardized multi-source dataset and the number of sample plots is first determined. Only when the feature dimension of the standardized multi-source dataset is greater than the number of sample plots is the standardized multi-source dataset pre-compressed, and the pre-compressed dataset is orthogonally decomposed.

[0098] When the feature dimension of the standardized multi-source dataset is no greater than the number of sample plots, pre-compression is not performed, and orthogonal decomposition is directly performed on the standardized multi-source dataset.

[0099] For example, define The feature dimensions after pre-compression. For the original feature dimension, For the number of sample plots, when At that time, pre-compression was performed using standard principal component analysis: the data matrix was... Perform singular value decomposition Take the front The principal component scores are used as the pre-compressed data matrix. ,in, For data matrix The left singular vector matrix, It is a diagonal matrix, that is, the diagonal elements are The singular values ​​are arranged in descending order; For data matrix The right singular vector matrix The transpose of .

[0100] Preferably, the above The selection rule is as follows: selection is based on a cumulative variance contribution rate of not less than 85%, and it is mandatory to meet the following conditions. If taken If the cumulative variance contribution rate of each principal component is still less than 85%, then... For the final value ( The hard constraints take precedence over the 85% threshold. This pre-compression step involves only the projection operation of standard PCA and does not involve the parameter construction of the subsequent selective inference framework. Therefore, it is methodologically decoupled from the subsequent statistical inference based on conditional principal component analysis and there is no circular dependency.

[0101] when At this time, no pre-compression is required, making That is, the pre-compressed data matrix is ​​equal to the original data matrix. .

[0102] In a preferred embodiment, after the pre-compression step is completed, the pre-compressed... The model is as follows:

[0103] ;

[0104] in, The signal mean matrix represents the true potential structure of the sample plot quality. : Noise matrix.

[0105] Considering that forestry multi-source data have significantly different measurement accuracy and noise levels from different sources, this embodiment will use pre-compressed data... The features are categorized according to data source. There are data blocks, and the noise elements within each block are independently and identically distributed on a matrix with a mean of 0 and a variance of . The normal distribution, i.e. ,in, ; Indicates the first The data block number to which each feature belongs ( ), For the first The noise variance of the data block to which each feature belongs.

[0106] Furthermore, the noise variance of each data block The estimation method is as follows: For the th Data block matrix , Calculate the number of features within this data block. A singular value ,but:

[0107] ;

[0108] in, , , For the first Data block matrix of The singular values ​​are arranged in descending order. This indicates the operation of taking the median. The median of the Marchenko-Pastur distribution. This represents the number of sample plots. It should be noted that this estimation method is applicable to... In scenarios where the mean matrix of the medium signal is approximately low-rank within each data block, such as when the rank does not exceed a certain threshold... When the assumption is not satisfied or In this case, empirical variance estimation based on repeated technical samples is used as an alternative.

[0109] The noise variance of each data block is combined to form the overall noise level. :

[0110] ;

[0111] When the noise variance of each data block When the ratio between them does not exceed 5, the noise covariance matrix in the pre-compressed projection direction can be approximated as: ,in, for The identity matrix, i.e., a square matrix with 1s on the diagonal and 0s elsewhere, allows the subsequent Wishart-Jacobian density formula for homoscedasticity to be directly applied. When the ratio exceeds 5, each data block needs to be independently normalized for noise: the feature of each data block is divided by the corresponding estimated standard deviation. This ensures that the standardized noise variance of each block is uniformly 1, i.e. Then merge them; at this time The Wishart-Jacobian formula for homoscedasticity holds strictly.

[0112] Through the above block heteroscedasticity noise modeling, the differentiated noise levels of each data source are effectively quantified and unified to a comparable noise scale, laying the foundation for accurate signal extraction in subsequent conditional principal component analysis.

[0113] As a preferred embodiment, conditional principal component analysis (C-PCA) includes: processing the pre-compressed data matrix... Perform singular value decomposition:

[0114] ;

[0115] in: : Represents the pre-compressed data matrix of These singular values, arranged from largest to smallest, reflect the data energy intensity in each principal component direction. Pre-compressed data matrix The left singular vector matrix; Pre-compressed data matrix The right singular vector matrix.

[0116] Definition of the first The principal component pairs of the signal mean matrix The conditional proportion of variance explained (C-PVE) is:

[0117] ;

[0118] in, and They are respectively and The Column, i.e., the first The left and right singular vectors corresponding to each principal component; The signal mean matrix The square of the Frobenius norm, when hour Therefore, the notation degenerates into .

[0119] Constructing the conditional principal component score matrix ,in, For the front A matrix consisting of principal component loading vectors, where m is the variable to be determined, is obtained by maximizing the log-likelihood function. To select the optimal value of m, denoted as :

[0120] ;

[0121] ;

[0122] in: The optimal number of principal components to retain, determined by the rank selection criterion. :by The log-likelihood function of the split point is used to evaluate the previous... Each principal component is considered a signal, and then... One is considered as the goodness of fit for noise. The index variable is used to iterate through the candidate values, and its value ranges from 1 to... ; It is a normal density function; and They are respectively the front One and after The sample mean of the squared singular values; For the front One and after Combined variance estimation of squared singular values; Pre-compressed data matrix of singular value vectors ; : No. Singular values ​​of squares To sum local index variables, max: Retrieves the parameter value corresponding to the maximum value.

[0123] Through the above three sub-steps, the multi-source data is adaptively reduced to [dimensionality]. With mutually orthogonal conditional principal component directions, effective signals are preserved and differential noise is separated, providing a low-noise, high signal-to-noise ratio data foundation for subsequent comprehensive quality scoring and anomaly identification of sample plots.

[0124] Step S130: Project the standardized multi-source dataset onto the directions of each principal component to obtain the scores of each principal component. Based on the noise estimation results of the dimensionality-reduced multi-source dataset, perform noise reduction correction on the scores of each principal component, and perform polarity correction on the corrected scores according to the index attributes of each principal component direction.

[0125] This step utilizes the comprehensive noise level estimated in step S120. The principal component scores are denoised to suppress noise components in the original scores. The corrected scores are then normalized and polarity corrected to ensure that the scores in each principal component direction are comparable in dimensionality and that a higher score indicates better sample land quality. This provides standardized input data for subsequent weighting and comprehensive scoring. The specific steps include:

[0126] (1) To facilitate the subsequent processing of sample plot scores according to the principal component direction, the principal component score matrix is ​​denoted as follows. transpose The elements are subscript Principal component numbering subscript Indicates the sample plot number All symbols used in the following text involving principal component indicators and sample plot indicators shall be based on this transposed convention.

[0127] (2) Denoising score, based on the comprehensive noise level estimated in step S120. Construct a shrinkage factor and define the denoised principal component score. for:

[0128] ;

[0129] in, For local index variables within the summation expression, For all The sum of squares of the singular values, i.e., the total observation energy; , For the pre-compressed data matrix The square of the Frobenius norm; The original principal component scores before denoising, i.e., the conditional principal component score matrix. Element; : Feature dimensions after pre-compression; Overall noise level.

[0130] (3) In order to eliminate the differences in dimensions and numerical ranges between the principal component directions and make the scores in each direction comparable before weighting, the range normalization is performed on the denoised scores, and the scores in each principal component direction are linearly mapped to the [0,1] interval to obtain the normalized scores. The specific formula is as follows:

[0131] ;

[0132] Wherein, if a certain indicator takes the same value in all sample plots, then it is defined as follows: Furthermore, this indicator is not included in subsequent entropy weight calculations; Number of sample plots; : in the In each principal component direction, all In each sample plot The minimum value; : in the In each principal component direction, all In each sample plot The maximum value; Indicates the sample plot number; The index variable represents the sample plot, used to traverse all sample plots to determine the global minimum and maximum values.

[0133] (4) Direction alignment and polarity correction: The principal component direction is a data-driven mathematical result, and its positive and negative directions are mathematically arbitrary—that is, if a vector is the principal component direction, its opposite vector is also an equivalent direction that satisfies the mathematical definition. Therefore, a high score may correspond to a high NDVI (positive indicator, representing good quality) or a high disease incidence rate (negative indicator, representing poor quality). The principal component direction itself does not automatically align with the direction of good ecological quality. Without polarity correction, direct weighted summation may lead to the paradox of a higher score resulting in a worse ecological quality, causing the assessment results to lose their business interpretability.

[0134] Therefore, this invention determines and corrects the polarity of each principal component direction after normalization, obtaining the final corrected principal component score. The specific method is as follows: For each principal component Examine the top 3 original features with the largest absolute values ​​in the load vector; if more than half of these features are known negative indicators, then the direction of the principal component is opposite to the quality direction, and the corresponding normalized score needs to be complemented. If there is a lack of domain prior knowledge of negative indicators, the complementation decision is postponed until domain experts confirm the directional meaning of each principal component.

[0135] Step S140: Determine the weight of each principal component based on the polarity-corrected principal component scores, and calculate the comprehensive quality score of each sample plot by weighting according to the weights of each principal component.

[0136] This step employs a data-driven entropy weighting method. It adaptively determines the weight of each principal component based on the dispersion of its scores among sample plots after polarity correction, and then calculates the overall quality score for each sample plot accordingly. The basic principle of the entropy weighting method is: the greater the difference in a principal component's score among different sample plots, the richer the information it contains for distinguishing between sample plots, and the higher its weight should be assigned in the overall evaluation; conversely, if a principal component's score tends to be consistent among sample plots, its distinguishing ability is weak, and it should be assigned a lower weight. Specifically, it includes the following steps:

[0137] (1) Convert the score of each sample plot in each principal component direction into the proportion of the total score of all sample plots in that direction, and use it as the probability distribution of the score of each principal component. The information entropy of each principal component among all sample plots is calculated based on the probability distribution of its score. :

[0138] ;

[0139] ;

[0140] in, Let the information entropy be the direction of the k-th principal component. The smaller the value, the more dispersed the principal component score is among the sample plots, the richer the discriminative information it carries, and the greater the weight it should be assigned. The score percentage of the j-th plot along the k-th principal component direction, i.e., the proportion of the score of the plot to the total score of all plots along the principal component direction, where k is the principal component number and j is the plot number. This represents the index variable of the sample plot, used to iterate through all sample plots; The sum of the scores of all plots along the direction of the k-th principal component is used as a normalization factor. This is the final score after polarity correction in step S130. This is a sample plot index variable used to iterate through all sample plots to calculate the sum. When, define .

[0141] (2) Determine the weights of each principal component based on the information entropy of each principal component's score:

[0142] ;

[0143] in, Let t be the weight of the k-th principal component, t be the summation index variable, and m be the number of optimal principal components to retain. The greater the redundancy, the more effective information the principal component contains in that direction, and the greater the weight it should be assigned.

[0144] (3) The scores of each plot in each principal component direction are weighted and summed according to their corresponding weights to obtain the comprehensive quality score of each plot:

[0145] ;

[0146] in, The overall quality score for the j-th sample plot.

[0147] Furthermore, based on the comprehensive quality score, the quality levels of the sample plots are divided according to preset grading rules. The [0,1] interval is divided into five equal parts using the equidistant method, corresponding to five levels: excellent, good, average, poor, and very poor. In an optional embodiment, the grading threshold can also be determined using the natural breakpoint method or the percentile method based on historical data to adapt to the quality distribution characteristics of different forest areas.

[0148] Through this step, the contributions of each principal component direction are objectively quantified and integrated into a single comprehensive score. The quality differences between sample plots can be compared and graded on the same scale, providing an intuitive and operable decision-making basis for forestry management and ecological monitoring.

[0149] Step S150: Construct the conditional probability distribution of the signal intensity of each principal component based on the orthogonal decomposition results, calculate the significance measure according to the conditional probability distribution, and determine whether there are anomalies in each sample plot in each principal component direction according to the significance measure.

[0150] This step utilizes the Selective Inference framework to perform a statistical significance test on the signal intensity of each principal component direction extracted in step S120, in order to determine whether there are statistically significant anomalies in each principal component direction for each sample.

[0151] It should be noted that, since the principal component directions are derived from the same dataset... The signal strength of principal components, estimated by singular value decomposition, becomes susceptible to selection bias if conventional statistical tests are performed directly on the same data for selecting a direction and then using the same data for inference. This leads to an inflated false positive rate and unreliable statistical significance of anomaly detection. The selective inference framework was designed to address this problem. Its conditional probability density incorporates the selection effect of principal component retention into the probability model, thereby inferring signal strength under the given selection event and correcting for selection bias.

[0152] (1) Construction of conditional probability distribution: Based on the singular values ​​and singular vectors in the orthogonal decomposition results, the signal intensity parameters of each principal component direction are determined; the data-driven screening process of the number of principal components retained is used as a conditional event, and the screening process includes the judgment of whether each principal component is retained; under the condition that the conditional event occurs, the conditional probability distribution of the signal intensity of each principal component is constructed.

[0153] For example, for the first Each principal component, given a selection event Under the condition that the k-th principal component is preserved by the rank selection criterion, the k-th principal component... The conditional density of the singular values ​​of the principal components is:

[0154]

[0155] in, : No. Singular values ​​of principal components; Except for the first The vector consisting of all singular values ​​other than the singular values ​​of the principal components, i.e. , For indicator functions, when Located in the interval The value is 1 if the time condition is met, and 0 otherwise. For indicator functions, when No more than the number of principal components retained The value is 1 if it is true, and 0 otherwise. : The left singular vector matrix; : The right singular vector matrix; : The index number of the principal component; The optimal number of principal components to retain, determined by a data-driven rank selection criterion; Combined noise variance, a weighted average of the noise variances of each data block; Left singular vector matrix The Column, corresponding to the first The left singular vectors of the principal components; : Signal mean matrix, dimension 1 , representing the true potential signal of the pre-compressed data; : Right singular vector matrix The Column, corresponding to the first The right singular vectors of the principal components; Number of sample plots; : Feature dimensions after pre-compression; : Summation index, traversal by All principal component numbers outside; : No. The singular values ​​of each principal component ; : No. The singular values ​​of the principal components are used as the lower bound of the interval; : No. The singular values ​​of the principal components are used as the upper bound of the interval; : Interval indicator function, when The value is 1 if the singular value is active and 0 otherwise, and is used to force the sorting of singular values. The Jacobian term generated by the Wishart distribution density transformation. It is the product of the absolute values ​​of the differences between the squares of all other singular values ​​and the square of the current singular value.

[0156] Crucially, because the present invention forces the satisfaction of the pre-compression step... Therefore, the index This ensured that hour The conditional density converges at the origin, thus making the selective inference framework mathematically integrable and engineeringally feasible in high-dimensional forestry data. This guarantee of mathematical convergence is the technical prerequisite for this invention to apply classical selective inference theory to high-dimensional forestry data.

[0157] (2) Null hypothesis and selectivity Value calculation:

[0158] Constructing the null hypothesis It is equivalent to If the null hypothesis is rejected, it means that there is a statistically significant signal in the direction of the kth principal component. The plot score in this direction can be used for anomaly identification.

[0159] Define selectivity Values ​​used for testing after considering the selection event. Significance:

[0160] ;

[0161] Among them, kernel function for:

[0162] ;

[0163] in, To make the first The singular values ​​of each principal component are replaced with The singular value vector after; These are the parameter values ​​in the hypothesis test, under the null hypothesis. ; : No. The conditional probability density of the singular values ​​of the principal components under other conditions; Selectivity value; The optimal number of singular values ​​to retain is determined by recalculating the rank based on the replaced singular value vector. : Kernel function, containing exponential, Jacobian, and singular value difference product terms, used to construct Integral the numerator and denominator of the value; : Integral variable, representing the first Candidate values ​​for the singular values ​​of each principal component; : No. The square singular values ​​are used to calculate item; From the observed values The integral interval to positive infinity; : Signal strength scalar, representing the first The signal projection along the direction of each principal component is the target parameter for null hypothesis testing. : The product of the absolute values ​​of the differences between the squares of all other singular values ​​and the square of the current singular value, derived from the Jacobian determinant of the Wishart distribution; : Jacobian term, derived from the density transformation of the Wishart distribution, with a positive exponent to guarantee integral convergence.

[0164] The construction principle of this p-value is as follows: kernel function Under the null hypothesis, the singular values ​​of the k-th principal component are described as taking different values. Relative likelihood at time; selection indicator function Ensure in every The selection event of re-evaluating whether the k-th principal component can still be retained is carried out under the following values; the molecule is the observed value. The probability of the tail reaching positive infinity, with the total probability of the entire space as the denominator, is the ratio of the two, which is the conditional p-value after considering the selection effect. The closer this p-value is to 0, the less likely the observed signal strength is to occur under the null hypothesis, thus providing a statistical basis for rejecting the null hypothesis.

[0165] (3) Anomaly detection:

[0166] Set significance level The present invention is preferred , calculate and Comparison: If Then we reject the null hypothesis. If a statistically significant signal exists in the direction of the k-th principal component, the sample plot should be marked as an anomaly in the corresponding principal component direction; otherwise, the null hypothesis is not rejected, and the sample plot has no significant anomaly in the corresponding principal component direction.

[0167] Through the above selective inference framework, this invention provides a rigorous statistical significance test guarantee for anomaly identification, and can distinguish between genuine ecological anomalies and random data fluctuations with a controllable Class I error rate, overcoming the inherent defects of traditional threshold methods or empirical rules in statistical inference.

[0168] Step S160: Classify the sample plots that are determined to be abnormal and output graded early warning signals, and output diagnostic prompt information by combining the comprehensive quality score of each sample plot with the abnormality determination results.

[0169] This step performs multi-level source tracing and classification on sample plots identified as abnormal, outputs graded early warning signals, and generates diagnostic prompts by combining the comprehensive quality score with the anomaly identification results. Simultaneously, this step also constructs confidence intervals for the signal intensity of each principal component based on heteroscedasticity data thinning technology to quantify the uncertainty of the assessment results and establishes a dynamic model update mechanism to ensure that the assessment system can adaptively evolve with the accumulation of monitoring data. Figure 3 As shown, the specific steps include:

[0170] (1) Criteria for classifying and quantifying anomaly types:

[0171] To facilitate tracing the root cause of anomalies and guide targeted on-site verification, this invention classifies anomalies into the following three levels:

[0172] Data layer anomalies: These are manifested as multiple statistical outliers appearing simultaneously within the same data block, while cross-source consistency verification fails. Cross-source consistency is evaluated using the Pearson correlation coefficient. When the residual between the remote sensing inversion biomass and the ground-measured biomass of a sample plot exceeds 2.5 times the standard deviation of the historical residual mean, it is judged as cross-source inconsistency.

[0173] Anomalies in the environmental layer: This manifests as the C-PVE of multiple data blocks simultaneously exhibiting deviations exceeding ±2 standard deviations from the historical baseline during the same period;

[0174] Ecosystem anomalies: This is manifested by the C-PVE of a specific principal component showing a monotonically decreasing trend for three or more consecutive periods, and the cumulative decrease exceeding the mean of historical changes plus twice the standard deviation, indicating a continuous degradation process.

[0175] Through the above three-level classification, the abnormality diagnosis is elevated from whether it is abnormal to what kind of abnormality and why it is abnormal, providing clear guidance for subsequent early warning classification and on-site verification.

[0176] (2) Graded Early Warning Judgment Rules: Based on the selective p-value and abnormal duration calculated in the above steps, graded early warning signals are output according to the preset multi-level corresponding rules. The core principle of graded judgment is: the grade judgment is based on the simultaneous satisfaction of the p-value range and duration conditions; if only the p-value condition is satisfied but the duration is insufficient, the grade corresponding to the p-value is downgraded by one level. Specifically, the early warning levels are divided into four levels from low to high: blue, yellow, orange, and red. The corresponding judgment conditions are shown in Table 1 below:

[0177] Table 1

[0178]

[0179] For example: in some way However, if only a single occurrence occurs, failing to meet the requirement of three or more consecutive occurrences, it is classified as an orange alert according to the downgrade rule. This downgrade rule is designed to prevent excessively high-level alerts from being triggered by occasional extreme observations, reflecting the robustness and fault tolerance of the early warning system.

[0180] (3) Anomaly Diagnosis Report Output: Based on the above anomaly classification results and graded early warning signals, an anomaly diagnosis report is generated. This report should include at least the following information: anomaly plot identifier, anomaly level, specific principal component direction of the anomaly and its corresponding ecological meaning, early warning level, and recommended key points for on-site verification. Through the above diagnosis report, decision-makers can quickly locate problematic plots and take targeted management measures.

[0181] In a preferred embodiment, the above-mentioned anomaly determination is a significant conclusion in the sense of point estimation, but the uncertainty interval of this conclusion is not provided. To quantify the uncertainty of the evaluation results, the present invention further constructs the confidence intervals of the signal intensity of each principal component, specifically including the following steps:

[0182] (1) Thinning of heteroscedastic data and independent SVD direction estimation.

[0183] For the The features of each data block are used to generate an independent noise matrix. Its elements obey .

[0184] For the split constant (Preferred version of the present invention) ),definition: ; ;but and They are statistically independent and Follows a matrix normal distribution:

[0185] ;

[0186] in, : Matrix normal distribution; : Signal mean matrix; An identity matrix, representing rows that are independent of each other; A diagonal matrix, where the diagonal elements are the variances of each feature; : No. The data block to which each feature belongs The noise variance; : Feature dimensions after pre-compression.

[0187] It should be emphasized that the above anomaly determination is based on The singular value decomposition serves hypothesis testing; the confidence interval construction in this step is entirely based on... Perform SVD again to obtain the direction vector. and .because and Independent of each other, based on The estimated direction and based on The estimated signal strength is strictly independent and does not depend on any approximation conditions. This confidence interval corresponds to the signal strength based on... Re-estimated C-PVE:

[0188] ;

[0189] in, The Column, that is, based on The estimated first The left singular vectors of the principal components; The Column, that is, based on The estimated first The right singular vectors of the principal components; :based on The re-estimated first The proportion of conditional variance explained by each principal component; :based on Estimated direction vector and signal mean matrix The projected intensity is The base of the squared term in the numerator; The square of the Frobenius norm.

[0190] (2) C-PVE confidence interval.

[0191] For a given confidence level Construct confidence intervals :

[0192] ;

[0193] in, The present invention is preferred , , for The confidence interval is obtained by numerical inversion of the signal strength, where This is the lower bound of the confidence interval. This is the upper bound of the confidence interval; for The confidence interval, where This is the lower bound of the confidence interval. This represents the upper bound of the confidence interval. As a preferred embodiment, forest ecosystems exhibit significant spatiotemporal dynamics, making it difficult for fixed-parameter models to adapt to the year-to-year changes in forest areas over the long term. Therefore, this invention sets two types of automatic update trigger conditions; satisfying either one initiates a full-process iterative update:

[0194] The number of newly added monitoring plots reaches 10% or more of the original total number of plots.

[0195] After the update is triggered, the system automatically re-executes all processing steps S110 to S150 in sequence, including data preprocessing, dimensionality reduction modeling, comprehensive score calculation, anomaly significance test, graded early warning judgment and confidence interval estimation, to complete the full-process iterative optimization of model parameters, so that the evaluation model can adaptively evolve with the expansion of the monitoring system and ensure the time consistency and comparability of long-term evaluation results.

[0196] Based on the above embodiments, the present invention achieves the following beneficial effects:

[0197] (1) This invention constructs a block heteroscedastic noise model, divides the features of multi-source forestry data into different data blocks according to their sources and estimates their noise variances separately, effectively distinguishing the differentiated noise levels of data from different sources such as remote sensing, ground surveys, soil, and microorganisms. On this basis, conditional principal component analysis is used to achieve deep fusion of multi-source data, overcoming the information distortion problem caused by traditional equal variance fusion methods, and significantly improving the accuracy and robustness of sample plot quality assessment.

[0198] (2) The present invention adopts a data-driven weighting mechanism based on information entropy. The weight of the index is determined adaptively by the degree of dispersion of the scores of each principal component among the sample plots. This avoids the human bias caused by subjective weighting methods such as traditional analytic hierarchy process or expert scoring. The weight allocation is entirely based on the statistical characteristics of the data itself, which enhances the objectivity and repeatability of the evaluation results.

[0199] (3) The present invention introduces a selective inference framework to construct the conditional probability distribution of the signal intensity of each principal component, and incorporates the data-driven screening process of the number of principal components retained as a conditional event into the probability model. This effectively corrects the selection bias caused by first using the same data to select the direction and then using the same data to make inference, so that the anomaly judgment has strict statistical significance test guarantee, and can distinguish between real ecological anomalies and random data fluctuations. Compared with the traditional threshold method, it significantly reduces the risk of misjudgment.

[0200] (4) This invention uses heteroscedasticity data thinning technology to split the dimensionality-reduced data into two independent sets of matrices, which are used for principal component direction estimation and signal intensity inference, respectively. It constructs a confidence interval for the conditional variance explanation ratio, realizes the quantitative characterization of the uncertainty of the evaluation results, and provides a statistical basis for reliable determination of the quality level of the sample plot and cross-period comparison.

[0201] (5) The present invention establishes an incremental model dynamic update mechanism. When the number of new time-series monitoring data or new sample plots reaches the preset ratio threshold, the whole process iterative update is automatically triggered, so that the evaluation model can adapt and evolve with the expansion of the monitoring system without the need for overall remodeling, thus ensuring the time consistency of long-term evaluation results.

[0202] Example 2

[0203] Based on Embodiment 1 above, this embodiment uses a forest area as the application scenario to verify the method of the present invention with actual forest area data, so as to prove the technical effect and practicality of the present invention.

[0204] Fixed monitoring plots were established in three core management areas of a certain forest area: Pudi, Jinpo, and Jiamu, totaling 20 monitoring plots (8 in Pudi, 6 in Jinpo, and 6 in Jiamu, i.e., the number of plots). In the above 20 sample plots, the following four types of raw data were collected:

[0205] UAV multispectral remote sensing imagery: 12 remote sensing features, including NDVI, EVI, and canopy height CHM, were extracted after preprocessing.

[0206] Field ground survey data includes 10 population structure traits such as plant height, ground diameter, crown width, and aboveground and belowground biomass, as well as stress indicators such as disease and pest incidence and insect population density.

[0207] Soil testing indicators include seven physicochemical parameters: pH, organic matter, total nitrogen, available phosphorus, available potassium, bulk density, total porosity, and soil moisture content.

[0208] Litter decomposition and microbial data: Litter decomposition rate and microbial high-throughput sequencing abundance data were used to extract the abundance characteristics of six functional microbial communities after variance stabilization transformation, which were used to characterize the material cycling function of the ecosystem.

[0209] Based on the above collection and processing, the total number of original features is... This constitutes the original data matrix. .

[0210] Following the data preprocessing procedure described in step S110 of implementation one, the four types of raw data are processed separately: for proportional indicators, a truncation threshold is set. A boundary-constrained logit transformation is performed to avoid divergence of the logarithmic transformation at 0 and 1 boundaries; for sequencing count features, a... Variance stabilization transformation is performed; for continuous features, the original numerical form is preserved. Subsequently, all data sources are uniformly matched with coordinates, extreme outliers are removed, and they are associated according to sample plot labels to form a standardized multi-source dataset.

[0211] Furthermore, the standardized multi-source dataset is pre-compressed and dimensionality reduced, because... If the number of sample plots is greater than 20, it falls under the category of high-dimensional, low-sample problem, requiring PCA pre-compression and selection of the compression dimension. Cumulative variance contribution rate Satisfying constraints If the cumulative variance contribution rate of the first 19 dimensions is less than 85%, then it will be forcibly taken. Compressed matrix .

[0212] After pre-compression The features were divided into four data blocks according to their data sources: remote sensing, ground survey, soil, and microorganisms. The noise variance of each data block was estimated separately, and the estimation results are as follows: Remote sensing block Ground survey block Soil blocks Microbial blocks Since the maximum ratio of variances among the groups is less than 3, no block standardization is required, and the overall noise level can be calculated directly. .right Conditional principal component analysis was performed, and the optimal number of principal components was determined using a data-driven rank selection criterion. The four principal components correspond to the vegetation growth comprehensive factor, soil fertility factor, litter decomposition function factor, and population structure factor, respectively, and explained a total of 78% of the effective signal variance, indicating that the four principal components have fully preserved the core information of the plot quality.

[0213] Further, the denoising correction is calculated according to the denoising shrinkage formula in step S130, and the range normalization is performed on the denoised score. Then, the polarity of each principal component direction is determined: the top three original features with the largest absolute values ​​in the loading vector of each principal component are checked one by one. It is determined that the dominant features of the first three principal components are all ecological positive indicators, and the scores are not reversed; the core traits of the fourth principal component are diameter at breast height and biomass, which are also positive attributes and do not require direction reversal.

[0214] Based on the scores of each principal component after polarity correction, the entropy weight method is used to calculate the weights of each principal component. The entropy weight calculation results are as follows: The weighted average quality score of each sample plot is obtained. The samples were divided into five levels at equal intervals: excellent, good, average, poor, and very poor. The statistics showed that there were 3 excellent samples, 7 good samples, 6 average samples, 3 poor samples, and 1 very poor sample.

[0215] Furthermore, using the selective inference framework described in step S150, a significance level is set. Calculate the selectivity of each principal component for each sample plot. Significantly anomalous sample plots were identified. The results are as follows:

[0216] (1) Second principal component of plot No. 7 Two consecutive periods of anomalies were detected, triggering an orange alert. Source analysis revealed that the anomaly stemmed from a significant drop in available phosphorus content in the soil compared to historical baselines, indicating an environmental stratum anomaly.

[0217] (2) Third principal component of plot No. 11 Two consecutive periods of abnormality have triggered a yellow alert. Source analysis revealed that the abnormality stemmed from a persistently low rate of litter decomposition, leading to a decline in nutrient cycling efficiency, which is considered an ecological layer anomaly.

[0218] (3) First principal component of plot No. 15 The three consecutive periods of abnormality have triggered a red alert. This abnormality stems from the continuous decline in canopy NDVI and CHM, indicating a trend of canopy degradation and declining growth, which is an ecological layer anomaly.

[0219] (4) Second principal component of plot No. 19 The alert level was downgraded to blue because the single-period alert was significant but the duration was insufficient. Source analysis showed that the soil phosphorus content in the sample plot was only slightly low and had not undergone substantial degradation.

[0220] Finally, following the heteroscedastic data thinning method described in Example 1 above, the splitting constant is taken. Construct independent samples ,right Singular value decomposition was performed separately, and the Scheme-conservative approximation was used to solve for the 90% confidence intervals of C-PVE corresponding to the above four anomalies: [0.001, 0.089], [0.012, 0.126], [0.001, 0.058], and [0.016, 0.099]. Since the two SVD steps are independent of each other, the slight difference in principal component ordering is normal and does not affect the functions of anomaly detection and uncertainty quantification.

[0221] To compare the technical effects of this invention, the mean ± 2 standard deviation threshold method, commonly used in traditional plot quality assessment, was employed to detect anomalies in the same dataset. The results showed that the traditional threshold method detected only two anomalies (plots 7 and 15), missed plots 11 and 19, and misclassified one normal plot as an anomaly.

[0222] In contrast, all four anomalies (No. 7, No. 11, No. 15, and No. 19) identified by the method of this invention were verified in the field and no false positives occurred. The above comparative experiments show that the present invention is significantly superior to the traditional threshold method in terms of both recall and precision in anomaly identification. Its core advantage lies in the fact that the selective inference framework provides strict Class I error control, while the confidence interval construction further enhances the reliability of the judgment in boundary cases.

[0223] In summary, this method successfully identified four anomalous plots, including one orange alert, one yellow alert, one red alert, and one blue alert, all of which were confirmed through on-site verification. The control experiment further demonstrates the significant advantages of this invention over the traditional threshold method in terms of recall and precision. This embodiment fully verifies the applicability and effectiveness of this invention under different forest areas and data conditions.

[0224] Example 3

[0225] Figure 4 This is a schematic diagram of the plot quality assessment and anomaly identification device based on multi-source forestry data fusion provided in Embodiment 3 of the present invention, as shown below. Figure 4 As shown, the device includes:

[0226] The data acquisition and preprocessing module 410 is used to collect multi-source forestry data of the target forest area and perform data preprocessing to form a standardized multi-source dataset.

[0227] The dimensionality reduction decomposition module 420 is used to perform dimensionality reduction processing on the standardized multi-source dataset and to perform orthogonal decomposition on the dimensionality-reduced multi-source dataset to extract the directions of each principal component.

[0228] The projection correction module 430 is used to project the standardized multi-source dataset onto the directions of each principal component to obtain the scores of each principal component, perform noise correction on the scores of each principal component based on the noise estimation results of the dimensionality-reduced multi-source dataset, and perform polarity correction on the corrected scores according to the index attributes of each principal component direction.

[0229] The comprehensive scoring module 440 is used to determine the weight of each principal component based on the scores of each principal component after polarity correction, and to calculate the comprehensive quality score of each sample plot by weighting according to the weights of each principal component.

[0230] Anomaly identification module 450 is used to construct a conditional probability distribution of the signal intensity of each principal component based on the orthogonal decomposition result, calculate a significance measure based on the conditional probability distribution, and determine whether there are anomalies in each sample plot in each principal component direction based on the significance measure.

[0231] The early warning and diagnosis output module 460 is used to classify the sample plots that are judged to be abnormal and output graded early warning signals, and output diagnostic prompt information based on the comprehensive quality score of each sample plot and the abnormality judgment result.

[0232] The plot quality assessment and anomaly identification device based on multi-source forestry data fusion provided in this embodiment of the invention can execute the plot quality assessment and anomaly identification method based on multi-source forestry data fusion provided in any of the above embodiments of the invention. It has the corresponding functions and beneficial effects of executing the plot quality assessment and anomaly identification method based on multi-source forestry data fusion. For detailed process, please refer to the relevant operations of the plot quality assessment and anomaly identification method based on multi-source forestry data fusion in the foregoing embodiments.

[0233] Example 4

[0234] Figure 5 This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0235] like Figure 5As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0236] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0237] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes the plot quality assessment and anomaly identification method based on multi-source forestry data fusion described above.

[0238] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0239] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for plot quality assessment and anomaly identification based on multi-source forestry data fusion, characterized in that, include: Collect multi-source forestry data from the target forest area and perform data preprocessing to form a standardized multi-source dataset; The standardized multi-source dataset is subjected to dimensionality reduction, and the dimensionality-reduced multi-source dataset is subjected to orthogonal decomposition to extract the directions of each principal component; The standardized multi-source dataset is projected onto the directions of each principal component to obtain the scores of each principal component. The scores of each principal component are then denoised and corrected based on the noise estimation results of the dimensionality-reduced multi-source dataset. Finally, the corrected scores are polarity-corrected according to the index attributes of each principal component direction, including: The standardized multi-source dataset is projected onto each principal component direction to obtain the original scores of each sample plot in each principal component direction. Based on the noise estimation results of the dimensionality-reduced multi-source dataset, the noise energy component is separated from the total observed energy of the principal component direction, and the original scores are shrunk and corrected according to the remaining signal energy to obtain the denoised principal component scores. Based on the index attributes of each principal component direction, it is determined whether the principal component direction is consistent with the positive trend of sample plot quality. If not, the polarity reversal correction is performed on the denoised principal component scores of each sample plot in the principal component direction. The noise estimation result is obtained by: dividing each feature of the dimensionality-reduced multi-source dataset into multiple data blocks according to the original data source; determining the noise variance within each data block using random matrix theory based on the singular value spectrum tail information of each data block; and merging the noise variances within each data block to obtain the noise estimation result. Determining whether a principal component direction is consistent with the positive trend of sample plot quality based on the index attributes of each principal component direction includes: obtaining the degree of correlation between each principal component direction and each original index; for any principal component direction, selecting several original indices according to the degree of correlation; determining whether the number of negative indices among the several original indices reaches a preset proportion or a preset number; if so, determining that the principal component direction is inconsistent with the positive trend of sample plot quality. The weights of each principal component are determined based on the scores of each principal component after polarity correction, and the overall quality score of each sample plot is calculated by weighting the principal components according to their weights. Based on the orthogonal decomposition results, conditional probability distributions of the signal intensities of each principal component are constructed. A significance metric is calculated based on these conditional probability distributions, and the presence of anomalies in each principal component direction is determined based on the significance metric. This includes: Based on the singular values ​​and singular vectors in the orthogonal decomposition results, the signal intensity parameters of each principal component direction are determined; the data-driven screening process based on the number of principal components retained is used as a conditional event, the screening process including the judgment of whether to retain each principal component; given the occurrence of the conditional event, a conditional probability distribution of the signal intensity of each principal component is constructed; according to the conditional probability distribution, the significance measure of the signal intensity in each principal component direction is calculated; the significance measure is compared with a preset significance level threshold; if the significance measure is less than or equal to the preset significance level threshold, it is determined that the sample plot has an anomaly in the corresponding principal component direction; the sample plots determined to be abnormal are classified and graded early warning signals are output, and diagnostic prompt information is output by combining the comprehensive quality score of each sample plot with the anomaly judgment result.

2. The method for plot quality assessment and anomaly identification based on multi-source forestry data fusion according to claim 1, characterized in that, The multi-source forestry data is preprocessed to form a standardized multi-source dataset, including: Data preprocessing is performed on the collected multi-source forestry data to obtain a dataset to be transformed with a unified spatial reference and dimensional standard. Based on the data type distribution characteristics of each data source in the dataset to be transformed, corresponding transformation processes are performed to transform each data source into a comparable data format. The data sources that have been transformed into comparable data formats are associated according to their sample site identifiers to form a standardized multi-source dataset.

3. The method for plot quality assessment and anomaly identification based on multi-source forestry data fusion according to claim 2, characterized in that, The standardized multi-source dataset is subjected to dimensionality reduction processing, and the dimensionality-reduced multi-source dataset is then orthogonally decomposed to extract the directions of each principal component, including: When the feature dimension of the standardized multi-source dataset is greater than the number of sample plots, pre-compression is performed on the standardized multi-source dataset, and the orthogonal decomposition is performed on the pre-compressed dataset to extract the directions of each principal component. When the feature dimension of the standardized multi-source dataset is not greater than the number of sample plots, the orthogonal decomposition is directly performed on the standardized multi-source dataset to extract the directions of each principal component.

4. The method for plot quality assessment and anomaly identification based on multi-source forestry data fusion according to claim 1, characterized in that, The weights of each principal component are determined based on the polarity-corrected principal component scores, and the overall quality score of each sample plot is calculated by weighting the principal component scores according to their respective weights, including: The scores of each plot in each principal component direction are converted into the proportion of the total scores of all plots in that direction, which is used as the probability distribution of the scores of each principal component. The information entropy of each principal component among all sample plots is calculated based on the probability distribution of its score, and the weight of each principal component is determined based on the information entropy of its score. The scores of each plot in each principal component direction are weighted and summed according to their corresponding weights to obtain the comprehensive quality score of each plot.

5. The method for plot quality assessment and anomaly identification based on multi-source forestry data fusion according to claim 1, characterized in that, The sample plots identified as abnormal are classified and graded early warning signals are output, including: Based on preset anomaly type determination criteria, sample plot anomalies are classified into data layer anomalies, environmental layer anomalies, and ecological layer anomalies. Based on the value range of the significance metric and the duration of the anomaly, the corresponding warning level is output according to the preset multi-level correspondence rules.

6. The method for plot quality assessment and anomaly identification based on multi-source forestry data fusion according to claim 5, characterized in that, The method further includes: When the significance metric meets the value range condition corresponding to a certain warning level, but the duration of the abnormality does not reach the duration condition corresponding to that warning level, the output is reduced by one level according to that warning level.

7. A device for plot quality assessment and anomaly identification based on multi-source forestry data fusion, characterized in that, The apparatus for performing the plot quality assessment and anomaly identification method based on multi-source forestry data fusion as described in any one of claims 1 to 6, the apparatus comprising: The data acquisition and preprocessing module is used to collect multi-source forestry data from the target forest area and perform data preprocessing to form a standardized multi-source dataset. The dimensionality reduction decomposition module is used to perform dimensionality reduction processing on the standardized multi-source dataset and to perform orthogonal decomposition on the dimensionality-reduced multi-source dataset to extract the directions of each principal component. The projection correction module is used to project the standardized multi-source dataset onto the directions of each principal component to obtain the scores of each principal component. Based on the noise estimation results of the dimensionality-reduced multi-source dataset, the scores of each principal component are denoised and corrected. The corrected scores are then polarized according to the index attributes of each principal component direction. The comprehensive scoring module is used to determine the weight of each principal component based on the scores of each principal component after polarity correction, and to calculate the comprehensive quality score of each sample plot by weighting the principal components according to their weights. An anomaly identification module is used to construct a conditional probability distribution of the signal intensity of each principal component based on the orthogonal decomposition results, calculate a significance measure based on the conditional probability distribution, and determine whether there are anomalies in each sample plot in each principal component direction based on the significance measure. The early warning and diagnosis output module is used to classify sample plots that are judged to be abnormal and output graded early warning signals, and output diagnostic prompts based on the comprehensive quality score of each sample plot and the abnormality judgment results.

Images