Tree health AI evaluation method and system fused with multispectral analysis

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a tree health AI evaluation method and system fused with multispectral analysis, and the method comprises the following steps: obtaining the multispectral reflectivity of a tree canopy, extracting a time sequence deviation and an extreme point, generating multispectral feature data, analyzing the change direction and extreme frequency of a key spectral segment, and screening a high-correlation spectral segment combination. The method comprises the following steps: forming a health spectral feature library, monitoring abnormal fluctuation, tracking a disturbance propagation chain, constructing feature vectors in combination with environmental parameters, training a health multi-classification model by adopting a support vector machine algorithm, obtaining a health state boundary, evaluating input data and outputting a tree health state result. According to the method, the health assessment model is constructed by analyzing the subtle change of the multi-time-sequence multi-spectral reflectivity, mining the dynamic association of the key wave bands and combining the environmental parameters, so that the health change sensitivity and classification precision are improved, early tree abnormal condition detection and time-space continuous analysis are realized, and a scientific basis is provided for forestry management.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI assessment method and system for tree health that integrates multispectral analysis. Background Technology

[0002] The field of artificial intelligence (AI) involves using computer algorithms and models to analyze and reason about complex data, and to achieve automated decision support in several core areas such as image recognition, speech processing, and data mining. It has particularly wide applications in multi-source data fusion, feature extraction, and automatic classification and recognition. This field relies on systematic technical processes such as large-scale data training, model building, and result inference. In the application of AI in tree health detection and assessment, related research focuses on fusing multispectral image data with environmental information, and using the analysis of spectral reflectance changes in different bands to assess and classify the physiological state of trees.

[0003] Traditional tree health assessment methods rely on manual observation or analysis based on single-spectral images to judge tree diseases, pests, physiological state, or growth status. They identify tree diseases by leaf color changes and morphological characteristics in visible light images or thermal distribution characteristics in infrared images. However, when faced with complex forest environments or multiple interferences, the assessment results are easily affected by factors such as changes in light intensity and species differences, leading to inaccuracies. These methods only extract crown texture from single-band images or use temperature distribution in near-infrared images to mark potential lesions, making it difficult to achieve a comprehensive identification and analysis of tree health status.

[0004] Existing technologies are limited by single-band static images and manual identification methods, lacking in-depth capture of the temporal response of key band reflection signals. This makes it difficult to show the continuous evolution of tree health status with environmental changes. When faced with environmental disturbances and tree species heterogeneity, signs of disease are easily masked or misjudged, resulting in weak dynamic perception capabilities and a tendency to miss early warning opportunities. The lack of quantitative characterization of health fluctuations limits the proactive identification and timely prevention and control of potential risks in practical applications. Summary of the Invention

[0005] To address the limitations of existing technologies, which rely on single-band static images and manual identification, lack in-depth capture of the temporal response of key band reflection signals, and struggle to demonstrate the continuous evolution of tree health status with environmental changes, this invention provides a tree health AI assessment method integrating multispectral analysis, comprising the following steps: To achieve the above objectives, the present invention adopts the following technical solution: a tree health AI assessment method integrating multispectral analysis, comprising the following steps: S1: Acquire multispectral reflectance data of tree canopy through multispectral sensors, label tree species ID and phenological period information, extract reflectance deviation between adjacent time points according to time series, record wavelength, rate of change and corresponding tree species ID and phenological period to form a multispectral feature dataset; S2: Call the multispectral feature dataset, analyze the direction of spectral band changes, count the frequency of extreme points of spectral bands, use the Pearson correlation coefficient algorithm to filter spectral band combinations whose correlation exceeds the correlation threshold, remove normal phenological change signals, and generate a tree health spectral feature library. S3: Based on the tree health spectral feature library, monitor abnormal fluctuations in reflectance within the time window, compare changes in response of other spectral bands, detect response delay, track the cross-spectral propagation of disturbance signals, record the disturbance source spectral band identifier, response spectral band identifier, and propagation delay parameters, and construct a disturbance propagation chain feature library; S4: Based on the disturbance source identifier and propagation delay parameter in the disturbance propagation chain feature library, and combined with the environmental temperature, humidity and light intensity corresponding to the tree species' ecological habits, construct a dynamic feature vector, use the support vector machine algorithm to train a multi-classification model of health status, and generate a tree health AI assessment model.

[0006] As a further embodiment of the present invention, the multispectral feature dataset includes extreme wavelengths, reflectance changes, and time indexes; the tree health spectral feature library includes spectral change trends, associated spectral band combinations, and extreme point occurrence frequencies; the disturbance propagation chain feature library includes disturbance propagation paths, associated spectral bands, and response delay times; and the tree AI intelligent assessment model includes health status grading boundaries, feature vector parameters, and support vector machine model parameters.

[0007] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect reflectance data of tree canopy bands with tree species ID and phenological period annotation at different observation time points using a multispectral sensor, arrange the reflectance values ​​at the same wavelength in chronological order, construct a time series according to wavelength, tree species ID and phenological period information and record the corresponding reflectance group data to generate a multi-temporal band reflectance sequence. S102: Based on the multi-temporal band reflectance sequence, calculate the difference of band reflectance data at adjacent time points one by one, extract the reflectance change amplitude and its time index according to the time point sequence corresponding to the difference, associate the tree species ID and phenological period information, and generate a band reflectance deviation dataset. S103: Based on the band reflectance deviation dataset, set a fixed reflectance variation range limit for the reflectance deviation value of each group of wavelengths, identify the wavelength and time index combination that exceeds the variation range limit, and combine the tree species ID and phenological period to record the corresponding reflectance variation value to generate a multispectral feature dataset.

[0008] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Call the multispectral feature dataset, extract the reflectance sequences of the red edge, near-infrared and green spectral bands, analyze the reflectance change trend of the spectral bands at different time points, record the increase or decrease of reflectance values ​​at continuous time points, and generate a sequence of spectral band change directions. S202: Based on the sequence of spectral change directions, count the time nodes when each local maximum and minimum value of each spectral segment appears in the time series, accumulate the frequency of occurrence of extreme points, calculate the distribution of the number of extreme points of the spectral segment in the complete time series, and obtain the extreme frequency of the spectral segment. S203: Based on the extreme frequencies of the spectral bands, the Pearson correlation coefficient algorithm is used to calculate the correlation of extreme frequency sequences between the red edge, near-infrared, and green spectral bands. Spectral band combinations with correlation greater than the correlation threshold and excluding normal phenological change signals are selected to establish a tree health spectral feature library. The correlation threshold of the Pearson correlation coefficient is set between 0.6 and 0.85, preferably 0.75. The threshold is obtained by analyzing multiple sets of multi-temporal multispectral experimental data to ensure that the selected spectral band combination has a moderate or higher linear correlation under typical physiological state changes, so as to enhance the stability and physiological interpretability of the model.

[0009] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the tree health spectral feature library, obtain the reflectance value of the observation node in the monitoring window, determine whether the reflectance difference of a specific band exceeds the corresponding tree species fluctuation benchmark value, filter the fluctuation over-limit band number, and generate abnormal reflectance band identifier. S302: Based on the abnormal reflectivity band identifier, calculate the time difference between the starting point of reflectivity change in the non-abnormal band and the starting point of the corresponding abnormal band change, construct the time response mapping between bands, extract the time interval value, and generate a spectral band response delay coefficient group. S303: Based on the spectral band response delay coefficient group, arrange the band correspondence in the delay order from the abnormal band to the response band, combine the original band number and its corresponding response band number, match the propagation time value between the two, extract the propagation direction and delay value, and establish a disturbance propagation chain feature library.

[0010] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Based on the disturbance source identifier and propagation delay parameter in the disturbance propagation chain feature library, combined with the environmental temperature, humidity and light intensity corresponding to the tree species' ecological habits, calculate the ratio of the change amplitude of the disturbance source parameter to the delay difference, screen samples that meet the set ratio range, and obtain the dynamic feature vector of the propagation response. S402: Based on the propagation response dynamic feature vector, classify and process sample labels, construct a nonlinear mapping using the RBF kernel function in the support vector machine, calculate the distance from the perturbed sample to the center point of the grade label, determine the four-class boundaries, and obtain the four-level classification boundary interval values. The four-level classification boundary interval value refers to the set of thresholds that divide the sample feature space into four non-overlapping and continuous health level discrimination intervals after nonlinear mapping of the perturbation sample features based on the support vector machine model. S403: Call the four-level classification boundary interval values ​​and the propagation response dynamic feature vector, input the perturbation sample into the support vector machine model, count the classification probability values ​​of the sample in the four levels and summarize them to generate a tree health AI assessment model.

[0011] As a further aspect of the present invention, the method further includes step S5: S5: The tree health AI assessment model is used to process the multispectral data of the tree to be assessed, calculate the similarity score between the input spectral features and the health level classification boundary, compare the similarity values ​​to determine the current health status level of the tree, and obtain the tree health status assessment result. The tree health status assessment results include health grade results and similarity scores.

[0012] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the tree health AI assessment model, process the multispectral data of the trees to be assessed, analyze the band spectral reflectance, vegetation index and local spectral gradient, perform feature normalization to adjust the scale and distribution of band values, and generate spectral embedded feature values. The normalization process uses the Min-Max normalization method to linearly map the wavelength reflectance of the differential spectral bands between 0 and 1 within the observation period, thereby eliminating interference from time and illumination differences. S502: Based on the spectral embedding feature value, extract the level feature center vector according to the health level classification boundary parameter, calculate the similarity value between the embedded feature and the center according to the cosine similarity method, and sort all level similarity results by size to obtain the health level similarity ranking value. The level feature center vector refers to the average embedding representation of the feature vectors corresponding to health levels in the training set; S503: Based on the health level similarity ranking value, extract the health level corresponding to the first similarity value in the ranking, combine the coordinate position of the spectral embedding feature value in the feature space, compare it with the boundary coordinates of the health level interval, classify the tree to be evaluated into the corresponding level interval, and generate the tree health status assessment result.

[0013] As a further aspect of the present invention, the spectral embedding feature value refers to the feature vector that represents the health status of trees, obtained by mapping the tree multispectral data after normalization and feature fusion. The health level similarity ranking value refers to the result calculated based on the spectral embedding feature value and the health level feature center vector and sorted according to the similarity.

[0014] A tree health AI assessment system integrating multispectral analysis includes: The multispectral acquisition module acquires multispectral reflectance data of tree canopy through multispectral sensors, arranges reflectance in time series, extracts reflectance deviation between adjacent time points, identifies extreme points, records wavelength and rate of change, and simultaneously collects and labels current phenological period and tree species identity information to form a multispectral feature dataset, which is then transmitted to the feature extraction module. The feature extraction module calls the multispectral reflectance dataset, analyzes the changes in red edge, near-infrared, and green light bands, counts extreme frequency, filters band combinations with Pearson correlation coefficients exceeding the threshold, distinguishes and removes signals that conform to the leaf color and reflectance changes of the tree species in the normal phenological stage, generates a tree health spectral feature library, and transmits it to the perturbation analysis module. The disturbance analysis module, based on the tree health spectral feature library, monitors abnormal fluctuations in reflectance within a set time window, compares the response time differences of different bands, detects signal propagation delay, tracks the path of the disturbance signal between bands, records the disturbance source, response band, and delay parameters, constructs a disturbance propagation feature library, and transmits it to the state classification module. The state classification module constructs a dynamic feature vector based on the disturbance source identifier and propagation delay parameter in the disturbance propagation feature library, combined with the environmental temperature, humidity and light intensity values ​​corresponding to the tree species' ecological habits. This vector is then input into the support vector machine algorithm for training, calculates the four-level classification boundary parameters, generates a tree health AI assessment model, and is transmitted to the assessment output module. The evaluation output module processes the spectral data of the trees to be evaluated through the tree health AI evaluation model, calculates the similarity score between the input spectral features and the health level classification boundary, compares the similarity values ​​to determine the current health status level of the trees, and outputs the tree health status evaluation result.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention analyzes the subtle trends in multi-temporal and multi-spectral reflectance to comprehensively uncover the transmission and dynamic correlation characteristics between key bands. It can distinguish the initial response, propagation path, and delay patterns of health status signals in differentiated spectral bands. By combining real-time environmental parameters such as temperature, humidity, and light intensity, a dynamic health assessment model is constructed. The model's sensitivity to changes in health status and classification accuracy are optimized, thereby effectively identifying early latent abnormalities and potential disease states in trees. This achieves spatiotemporal continuity and differentiated expression of health analysis results, improves the timeliness and reliability of anomaly detection and early warning, and provides more guiding scientific basis and data support for forestry management decisions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0020] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0021] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] Please see Figure 1 This invention provides an AI-based tree health assessment method that integrates multispectral analysis, comprising the following steps: S1: Acquire multispectral reflectance data of tree canopy through multispectral sensors, label tree species ID and phenological period information, extract reflectance deviation between adjacent time points according to time series, record wavelength, rate of change and corresponding tree species ID and phenological period to form a multispectral feature dataset; S2: Call the multispectral feature dataset, analyze the direction of spectral band changes, count the frequency of extreme points of spectral bands, use the Pearson correlation coefficient algorithm to filter spectral band combinations whose correlation exceeds the correlation threshold, remove normal phenological change signals, and generate a tree health spectral feature library. S3: Based on the tree health spectral feature library, monitor abnormal fluctuations in reflectance within a time window, compare changes in response in other spectral bands, detect response delay, track the cross-spectral propagation of disturbance signals, record the disturbance source spectral band identifier, response spectral band identifier, and propagation delay parameters, and construct a disturbance propagation chain feature library; S4: Based on the disturbance source identifier and propagation delay parameters in the disturbance propagation chain feature library, and combined with the environmental temperature, humidity and light intensity corresponding to the ecological habits of tree species, a dynamic feature vector is constructed. The support vector machine algorithm is used to train a multi-classification model of health status to generate a tree health AI assessment model. S5: Process the multispectral data of the trees to be evaluated using the tree health AI assessment model, calculate the similarity score between the input spectral features and the health level classification boundary, compare the similarity values ​​to determine the current health status level of the trees, and obtain the tree health status assessment results.

[0024] The multispectral feature dataset includes extreme wavelengths, reflectance changes, and time indices; the tree health spectral feature library includes spectral change trends, associated spectral band combinations, and extreme point frequency; the perturbation propagation chain feature library includes perturbation propagation paths, associated spectral bands, and response delay times; the tree AI intelligent assessment model includes health status grading boundaries, feature vector parameters, and support vector machine model parameters; and the tree health status assessment results include health level results and similarity scores.

[0025] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect reflectance data of tree canopy bands with tree species ID and phenological period annotation at different observation time points using a multispectral sensor, arrange the reflectance values ​​at the same wavelength in chronological order, construct a time series according to wavelength, tree species ID and phenological period information and record the corresponding reflectance group data to generate a multi-temporal band reflectance sequence. By collecting reflectance data of tree canopy bands at differentiated observation time points using a multispectral sensor, the strategy for setting these differentiated observation time points must first be clearly defined. For example, observation times can be set at 9:00 AM, 12:00 PM, and 3:00 PM daily, three times a day, for 10 consecutive days, generating 30 observation time points. Each observation must record the corresponding tree species ID and phenological period information to ensure the traceability and classification of subsequent data analysis. Each time point should be named accordingly. Then, the reflectance data collected from each observation were categorized and summarized, mainly focusing on the visible light band (e.g., red light 660nm, green light 560nm, blue light 470nm) and the near-infrared band (800nm), with each band denoted as [missing information]. , respectively Each band corresponds to 30 time points. Reflectance data labeled under different tree species and phenological stages are arranged in chronological order to generate a time series, producing a format like... The reflectance sequence, where Indicates at wavelength Below, under the tree species ID and phenological period P annotation, the first The reflectance value at each time point, for example This indicates that the reflectance of the pine tree sample during the budding stage was 0.37 in the green light band during the 5th observation. Reflectance data from different time points within the same band were sequenced according to tree species ID and phenological period information. For example, the sequence corresponding to the green light band is... Then, the band sequence is arranged according to... After being processed and summarized, they are compiled into a multi-temporal band reflectance sequence set. This set is used to record the reflectance change process of each band at the observation time point. For example, taking the test results of a certain pine tree in the experimental forest as an example, the following data were obtained from 30 observations over 10 days: Table 1: Multi-temporal reflectivity data (partial)

[0026] As shown in Table 1, the reflectance of different wavelengths forms a series of reflectance sequences on the time axis. In the 470nm wavelength band, the reflectance sequence of pine tree ID001 during the budding stage is [0.24, 0.26, 0.27, 0.25, 0.26], which reflects the blue light reflection characteristics of the canopy during this period. Corresponding sequences are established for the other wavelength bands in the same way. During the execution process, the data acquisition steps must ensure that the wavelength acquisition time is synchronized, and the equipment response time should be controlled within ±1 second. The wavelength range of reflectance acquisition can also be combined with the canopy structure of different tree species. Adjustments are made, for example, conifers are sensitive to the red-edge band of reflectance, so the 720nm band can be introduced for expansion, resulting in new data sequences at the same acquisition frequency; the acquired data needs to be processed for consistency, such as removing outliers (e.g., a data value that is significantly lower than the average of three consecutive points by more than 30% is defined as an outlier), and after outlier removal, it needs to be filled with the average of nearby time points. After cleaning and calibration, the band sequences are aligned according to observation time, tree species ID and phenological period information to form a complete and clearly labeled multi-temporal band reflectance sequence dataset.

[0027] The results indicate that, through the above data collection and processing process, the reflectance sequence of the band at multiple time points has been constructed, forming a complete multi-temporal band reflectance sequence dataset.

[0028] S102: Based on the multi-temporal band reflectance sequence, the difference of band reflectance data at adjacent time points is calculated item by item. The reflectance change amplitude and its time index are extracted according to the time point sequence corresponding to the difference. The tree species ID and phenological period information are associated and recorded to generate a band reflectance deviation dataset. Based on multi-temporal band reflectivity sequences, a specific band is first selected. With green light band For example, for a sample of a certain tree species (e.g., tree species ID001) at a certain phenological stage (e.g., budding stage P1), from the time point... to A total of 30 reflectance values ​​were obtained. to The difference in reflectance between adjacent time points is calculated item by item, that is, for arrive ,calculate ,like and Then there is This step is performed for each band sequence, such as the near-infrared band. ,like , The corresponding difference During execution, the difference should be kept with a precision of two decimal places, using the difference sequence. Combined with the corresponding time index Constructing band reflectivity variation data for Each band generates a set of 29 reflectance variation records. This process is then extended to all bands. After performing the same operation, the reflectance difference data corresponding to each band is integrated and summarized into a multi-band reflectance deviation dataset. Each record contains a point-in-time index. Reflectivity variation range and their corresponding wavelengths With tree species ID and phenological period labeling information, such as a record as This indicates that the green light band of tree species ID001 during the budding stage is... and The reflectance variation between intervals is 0.02. To avoid abnormal changes, an amplitude filtering rule needs to be implemented, for example, for three consecutive intervals. Cases exceeding 0.03 are marked as drastic change segments, and their starting time index and wavelength are recorded. When calculating the difference, time point data marked as abnormal fill should also be excluded to avoid amplification errors. Finally, the constructed band reflectivity deviation dataset is output.

[0029] S103: Based on the band reflectance deviation dataset, set a fixed reflectance variation range limit for the reflectance deviation value of each group of wavelengths, identify the wavelength and time index combination that exceeds the variation range limit, and combine the tree species ID and phenological period to record the corresponding reflectance variation value to generate a multispectral feature dataset. Based on the band reflectivity deviation dataset First, for each record in the set... To make a judgment, a fixed limit for the range of reflectivity variation needs to be set. This threshold value should be set differently for different wavelengths, such as blue light. set up Green light set up Red light set up Near-infrared set up The set values ​​are derived from the statistical difference analysis results of measured data of healthy trees and trees affected by Fusarium wilt. For example, by statistically analyzing the range of reflectance changes of 500 tree samples over a week, the 90th percentile is taken as the upper limit of the band variation. Perform a judgment operation on each piece of data; if... The record is then considered an anomalous change point and is retained. Otherwise, it will be removed, such as a certain record. ,because Therefore, this data point is retained, and the execution process needs to iterate through it. Enter the entries and create the corresponding output sets: ; This involves filtering records of reflectance changes exceeding the corresponding band threshold. During the process, it is crucial to ensure that the original band marker and time index accurately correspond to each judgment operation to avoid mismatches. The final output set is as follows. This is the multispectral feature dataset, which serves as the input source for subsequent tree health assessment. This set does not contain data that does not exceed the amplitude limit and represents the time points and magnitudes of significant changes across multiple bands.

[0030] Please see Figure 3 The specific steps of S2 are as follows: S201: Call the multispectral feature dataset, extract the reflectance sequences of the red edge, near-infrared and green spectral bands, analyze the reflectance change trend of the spectral bands at different time points, record the increase or decrease of reflectance values ​​at continuous time points, and generate a sequence of spectral band change directions. Using the reflectance sequences of the red edge (720nm), near-infrared (800nm), and green (560nm) bands from the multispectral feature dataset, the reflectance is first extracted from the red edge band. to Reflectance sequence: ; Near-infrared spectral extraction ; Green spectrum extraction ; For each spectral band, calculate from to Direction change sequence ,in Indicates an increase, Indicates a decrease. It indicates that it remains unchanged, for example, in the red-edge spectral section if , ,but In the green light , ,but After extracting the direction sequence for each time series, the combination of spectral patterns in three consecutive time periods is recorded, such as near-infrared in... to The presence of the direction sequence [1, -1] indicates an upward followed by a downward sequence. Each group of consecutive direction sequences is extracted by sliding a window, for example, by shifting the window by a length of 3 to form a sequence fragment. This process involves statistically analyzing the types and frequencies of spectral variation trend patterns, and finally merging and organizing the variation direction sequences of red edge, near-infrared, and green light into a set of spectral variation direction sequences. ,in Indicates the band. For time point indexing, To illustrate the directional changes, a complete record of the trend data for each band is generated, as shown in Table 2.

[0031] Table 2: Sequence of Spectral Reflectance Direction Changes

[0032] Table 2 lists the specific directional information of reflectance changes under the selected spectral band and time index.

[0033] S202: Based on the sequence of spectral change directions, count the time nodes when each local maximum and minimum value of each spectral segment appears in the time series, accumulate the frequency of occurrence of extreme points, calculate the distribution of the number of extreme points of the spectral segment in the complete time series, and obtain the extreme frequency of the spectral segment. Based on the set of spectral band change direction sequences, local extreme points appearing in each spectral band are identified one by one. The maximum value is determined when the current value in the sequence is greater than the reflectance values ​​of the two adjacent time points, i.e., if... and Then record If it is a maximum point, and Then record For example, in the green spectral band, if the minimum value is... , , ,but To find the maximum value, construct an index sequence of extreme points. and For data at 30 time points, using Each time point is analyzed to identify the maximum and minimum values, and their band labels and extreme values ​​are recorded. The frequency of extreme points in each spectral band is calculated, and the frequency is defined as: ; in The number of extreme points in a certain waveband. This represents the total number of time points. For example, if the red-edge spectral band identifies 4 maxima and 3 minima out of 30 time points, then its total number of extrema is 7, corresponding to a frequency of... By comparing the number of extreme points and frequency distribution in different bands, a frequency distribution vector is constructed. This will be used for the next step of correlation analysis.

[0034] S203: Based on the extreme frequency of the spectral bands, the Pearson correlation coefficient algorithm is used to calculate the correlation of extreme frequency sequences between the red edge, near-infrared, and green spectral bands. Spectral band combinations with correlation greater than the correlation threshold and excluding normal phenological change signals are selected to establish a tree health spectral feature library. The correlation threshold of the Pearson correlation coefficient was set between 0.6 and 0.85, preferably 0.75. The threshold was obtained by analyzing multiple sets of multi-temporal multispectral experimental data to ensure that the selected spectral band combination has a moderate or higher linear correlation under typical physiological state changes, so as to enhance the stability and physiological interpretability of the model. Using frequency distribution vector Corresponding to the extreme frequencies of the green, red-edge, and near-infrared spectral bands respectively, the Pearson correlation coefficient is used to perform correlation analysis of the extreme frequencies between the spectral bands. For any two bands... and Let their extreme frequency vectors be respectively. and Use the formula: ; Perform calculations, assuming Group sampling data, extract the green light frequency vector The red border is ,but: ; ; ; ;

[0035] If the preset correlation threshold is 0.75, then If the screening criteria are not met, the spectral band combination will not be retained and will be recalculated. For example, the correlation coefficient between the red edge and the near-infrared band is... Therefore, This change does not fall within the normal fluctuation range of the phenological period (which can be compared with a control sample database) and meets the screening criteria, therefore it will be included. and The combinations selected are those with high frequency correlation of spectral bands. Finally, the combinations of spectral bands that meet the correlation conditions are added to the tree health spectral feature library, along with their extreme frequency vectors and time index information, for use in subsequent discrimination model calls.

[0036] Please see Figure 4 The specific steps of S3 are as follows: S301: Based on the tree health spectral feature library, obtain the reflectance value of the observation node within the monitoring window, determine whether the reflectance difference of a specific band exceeds the corresponding tree species' fluctuation benchmark value, filter the fluctuation exceeding the limit band number, and generate abnormal reflectance band identifiers. Based on the tree health spectral feature database, the process of obtaining reflectance values ​​of observation nodes within a monitoring window first requires continuous spectral data collection at fixed monitoring nodes within a specific time window. For example, reflectance data can be collected every 2 hours, forming 12 sets of data per day. Reflectance values ​​in different bands are extracted from each set as the initial data array. For the spectral reflectance value collected by observation node 0001 within a monitoring period (May 1st to May 7th, 2025), its red-edge band (e.g., 720nm wavelength) and near-infrared band (e.g., 860nm wavelength) are compared to determine if the reflectance difference exceeds the tree species' volatility benchmark value. The volatility benchmark value should be set in conjunction with historical health sample data. For example, assuming the normal volatility range for a certain tree species in the 720nm band is ±0.015, if node 0001... The reflectance value collected at 10:00 AM on May 2nd was 0.324, while the historical average was 0.310, resulting in a difference of 0.014, which is within the baseline value and not marked as abnormal. However, if the value collected in the 860nm band was 0.428, while the baseline average was 0.401, the difference would be 0.027, exceeding the volatility threshold of 0.015. This band is therefore considered an abnormal band. When performing the difference judgment, the current value of each band's sampled data needs to be directly calculated from the baseline value, and then the absolute value is compared with the volatility baseline to determine whether it exceeds the limit. Band numbers such as "B860" are selected and added to the abnormal identifier sequence array, ultimately generating the abnormal reflectance band identifier "{B860}". The baseline value needs to be obtained by averaging data from at least 20 healthy tree samples at the same time over 10 consecutive days. The setting process is shown in Table 3. Table 3: Reference Values ​​for Reflectance Fluctuation in Typical Wavebands of Tree Species A

[0037] As shown in Table 3, the table lists the band number, center wavelength, average reflectance value and corresponding volatility threshold of typical healthy samples, which are used as the basis for judging whether the current observation of a certain monitoring node exceeds the baseline volatility range.

[0038] S302: Based on the anomalous reflectance band identifier, calculate the time difference between the starting point of reflectance change in the non-abnormal band and the starting point of change in the corresponding anomalous band, construct the time response mapping between bands, extract the time interval values, and generate a group of spectral band response delay coefficients. Based on the anomalous reflectance band identifiers, the time difference between the starting point of reflectance change in non-abnormal bands and the starting point of change in corresponding anomalous bands is found in the current data sequence. Specifically, for each pair of anomalous and non-abnormal bands, the time point at which their reflectance first shows a sustained change is located. For example, node 0001 shows a continuous increase in reflectance in band B720 starting at 12:00 on May 2nd, while the anomalous band B860 shows a significant jump at 10:00 on the same day. The starting time of the change in B720 is recorded as [time point missing]. The change time of B860 is The difference is 2 hours. Then, similar data points are taken from multiple monitoring cycles. Assuming a total of 5 monitoring cycles, then... The reflectance data sequences of the monitoring points in the two bands are extracted as follows: ; ; Time series difference is The average time difference is: The spectral response delay coefficient is calculated using the following formula: ; in, This represents the spectral band response delay coefficient between the i-th non-abnormal band and the j-th anomalous band. This represents the time of the start point of the reflectance change in the i-th non-abnormal band at the k-th monitoring time. This represents the time of the start point of the reflectance change in the j-th anomalous band at the k-th monitoring time. This represents the reflectance value of the i-th non-abnormal band at the k-th monitoring time. This represents the reflectance value of the j-th anomalous band at the k-th monitoring time. This represents the number of times the i-th non-abnormal band and the j-th anomalous band co-occurred during monitoring. This represents the average time difference between the i-th non-abnormal band and the j-th anomalous band.

[0039] ; The advantage of this formula lies in its ability to reflect the relative rhythm of change between the two bands over different observation periods by incorporating the sum of the time difference and reflectance weights, along with a time variance term. This allows it to numerically represent the degree of response delay between them. The result shows that the response delay coefficient between B720 and B860 is 2.2797 hours.

[0040] S303: Based on the spectral band response delay coefficient group, the band correspondence is arranged in the delay order from the abnormal band to the response band. Combined with the original band number and its corresponding response band number, the propagation time value between the two is matched, the propagation direction and delay value are extracted, and a disturbance propagation chain feature library is established. After obtaining the response delay coefficient array between multiple spectral bands, the anomalous band is used as the propagation starting point. The delay coefficients of each anomalous band number and its corresponding non-anomalous band number are sorted. For example, if the anomalous band is B860, the corresponding non-anomalous bands B720, B680, and B540 all have calculated delay coefficients, arranged in ascending order of delay as B720 (2.28 hours), B680 (2.96 hours), and B540 (3.52 hours). Therefore, the propagation direction of the disturbance from the anomalous band to the other bands is established as B860→B720→B680→B540, and the corresponding propagation time value sequence is recorded using an index as [2.28, 2.96...]. [3.52] hours. Normalize or weight this sequence and extract the delay distribution features. For example, set the propagation time in the propagation chain to less than 2.5 hours as "fast response", 2.5~3.0 hours as "medium-speed response" and more than 3.0 hours as "lagging response". Then the above sequence can be mapped to "fast-medium-slow" respectively. This propagation feature can be uniformly encapsulated as chain structure data. For example, the chain identifier "C01" corresponds to the path "B860→B720 (fast) →B680 (medium) →B540 (slow)" as an instance record of the disturbance propagation chain feature library. It can be used to compare the propagation response features of new data samples and determine whether it belongs to a certain known abnormal diffusion mode.

[0041] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the disturbance source identifier and propagation delay parameter in the disturbance propagation chain feature library, combined with the environmental temperature, humidity and light intensity values ​​corresponding to the tree species' ecological habits, calculate the ratio of the change amplitude of the disturbance source parameter to the delay difference, screen samples that meet the set ratio range, and obtain the dynamic feature vector of the propagation response. Based on the disturbance source identifier and propagation delay parameters in the disturbance propagation chain feature database, the environmental temperature, humidity, and light intensity values ​​are obtained. To further detail this process, we assume the disturbance source identifier is... The propagation delay is Data will be collected based on different environmental factors. For a specific environment, the temperature value... Based on sensor monitoring, assuming the ambient temperature is... Humidity value The humidity is obtained from a humidity sensor, assuming the humidity is... Light intensity value This is obtained through a light sensor, assuming the value is... By combining tree species ecological habit parameters with current environmental data, the accuracy of disturbance feature screening can be further improved. The data will be used to calculate the ratio of the disturbance source's change amplitude to the delay difference. The specific calculation formula is as follows: ratio of disturbance amplitude change ; in, This indicates the magnitude of change in the disturbance source parameters. The unit should be determined based on the specific physical quantity of the disturbance parameter; for example, temperature is measured in °C, humidity in %, and light intensity in lx. Ensure that the units are clearly defined and consistent when calculating ratios. This indicates a propagation delay. For example, if the disturbance source is in a time period... arrive The range of change within is The propagation delay is Then the ratio of the change in disturbance amplitude is: ; Next, samples that meet the set ratio range are selected for analysis. Let's assume the set ratio range is... Based on the above example, the perturbation amplitude change ratio is 0.5, which meets the requirements of this interval, so it can be selected as a sample. Finally, by combining the tree species' ecological habits and perturbation characteristic data, a dynamic feature vector of the propagation response is generated. The feature vector will represent the propagation response characteristics of the perturbation source in this environment.

[0042] S402: Based on the dynamic feature vector of the propagation response, classify the sample labels, construct a nonlinear mapping using the RBF kernel function in the support vector machine, calculate the distance from the perturbed sample to the center point of the grade label, determine the four-class boundaries, and obtain the four-level classification boundary interval values. The four-level classification boundary interval value refers to the set of thresholds that divide the sample feature space into four non-overlapping and continuous health level discrimination intervals after nonlinear mapping of the perturbation sample features based on the support vector machine model. Sample labels are classified based on the dynamic feature vector of the propagation response. To ensure classification accuracy, the RBF kernel function in Support Vector Machine (SVM) is used for non-linear mapping. Specifically, the dynamic feature vector of the propagation response is... The vectors will be input into the support vector machine model. Through the mapping of the RBF kernel function, the mapped feature space can be obtained. The specific nonlinear mapping formula is as follows: ; in, and This represents the feature vectors of two samples. These are the parameters of the RBF kernel function, which are adjusted based on the dataset. Assume... For two feature vectors and The mapping result is: ; Next, the distance from the perturbed sample to the center point of the grade label is calculated to determine the boundaries of the four categories. Assume the sample feature space is divided into four non-overlapping intervals, each corresponding to a different health grade. Based on the output of the support vector machine, the center points of the four categories can be obtained. For example, the center point of category 1 is... The center point of category 2 is And so on.

[0043] By calculating the distance between the sample and the center point, the Euclidean distance formula is used: ; For the sample The distance from the center point of category 1 is: ; This will be used to determine the health level to which the perturbed sample belongs. Finally, the boundary interval values ​​for the four categories are obtained.

[0044] S403: Call the four-level classification boundary interval values ​​and the propagation response dynamic feature vector, input the perturbation sample into the support vector machine model, count the classification probability values ​​of the sample in the four levels and summarize them to generate the tree health AI assessment model. By calling the four-level classification boundary interval values ​​and the propagation response dynamic feature vectors determined above, the data is input into the Support Vector Machine (SVM) model to calculate and statistically analyze the classification probability values ​​of the samples in the four health levels. The SVM model outputs a probability value for each category based on the sample features, representing the probability that the sample belongs to each category. For example, if a sample is input into the SVM model, the resulting classification probability value is: , Based on the probability values, the generated tree health AI assessment model will indicate that the sample belongs to category 1. By summarizing the classification probability values ​​of the samples, the health status of the entire dataset can be further assessed, ultimately yielding the output of the tree health AI assessment model.

[0045] Table 4: Examples of Disturbance Source Parameters and Environmental Data

[0046] As shown in Table 4, environmental data of differentiated disturbance sources and their disturbance amplitude change ratios are given. The data are used to screen samples that meet the ratio range, and then generate dynamic feature vectors of propagation response for subsequent classification and modeling analysis.

[0047] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the tree health AI assessment model, the multispectral data of the trees to be assessed is processed, the band spectral reflectance, vegetation index and local spectral gradient are analyzed, feature normalization is performed to adjust the scale and distribution of the band values, and spectral embedded feature values ​​are generated. The normalization process uses the Min-Max normalization method, which linearly maps the wavelength reflectance of the differential spectral bands between 0 and 1 within the observation period to eliminate interference from time and illumination differences. The system processes multispectral data of the trees to be evaluated based on a tree health AI assessment model. In the first step, the system acquires raw reflectance data of multiple bands within a specified time period from a multispectral imaging device installed in the monitoring area. For example, the raw reflectances of the five bands (450nm, 550nm, 670nm, 800nm, and 900nm) are 0.12, 0.24, 0.35, 0.48, and 0.53, respectively. This data is stored in array form, with the multichannel reflectance input corresponding to each sampled pixel being... Subsequently, vegetation indices such as NDVI and EVI are calculated from the bands based on preset combinations. If NDVI is constructed using the 800nm ​​and 670nm channels, the corresponding calculation is as follows: ; Besides NDVI, EVI can also be constructed and calculated by combining 800nm ​​and 550nm channels. For local spectral gradient extraction of spectral reflectance, the model performs first-order difference calculations on the reflectance differences between continuous bands, such as... , , forming a local spectral gradient vector The system combines the original reflectance array, vegetation index values, and gradient vectors to form a complete set of original features. Subsequently, the system normalizes the feature parameters using minimum-maximum standardization, linearly mapping each reflectance value to the minimum and maximum values ​​of that band within the current observation period. For example, if the reflectance of the 550nm band currently ranges from [0.18, 0.46] within the observation period, then the current normalized value for that band is: ; Normalization is applicable to both band and exponential parameters, resulting in an input vector with a uniform scale. ; The values ​​are all in the range [0, 1], ensuring consistent interpretability under different sampling periods and lighting conditions. Finally, this normalized vector is input into the model's embedding layer, where a deep neural network extracts high-dimensional spectral embedding features, outputting a uniformly dimensional embedding vector. ,in Using the feature space dimension, we ultimately obtain the representation vector of the current tree sample in the embedding space. .

[0048] S502: Based on spectral embedding feature values, extract the level feature center vector according to the health level classification boundary parameters, calculate the similarity value between the embedded feature and the center according to the cosine similarity method, and sort all level similarity results by size to obtain the health level similarity ranking value. The level feature center vector refers to the average embedding representation of the feature vectors corresponding to the health levels in the training set; Based on spectral embedding eigenvalues The next step is to classify the health levels. First, the embedding feature center vector for each health level is calculated from the level tree samples in the training set. For example, if the levels are divided into three categories [healthy (H), sub-healthy (M), unhealthy (U)], the corresponding center vectors are as follows: Each center vector is obtained by averaging the embedding vectors of that level in the training samples. For example, the samples for health level H are... Each, feature dimension If the sample embedding vector is: Then its center vector is: ; The system then analyzes the embedding features of the current sample to be evaluated. The cosine similarity is calculated between the vector and the center vector, and the specific calculation method is as follows: ; set up Substitute into the above formula and calculate its relationship with... Cosine similarity: Vector dot product: ; Modulus calculation: ; ; Similarity results: ; Calculate the similarity between each level and the level center using this method, and sort the results from high to low. For example, you can get [Healthy (0.999), Sub-healthy (0.86), Unhealthy (0.72)], which will eventually form the similarity ranking value of the health level.

[0049] S503: Based on the similarity ranking value of health level, extract the health level corresponding to the first similarity value in the ranking, combine the coordinate position of the spectral embedding feature value in the feature space, compare it with the boundary coordinates of the health level interval, classify the tree to be evaluated into the corresponding level interval, and generate the tree health status assessment result. Based on the similarity scores of health levels, the system extracts the corresponding level label from the top-ranked sample as the initial level assessment, and simultaneously embeds the spectral features of the sample to be evaluated. The vector position mapped to the feature space is compared with the spatial coordinates defined by the level boundary threshold. For example, the embedding space is defined as a block of intervals in the three-dimensional principal component space (PC1-PC2-PC3), such as the health level boundary range being defined as... , , The principal component projection values ​​of the sample to be evaluated after PCA dimensionality reduction are... , , Since the point falls entirely within the coordinate range of the health level space, its health level is ultimately determined to be healthy (H). Based on this classification, the system outputs the final health status assessment result as "healthy". If multiple level boundaries overlap, the system will further distinguish based on the geometric distance difference of the sample within the boundary space and select the closest level space as the final classification basis.

[0050] Table 5: Comparison of Sample Feature Vectors and Rank Centers

[0051] As shown in Table 5, Sample 1 points to the health level in both similarity calculation and spatial coordinate comparison, and its tree health status is ultimately assessed as healthy.

[0052] Please see Figure 7 A tree health AI assessment system integrating multispectral analysis includes: The multispectral acquisition module acquires multispectral reflectance data of tree canopy through multispectral sensors, arranges reflectance in time series, extracts reflectance deviation between adjacent time points, identifies extreme points, records wavelength and rate of change, and simultaneously collects and labels current phenological period and tree species identity information to form a multispectral feature dataset, which is then transmitted to the feature extraction module. The feature extraction module calls the multispectral reflectance dataset to analyze the changes in red edge, near-infrared, and green light bands, counts extreme frequency, filters band combinations with Pearson correlation coefficients exceeding the threshold, distinguishes and removes signals that conform to the leaf color and reflectance changes of the tree species in the normal phenological stage, generates a tree health spectral feature library, and transmits it to the perturbation analysis module. The disturbance analysis module, based on the tree health spectral feature library, monitors abnormal fluctuations in reflectance within a set time window, compares the response time difference of different bands, detects signal propagation delay, tracks the path of disturbance signals between bands, records the disturbance source, response band, and delay parameters, constructs a disturbance propagation feature library, and transmits it to the state classification module. The state classification module constructs a dynamic feature vector based on the disturbance source identifier and propagation delay parameter in the disturbance propagation feature library, combined with the environmental temperature, humidity and light intensity values ​​corresponding to the tree species' ecological habits. This vector is then input into the support vector machine algorithm for training, calculates the four-level classification boundary parameters, generates a tree health AI assessment model, and is passed to the assessment output module. The evaluation output module processes the spectral data of the trees to be evaluated using a tree health AI evaluation model, calculates the similarity score between the input spectral features and the health level classification boundary, compares the similarity values ​​to determine the current health status level of the trees, and outputs the tree health status evaluation results.

[0053] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A tree health AI assessment method integrating multispectral analysis, characterized in that, Includes the following steps: S1: Acquire multispectral reflectance data of tree canopy through multispectral sensors, label tree species ID and phenological period information, extract reflectance deviation between adjacent time points according to time series, record wavelength, rate of change and corresponding tree species ID and phenological period to form a multispectral feature dataset; S2: Call the multispectral feature dataset, analyze the direction of spectral band changes, count the frequency of extreme points of spectral bands, use the Pearson correlation coefficient algorithm to filter spectral band combinations whose correlation exceeds the correlation threshold, remove normal phenological change signals, and generate a tree health spectral feature library. S3: Based on the tree health spectral feature library, monitor abnormal fluctuations in reflectance within the time window, compare changes in response of other spectral bands, detect response delay, track the cross-spectral propagation of disturbance signals, record the disturbance source spectral band identifier, response spectral band identifier, and propagation delay parameters, and construct a disturbance propagation chain feature library; S4: Based on the disturbance source identifier and propagation delay parameter in the disturbance propagation chain feature library, and combined with the environmental temperature, humidity and light intensity corresponding to the tree species' ecological habits, construct a dynamic feature vector, use the support vector machine algorithm to train a multi-classification model of health status, and generate a tree health AI assessment model.

2. The tree health AI assessment method integrating multispectral analysis according to claim 1, characterized in that, The multispectral feature dataset includes extreme wavelengths, reflectance changes, and time indexes; the tree health spectral feature library includes spectral change trends, associated spectral band combinations, and extreme point occurrence frequencies; the disturbance propagation chain feature library includes disturbance propagation paths, associated spectral bands, and response delay times; and the tree AI intelligent assessment model includes health status grading boundaries, feature vector parameters, and support vector machine model parameters.

3. The tree health AI assessment method integrating multispectral analysis according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect reflectance data of tree canopy bands with tree species ID and phenological period annotation at different observation time points using a multispectral sensor, arrange the reflectance values ​​at the same wavelength in chronological order, construct a time series according to wavelength, tree species ID and phenological period information and record the corresponding reflectance group data to generate a multi-temporal band reflectance sequence. S102: Based on the multi-temporal band reflectance sequence, calculate the difference of band reflectance data at adjacent time points one by one, extract the reflectance change amplitude and its time index according to the time point sequence corresponding to the difference, associate the tree species ID and phenological period information, and generate a band reflectance deviation dataset. S103: Based on the band reflectance deviation dataset, set a fixed reflectance variation range limit for the reflectance deviation value of each group of wavelengths, identify the wavelength and time index combination that exceeds the variation range limit, and combine the tree species ID and phenological period to record the corresponding reflectance variation value to generate a multispectral feature dataset.

4. The tree health AI assessment method integrating multispectral analysis according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Call the multispectral feature dataset, extract the reflectance sequences of the red edge, near-infrared and green spectral bands, analyze the reflectance change trend of the spectral bands at different time points, record the increase or decrease of reflectance values ​​at continuous time points, and generate a sequence of spectral band change directions. S202: Based on the sequence of spectral change directions, count the time nodes when each local maximum and minimum value of each spectral segment appears in the time series, accumulate the frequency of occurrence of extreme points, calculate the distribution of the number of extreme points of the spectral segment in the complete time series, and obtain the extreme frequency of the spectral segment. S203: Based on the extreme frequencies of the spectral bands, the Pearson correlation coefficient algorithm is used to calculate the correlation between the extreme frequency sequences of the red edge, near-infrared, and green spectral bands. Spectral band combinations with correlation greater than the correlation threshold and excluding normal phenological change signals are selected to establish a tree health spectral feature library.

5. The tree health AI assessment method integrating multispectral analysis according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: Based on the tree health spectral feature library, obtain the reflectance value of the observation node in the monitoring window, determine whether the reflectance difference of a specific band exceeds the corresponding tree species fluctuation benchmark value, filter the fluctuation over-limit band number, and generate abnormal reflectance band identifier. S302: Based on the abnormal reflectivity band identifier, calculate the time difference between the starting point of reflectivity change in the non-abnormal band and the starting point of the corresponding abnormal band change, construct the time response mapping between bands, extract the time interval value, and generate a spectral band response delay coefficient group. S303: Based on the spectral band response delay coefficient group, arrange the band correspondence in the delay order from the abnormal band to the response band, combine the original band number and its corresponding response band number, match the propagation time value between the two, extract the propagation direction and delay value, and establish a disturbance propagation chain feature library.

6. The tree health AI assessment method based on multispectral analysis according to claim 5, characterized in that, The specific steps of S4 are as follows: S401: Based on the disturbance source identifier and propagation delay parameter in the disturbance propagation chain feature library, combined with the environmental temperature, humidity and light intensity corresponding to the tree species' ecological habits, calculate the ratio of the change amplitude of the disturbance source parameter to the delay difference, screen samples that meet the set ratio range, and obtain the dynamic feature vector of the propagation response. S402: Based on the propagation response dynamic feature vector, classify and process sample labels, construct a nonlinear mapping using the RBF kernel function in the support vector machine, calculate the distance from the perturbed sample to the center point of the grade label, determine the four-class boundaries, and obtain the four-level classification boundary interval values. S403: Call the four-level classification boundary interval values ​​and the propagation response dynamic feature vector, input the perturbation sample into the support vector machine model, count the classification probability values ​​of the sample in the four levels and summarize them to generate a tree health AI assessment model.

7. The tree health AI assessment method integrating multispectral analysis according to claim 1, characterized in that, The method also includes step S5: S5: The tree health AI assessment model is used to process the multispectral data of the tree to be assessed, calculate the similarity score between the input spectral features and the health level classification boundary, compare the similarity values ​​to determine the current health status level of the tree, and obtain the tree health status assessment result. The tree health status assessment results include health grade results and similarity scores.

8. The tree health AI assessment method integrating multispectral analysis according to claim 7, characterized in that, The specific steps of S5 are as follows: S501: Based on the tree health AI assessment model, process the multispectral data of the trees to be assessed, analyze the band spectral reflectance, vegetation index and local spectral gradient, perform feature normalization to adjust the scale and distribution of band values, and generate spectral embedded feature values. S502: Based on the spectral embedding feature value, extract the level feature center vector according to the health level classification boundary parameter, calculate the similarity value between the embedded feature and the center according to the cosine similarity method, and sort all level similarity results by size to obtain the health level similarity ranking value. S503: Based on the health level similarity ranking value, extract the health level corresponding to the first similarity value in the ranking, combine the coordinate position of the spectral embedding feature value in the feature space, compare it with the boundary coordinates of the health level interval, classify the tree to be evaluated into the corresponding level interval, and generate the tree health status assessment result.

9. The tree health AI assessment method integrating multispectral analysis according to claim 8, characterized in that, The spectral embedding feature value refers to the feature vector that represents the health status of trees, obtained by normalizing and fusing the multispectral data of trees with features. The health level similarity ranking value refers to the result calculated based on the spectral embedding feature value and the health level feature center vector and sorted according to the similarity.

10. A tree health AI assessment system integrating multispectral analysis, characterized in that, The system is used to implement the tree health AI assessment method fused with multispectral analysis as described in any one of claims 1-9, the system comprising: The multispectral acquisition module acquires multispectral reflectance data of tree canopy through multispectral sensors, arranges reflectance in time series, extracts reflectance deviation between adjacent time points, identifies extreme points, records wavelength and rate of change, and simultaneously collects and labels current phenological period and tree species identity information to form a multispectral feature dataset, which is then transmitted to the feature extraction module. The feature extraction module calls the multispectral reflectance dataset, analyzes the changes in red edge, near-infrared, and green light bands, counts extreme frequency, filters band combinations with Pearson correlation coefficients exceeding the threshold, distinguishes and removes signals that conform to the leaf color and reflectance changes of the tree species in the normal phenological stage, generates a tree health spectral feature library, and transmits it to the perturbation analysis module. The disturbance analysis module, based on the tree health spectral feature library, monitors abnormal fluctuations in reflectance within a set time window, compares the response time differences of different bands, detects signal propagation delay, tracks the path of the disturbance signal between bands, records the disturbance source, response band, and delay parameters, constructs a disturbance propagation feature library, and transmits it to the state classification module. The state classification module constructs a dynamic feature vector based on the disturbance source identifier and propagation delay parameter in the disturbance propagation feature library, combined with the environmental temperature, humidity and light intensity values ​​corresponding to the tree species' ecological habits. This vector is then input into the support vector machine algorithm for training, calculates the four-level classification boundary parameters, generates a tree health AI assessment model, and is transmitted to the assessment output module. The evaluation output module processes the spectral data of the trees to be evaluated through the tree health AI evaluation model, calculates the similarity score between the input spectral features and the health level classification boundary, compares the similarity values ​​to determine the current health status level of the trees, and outputs the tree health status evaluation result.

Images