Spectral data acquisition and screening method for blue carbon data inversion

Abstract

The invention discloses a spectral data acquisition and screening method for blue carbon data inversion, and relates to the field of spectral data acquisition. The method comprises the following steps: (1) acquiring original spectral data; (2) constructing a hyperspectral unmixing model based on a spectrum modulation module and spectrum variability; (3) utilizing the constructed hyperspectral unmixing model to realize vegetation spectrum matching; (4) screening the original data based on a spectrum matching result; (5) realizing data re-verification based on dispersion estimation; and (6) according to a dispersion estimation result, obtaining a regional representative spectrum, and realizing spectral data acquisition and screening for blue carbon data inversion. The hyperspectral unmixing model based on the spectrum modulation module and the spectrum variability can adapt to heterogeneity of data obtained in different scenes, the spectrum matching precision is improved, and the comprehensive performance of spectrum data screening is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of spectral processing, and particularly relates to a spectral data acquisition and screening method for blue carbon data inversion. BACKGROUND

[0002] Hyperspectral imaging technology can simultaneously obtain spatial information and continuous, narrow-band spectral information of a target, thereby forming a "data cube" containing rich material composition data. This technology has been widely used in remote sensing, agriculture, food detection, medicine and environmental monitoring, and many other fields. The quality of spectral data directly determines the accuracy of subsequent quantitative analysis and model prediction. Therefore, before modeling using hyperspectral data, the original spectral data collected must be strictly verified and screened to eliminate invalid, abnormal or low-quality data. In particular, under the background of global climate change, accurate monitoring of blue carbon ecosystems (such as coastal salt marshes, mangroves, and seagrass beds) becomes crucial. These ecosystems have fixed and sequestered a large amount of carbon dioxide through photosynthesis, and their carbon sequestration efficiency and potential are much higher than those of terrestrial forests. Accurate inversion of key biophysical parameters (such as vegetation coverage, biomass, chlorophyll concentration, primary productivity, and soil carbon content) of blue carbon ecosystems has significant scientific value and economic significance for assessing global carbon cycles, accounting for carbon sink transactions, and guiding ecological restoration.

[0003] Currently, the screening of spectral data mainly relies on some basic quality evaluation methods. For example, the signal quality is evaluated by calculating the signal-to-noise ratio of the spectrum; the noise is identified by checking the smoothness of the spectral curve or whether there are abnormal sharp peaks; or the data points with obvious distortion are removed by setting a reflectance threshold. In addition, some methods also use multivariate scatter correction, standard normal variable transformation and other preprocessing methods to weaken the influence of physical scattering effects, thereby indirectly improving the usability of the data. However, the above traditional methods have obvious limitations. First, they mainly focus on the physical signal quality of the spectrum, but lack evaluation of the reliability of the spectrum. The spectrum with high signal-to-noise ratio and reasonable reflectance value may still be a mixed spectrum composed of multiple substances. Directly using such a mixed spectrum for qualitative or quantitative analysis will introduce significant errors, leading to decreased model accuracy and unreliable prediction results. Second, traditional methods cannot effectively identify and quantify the contribution of each component in a mixed pixel, making it difficult to determine whether a spectrum is sufficient to represent a "pure" endmember substance.

[0004] Hyperspectral unmixing technology is a key means to solve the above-mentioned mixed pixel problem. The core idea is to decompose the mixed spectrum into a series of end members (i.e. pure material spectrum) and their corresponding abundance ratio. Through the unmixing algorithm, the potential spectral end members can be extracted from the mixed data, and the composition of each end member in each pixel can be quantified. In addition, the role of hyperspectral unmixing technology should not be limited to the final data analysis stage, and its more important value can be applied in the quality verification and screening process of spectral data in advance. Through the abundance information and other indicators obtained by unmixing, it can provide a crucial quantitative basis for judging the reliability and representativeness of a spectral data. However, there is still a lack of a systematic method to deeply integrate hyperspectral unmixing into the verification and screening link of spectral data to realize the reliability evaluation and screening of data in the prior art.

[0005] Therefore, there is an urgent need for an innovative spectral data screening method to overcome the shortcomings of the prior art, ensure the high quality and high reliability of the data used to build the spectral library and inversion model, and ultimately improve the accuracy and reliability of blue carbon monitoring and evaluation. SUMMARY

[0006] To solve the problems in the prior art, the present application proposes a spectral data acquisition and screening method for blue carbon data inversion.

[0007] The technical scheme adopted by the present application is as follows: The present application discloses a hyperspectral image unmixing method based on end member oriented Transformer and end member beam, comprising the following steps: 1) Taking coastal blue carbon ecosystem vegetation as the collection object, dividing the collection object into collection regions, and collecting data in regions by using a spectral collection device to obtain original spectral data of each region; 2) Constructing a hyperspectral unmixing model based on a spectral modulation module and spectral variability, training the hyperspectral unmixing model based on a data set constructed based on pre-collected blue carbon resource standard spectra, and the hyperspectral unmixing model comprising a spectral modulation module, an abundance encoder module, a perturbation decoder module and a reconstruction decoder module; 3) Using the pre-trained hyperspectral unmixing model based on the spectral modulation module and the spectral variability to unmix the collected original spectral data; 4) Screening the data collected in each region based on the unmixing result, and deciding whether to collect original data again according to the screening result; 5) Calculating the discreteness between the spectra left after screening in each region according to step 4), judging the data validity according to a threshold value, and if the data validity is sufficient, executing step 6), otherwise re-collecting data; 6) Regionally averaging the remaining spectra as representative spectra of each region to obtain the acquisition and screening result of the spectral data.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) This invention proposes a spectral modulation module and designs a one-dimensional adaptive aligned convolutional layer that can adapt to spectral data collected in different outdoor scenarios. It adaptively adjusts the spectral feature acquisition area according to the position of the feature peaks, adapts to changes in the shape of the spectral peaks, and reduces the dependence on preprocessing.

[0009] 2) This invention proposes a perturbation decoder module, which models the differences in imaging conditions under natural conditions as perturbations to reduce the impact of spectral variability on the unmixing results. 3) This invention proposes for the first time a method for spectral data acquisition and screening based on a pre-trained hyperspectral unmixing model for blue carbon data inversion. Compared with traditional methods, it can adapt to various spectral peak changes and disturbances caused by different environments, thereby improving the reliability and robustness of the spectral data screening method. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the basic steps of the spectral data acquisition and screening method of the present invention; Figure 2 This is a schematic diagram of the hyperspectral unmixing neural network of the present invention; Figure 3 This is a schematic diagram of the one-dimensional adaptive convolutional layer function of the present invention; Figure 4 This is a schematic diagram illustrating the overall effect of the present invention. Detailed Implementation

[0011] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0012] like Figure 1 The diagram shown is a flowchart of the basic steps of the spectral data acquisition and screening method for blue carbon data inversion according to the present invention, which mainly includes: Step 1) In this embodiment, a single-point sampling spectrometer is used as an example. For the target to be collected, the collection area is divided into a nine-square grid, and samples are collected evenly in each collection area. By analyzing the spectra, the raw spectral data of each region can be obtained. ,in The number represents the spectral bands, and 9 represents the number of grid cells. In this embodiment... .

[0013] Step 2) Construct a spectral unmixing neural network (hyperspectral unmixing model) based on a spectral modulation module and spectral variability, the neural network is constructed based on a pre-acquired blue carbon resource standard spectrum dataset, and the constructed dataset is trained, as shown in Figure 2 The spectral unmixing neural network of the present application includes a spectral modulation module, an abundance encoder module, a perturbation decoder module and a reconstruction decoder module; wherein the spectral modulation module is used to modulate and preprocess the data acquired under different imaging conditions and preliminarily realize feature extraction to obtain spectral features; the abundance encoder module mines the information corresponding to each endmember in the spectral features and outputs an abundance map; the perturbation decoder module analyzes the influence of different imaging conditions on the model and the perturbation existing in the spectrum, and outputs the amount of perturbation existing in the current spectral data; the reconstruction decoder module linearly mixes the endmember spectrum according to the abundance map, and introduces the estimated perturbation amount to obtain the final reconstruction result. The present embodiment uses a pre-constructed dataset to train the initial weights of the network.

[0014] As Figure 2 shown, the specific working process of each module in the hyperspectral unmixing neural network based on a spectral modulation module and spectral variability is as follows: Step 21) Since the hyperspectral unmixing model based on a neural network relies on training data to work, a hyperspectral unmixing dataset for blue carbon resources is first constructed. The present application mainly faces coastal blue carbon resources such as mangrove forests, tidal flats and uninhabited islands, so the spectral curves of typical carbon storage plants such as mangrove plants, diatoms and seaweed are selected as reference spectra, i.e. endmembers of the dataset. The originally acquired spectrum needs to be converted into reflectivity before being used to construct the unmixing dataset, and the formula is as follows: Wherein, is the reflectivity of the converted material, is the radiance value obtained by measuring the target material with a spectrometer, is the dark current value during measurement, is the radiance value obtained by measuring a standard reference white plate with a spectrometer under almost the same time and imaging conditions, is the calibrated reflectivity of the standard reference white plate. Thus, a spectral curve library of reference materials , is obtained, which represents the number of reference materials in the spectral curve library, i.e. the number of endmembers.

[0015] Then, in order to ensure the reliability of the simulation dataset and to conform to the material distribution and spectral mixing rules of the real scene, the present embodiment mixes the material spectral curves according to the augmented linear mixing model under different proportions and spectral variability, and the formula of the augmented linear mixing model is as follows: Wherein to the mixed spectrum according to the model, and are the abundance and endmember respectively when mixed, , and represent the scaling, perturbation and noise effects respectively when the spectrum is mixed. In order to make the abundance of the simulation dataset close to the real material abundance distribution, the Dirichlet distribution is used to generate the simulation abundance in this embodiment, the probability density function of which is: wherein is used to control the sparsity of the abundance, represents dimensional simplex, is the multivariate generalization of the Beta function, which is used as a normalization constant, wherein is the Gamma function. In this embodiment, , depends on the number of endmembers .

[0016] In addition, the noise and various perturbation terms are generated based on two-dimensional Gaussian noise, and the distribution function is wherein, and represent the mean and variance of the distribution respectively.

[0017] Based on the above mixing model and distribution, a simulation dataset is generated, which is used for subsequent unmixed neural network training.

[0018] Step 22): Constructing a spectral modulation module of hyperspectral unmixing neural network based on spectral modulation module and spectral variability, which is used to reduce the dependence of the model on preprocessing and improve the reliability of screening according to the spectral peak changes of the spectrum collected in different natural scenes. The spectral modulation module is obtained by stacking three layers of submodules, and each layer of submodule includes a one-dimensional adaptive convolution layer and a GELU activation function layer.

[0019] As shown in Figure 3 , the one-dimensional adaptive convolution layer can adaptively adjust the key area of feature extraction according to the changes of the spectral peak of the input data, which is used for subsequent unmixing. Specifically, the one-dimensional adaptive convolution layer includes an offset layer, a modulation layer and a dynamic convolution layer, wherein the offset layer is used to learn the key spectral peak position and output the offset of the sampling point in the one-dimensional spectral direction, which is composed of a one-dimensional convolution layer and a fully connected layer, and the process is represented as follows: wherein, This represents the offset output by the module. Represents the size of the one-dimensional convolution kernel. This represents the length of the one-dimensional output sequence. The amplitude is controlled within , is used to describe the amount of movement of the convolution kernel in a one-dimensional direction.

[0020] Similarly, the modulation layer consists of a one-dimensional convolutional layer and a fully connected layer, and the process can be represented as follows: in, This represents the modulation amount output by the module. The amplitude range is This can directly enhance the weight of characteristic bands and emphasize the importance of key bands. This represents the number of one-dimensional spectral samples.

[0021] Based on the offset and modulation estimates of the offset layer and the modulation layer of the one-dimensional adaptive convolutional layer, the one-dimensional adaptive convolutional layer dynamically adjusts the position of interest when the one-dimensional convolutional kernel moves along the one-dimensional spectrum based on the offset, and adjusts the importance of each band in combination with the modulation. The specific formula can be expressed as: in, Representing the The result of the next one-dimensional adaptive convolution Enumerated the first The positional offset of all neighboring sampling points, when the kernel size is... hour, . The first one representing network prediction The sampling point at the th sampling point The learnable offset during each convolution. Since the offset is usually a decimal, one-dimensional adaptive convolution requires sampling values ​​at the decimal position index. Therefore, this invention introduces bilinear interpolation to determine the value of the sampling point. The index after introducing the offset can be expressed as: because Since the value is usually a decimal, the value of the corresponding index sampling point is calculated using the following formula: in To enumerate the set of all data locations, This is a bilinear interpolation function in one dimension. The above calculation process is the specific operation of a one-dimensional adaptive aligned convolutional layer. In actual operation, in order to introduce the model's ability to extract nonlinear features, this invention combines the Gaussian error linear function (GELU) as an activation function for feature extraction. The specific formula is as follows: Each basic submodule in the spectral modulation module can be represented as: AAC represents a one-dimensional adaptive convolutional layer. This is the output of the first-layer submodule. The spectral modulation module has three layers of submodules, therefore the final output modulated features are: .

[0022] Step 23): The abundance encoder module aims to utilize the features modulated by the spectral modulation module. This invention mines information from endmembers of various categories and interprets the abundance corresponding to each category. The abundance encoder module in this invention combines one-dimensional convolution and a multilayer perceptron. The one-dimensional convolution integrates features from the spectral vector, while the multilayer perceptron maps the features to the number of endmember dimensions and encodes them as abundance. This can be expressed by the formula: in, This is the abundance map output by the abundance encoder module. The Softmax function is used to ensure that the sum of the abundance values ​​of each endmember within the same spectrum is 1.

[0023] Step 24): The perturbation decoder module analyzes and estimates the perturbations under the current spectral imaging conditions compared to the standard measurement environment, and reduces their impact on network reconstruction. The perturbation decoder module introduces the idea of ​​a perturbation dictionary and is implemented based on a neural network. First, the modulated features... The coefficients are transformed into a perturbation dictionary through two sets of one-dimensional convolutions and a multilayer perceptron. To ensure that the sum of the perturbation coefficients is 1, the Softmax activation function is also used, which is expressed as follows: in, These are the coefficients of the perturbation dictionary. To use a neural network to model perturbations present in spectral curves of real-world scenes using a perturbation dictionary, this invention simulates the perturbation dictionary using a high-dimensional single-layer fully connected layer, and optimizes the perturbation library within the dictionary through training. The estimation process for the perturbation quantity can be expressed as: in The amount of perturbation estimated by the perturbation decoder is used to characterize the perturbations present in the current data.

[0024] Step 25): The reconstruction decoder module is used to reconstruct the original spectrum, thereby enabling self-supervised training of the network. The reconstruction decoder consists of a linear reconstruction part and a spectral variable part, which can be specifically represented as: Among them, FC is a fully connected layer without offset, which is used to map the low-dimensional abundance map to the original spectral band dimension, and its weights can be regarded as endmember curves.

[0025] Step 26): To directly utilize the aforementioned neural network in subsequent spectral dataset selection, this step trains the constructed network based on a manually constructed hyperspectral unmixed dataset for blue carbon resources. To ensure the trained network closely reflects real-world scenarios, this invention designs a training scheme based on the concept of knowledge distillation, as follows: Figure 2 As shown. During training, the present invention uses a real dataset to train the above neural network as the teacher model in the training step; the above neural network trained using a simulated dataset serves as the student model in the training step; the two models extract spectral features of meaningful band spectra based on weight sharing of the spectral modulation module, which facilitates the demixing of subsequent measured spectral data.

[0026] During training, both the teacher network and the student network use the entire hyperspectral image as training samples. The network adopts a self-supervised training method. The samples are input into the network model, and the root mean square error of reconstruction and the reconstructed spectral angular distance are used as loss functions. The network weights are updated based on the Adam gradient descent method with adaptive adjustment of the learning rate. In this embodiment, the training is iterated 200 times.

[0027] Step 3): Based on the trained student network, the original spectra of each region are unmixed to obtain the abundance of typical materials in the spectrum, which serves as the basis for subsequent screening.

[0028] Step 4): Based on the abundance values ​​of various materials estimated in Step 3) in the currently acquired raw spectral data, combined with the types of materials contained in the imaging scene, and according to the threshold... To initially determine the effectiveness of the current material, if the abundance value of the current material in the spectrum is greater than a threshold... If the current spectrum is not found, it is considered a valid spectrum and is retained. This can be expressed by the formula: in For an effective spectral set, A collection of materials in the scene. Representative spectrum middle The abundance value of the material, in this embodiment It was set to 0.8.

[0029] After initial screening of spectra, based on the screening results Count the number of spectra retained in each region. If the total number of spectra retained in a corresponding region is less than... If the condition is met, then the operation will be repeated from step 1) until the condition is met.

[0030] Step 5): To ensure the effectiveness of the acquired spectra, this invention performs further filtering based on the data dispersion. Specifically, this embodiment uses the covariance ellipsoid volume to calculate the dispersion within each region's data. First, the covariance matrix is ​​calculated for the spectral data of each region. : in, This represents the covariance matrix of the current region's data. This represents the number of spectra remaining in each region after the initial screening. Eigenvalue decomposition is performed on the covariance matrix: in, The feature vector representing the current data. It is obtained by arranging the eigenvalues ​​of the covariance matrix diagonally. Therefore, the volume of the covariance ellipsoid formed by the data of the current region can be calculated. : Where det represents the determinant of the matrix. Representative and covariance matrix Identity matrices of the same dimension This is used to avoid the argument being less than 0. Since the covariance ellipsoid volume reflects the compactness of the ellipsoid formed by the spectral data, a larger ellipsoid volume in high-dimensional space indicates higher data dispersion. Based on a threshold... This further determines the validity of the data for the current region. If the volume of the covariance ellipsoid of the current region is smaller than... If the data is valid, then the collected data is considered valid.

[0031] Based on the filtered data, if the remaining number of spectra in each region is less than If the data is not collected, return to step 1 and re-collect and filter the data.

[0032] Step 6): Based on the data after the initial screening based on unmixing and the second screening based on dispersion, the average spectrum retained by each region according to the divided nine-square grid is used as the representative spectrum of each region, thus obtaining the results of spectral data acquisition and screening.

[0033] Taking the spectral imaging of the Suaeda salsa region as an example,Figure 4 This paper illustrates the overall process and effects of the present invention. During field spectral imaging, it is often unavoidable to collect spectral information about the surrounding environment of vegetation, such as soil and water. Therefore, the imaging spectra cannot be directly used as research samples and spectral screening is required. Spectral data screening based on the abundance of feature endmembers can effectively remove spectral samples with low vegetation information that are affected by spectral mixing. After further spectral screening based on dispersion, the spectra with the highest correlation to Suaeda salsa information and minimal influence from the imaging environment are retained as the final screening result of this method.

[0034] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for spectral data acquisition and screening for blue carbon data inversion, characterized in that, Includes the following steps: 1) Taking the coastal blue carbon ecosystem vegetation as the data collection object, the data collection area was divided into collection areas, and the original spectral data of each area was obtained by using spectral acquisition equipment to collect data in different areas. 2) Construct a hyperspectral unmixing model based on a spectral modulation module and spectral variability. The hyperspectral unmixing model is trained on a pre-collected standard spectral dataset of blue carbon resources. The hyperspectral unmixing model includes a spectral modulation module, an abundance encoder module, a perturbation decoder module, and a reconstruction decoder module. 3) The collected raw spectral data is unmixed using a pre-trained hyperspectral unmixing model based on spectral modulation module and spectral variability; 4) Based on the unmixing results, filter the data collected in each region, and decide whether to collect the original data again based on the filtering results; 5) Calculate the dispersion between the spectra remaining after filtering in step 4) in each region, and judge the validity of the data according to the threshold. If the data validity is sufficient, proceed to step 6); otherwise, re-collect the data. 6) The spectra retained by averaging across regions are used as representative spectra for each region to obtain the results of spectral data collection and screening.

2. The spectral data acquisition and screening method for blue carbon data inversion according to claim 1, characterized in that, Step 1) includes: For the object to be collected, the collection area is divided according to a nine-square grid, and the data is collected evenly in each collection area. By analyzing the spectra, the raw spectral data of each region can be obtained. ,in This represents the number of spectral bands.

3. The spectral data acquisition and screening method for blue carbon data inversion according to claim 1, characterized in that, Step 2) The dataset is constructed based on the standard spectra of plants corresponding to blue carbon resources combined with a random noise distribution function; the spectral modulation module is used to modulate and preprocess data collected under different imaging conditions and initially extract features to obtain spectral features; the abundance encoder module mines the information corresponding to each endmember in the spectral features and outputs an abundance map; the perturbation decoder module analyzes the influence of different imaging conditions on the model and the perturbations in the spectrum, and outputs the amount of perturbation in the current spectral data; the reconstruction decoder module linearly mixes the endmember spectra according to the abundance map and introduces the estimated perturbation to obtain the final reconstruction result; the hyperspectral unmixing model is divided into a teacher network and a student network according to the differences in the input training dataset, and the hyperspectral unmixing model is trained based on the knowledge distillation method.

4. The spectral data acquisition and screening method for blue carbon data inversion according to claim 3, characterized in that, The dataset is constructed based on the standard spectra of plants corresponding to blue carbon resources combined with a random noise distribution function, and includes: The spectral curves of typical carbon-storing plants were selected as reference spectra, i.e., the endmembers of the dataset; the original collected spectra were converted into reflectance to construct the dataset, as shown in the following formula: ； in, It is the reflectivity of the material after conversion. The radiance value of the target material is obtained by measuring it with a spectrometer. This is the dark current value during measurement. To measure the radiance values ​​obtained by the spectrometer from a standard reference white plate under the same time and imaging conditions, The calibrated reflectance of the standard reference white board is used to obtain the spectral curve library of the reference material. , This represents the number of reference materials in the spectral curve library, i.e., the number of endmembers. An augmented linear mixing model was used to analyze the spectral curves of the mixed materials under different proportions and spectral variability conditions. The formula for the augmented linear mixing model is as follows: ； in The spectrum obtained after mixing according to the model, and These represent the abundance and endmembers during mixing, respectively. , as well as These represent scaling, perturbation, and noise effects during spectral mixing, respectively.

5. The spectral data acquisition and screening method for blue carbon data inversion according to claim 3, characterized in that, The spectral modulation module is composed of three stacked sub-modules, each of which includes a one-dimensional adaptive convolutional layer and a GELU activation function layer. One-dimensional adaptive convolutional layers adaptively adjust the focus region of feature extraction based on changes in the spectral peaks of the input data for subsequent demixing; one-dimensional adaptive convolutional layers include offset layers, modulation layers, and dynamic convolutional layers. Based on the offset and modulation layers in the one-dimensional adaptive convolutional layer, the offset and modulation amount are estimated. The dynamic convolutional layer dynamically adjusts the position of interest when the one-dimensional convolutional kernel moves along the one-dimensional spectrum based on the offset, and adjusts the importance of each band in combination with the modulation amount. The specific formula is expressed as follows: ； in, Representing the The result of the next one-dimensional adaptive convolution Represents the first The weights of an adaptive convolutional kernel, Representing the The next one-dimensional adaptive convolution in the th... The modulation amount at each sampling point location, Enumerated the first The position offset of all the neighborhoods of each sampling point The kernel size is [size]. The first one representing network prediction The sampling point at the th sampling point Learnable offset during each convolution; Each submodule in the spectral modulation module is represented as follows: ； AAC represents a one-dimensional adaptive convolutional layer. For GELU activation function, The output of the first submodule is shown. The spectral modulation module has three submodules, and the final output modulated features are as follows: .

6. The spectral data acquisition and screening method for blue carbon data inversion according to claim 5, characterized in that, The abundance encoder module uses the features modulated by the spectral modulation module to mine the information of each type of endmember and interpret the abundance corresponding to each type of endmember. The abundance encoder module is implemented by combining one-dimensional convolution and multilayer perceptron. One-dimensional convolution is used to integrate features in the spectral vector, and multilayer perceptron is used to map the features to the endmember dimension and encode them as abundance.

7. The spectral data acquisition and screening method for blue carbon data inversion according to claim 6, characterized in that, The perturbation decoder module modulates the features. The coefficients are transformed into a perturbation dictionary through two sets of one-dimensional convolutions and a multilayer perceptron. To ensure that the sum of the perturbation coefficients is 1, the Softmax activation function is used, which is formally expressed as: ； in, The coefficients of the perturbation dictionary; The estimation process of the perturbation quantity is represented by simulating a perturbation dictionary based on a high-dimensional single-layer fully connected layer, and optimizing the perturbation library in the perturbation dictionary through training: ； in The amount of perturbation estimated by the perturbation decoder is used to characterize the perturbations present in the current data.

8. The spectral data acquisition and screening method for blue carbon data inversion according to claim 7, characterized in that, The reconstruction decoder module is used to reconstruct the original spectrum, thereby enabling self-supervised training of the network. The reconstruction decoder module consists of a linear reconstruction part and a spectral variable part, specifically represented as follows: ； Here, FC is a fully connected layer without offset, used to map low-dimensional abundance maps to the original spectral band dimension, and its weights are regarded as endmember curves.

9. The spectral data acquisition and screening method for blue carbon data inversion according to claim 1, characterized in that, Step 3) includes: based on the trained student network, unmixing the original spectra of each region to obtain the abundance of typical materials in the spectrum, which serves as the basis for subsequent screening; Step 4) includes: based on the abundance values ​​of various materials estimated in step 3) in the currently acquired raw spectral data, combined with the types of materials contained in the imaging scene, and according to a threshold... To initially determine the effectiveness of the current material, if the abundance value of the current material in the spectrum is greater than a threshold... If the spectrum is not found in the specified range, it is considered a valid spectrum and is retained. After filtering the spectra, the number of retained spectra in each region is counted based on the filtering results. If the total number of retained spectra in a corresponding region is less than [a certain value], the spectrum is considered valid. If the condition is met, then the operation will be repeated from step 1) until the condition is met.

10. The spectral data acquisition and screening method for blue carbon data inversion according to claim 1, characterized in that, Step 5) includes: The dispersion of data within each region is calculated using the covariance ellipsoid volume; firstly, the covariance matrix is ​​calculated for the spectral data of each region. Perform eigenvalue decomposition on the covariance matrix; thereby calculate the volume of the covariance ellipsoid formed by the data of the current region. Since the volume of the covariance ellipsoid reflects the compactness of the ellipsoid formed by the spectral data, a larger volume of the ellipsoid formed by the data in high-dimensional space indicates a higher degree of data dispersion. Based on a threshold... This further determines the validity of the current region's data; if the volume of the covariance ellipsoid of the current region is smaller than... If the collected data is valid, then based on the filtered data, if the remaining number of spectra in each region is less than [a certain value], the data is considered valid. If the data is not collected, return to step 1 and re-collect and filter the data.

Images