Spectral correction method for remote sensing image based on mixed pixel decomposition

Abstract

The application belongs to the technical field of image spectrum correction, and particularly relates to a remote sensing image spectrum correction method based on mixed pixel decomposition, which comprises the following steps: acquiring remote sensing image data and end member spectrum data of photosynthetic vegetation, non-photosynthetic vegetation and bare soil; performing mixed pixel decomposition on the remote sensing image data by using the end member spectrum data to obtain spectral components of the photosynthetic vegetation, non-photosynthetic vegetation and bare soil in each mixed pixel; taking the spectral components as prior information and jointly constructing multi-source input features with original image spectra of the remote sensing image data; training a spectrum correction model, learning a nonlinear mapping relationship between the original image spectra of the remote sensing image data and target soil spectra; and processing to-be-corrected remote sensing images by using the trained spectrum correction model to obtain corrected spectral images. The method effectively separates the spectral contributions of different ground object components in the mixed pixel and reduces the interference of non-soil components on the soil spectrum signal.

Description

Technical Field

[0001] This application belongs to the field of image spectral correction technology, specifically a remote sensing image spectral correction method based on hybrid pixel decomposition. Background Technology

[0002] In the field of remote sensing technology, satellite remote sensing can provide rich information on the spectral reflectance of the Earth's surface, which has important application value in agricultural monitoring and ecological environment assessment. However, due to the limited spatial resolution of sensors, in low- and medium-resolution remote sensing images, a single pixel often contains multiple surface objects, resulting in mixed pixel phenomena. Especially in arid and semi-arid regions, the surface is usually composed of multiple components such as bare soil (BS), photosynthetic vegetation (PV), and non-photosynthetic vegetation (NPV), and its pixel spectral information is a comprehensive response of the spectra of multiple ground objects. Directly using mixed spectra for analysis and modeling can easily introduce interference information from non-target components, thereby reducing the accuracy and stability of soil parameter inversion. To alleviate the above problems, existing technologies usually use mixed pixel decomposition methods to process remote sensing images, with linear spectral demixing methods being the main approach. However, under actual surface conditions, due to the influence of factors such as soil particle structure, vegetation shading, scattering effects, and multiple reflections, the spectral mixing process between different ground objects often exhibits obvious nonlinear characteristics. Traditional linear decomposition methods are difficult to accurately describe this complex spectral coupling relationship, thus making it difficult to accurately extract the true spectral signals of the target components. Therefore, considering the nonlinear spectral coupling mechanism of mixed pixels, how to effectively suppress the interference of non-target components and accurately extract soil spectral information has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0003] This application provides a remote sensing image spectral correction method based on hybrid pixel decomposition, which solves the problem that traditional linear decomposition methods are difficult to accurately describe such complex spectral coupling relationships, thus making it difficult to accurately extract the true spectral signals of target components.

[0004] This application provides a remote sensing image spectral correction method based on hybrid pixel decomposition, including: Acquire remote sensing image data as well as endmember spectral data of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil; The remote sensing image data is decomposed into mixed pixels using the endmember spectral data to obtain the spectral components of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil in each mixed pixel. The spectral components are used as prior information and together with the original image spectrum of the remote sensing image data to construct multi-source input features. The multi-source input features are input into a pre-constructed spectral correction model, and the spectral correction model is trained to learn the nonlinear mapping relationship between the original image spectrum of the remote sensing image data and the target soil spectrum. The trained spectral correction model is used to process the remote sensing image to be corrected, resulting in a corrected spectral image, which is then associated with georeferenced information.

[0005] Further, the remote sensing image data is decomposed into mixed pixels using the endmember spectral data, including: Multiple candidate mixed linear spectral models are constructed for each mixed pixel. Each candidate mixed linear spectral model corresponds to a different combination of endmembers. The candidate mixed linear spectral model represents the spectral reflectance of the mixed pixel as a linear combination of the endmember spectra weighted according to their corresponding abundance ratios. The abundance of each endmember in each candidate mixed linear spectral model is solved using the fully constrained least squares method. Based on the abundance of each endmember obtained from the solution, the pixel spectra of each candidate mixed linear spectral model are reconstructed to obtain the reconstructed spectra. The root mean square error and spectral angular distance between the reconstructed spectra and the original pixel spectra are calculated respectively. A dual-constraint error evaluation function is constructed using the root mean square error and spectral angular distance; the root mean square error and spectral angular distance are normalized and then weighted and summed to obtain the comprehensive error evaluation value of each candidate mixed linear spectral model; By comparing the comprehensive error evaluation values ​​of each candidate mixed linear spectral model, the candidate mixed linear spectral model with the smallest error evaluation value is selected as the best decomposition result of the mixed pixels.

[0006] Further, the root mean square error and spectral angular distance between the reconstructed spectrum and the original pixel spectrum are calculated, including: , , In the formula For original pixels Band reflectivity; To reconstruct pixels Band reflectivity; For the number of bands, The root mean square error, Spectral angular distance for The transpose of .

[0007] Furthermore, the candidate mixed linear spectral model is configured to: for each band of each mixed pixel, multiply the reflectance of each endmember in that band by the corresponding abundance and sum them, and add the residual error to reconstruct the reflectance of the mixed pixel in that band.

[0008] Furthermore, the abundance of each endmember in each candidate mixed linear spectral model is solved using the fully constrained least squares method, with non-negative abundance and an abundance sum of 1 as physical constraints.

[0009] Furthermore, the abundance of each endmember in each candidate mixed linear spectral model is solved using the fully constrained least squares method, including: The product of endmember spectra and abundance is approximated to the pixel spectra, and a function is established with the objective function of minimizing the sum of squared approximation errors. The convex quadratic programming problem is subject to a non-negativity constraint and a sum-to-one constraint, wherein the non-negativity constraint requires that the abundance of each endmember is greater than or equal to zero, and the sum-to-one constraint requires that the sum of the abundance of all endmembers is equal to 1. The convex quadratic programming problem is solved using a constrained optimization algorithm to obtain the optimal abundance that satisfies all constraints.

[0010] Furthermore, the pixel spectra of each candidate mixed linear spectral model are reconstructed to obtain the reconstructed spectra, including: The product of the endmember spectrum of photosynthetic vegetation and its abundance is taken as the spectral contribution of photosynthetic vegetation components. The product of the endmember spectra of non-photosynthetic vegetation and its abundance is taken as the spectral contribution of non-photosynthetic vegetation components. The product of the end-member spectrum of bare soil and its abundance is taken as the spectral contribution of bare soil components. The spectral contributions of the photosynthetic vegetation component, non-photosynthetic vegetation component, and bare soil component are accumulated band by band to obtain the reconstructed spectrum.

[0011] Furthermore, the spectral correction model employs a convolutional neural network, comprising three parallel feature extraction branches that respectively process the spectral components of photosynthetic vegetation, the spectral components of non-photosynthetic vegetation, and the original image spectrum. Each branch contains at least one convolutional layer, one batch normalization layer, and one pooling layer in sequence. The features extracted from the three branches are concatenated and fused, and then input into multiple fully connected layers for nonlinear transformation. A Dropout layer is introduced between fully connected layers to suppress overfitting; The output layer uses the Sigmoid activation function to constrain the predicted values ​​to the interval [0,1].

[0012] Compared with the prior art, the embodiments of this application have at least the following beneficial effects: This application effectively separates the spectral contributions of different land cover components in mixed pixels by introducing the results of mixed pixel decomposition as prior information, thereby reducing the interference of non-soil components on soil spectral signals. By constructing a nonlinear spectral correction model, complex nonlinear spectral coupling relationships can be effectively explored, improving the accuracy of soil spectral extraction.

[0013] By fusing the original image spectrum with the spectral components of the components to construct multi-source input features, the model's ability to express spectral information is improved, and the stability of the spectral correction results is enhanced. By introducing component information with clear physical meaning as input to the spectral correction model, the dependence of the spectral correction model on the data distribution of a specific region is reduced, thereby improving the applicability and generalization ability of the method in different regions. Attached Figure Description

[0014] Figure 1 A flowchart illustrating a remote sensing image spectral correction method based on hybrid pixel decomposition, provided for embodiments of this application; Figure 2 A schematic diagram of the spectral curves of (a) the original image and (b) the corrected spectral image provided for embodiments of this application; Figure 3 This is a schematic diagram illustrating the correction accuracy of the spectral correction model provided in the embodiments of this application across various image bands. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0016] To address the influence of factors such as soil particle structure, vegetation shading, scattering effects, and multiple reflections under actual surface conditions, as mentioned in the background section, the spectral mixing process between different land features exhibits significant nonlinear characteristics. Traditional linear spectral mixing decomposition methods struggle to accurately describe this nonlinear coupling relationship, making it difficult to accurately extract the true spectral signals of the target components.

[0017] This application discloses a remote sensing image spectral correction method based on mixed pixel decomposition. First, remote sensing image data and endmember spectral data of photosynthetic vegetation (PV), non-photosynthetic vegetation (NPV), and bare soil (BS) are acquired. Then, a multi-endmember spectral fusion analysis method is used to decompose the image into mixed pixels, obtaining the spectral components of different components in each mixed pixel. Based on this, multi-source input features are constructed by combining the image spectrum with the spectral components of different components, and then input into a spectral correction model based on a convolutional neural network to establish a nonlinear mapping relationship between the image spectrum and the target soil spectrum. Finally, the corrected spectral image is output while retaining its georeferenced information. This application effectively suppresses spectral coupling interference from non-soil components in mixed pixels and improves the stability of soil spectral information by using the mixed pixel decomposition results as prior information and combining them with a convolutional neural network for spectral correction. Simultaneously, by introducing component information with clear physical meaning as input to the spectral correction model, the dependence of the spectral correction model on the data distribution of a specific region is reduced, thereby improving the applicability and generalization ability of the method across different regions. This application can be widely applied in fields such as agricultural remote sensing, soil environmental monitoring, and geographic information processing.

[0018] See Figure 1 As shown in the figure, this application provides a remote sensing image spectral correction method based on hybrid pixel decomposition, including: S101, acquire remote sensing image data as well as endmember spectral data of photosynthetic vegetation, non-photosynthetic vegetation and bare soil; S102, using the end-member spectral data to perform mixed pixel decomposition on the remote sensing image data to obtain the spectral components of photosynthetic vegetation, non-photosynthetic vegetation and bare soil in each mixed pixel; S103, the spectral components are used as prior information and together with the original image spectrum of the remote sensing image data to construct multi-source input features; S104, The multi-source input features are input into the pre-constructed spectral correction model, and the spectral correction model is trained to learn the nonlinear mapping relationship between the original image spectrum of the remote sensing image data and the target soil spectrum. S105, the trained spectral correction model is used to process the remote sensing image to be corrected to obtain the corrected spectral image, and the corrected spectral image is associated with georeferenced information.

[0019] In step S101, the remote sensing image data is a multispectral remote sensing image. After acquiring the original remote sensing image data, a series of preprocessing operations are required to ensure the accuracy of subsequent mixed pixel decomposition and spectral correction. These preprocessing operations specifically include the following steps: radiometric calibration, converting the original digital quantization values ​​(DN values) recorded in the remote sensing image into radiance values ​​with actual physical meaning. The radiometric calibration process utilizes the radiometric calibration coefficients calibrated at the sensor's factory, converting the DN values ​​into radiance through linear or nonlinear transformations; atmospheric correction, further converting the radiance values ​​into true surface reflectance to eliminate the absorption and scattering effects of atmospheric molecules, aerosols, water vapor, etc., on solar radiation. Various mature models can be used for atmospheric correction. After atmospheric correction, the output surface reflectance data ranges from 0 to 1, possessing clear physical meaning; geometric correction, performing geometric fine correction on the remote sensing image to eliminate geometric distortions caused by factors such as sensor attitude changes, Earth curvature, and terrain undulations. Geometric correction typically utilizes ground control points or high-precision digital elevation models, combined with sensor orbit parameters, to register the image to a standard geographic coordinate system. Cloud and shadow masking involves identifying and marking cloud-covered areas, cloud shadows, and terrain shadows in the image. The pixel spectra of these areas are affected by cloud interference or insufficient lighting, making them unsuitable for subsequent soil spectral correction. Automatic identification can be achieved using thresholding, spectral indexing, or machine learning-based cloud detection algorithms, generating corresponding mask files for removal or separate processing of these areas.

[0020] After the above preprocessing, the remote sensing image data is output as surface reflectance data. Each mixed pixel contains spectral reflectance values ​​of B bands, forming a B-dimensional spectral vector. These spectral vectors are the original observation data in the subsequent mixed pixel decomposition.

[0021] Photosynthetic vegetation refers to living green vegetation capable of photosynthesis, including forests, grasslands, and farmland crops. End-member spectral data of photosynthetic vegetation characterizes the standard spectral reflectance curve of pixels with pure photosynthetic vegetation cover. This end-member spectrum has the following typical characteristics: in the blue and red light regions of the visible light band, reflectance is low due to the strong absorption of chlorophyll; a small reflectance peak exists in the green light band, presenting a green visual effect; in the near-infrared band, reflectance significantly increases due to multiple scattering by the internal cell structures of the leaves, forming the typical red-edge feature of the vegetation spectrum. End-member spectral data of photosynthetic vegetation can be obtained through one or more of the following methods: selection from publicly available spectral libraries, automatic extraction from remote sensing images, and acquisition through field or laboratory measurements.

[0022] Non-photosynthetic vegetation refers to vegetation that has withered and died, losing its photosynthetic capacity, including crop straw, withered grass, fallen leaves, and dead branches. End-member spectral data of non-photosynthetic vegetation characterizes the standard spectral reflectance curve of pixels covered by pure non-photosynthetic vegetation. Compared with photosynthetic vegetation, non-photosynthetic vegetation exhibits significantly different spectral characteristics: in the visible light band, there is no strong absorption of chlorophyll, resulting in higher overall reflectance; in the short-wave infrared band (especially near 2100 nm), characteristic absorption peaks exist due to the absorption by organic components such as cellulose, lignin, and hemicellulose; in the near-infrared band, the reflectance is lower than that of photosynthetic vegetation, and there is no obvious red-edge characteristic.

[0023] Bare soil refers to the exposed soil surface that is not covered by vegetation. End-member spectral data of bare soil characterizes the standard spectral reflectance curve of pixels covered by pure bare soil. The spectral characteristics of bare soil are influenced by various factors, including soil texture (sand, loam, clay), soil mineral composition (iron oxides, carbonates, clay minerals, etc.), soil organic matter content, soil moisture content, and soil surface roughness. Generally, dry sandy soil has high reflectance that increases slowly with wavelength; dry clay has moderate reflectance and exhibits water absorption characteristics in the shortwave infrared band; organic-rich black soil has low overall reflectance; and soils containing iron oxides exhibit a reddish-yellow hue in the visible light band.

[0024] The acquired endmember spectral data of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil are stored and represented in vector form. Assuming the remote sensing image has B bands, each endmember spectrum can be represented as a B-dimensional vector.

[0025] In step S102, for remote sensing imagery, due to spatial resolution limitations, a single pixel often covers an area of ​​several square meters to several square kilometers, which typically contains multiple land cover types. Therefore, most pixels are not composed of a single pure land cover, but rather are mixed pixels formed by a certain area ratio of photosynthetic vegetation, non-photosynthetic vegetation, bare soil, and other land cover. The purpose of mixed pixel decomposition is precisely to separate the pure spectral contribution of each component from the mixed spectrum of each pixel.

[0026] Step S102 utilizes the end-member spectral data of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil obtained in step S101 to decompose each mixed pixel in the remote sensing image and quantitatively calculate the spectral components of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil in each mixed pixel. The spectral component refers to the spectral contribution value of each end-member component in the mixed pixel, which is obtained by multiplying the reflectance of the standard spectrum of the end-member by the area proportion (i.e., abundance) it occupies in the pixel.

[0027] This application employs a linear spectral mixing model as the basis for mixed pixel decomposition. This model assumes that incident light undergoes only a single reflection between the surface components, and that there is no interaction between the components. Therefore, the reflectance of a mixed pixel in each band is equal to the weighted sum of the reflectances of each endmember in that band according to their respective abundances, plus a residual error term. Mathematically, for any pixel, its reflectance in band B can be expressed as the sum of the products of the reflectance and abundance of photosynthetic vegetation endmembers, the reflectance and abundance of non-photosynthetic vegetation endmembers, and the reflectance and abundance of bare soil endmembers, plus the residual error for that band.

[0028] Linear spectral mixing models have advantages such as clear physical meaning, high computational efficiency, and ease of imposing constraints, making them a commonly used method in this field for mixing pixel decomposition. Although nonlinear mixing phenomena may exist on the real earth's surface, linear models can provide sufficiently accurate decomposition results in most cases and provide a good prior information foundation for the establishment of subsequent nonlinear spectral correction models.

[0029] As a preferred embodiment of this application, a multi-endmember spectral fusion analysis method is used for fusion pixel decomposition. This method is an improvement on the traditional linear spectral fusion analysis method. Its core idea is to construct multiple candidate fusion linear spectral models for each fusion pixel, each candidate fusion linear spectral model using different endmember combinations, and then selecting the optimal decomposition result. The candidate fusion linear spectral models are configured such that, for each band of each fusion pixel, the reflectance of each endmember in that band is multiplied by the corresponding abundance, summed, and then the residual error is added to reconstruct the reflectance of the fusion pixel in that band. The purpose of this is to address the problem of spatial heterogeneity of land cover, that is, even if different regions belong to the same photosynthetic vegetation or bare soil, their spectral characteristics may differ. For example, the spectra of different forest types and different soil types are different.

[0030] The specific implementation of the multi-endmember spectral mixture analysis method includes: An endmember spectral library was constructed. Based on the endmember spectral data obtained in step S101, endmember libraries for photosynthetic vegetation, non-photosynthetic vegetation, and bare soil were established. The photosynthetic vegetation endmember library contains spectra of various vegetation types, such as broadleaf forest, coniferous forest, grassland, and crop spectra. The non-photosynthetic vegetation endmember library contains endmember spectral data of various types of dead vegetation, such as wheat straw, corn straw, fallen leaves, and hay. The bare soil endmember library contains spectra of various soil types, such as sandy soil, clay, black soil, red soil, and saline-alkali soil. The number of endmembers in each endmember library can be determined according to the complexity of the land cover in the study area.

[0031] A candidate set of mixed linear spectral models is generated. For each mixed pixel, one endmember spectrum is selected from the photosynthetic vegetation endmember library, the non-photosynthetic vegetation endmember library, and the bare soil endmember library. These three endmembers are combined to form a candidate linear spectral mixing model. All possible endmember combinations are iterated to generate the candidate set of mixed linear spectral models. To reduce computational load, intelligent filtering strategies can also be used. For example, the spectral similarity between endmembers can be calculated, and redundant endmembers with highly similar spectra can be removed; or, based on prior knowledge of the study area, only a subset of endmembers that may be related to the current pixel can be retained; or, a random sampling method can be used to extract representative combinations from a large endmember library.

[0032] The abundance of endmembers in each candidate mixed linear spectral model is determined. For each candidate mixed linear spectral model, the area proportions of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil are calculated using the fully constrained least squares method, i.e., the abundance. Two physical constraints must be strictly satisfied during the solution process: first, a non-negativity constraint, meaning the abundance of all endmembers must be greater than or equal to zero, since the area proportion of any land cover cannot be negative; second, a sum-of-one constraint, meaning the sum of the abundances of all endmembers must equal one, because this represents that the total area of ​​all land cover types within the mixed pixel is equal to the area of ​​the mixed pixel itself.

[0033] The core of fully constrained least squares is to minimize the sum of squared errors between the original image spectrum and the reconstructed spectrum, while satisfying the two constraints mentioned above. The reconstructed spectrum refers to the spectrum recalculated using the abundance and reflectance of the endmember spectra obtained from the solution, and ideally, it should be very close to the original image spectrum.

[0034] In one example, the abundance of each endmember in each candidate mixed linear spectral model is solved using the fully constrained least squares method, including: The product of reflectance and abundance of the endmember spectrum is used to approximate the pixel spectrum, and a function is established to minimize the sum of squares of the approximation error. The objective function is the sum of squares of the differences between the original image spectrum and the reconstructed spectrum. Apply a nonnegativity constraint and a sum-to-one constraint, where the nonnegativity constraint requires that the abundance of each endmember be greater than or equal to zero, and the sum-to-one constraint requires that the sum of the abundance of all endmembers be equal to 1. The objective function problem is solved using a constrained optimization algorithm to obtain the optimal abundance that satisfies all constraints.

[0035] Based on the abundance obtained from the solution, the pixel spectra of each candidate mixed linear spectral model are reconstructed according to the linear spectral mixing model. Specifically, for a candidate mixed linear spectral model, its photosynthetic vegetation endmember spectrum is multiplied by the photosynthetic vegetation abundance to obtain the spectral contribution of the photosynthetic vegetation component; its non-photosynthetic vegetation endmember spectrum is multiplied by the non-photosynthetic vegetation abundance to obtain the spectral contribution of the non-photosynthetic vegetation component; and its bare soil endmember spectrum is multiplied by the bare soil abundance to obtain the spectral contribution of the bare soil component. Then, the spectral contributions of these three components are added band by band to obtain the reconstructed spectrum of the candidate mixed linear spectral model.

[0036] Normalization and Comprehensive Error Calculation: Since the root mean square error (RMSE) and spectral angular distance have different dimensions and numerical ranges, they cannot be directly compared or weighted. Therefore, for all candidate mixed linear spectral models of the current mixed pixels, minimum-maximum normalization is applied to both the RMSE and spectral angular distance. Specifically, the minimum and maximum RMSE values ​​of all candidate mixed linear spectral models are found. The minimum RMSE value is subtracted from the RMSE of each candidate mixed linear spectral model, and then divided by the range to obtain the normalized RMSE. Similarly, the spectral angular distance is also normalized in the same way.

[0037] After normalization, the normalized root mean square error and the normalized spectral angular distance are weighted and summed to obtain the comprehensive error evaluation value of each candidate mixed linear spectral model. The weighting coefficients can be adjusted according to actual application requirements. When the accuracy requirement for the absolute value of spectral reflectance is higher than the waveform matching accuracy, the weight of the root mean square error can be increased; when the accuracy requirement for spectral waveform matching is higher, the weight of the spectral angular distance can be increased.

[0038] After the reconstructed spectrum is calculated, it is compared with the original pixel spectrum. A dual-constraint error evaluation function is constructed using the root mean square error (RMSE) and spectral angular distance. The RMSE and spectral angular distance are normalized and then weighted and summed to obtain the comprehensive error evaluation value for each candidate mixed linear spectral model. Two error evaluation indices are calculated: the first is the RMSE, which measures the difference in numerical magnitude between the reconstructed spectrum and the original image spectrum. A smaller RMSE indicates that the reconstructed spectrum is numerically closer to the original image spectrum. The second is the spectral angular distance, which measures the similarity in vector direction between the reconstructed spectrum and the original image spectrum, i.e., the similarity in shape of their spectral curves. A smaller spectral angular distance indicates that the shapes of the two spectral curves are more consistent. These two indices evaluate the model's fitting quality from different perspectives and are complementary. The RMSE is sensitive to the absolute value difference of reflectance, while the spectral angular distance is sensitive to the shape difference of the spectral waveform.

[0039] By comparing the comprehensive error evaluation values ​​of all candidate mixed linear spectral models for the current pixel, the candidate mixed linear spectral model with the smallest error is selected as the best decomposition result for that mixed pixel. This selection method ensures that each mixed pixel can find the most suitable spectral combination of the actual land cover composition from the endmember library, thereby improving the accuracy and adaptability of mixed pixel decomposition.

[0040] When multiple candidate mixed linear spectral models have equal and minimum combined errors, the candidate mixed linear spectral model with the simplest endmember combination or the smallest difference between endmember spectra can be selected as the best decomposition result. When the combined error of all candidate mixed linear spectral models exceeds a preset threshold, it indicates that no candidate mixed linear spectral model can fit the spectrum of the mixed pixel well. In this case, the mixed pixel can be marked as unreliable for decomposition, and the default candidate mixed linear spectral model or the decomposition result of the neighboring pixel can be used instead.

[0041] The abundance obtained from solving the optimal candidate mixed linear spectral model is used as the final abundance output for that pixel. Then, based on the abundance and endmember spectra, the spectral components of each endmember are calculated. The spectral components are calculated by multiplying the spectral vector of each endmember by its corresponding abundance. For photosynthetic vegetation, its spectral components are equal to the photosynthetic vegetation abundance multiplied by its endmember spectral vector; for non-photosynthetic vegetation, its spectral components are equal to the non-photosynthetic vegetation abundance multiplied by its endmember spectral vector; for bare soil, its spectral components are equal to the bare soil abundance multiplied by its endmember spectral vector.

[0042] These three spectral components have the same vector length as the original pixel spectrum (all in B bands), representing the spectral contributions of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil to the mixed pixel, respectively. The original pixel spectrum should theoretically be equal to the sum of these three spectral components plus a residual error term. The spectral components are the key bridge connecting the mixed pixel decomposition with the subsequent spectral correction model. As prior information with clear physical meaning, they will be input into the subsequent convolutional neural network to guide the correction of the soil spectrum.

[0043] The spectral components obtained through the above-described mixed pixel decomposition have clear physical meanings. The photosynthetic vegetation spectral component represents the contribution of living green vegetation in the mixed pixel to the observed spectrum. Due to the high reflectivity of photosynthetic vegetation in the near-infrared band, this component dominates in pixels with high vegetation cover. The non-photosynthetic vegetation spectral component represents the contribution of dead vegetation to the observed spectrum, and is more significant in pixels after crop harvest or in arid regions. The bare soil spectral component represents the contribution of soil to the observed spectrum, which is the target signal that this method ultimately aims to extract and correct.

[0044] In traditional remote sensing processing methods, soil signals in mixed pixels are masked by vegetation signals, making it difficult to directly obtain pure soil spectra. However, through the mixed pixel decomposition of this application, the spectral contributions of different components are successfully separated, and the spectral components of bare soil are extracted from the mixed spectrum, providing crucial prior information for subsequent spectral correction.

[0045] In one example, the formula for blended pixel decomposition is as follows: , , , In the formula For mixed pixels Band reflectivity; For end element In the band Reflectivity; For end element The proportion of pixels occupied, i.e., abundance; The number of endmembers; This refers to residual error; This is a double-constraint error evaluation function; These are the weighting coefficients; Indicates the first The root mean square error of each candidate mixed linear spectral model They represent the first The spectral angular distance between the candidate mixed linear spectral models; and These represent the maximum and minimum root mean square errors of all candidate mixed linear spectral models for the current mixed pixel, respectively. and These represent the maximum and minimum spectral angular distances of all candidate mixed linear spectral models for the current mixed pixel, respectively.

[0046] Calculate the root mean square error and spectral angular distance between the reconstructed spectrum and the original pixel spectrum, including: , , In the formula For original pixels Band reflectivity; To reconstruct pixels Band reflectivity; For the number of bands, The root mean square error, Spectral angular distance for The transpose of .

[0047] The formulas for calculating spectral components are as follows: , In the formula For end element spectral components; For end element Spectral reflectance.

[0048] In step S103, the spectral components obtained from the hybrid pixel decomposition are used as prior information with clear physical meaning and organically fused with the original image spectral data of the remote sensing image to jointly construct multi-source input features for deep learning model training and prediction. The advantage of this fusion strategy is that the original image spectral data comes directly from sensor observations and contains comprehensive information on all surface features, but the information is mixed, making it difficult to directly separate the pure signal of the soil; while the spectral components, through hybrid pixel decomposition, extract the independent contribution of each component from the original image spectrum, which has a clear physical meaning.

[0049] Multi-source input features include at least two or all three of the following three types of information. As a preferred embodiment of this application, it is recommended to include all of the following types of information simultaneously: Raw image spectral features. These are vectors composed of the raw reflectance values ​​of each mixed pixel in each band of the remote sensing image data. They represent the raw data observed by the sensor, containing comprehensive spectral information of all ground cover components within the mixed pixel, and form the basis for subsequent processing.

[0050] Photosynthetic vegetation spectral component. This refers to the spectral contribution of the photosynthetic vegetation component obtained through the decomposition of the mixed pixels. It is calculated by multiplying the photosynthetic vegetation abundance by the spectral reflectance of the photosynthetic vegetation endmembers. This component represents the degree of contribution of living green vegetation in the mixed pixels to the observed spectrum. Its magnitude reflects the vegetation coverage, and its spectral shape reflects the type and growth status of the vegetation.

[0051] Non-photosynthetic vegetation spectral component. This refers to the spectral contribution of the non-photosynthetic vegetation component obtained through mixed pixel decomposition. It is calculated by multiplying the abundance of non-photosynthetic vegetation by the spectral reflectance of the non-photosynthetic vegetation endmembers. This component represents the degree of contribution of dead vegetation in the mixed pixels to the observed spectrum, and is more significant after crop harvest or in arid and semi-arid regions.

[0052] The bare soil spectral component refers to the spectral contribution of the bare soil component obtained through mixed pixel decomposition. It is calculated by multiplying the bare soil abundance by the bare soil endmember spectral reflectance. This component represents the degree to which the soil in a pixel contributes to the observed spectrum, which is precisely the approximate representation of the target signal that this method ultimately aims to extract and correct.

[0053] Abundance is a vector composed of the abundance of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil, obtained from the decomposition of mixed pixels. Abundance is a scalar value between 0 and 1, representing the area proportion of each vegetation type in the pixel. Compared to spectral components, abundance information is simpler, containing only proportional information and not spectral shape information.

[0054] The construction methods for multi-source input features include, but are not limited to, the following options: Channel stitching method. This is the simplest and most direct fusion method, which involves stitching various features together along the channel dimension to form a longer feature vector. Specifically, the original image spectral vector, photosynthetic vegetation spectral component, non-photosynthetic vegetation spectral component, and bare soil spectral component are concatenated end-to-end to form a multi-source input feature vector with a dimension of 4 x B. The first B elements of this feature vector are the original image spectral values, the next B elements are the photosynthetic vegetation spectral component values, the following B elements are the non-photosynthetic vegetation spectral component values, and the last B elements are the bare soil spectral component values. The advantage of this method is its simplicity and ease of implementation, without introducing additional parameters, and with complete and lossless information.

[0055] Multi-channel image array method. This method organizes various features into a multi-channel data structure, similar to multi-channel images in image processing. In this structure, the input data for each mixed pixel can be viewed as a B x 4 matrix, or a two-dimensional array with a height of B and a width of 4. Each column corresponds to a feature type, and each row corresponds to a band. This organization facilitates subsequent convolutional neural network operations, especially when using one-dimensional convolution, as it can slide along the spectral dimension and process all feature channels simultaneously.

[0056] Before being input into the spectral correction model, the constructed multi-source input features require a series of preprocessing operations to ensure the model's training stability and convergence speed. These include: Normalization: Scaling the value range of various features to a uniform range, typically 0 to 1 or -1 to 1. Since the original image spectral reflectance is already within the 0 to 1 range, and the spectral components are the product of abundance and endmember reflectance, also within the 0 to 1 range, the original features already have a relatively uniform value range. Outlier removal: Identifying and eliminating obviously unreasonable values ​​from the multi-source input features. For example, due to potential numerical errors during the mixed pixel decomposition process, the calculated spectral components may have small negative values ​​or values ​​slightly greater than 1. For these slightly out-of-range values, truncation can be used, setting values ​​less than 0 to 0 and values ​​greater than 1 to 1. For values ​​severely out of range, such as those with orders of magnitude errors or missing data, the pixel should be marked as an outlier and excluded in subsequent processing, or a neighborhood filling strategy should be used. Data augmentation. When the number of training samples is limited, data augmentation operations can be performed on multi-source input features to expand the training dataset and improve the model's generalization ability. Specific methods of data augmentation include, but are not limited to: adding small random noise to features to simulate the effect of sensor noise; randomly scaling features to simulate reflectivity changes under different lighting conditions; randomly discarding bands in features to simulate band loss; and simultaneously applying the same random spectral perturbation to the original image spectrum and spectral components to maintain physical consistency.

[0057] In step S104, the multi-source input features are input into the pre-constructed spectral correction model, and the spectral correction model is trained to learn the nonlinear mapping relationship between the original image spectrum of the remote sensing image data and the target soil spectrum. The spectral correction model adopts a convolutional neural network, which includes three parallel feature extraction branches to process the photosynthetic vegetation spectral components, non-photosynthetic vegetation spectral components and the original image spectrum, respectively. Each branch contains at least one convolutional layer, one batch normalization layer, and one pooling layer in sequence. The features extracted from the three branches are concatenated and fused, and then input into multiple fully connected layers for nonlinear transformation. A Dropout layer is introduced between fully connected layers to suppress overfitting; The output layer uses the Sigmoid activation function to constrain the predicted values ​​to the interval [0,1].

[0058] The kernel size can be set according to the actual situation, preferably 3×3, and the dropout rate can be any value between 0 and 0.4.

[0059] The spectral correction model adopts a deep learning architecture based on convolutional neural networks. Taking into account the fact that different types of information in multi-source input features have different physical meanings, it uses a multi-branch parallel structure to decouple and extract each type of feature, and then fuses them in the deep layers of the network. Through multi-layer nonlinear transformation, it completes the complex mapping from mixed spectrum to pure soil spectrum.

[0060] Compared with traditional fully connected neural networks or simple convolutional networks, it has the following advantages: the multi-branch structure allows the network to learn the feature patterns of photosynthetic vegetation, non-photosynthetic vegetation, and the original image spectrum in a targeted manner, avoiding mutual interference between different types of information in the shallow layers of the network; convolutional layers can effectively capture the local correlation between spectral bands, which is a key characteristic in spectral data processing; batch normalization layers accelerate network convergence and improve training stability; pooling layers reduce feature dimensionality and enhance the translation invariance of the network; dropout layers effectively suppress overfitting and improve the model's generalization ability.

[0061] The spectral correction model employs three parallel feature extraction branches, each processing a specific type of input feature. The three branches have a symmetrical network structure, but their network parameters are set independently and not shared. This allows for the learning of different filters based on the statistical properties of different types of input features.

[0062] The first branch is the photosynthetic vegetation spectral component processing branch. This branch receives the photosynthetic vegetation spectral component as input, which is the result of multiplying the photosynthetic vegetation abundance obtained from mixed pixel decomposition by its endmember spectrum. Typical characteristics of the photosynthetic vegetation spectral component include: low reflectance regions in the visible light band due to chlorophyll absorption, and high reflectance regions in the near-infrared band due to scattering by leaf cell structures. The goal of this branch is to extract the spectral features of photosynthetic vegetation, particularly how these features interfere with the spectral signal of the underlying soil.

[0063] The second branch is the non-photosynthetic vegetation spectral component processing branch. This branch receives the non-photosynthetic vegetation spectral component as input, which is the result of multiplying the non-photosynthetic vegetation abundance obtained from mixed pixel decomposition by its endmember spectrum. Typical characteristics of the non-photosynthetic vegetation spectral component include relatively high reflectance in the visible light band and characteristic absorption peaks in the short-wave infrared band caused by cellulose and lignin absorption. The goal of this branch is to extract the spectral features of non-photosynthetic vegetation, especially the interference patterns of dead vegetation on the soil spectrum.

[0064] The third branch is the raw image spectral processing branch. This branch receives the raw remote sensing image spectral vector as input. This vector is the mixed spectrum actually observed by the sensor, containing comprehensive information on photosynthetic vegetation, non-photosynthetic vegetation, bare soil, and other possible ground features. The goal of this branch is to extract the global features of the raw mixed spectrum, preserving complete observation information as a supplement to the information from the other two branches.

[0065] Each branch contains, in turn, a convolutional layer, a batch normalization layer, and a pooling layer. The specific implementation of each layer will be described in detail below.

[0066] Convolutional layers are the core component of convolutional neural networks. Their function is to perform local filtering on input features in the spectral dimension, extracting spectral correlation features between adjacent bands. Unlike two-dimensional convolution used to process two-dimensional images, this application uses one-dimensional convolution because spectral data is a one-dimensional sequence in the wavelength dimension. The data in each band has a strong correlation with the data in adjacent bands, while the correlation with non-adjacent bands is relatively weak.

[0067] In a convolutional layer, a learnable filter (also called a convolutional kernel) slides along the spectral dimension, multiplying element-wise with local regions of the input features and summing the results to generate an output value. Specifically, for a given convolutional kernel, as it moves along the spectral dimension, it generates a series of output values ​​that constitute the feature map extracted by that kernel. Using multiple convolutional kernels side-by-side allows for the extraction of various feature types.

[0068] The size of the convolution kernel can be set according to the actual application scenario and data characteristics. Smaller convolution kernels (such as 1×1) can only learn the feature transformations of a single band and cannot capture the correlation between bands; medium-sized convolution kernels (such as 3×3) can capture the local correlation between three adjacent bands, which is the most basic feature pattern in spectral data; larger convolution kernels (such as 5×5 or 7×7) can capture spectral features over a wider range, but will introduce more parameters and increase the risk of overfitting. After comprehensive consideration, in the preferred embodiment of the present invention, the convolution kernel size is set to 3×3. This size can effectively capture the local correlation between adjacent bands without introducing too many parameters, achieving a good balance between feature extraction capability and computational efficiency.

[0069] The stride of the convolutional layer is set to 1, meaning the convolutional kernel moves one band at a time to ensure no band information is missed. The padding method uses identical padding, i.e., zero padding at both ends of the sequence formed by the spectral data, ensuring the output feature map and the input feature map have the same length in the spectral dimension, facilitating connections to subsequent network layers.

[0070] The batch normalization layer, located after the convolutional layers, accelerates network convergence, improves training stability, and allows for larger learning rates. The basic principle of batch normalization is to standardize the input of each mini-batch of training samples to have zero mean and unit variance, and then restore the network's expressive power through learnable scaling and translation parameters.

[0071] Specifically, for a batch of input data, the batch normalization layer first calculates the mean and variance of the batch, then subtracts the mean from each sample and divides by the standard deviation of the variance to obtain a standardized distribution. This process effectively reduces the internal covariate shift problem, i.e., the problem of input distribution variations between layers in the network. In the early stages of training, the network parameters are randomly initialized, and the distribution of inputs to each layer may change drastically, leading to training difficulties. Batch normalization forces each layer's input to maintain a stable distribution, making the network easier to train.

[0072] Furthermore, batch normalization also has a regularization effect because the mean and variance of each batch are randomly estimated, introducing slight noise, similar to the regularization effect of Dropout. This makes the network less sensitive to small changes in the input data, helping to suppress overfitting.

[0073] In this spectral correction model, after the batch normalization layer is applied to each convolutional layer, each feature channel is normalized separately.

[0074] The role of pooling layers is to perform dimensionality reduction sampling on feature maps, reducing the spatial dimension of features and thus reducing the number of parameters in subsequent fully connected layers. This also enhances the network's invariance to small input displacements. For spectral data, pooling operations are equivalent to compressing and abstracting information from adjacent bands, extracting salient features from local regions.

[0075] There are two main types of pooling layers: max pooling and average pooling. Max pooling takes the maximum value within the pooling window as the output, and its advantage is that it can retain the most significant features in the local region and has a certain degree of robustness to noise. Average pooling takes the average value within the pooling window as the output, and its advantage is that it can retain the overall information of the local region and the feature response is smoother.

[0076] In one embodiment, the pooling window size is set to 2, and the stride is set to 2, meaning that pooling is performed on features from every two adjacent bands, outputting a single value. This setting halves the spectral dimension of the feature map, effectively reducing the computational burden on subsequent network layers. The pooling method can be max pooling or average pooling; preferably, max pooling is used to highlight salient features in each local spectral region.

[0077] After processing by convolutional layers, batch normalization layers, and pooling layers, the feature maps output by each branch are reduced in the spectral dimension and increased in the channel dimension compared to the original input, forming a higher-level and more abstract feature representation.

[0078] After the three parallel branches complete their feature extraction, they fuse the extracted features to form a comprehensive feature representation. Feature fusion is a key step in the network architecture, and its goal is to integrate information from different branches to provide a rich feature foundation for subsequent nonlinear mapping.

[0079] Feature fusion is achieved through a concatenation method. Specifically, the feature vectors output from the first branch, the second branch, and the third branch are concatenated end-to-end along the channel dimension to form a longer feature vector. This feature vector simultaneously includes the spectral features of photosynthetic vegetation, the spectral features of non-photosynthetic vegetation, and the comprehensive features of the original mixed spectrum.

[0080] After feature stitching is completed, an optional feature interaction layer can be introduced to learn the interaction relationships between different features. Feature interaction can employ outer product operations or multi-head attention mechanisms, enabling the network to capture pairwise interactions and synergistic effects among photosynthetic vegetation, non-photosynthetic vegetation, and bare soil features. This is particularly important for accurately modeling nonlinear spectral mixing processes.

[0081] The fused feature vector is input into multiple fully connected layers for nonlinear transformation. Fully connected layers, also known as densely connected layers, have connections between each input neuron and each output neuron, with the number of weight parameters equal to the input dimension multiplied by the output dimension.

[0082] The role of fully connected layers is to globally combine the extracted local features, learn the high-order interactions between features, and complete the mapping from abstract features to the target spectrum. The number of fully connected layers can be adjusted according to the complexity of the task. For spectral correction tasks, two to three fully connected layers are usually sufficient to achieve good results. The output dimension of the first fully connected layer is typically set to 256 or 128, the output dimension of the second fully connected layer is set to 128 or 64, and the output dimension of the last fully connected layer is equal to the number of bands B of the target soil spectrum.

[0083] Dropout layers are introduced between fully connected layers to suppress overfitting.

[0084] The final layer of the network is the output layer, which outputs the target soil spectrum, i.e., the pure soil spectrum after correction to remove or suppress interference from photosynthetic and non-photosynthetic vegetation. The target soil spectrum has the same number of bands B as the original image spectrum, and the output value of each band represents the predicted value of the soil reflectance in that band.

[0085] The output layer employs a fully connected structure, mapping the output of the last fully connected layer to a B-dimensional vector. The activation function of the output layer is the sigmoid function, constraining the predicted values ​​within the interval [0,1].

[0086] The Adam optimizer is employed to minimize the mean square error (MSE) between the original image spectrum and the target soil spectrum. A cosine annealing strategy is used to periodically adjust the learning rate to prevent the spectral correction model from getting trapped in local optima. GPU parallel computing is used to improve computational efficiency, and an early stopping strategy is introduced to prevent model overfitting.

[0087] , In the formula This is the current iteration learning rate; Minimum learning rate; This is the maximum learning rate (initial learning rate). This represents the number of iterations already completed in the current round. For the first Total number of steps in an annealing cycle.

[0088] The initial learning rate was set to 0.0001, and the early stopping strategy was set to patiently pause for 50 rounds. After training, the spectral correction model was used to process the remote sensing imagery to obtain spectrally corrected spectral images. New image files were created using drivers from the GDAL (Geospatial Data Abstraction Library), and the geographic transformation and projection information of the images were set to ensure the usability of the fused imagery in geospatial context.

[0089] See Figure 2 As shown, Figure 2 In the image (a), the spectral curve of the original image is shown. Figure 2 As shown in the spectral curve of the corrected spectral image in (b), the corrected spectral image effectively suppresses spectral coupling interference from non-soil components in the mixed pixels. See also Figure 3As shown in the diagram, on the horizontal axis, B2 refers to the blue band of satellite data, B3 to the green band, B4 to the red band, B5 to the far-infrared band, B6 to the first shortwave infrared band (1.57-1.65 µm), and B7 to the second shortwave infrared band (2.11-2.29 µm). To verify the effectiveness of the remote sensing image spectral correction method in this application, mean square error (MSE) and spectral angle matching (SAM) are introduced as verification indicators to quantitatively evaluate the spectral correction effect. MSE measures the numerical deviation between two spectral vectors; a smaller value indicates a smaller error. Spectral angle matching measures the similarity between two spectral vectors; a smaller angle indicates lower spectral distortion.

[0090] The results show that, compared with the original image spectrum, the reflectance of the corrected spectrum is improved by eliminating the spectral coupling interference between photosynthetic vegetation (PV) and non-photosynthetic vegetation (NPV), thus enhancing the expression of soil spectral characteristics. Furthermore, the spectral correction model performs well across the entire image band. The low mean square error and spectral angle matching values ​​indicate that the corrected spectral image closely matches the target soil spectrum in both numerical values ​​and spectral shape. This verifies the good reliability of the method proposed in this application.

[0091] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for spectral correction of remote sensing images based on hybrid pixel decomposition, characterized in that, include: Acquire remote sensing image data as well as endmember spectral data of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil; The remote sensing image data is decomposed into mixed pixels using the endmember spectral data to obtain the spectral components of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil in each mixed pixel, including: Multiple candidate mixed linear spectral models are constructed for each mixed pixel. Each candidate mixed linear spectral model corresponds to a different combination of endmembers. The candidate mixed linear spectral model represents the spectral reflectance of the mixed pixel as a linear combination of the endmember spectra weighted according to their corresponding abundance ratios. The abundance of each endmember in each candidate mixed linear spectral model is solved using the fully constrained least squares method. Based on the abundance of each endmember obtained from the solution, the pixel spectra of each candidate mixed linear spectral model are reconstructed to obtain the reconstructed spectra. The root mean square error and spectral angular distance between the reconstructed spectra and the original pixel spectra are calculated respectively. A dual-constraint error evaluation function is constructed using the root mean square error and spectral angular distance; the root mean square error and spectral angular distance are normalized respectively and then weighted and summed to obtain the comprehensive error evaluation value of each candidate mixed linear spectral model; Compare the comprehensive error evaluation values ​​of each candidate mixed linear spectral model, and select the candidate mixed linear spectral model with the smallest error evaluation value as the best decomposition result of the mixed pixels; The spectral components are used as prior information and together with the original image spectrum of the remote sensing image data to construct multi-source input features. The multi-source input features are input into a pre-constructed spectral correction model, and the spectral correction model is trained to learn the nonlinear mapping relationship between the original image spectrum of the remote sensing image data and the target soil spectrum. The trained spectral correction model is used to process the remote sensing image to be corrected, resulting in a corrected spectral image, which is then associated with georeferenced information.

2. The remote sensing image spectral correction method based on hybrid pixel decomposition according to claim 1, characterized in that, Calculate the root mean square error and spectral angular distance between the reconstructed spectrum and the original pixel spectrum, including: ， ， In the formula For original pixels Band reflectivity; To reconstruct pixels Band reflectivity; For the number of bands, The root mean square error, Spectral angular distance for The transpose of .

3. The remote sensing image spectral correction method based on hybrid pixel decomposition according to claim 2, characterized in that, The candidate mixed linear spectral model is configured such that, for each band of each mixed pixel, the reflectance of each endmember in that band is multiplied by its corresponding abundance and summed, and then the residual error is added to reconstruct the reflectance of the mixed pixel in that band.

4. The remote sensing image spectral correction method based on hybrid pixel decomposition according to claim 1, characterized in that, The abundance of each endmember in each candidate mixed linear spectral model is solved using the fully constrained least squares method, with non-negative abundance and an abundance sum of 1 as physical constraints.

5. The remote sensing image spectral correction method based on hybrid pixel decomposition according to claim 4, characterized in that, The abundance of each endmember in each candidate mixed linear spectral model is solved using the fully constrained least squares method, including: The product of the reflectance and abundance of the endmember spectrum is used to approximate the pixel spectrum, and a function is established with the objective function of minimizing the sum of squared approximation errors. Physical constraints are applied: the nonnegativity constraint requires that the abundance of each endmember be greater than or equal to zero, and the abundance sum of 1 constraint requires that the sum of the abundance of all endmembers be equal to 1. The objective function is solved using a constrained optimization algorithm to obtain the optimal abundance that satisfies all constraints.

6. The remote sensing image spectral correction method based on hybrid pixel decomposition according to claim 1, characterized in that, The reconstructed spectra are obtained by reconstructing the pixel spectra of each candidate mixed linear spectral model, including: The product of the endmember spectrum of photosynthetic vegetation and its abundance is taken as the spectral contribution of photosynthetic vegetation components. The product of the endmember spectra of non-photosynthetic vegetation and its abundance is taken as the spectral contribution of non-photosynthetic vegetation components. The product of the end-member spectrum of bare soil and its abundance is taken as the spectral contribution of bare soil components. The spectral contributions of the photosynthetic vegetation component, non-photosynthetic vegetation component, and bare soil component are accumulated band by band to obtain the reconstructed spectrum.

7. The remote sensing image spectral correction method based on hybrid pixel decomposition according to claim 1, characterized in that, The spectral correction model employs a convolutional neural network, comprising three parallel feature extraction branches that process the spectral components of photosynthetic vegetation, non-photosynthetic vegetation, and the original image spectrum, respectively. Each branch contains at least one convolutional layer, one batch normalization layer, and one pooling layer in sequence. The features extracted from the three branches are concatenated and fused, and then input into multiple fully connected layers for nonlinear transformation. Dropout layers are introduced between fully connected layers to suppress overfitting; The output layer uses the Sigmoid activation function to constrain the predicted values ​​to the interval [0,1].

Images