A laboratory spectral reference constrained satellite hyperspectral pixel-by-pixel correction and soil carbon inversion method and system

Abstract

The application discloses a laboratory spectral reference constrained satellite hyperspectral per-pixel correction and soil carbon inversion method, and comprises the following steps: acquiring geographical coordinates of soil sampling points, measured values of soil carbon, laboratory spectral reflectance and original satellite hyperspectral images; converting the laboratory spectral reflectance into a laboratory equivalent wide-band spectral matrix; constructing a satellite hyperspectral matrix of the sampling points according to the geographical coordinates of the soil sampling points; taking the satellite hyperspectral matrix of the sampling points as input and taking the laboratory equivalent wide-band spectral matrix as output, establishing a per-band nonlinear spectral correction model, and applying a target band of the model to each to-be-inverted pixel in the original satellite hyperspectral image to generate a corrected satellite hyperspectral image per pixel; and based on the corrected satellite hyperspectral image and the measured values of the soil carbon, establishing a soil carbon inversion model to obtain a spatial distribution result of soil carbon in a research area. Based on the method, the soil carbon inversion precision, spatial continuity and spatial reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of soil remote sensing monitoring technology, specifically to a method and system for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral reference. Background Technology

[0002] Soil carbon is a crucial indicator of carbon storage, soil health, and long-term carbon sustainability in farmland ecosystems. Traditional soil carbon monitoring relies primarily on field sampling and laboratory physicochemical analysis, which struggles to meet the demands for large-scale, continuous, and rapid mapping. Satellite hyperspectral remote sensing provides continuous spatial coverage and rich spectral information, offering a vital data source for regional-scale soil carbon retrieval. However, directly utilizing raw satellite hyperspectral imagery for retrieval is susceptible to influences from atmospheric scattering and absorption, sensor response differences, surface moisture, and vegetation residue, leading to systematic deviations between pixel reflectance and the true soil spectrum. Conventional spectral preprocessing, feature band selection, and machine learning modeling can improve the relationship between model input and prediction to some extent, but they typically operate at the sample point or local feature level, failing to uniformly correct for pixel-by-pixel systematic biases across the entire raw satellite hyperspectral image. Laboratory visible-near-infrared spectroscopy, acquired under controlled illumination and observation conditions, can stably characterize absorption features caused by soil components, exhibiting high physical fidelity. However, laboratory spectra typically correspond to only a limited number of sampling points, lacking spatial continuity and unable to directly generate regional-scale soil carbon distribution results. Summary of the Invention

[0003] To address the problems of large spectral bias, obvious spatial mosaicking marks, spatial clustering of inversion residuals, and insufficient prediction stability when directly using satellite hyperspectral imagery for soil carbon inversion, this invention provides a method and system for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral references. This method combines the physical reference role of laboratory spectra with the spatial continuity of satellite imagery, enabling laboratory equivalent broadband spectra to constrain band-by-band and pixel-by-pixel correction of the original satellite hyperspectral imagery. This provides corrected satellite hyperspectral imagery for soil carbon inversion, reducing systematic spectral biases caused by atmospheric conditions, sensor response, surface conditions, and image mosaicking differences, thereby improving the accuracy, spatial continuity, and spatial reliability of soil carbon inversion.

[0004] According to one aspect of the present invention, a method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral references is provided, comprising: acquiring the geographic coordinates, measured soil carbon values, and laboratory spectral reflectance of multiple soil sampling points within a study area; acquiring satellite hyperspectral images covering the study area and preprocessing the satellite hyperspectral images to obtain raw satellite hyperspectral images; performing narrow-band to wide-band conversion on the laboratory spectral reflectance based on the center wavelength, bandwidth, and spectral response function of the satellite hyperspectral sensor to obtain a laboratory equivalent wide-band spectral matrix corresponding to the effective band of the satellite; and extracting the corresponding values ​​of the soil sampling points from the raw satellite hyperspectral images based on the geographic coordinates of the soil sampling points. Based on the satellite hyperspectral reflectance of the pixels to be inverted, a satellite hyperspectral matrix of sampling points is constructed. Using the satellite hyperspectral matrix of sampling points as input and the laboratory equivalent wideband spectral matrix as the output target, a nonlinear mapping relationship is established for each target band to form a band-by-band nonlinear spectral correction model. Each target band is modeled by all effective satellite bands at the soil sampling point. The band-by-band nonlinear spectral correction model is applied to each pixel to be inverted in the original satellite hyperspectral image according to the target band, generating a corrected satellite hyperspectral image pixel by pixel. Based on the corrected satellite hyperspectral image and the measured soil carbon values, a soil carbon inversion model is established to obtain the spatial distribution results of soil carbon in the study area.

[0005] Furthermore, the construction of the laboratory equivalent wideband spectral matrix includes: performing normalized convolution on the laboratory spectral reflectance and spectral response function within the effective band of each satellite to obtain the laboratory equivalent wideband spectrum corresponding to the effective band of each satellite; and constructing the laboratory equivalent wideband spectral matrix based on the laboratory equivalent wideband spectra corresponding to the effective bands of each satellite.

[0006] Furthermore, the acquisition of the spectral response function includes: when the satellite hyperspectral sensor provides a measured spectral response function, using the measured spectral response function; when the satellite hyperspectral sensor does not provide a measured spectral response function, constructing a Gaussian approximate response function based on the center wavelength and bandwidth of the satellite hyperspectral sensor.

[0007] Furthermore, the expression corresponding to the band-by-band nonlinear spectral correction model is:

[0008] ,

[0009] In the formula, L is the laboratory equivalent wideband spectral matrix, X is the satellite hyperspectral matrix of the sampling point, n is the number of soil sampling points, G is the spectral correction nonlinear mapping function, and p is the number of effective satellite bands.

[0010] Furthermore, the spectral correction nonlinear mapping function is one or more of the following: random forest regression, gradient boosting tree regression, support vector regression, neural network regression, and equivalent regression models capable of establishing nonlinear mapping relationships.

[0011] Furthermore, the extraction of satellite hyperspectral reflectance includes: extracting the satellite hyperspectral reflectance of the pixel to be inverted where the soil sampling point is located in the original satellite hyperspectral image, and using it as the satellite hyperspectral reflectance of the soil sampling point; or extracting the satellite hyperspectral reflectance of all pixels to be inverted within a neighborhood window centered on the soil sampling point in the original satellite hyperspectral image, calculating the mean of the satellite hyperspectral reflectance of all pixels to be inverted, and using it as the satellite hyperspectral reflectance of the soil sampling point.

[0012] Furthermore, the wavelength range of the laboratory spectral reflectance is 400-2500 nm; the satellite hyperspectral image includes multiple effective satellite bands in the 400-2500 nm range, and the multiple effective satellite bands correspond one-to-one with the laboratory equivalent wideband spectrum.

[0013] Furthermore, the soil sampling points are divided into model training points and independent validation points. The model training points are used for training the band-by-band nonlinear spectral correction model and the soil carbon inversion model, while the independent validation points are used to evaluate the generalization ability of the band-by-band nonlinear spectral correction model and the soil carbon inversion model.

[0014] Furthermore, the preprocessing includes radiometric calibration, atmospheric correction, geometric correction, image mosaicking, masking, and effective band selection. The masking includes removing pixels with clouds, shadows, water bodies, strong vegetation cover, and anomalous reflectivity, while retaining pixels with bare soil or low vegetation cover.

[0015] According to one aspect of the present invention, a satellite hyperspectral pixel-by-pixel correction and soil carbon inversion system constrained by laboratory spectral reference is provided, comprising: a soil sampling point data acquisition module for acquiring the geographic coordinates, measured soil carbon values, and laboratory spectral reflectance of multiple soil sampling points within the study area; a satellite hyperspectral image acquisition module for acquiring satellite hyperspectral images covering the study area and preprocessing the satellite hyperspectral images to obtain raw satellite hyperspectral images; a spectral response matching module for performing narrow-band to wide-band conversion on the laboratory spectral reflectance based on the center wavelength, bandwidth, and spectral response function of the satellite hyperspectral sensor to obtain a laboratory equivalent wide-band spectral matrix corresponding to the effective band of the satellite; and a sample-point-level spectral matching data construction module for constructing data from the original satellite hyperspectral images based on the geographic coordinates of the soil sampling points. The system extracts the satellite hyperspectral reflectance of the corresponding pixels to be inverted and constructs a satellite hyperspectral matrix for the sampling points. A band-by-band nonlinear spectral correction model establishment module is used to establish a nonlinear mapping relationship for each target band, taking the satellite hyperspectral matrix of the sampling points as input and the laboratory equivalent wideband spectral matrix as the output target, to form a band-by-band nonlinear spectral correction model. Each target band is modeled by all effective satellite bands at the soil sampling points. A corrected satellite hyperspectral image generation module applies the band-by-band nonlinear spectral correction model to each pixel to be inverted in the original satellite hyperspectral image according to the target band, generating a corrected satellite hyperspectral image pixel by pixel. A soil carbon inversion and spatial evaluation module is used to establish a soil carbon inversion model based on the corrected satellite hyperspectral image and the measured soil carbon values ​​to obtain the spatial distribution results of soil carbon in the study area.

[0016] The above technical solution addresses the problem of insufficient accuracy and spatial reliability in soil carbon inversion due to systematic discrepancies between raw satellite hyperspectral imagery and laboratory soil spectra. It discloses a pixel-by-pixel correction method for satellite hyperspectral imagery and soil carbon inversion constrained by laboratory spectral references, belonging to the field of soil remote sensing monitoring and quantitative hyperspectral inversion technology. This method acquires the geographic coordinates of soil sampling points, measured soil carbon values, laboratory visible-near-infrared spectra, and satellite hyperspectral imagery covering the study area. Based on the spectral response function of the satellite hyperspectral sensor, the narrow-band laboratory spectrum is converted into a laboratory equivalent wide-band spectrum corresponding to the satellite bands. The original satellite hyperspectral reflectance is extracted at the soil sampling point locations. Using all effective satellite hyperspectral bands as input and the laboratory equivalent wide-band reflectance of each target band as output, a band-by-band nonlinear spectral correction model is established. This model is applied pixel-by-pixel to the entire raw satellite hyperspectral imagery to generate a corrected satellite hyperspectral image. A soil carbon inversion model is established based on the corrected imagery, outputting the spatial distribution results of soil carbon. This invention can reduce systematic spectral biases caused by differences in atmosphere, sensor response, surface conditions, and image mosaic, thereby improving the accuracy, spatial continuity, and spatial reliability of soil carbon retrieval.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] (1) The present invention first uses the spectral response function of satellite sensor to match the response of laboratory high-resolution spectrum, so that the laboratory spectrum and satellite hyperspectral image are at the same band scale, reducing data inconsistency caused by differences in band center and bandwidth.

[0019] (2) The present invention uses the laboratory equivalent wideband spectrum as a physical reference to establish a band-by-band nonlinear spectral correction model, and applies the model pixel by pixel to the entire original satellite hyperspectral image, which can reduce the systematic spectral deviation caused by atmospheric conditions, sensor response and image mosaic differences.

[0020] (3) The corrected satellite hyperspectral image generated by the present invention retains both the spatial continuity of the satellite image and the physical fidelity of the laboratory spectrum, providing more reliable input data for soil carbon inversion. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1A flowchart illustrating the overall method of a laboratory spectral reference-constrained satellite hyperspectral pixel-by-pixel correction and soil carbon inversion method provided in this embodiment of the invention;

[0023] Figure 2 A schematic diagram of laboratory spectral response matching and band-by-band nonlinear correction provided in an embodiment of the present invention;

[0024] Figure 3 A flowchart illustrating the use of corrected satellite hyperspectral imagery for soil carbon inversion and spatial reliability assessment, provided in this embodiment of the invention. Detailed Implementation

[0025] It should be noted that:

[0026] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0029] Please refer to Figure 1 This invention provides a method for satellite hyperspectral pixel-by-pixel correction and soil carbon retrieval constrained by laboratory spectral standards, specifically including the following steps:

[0030] Step S1: Obtain the geographic coordinates, measured soil carbon values, and laboratory spectral reflectance of multiple soil sampling points within the study area.

[0031] Step S1 aims to obtain soil sampling data for the study area. Specifically, surface soil samples from cultivated land in the study area are collected, the geographical coordinates of the sampling points are recorded, and the soil carbon content (i.e., measured soil carbon value) of each soil sample is determined. At the same time, the visible-near-infrared spectra (i.e., laboratory spectral reflectance) of the soil samples are measured under laboratory conditions to form a high-resolution laboratory spectral dataset.

[0032] Step S2: Acquire satellite hyperspectral images covering the study area and preprocess the satellite hyperspectral images to obtain raw satellite hyperspectral images.

[0033] Step S2 aims to perform satellite hyperspectral image preprocessing. Preprocessing includes: radiometric calibration, atmospheric correction, geometric correction, image mosaicking, masking, and effective satellite band selection on the satellite hyperspectral image covering the study area, resulting in a raw satellite hyperspectral image that matches the spatial location of the sampling points. The masking process includes removing pixels with clouds, shadows, water bodies, strong vegetation cover, and anomalous reflectance, while retaining pixels with bare soil or low vegetation cover for soil carbon retrieval.

[0034] Step S3: Based on the center wavelength, bandwidth, and spectral response function of the satellite hyperspectral sensor, perform narrow-band to wide-band conversion on the laboratory spectral reflectance to obtain the laboratory equivalent wide-band spectral matrix corresponding to the satellite hyperspectral band.

[0035] Step S3 aims to perform laboratory spectral response matching (see Appendix). Figure 2 Specifically, the spectral response function of the satellite hyperspectral sensor is used to match the response of the laboratory spectral reflectance, so that the laboratory spectral reflectance and the satellite hyperspectral image are at the same band scale, reducing data inconsistencies caused by differences in band center and bandwidth.

[0036] Furthermore, when the satellite hyperspectral sensor provides a measured spectral response function, the measured spectral response function should be used preferentially; when the satellite hyperspectral sensor does not provide a measured spectral response function, a Gaussian approximate response function can be constructed based on the center wavelength and bandwidth.

[0037] (1)

[0038] In the formula, For the first Spectral response functions of each satellite's effective band. The response intensity normalization coefficient, For the first The center wavelength of each satellite's effective band. For the first The bandwidth parameter that determines the bandwidth of each satellite's effective band. This is a wavelength variable. When the bandwidth of the satellite hyperspectral sensor is expressed in full width at half maximum (FWHM),... The relationship with FWHM is as follows:

[0039] (2)

[0040] Furthermore, normalized convolutions are performed on the laboratory spectral reflectance and spectral response function within the effective band of each satellite to obtain the laboratory equivalent broadband spectrum corresponding to each satellite's effective band. A laboratory equivalent broadband spectral matrix is ​​then constructed based on the laboratory equivalent broadband spectra corresponding to each satellite's effective band. The formula for calculating the laboratory equivalent broadband spectrum is as follows:

[0041] (3)

[0042] In the formula, The laboratory equivalent wideband reflectivity corresponding to the effective band of the i-th satellite. Laboratory spectral reflectance, and Let be the integral range of the effective band of the i-th satellite. When the spectral response function has been normalized, the denominator of equation (3) is 1.

[0043] Furthermore, when the spectral interval of the laboratory spectral reflectance is 1 nm, equation (3) can be calculated in a discrete form. The laboratory spectral reflectance and spectral response function within each satellite's effective band are discretely weighted and averaged to obtain the equivalent wideband spectral data of the laboratory corresponding to each satellite's effective band. The corresponding calculation formula is as follows:

[0044] (4)

[0045] In the formula, For the effective band of the i-th satellite, the th The summation range is: [Number] laboratory spectral wavelengths. to All 1 nm sampling wavelengths between. This represents the number of laboratory spectral wavelengths within the effective band of the i-th satellite.

[0046] Step S4: Based on the geographic coordinates of the soil sampling points, extract the satellite hyperspectral reflectance of the corresponding pixels to be inverted from the original satellite hyperspectral image, and construct the satellite hyperspectral matrix of the sampling points.

[0047] Step S4 aims to construct the satellite hyperspectral matrix of the sampling points. Specifically, based on the geographic coordinates of the soil sampling points, the satellite hyperspectral reflectance of the pixels to be inverted at the soil sampling points is extracted from the original satellite hyperspectral image and used as the satellite hyperspectral reflectance of the soil sampling points. Alternatively, the satellite hyperspectral reflectance of all pixels to be inverted within a neighborhood window centered on the soil sampling point is extracted, the mean of the satellite hyperspectral reflectance of all pixels to be inverted is calculated, and this mean is used as the satellite hyperspectral reflectance of the soil sampling point. This yields the satellite hyperspectral reflectance corresponding to each soil sampling point. The satellite hyperspectral reflectances corresponding to all soil sampling points together form the sampling point satellite hyperspectral matrix. The sampling point satellite hyperspectral matrix, together with the laboratory equivalent broadband spectral matrix, forms sample-level correction data, which is used to correct the original satellite hyperspectral image.

[0048] Step S5: Using the satellite hyperspectral matrix of the sampling points as input and the laboratory equivalent wideband spectral matrix as the output target, establish a nonlinear mapping relationship for each target band to form a band-by-band nonlinear spectral correction model.

[0049] Step S5 aims to establish a band-by-band nonlinear spectral correction model (please refer to the appendix). Figure 2 Each target band is modeled using all effective satellite bands at the soil sampling point, and the overall band-by-band nonlinear spectral correction model is expressed as follows:

[0050] (5)

[0051] In the formula, L is the laboratory equivalent broadband spectral matrix, X is the satellite hyperspectral matrix of the sampling points, n is the number of soil sampling points, p is the number of effective satellite bands, and G is the spectral correction nonlinear mapping function. For the j-th target band, the following mapping relationship is adopted:

[0052] (6)

[0053] In the formula, For the first A spectral correction nonlinear mapping function for each target band. Without changing the band-by-band modeling and pixel-by-pixel application method, the spectral correction nonlinear mapping function can be selected from one or more of the following: random forest regression, gradient boosting tree regression, support vector regression, neural network regression, and equivalent regression models capable of establishing nonlinear mapping relationships. Preferably, the spectral correction nonlinear mapping function uses a random forest regression model, and cross-validation is used to optimize the model hyperparameters.

[0054] Furthermore, to prevent data leakage, soil sampling points were divided into model training points and independent validation points. Both the band-by-band nonlinear spectral correction model and the soil carbon inversion model were established using only the model training points; the independent validation points did not participate in the training of the band-by-band nonlinear spectral correction model or the soil carbon inversion model, and were only used to evaluate the generalization ability of the corrected satellite hyperspectral imagery and the soil carbon inversion model.

[0055] Step S6: Apply the band-by-band nonlinear spectral correction model to each pixel to be inverted in the original satellite hyperspectral image according to the target band, and generate the corrected satellite hyperspectral image pixel by pixel.

[0056] Step S6 aims to perform pixel-by-pixel spectral correction. Specifically, the nonlinear mapping relationship of each target band is applied to each pixel to be inverted in the original satellite hyperspectral image, and the corrected reflectance of each pixel is calculated band by band to obtain the corrected satellite hyperspectral image. It should be noted that the corrected satellite hyperspectral image maintains the spatial resolution, pixel location, and band structure of the original satellite hyperspectral image, but its spectral reflectance is obtained by laboratory equivalent wideband spectral constraints, which can reduce systematic spectral biases caused by atmospheric conditions, sensor response, and image mosaic differences.

[0057] Step S7: Based on the corrected satellite hyperspectral image and the measured soil carbon values, establish a soil carbon inversion model to obtain the spatial distribution results of soil carbon in the study area.

[0058] Step S7 aims to perform soil carbon inversion and spatial assessment (please refer to the appendix for details). Figure 3 A soil carbon inversion model was established based on corrected satellite hyperspectral imagery and measured soil carbon values. The model outputs the spatial distribution of soil carbon in the study area (i.e., a soil carbon spatial distribution map), and the inversion results are evaluated through accuracy, uncertainty, and residual spatial autocorrelation analysis. The expression for the soil carbon inversion model is as follows:

[0059] (7)

[0060] In the formula: Predicted soil carbon values; This is the original satellite hyperspectral matrix (i.e., the original satellite hyperspectral image). For band-by-band spectral correction functions; ( ) is the soil carbon inversion function, which can be established using random forest regression, partial least squares regression, gradient boosting tree regression, support vector regression, or neural networks.

[0061] It should be noted that the soil carbon inversion model is not a fixed, known model. It is derived by finding the relationship between soil carbon and related environmental variables, based on the concepts of linear regression, machine learning, and other models. Understandably, measured soil carbon values ​​are sparse, while the goal of this invention is to obtain soil carbon values ​​at any location. Therefore, this invention utilizes these sparse measured soil carbon values ​​to establish a soil carbon inversion model. The spatial distribution of soil carbon output by this model can reflect the soil carbon value at any location.

[0062] The technical solution of this invention will be further illustrated below using the example of soil carbon inversion in the Songliao Plain.

[0063] The study area is located in the Songliao Plain, spanning approximately 42°49′–43°46′N and 123°45′–124°53′E, covering an area of ​​about 3232 km². This region is dominated by agro-economy ecosystems, and exhibits spatial heterogeneity in land cover, soil types, and topography, making it suitable for validating satellite hyperspectral soil carbon retrieval methods.

[0064] 154 topsoil samples (0-20 cm) were collected from the study area. The latitude and longitude coordinates of each soil sampling point were recorded, and the soil carbon content was determined. After natural air drying, grinding, sieving, and homogenization, the laboratory spectral reflectance in the 350-2500 nm range was measured using a visible-near-infrared spectrometer under laboratory conditions. In this embodiment, laboratory spectral reflectance in the 400-2500 nm range was selected for subsequent processing. GF5 satellite hyperspectral imagery covering the study area was acquired. GF5 satellite hyperspectral imagery includes multiple effective satellite bands in the 400-2500 nm range. GF5 satellite hyperspectral imagery with low cloud cover, stable imaging quality, and in the bare soil or low vegetation cover stage was selected to reduce the influence of cloud shadows, vegetation residue, and crop canopy on the soil spectrum. The GF5 satellite hyperspectral imagery has visible-near-infrared and short-wave infrared bands, with a spatial resolution of 30 m.

[0065] Radiometric calibration, atmospheric correction, geometric correction, and image mosaicking were performed on GF5 satellite hyperspectral imagery. Masking was then applied based on quality labels, spectral thresholds, or auxiliary land cover data to remove pixels representing clouds, shadows, water bodies, strong vegetation cover, and anomalous reflectance, retaining only bare soil or low-vegetation-cover pixels. Further, low signal-to-noise ratio bands, bands with strong water vapor absorption, edge noise bands, and anomalous bands were removed, resulting in 237 effective satellite bands in the 400-2500 nm range, forming the original satellite hyperspectral imagery. Based on the center wavelength, bandwidth, and spectral response function of each effective band from the GF5 sensor, normalized convolutions were performed on high-resolution soil spectra at 1 nm intervals in the laboratory. If the GF5 sensor provides a measured spectral response function, it is preferentially used; otherwise, a Gaussian approximate response function is constructed based on the center wavelength and bandwidth to obtain laboratory equivalent wideband spectra corresponding one-to-one with the effective bands of the GF5 satellite.

[0066] Based on the geographic coordinates of the soil sampling points, the satellite hyperspectral reflectance of the pixels containing the soil sampling points is extracted from the original satellite hyperspectral imagery. To reduce positioning errors and pixel noise, the mean spectral reflectance of a 3×3 neighborhood window centered on the soil sampling point can also be extracted. The extracted satellite hyperspectral reflectance, together with the laboratory equivalent broadband spectrum, forms the sample-level correction data. During modeling, the satellite hyperspectral matrix X of the sampling point is used as input, and the laboratory equivalent broadband spectral matrix L is used as the output target. A nonlinear mapping relationship is established for each target band; each target band is modeled by all effective GF5 satellite bands at the corresponding soil sampling location.

[0067] In this embodiment, a random forest regression model is preferably used for the nonlinear mapping relationship. During training, hyperparameters such as the number of trees, maximum tree depth, minimum number of leaf node samples, and maximum number of features can be optimized through ten-fold cross-validation or grid search. In other embodiments, gradient boosting tree regression, support vector regression, neural network regression, or other equivalent regression models capable of establishing nonlinear mapping relationships can also be used to establish a band-by-band nonlinear spectral correction model. Using the trained band-by-band nonlinear spectral correction model, band-by-band calculations are performed on each pixel to be inverted in the original satellite hyperspectral image to obtain the corrected satellite hyperspectral image. A soil carbon inversion model is established based on the corrected satellite hyperspectral image and measured soil carbon values. During modeling, sample points can be divided into a model training set and a model validation set. Partial least squares regression, random forest regression, or gradient boosting decision tree models are used for training, and the accuracy of the soil carbon inversion model is evaluated using the coefficient of determination R², root mean square error RMSE, and consistency correlation coefficient CCC.

[0068] In this embodiment, the soil carbon inversion results based on the corrected satellite hyperspectral image and the random forest model are R²=0.52, RMSE=2.64 g / kg, and CCC=0.72. Compared with the original satellite hyperspectral image, R² increased by 142.9%, RMSE decreased by 0.9 g / kg, and CCC increased by 118.2% (see Table 1 for reference). Furthermore, the spatial continuity of the inversion results is enhanced, and the spatial autocorrelation of the residuals is reduced. Simultaneously, the splicing boundary effect of the soil carbon spatial distribution map generated from the corrected satellite hyperspectral image is weakened, and spatial continuity is enhanced.

[0069] Table 1. Accuracy Evaluation of Inversion Model

[0070]

[0071] Furthermore, soil carbon inversion uncertainty is calculated using a bootstrap sampling method. Preferably, the bootstrap sampling is repeated 100 times, with the soil carbon inversion model retrained and prediction results generated each time. The standard deviation of the 100 prediction results is used as the pixel-scale uncertainty to form a soil carbon inversion uncertainty map. The spatial autocorrelation of the residual Moran's α index is used to evaluate the residuals of soil carbon inversion. Spatial weights can be constructed using distance threshold weights, k-nearest neighbor weights, or adjacency weights; in this embodiment, a spatial weight matrix is ​​preferably constructed using a distance threshold between sampling points or k-nearest neighbor relationships, and Moran's α index is calculated for the residuals of the model validation sampling points to evaluate whether spatial clustering of the residuals exists.

[0072] It should be noted that the above embodiments use GF5 satellite hyperspectral imagery as an example, but the present invention is not limited to GF5 satellite; for other satellites or airborne hyperspectral sensors with a defined center wavelength, bandwidth, and spectral response function, laboratory spectral response matching, pixel-by-pixel spectral correction, and soil carbon inversion can all be performed according to the method described in this invention. This method can also be extended to the inversion of soil carbon components such as soil organic carbon and inorganic carbon, but the main protected object of this invention is soil carbon inversion.

[0073] Based on the same technical concept as the aforementioned embodiments, this invention also provides a satellite hyperspectral pixel-by-pixel correction and soil carbon inversion system constrained by laboratory spectral reference, comprising: a soil sampling point data acquisition module for acquiring the geographic coordinates, measured soil carbon values, and laboratory spectral reflectance of multiple soil sampling points within the study area; a satellite hyperspectral image acquisition module for acquiring satellite hyperspectral images covering the study area and preprocessing the satellite hyperspectral images to obtain raw satellite hyperspectral images; a spectral response matching module for performing narrow-band to wide-band conversion on the laboratory spectral reflectance based on the center wavelength, bandwidth, and spectral response function of the satellite hyperspectral sensor to obtain a laboratory equivalent wide-band spectral matrix corresponding to the effective band of the satellite; and a sample-point-level spectral matching data construction module for constructing data based on the geographic coordinates of the soil sampling points and the raw satellite hyperspectral image. The system extracts the satellite hyperspectral reflectance of the corresponding pixels to be inverted from the spectral image and constructs a satellite hyperspectral matrix for the sampling points. A band-by-band nonlinear spectral correction model establishment module is used to establish a nonlinear mapping relationship for each target band, taking the satellite hyperspectral matrix of the sampling points as input and the laboratory equivalent wideband spectral matrix as the output target, to form a band-by-band nonlinear spectral correction model. Each target band is modeled by all effective satellite bands at the soil sampling points. A corrected satellite hyperspectral image generation module applies the band-by-band nonlinear spectral correction model to each pixel to be inverted in the original satellite hyperspectral image according to the target band, generating a corrected satellite hyperspectral image pixel by pixel. A soil carbon inversion and spatial evaluation module is used to establish a soil carbon inversion model based on the corrected satellite hyperspectral image and the measured soil carbon values ​​to obtain the spatial distribution results of soil carbon in the study area.

[0074] In summary, this invention addresses the problem of insufficient accuracy and spatial reliability in soil carbon inversion due to systematic discrepancies between raw satellite hyperspectral imagery and laboratory soil spectra. It discloses a pixel-by-pixel correction method for satellite hyperspectral imagery and soil carbon inversion constrained by laboratory spectral references, belonging to the field of soil remote sensing monitoring and quantitative hyperspectral inversion technology. This method acquires the geographic coordinates of soil sampling points, measured soil carbon values, laboratory visible-near-infrared spectra, and satellite hyperspectral imagery covering the study area. Based on the spectral response function of the satellite hyperspectral sensor, the narrow-band laboratory spectrum is converted into a laboratory equivalent wide-band spectrum corresponding to the satellite bands. The original satellite hyperspectral reflectance is extracted at the sampling point locations. Using all effective satellite hyperspectral bands as input and the laboratory equivalent wide-band reflectance of each target band as output, a band-by-band nonlinear spectral correction model is established. This model is applied pixel-by-pixel to the entire raw satellite hyperspectral imagery to generate a corrected satellite hyperspectral image. A soil carbon inversion model is established based on the corrected imagery, outputting the spatial distribution results of soil carbon. This invention can reduce systematic spectral biases caused by differences in atmosphere, sensor response, surface conditions, and image mosaic, thereby improving the accuracy, spatial continuity, and spatial reliability of soil carbon retrieval.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral standards, characterized in that, include: The geographic coordinates, measured soil carbon values, and laboratory spectral reflectance of multiple soil sampling points within the study area were obtained. Acquire satellite hyperspectral images covering the study area and preprocess the satellite hyperspectral images to obtain raw satellite hyperspectral images; Based on the center wavelength, bandwidth and spectral response function of the satellite hyperspectral sensor, the laboratory spectral reflectance is converted from narrow band to wide band to obtain the laboratory equivalent wide band spectral matrix corresponding to the effective band of the satellite. Based on the geographic coordinates of the soil sampling points, the satellite hyperspectral reflectance of the corresponding pixels to be inverted is extracted from the original satellite hyperspectral image, and a satellite hyperspectral matrix of the sampling points is constructed. Using the satellite hyperspectral matrix of the sampling points as input and the laboratory equivalent wideband spectral matrix as output target, a nonlinear mapping relationship is established for each target band to form a band-by-band nonlinear spectral correction model; wherein, each target band is jointly modeled by all effective satellite bands at the soil sampling points; The band-by-band nonlinear spectral correction model is applied to each pixel to be inverted in the original satellite hyperspectral image according to the target band, and the corrected satellite hyperspectral image is generated pixel by pixel. A soil carbon inversion model was established based on the corrected satellite hyperspectral imagery and the measured soil carbon values ​​to obtain the spatial distribution results of soil carbon in the study area.

2. The method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral reference as described in claim 1, characterized in that, The construction of the laboratory equivalent broadband spectral matrix includes: Normalized convolutions are performed on the spectral reflectance and spectral response function of each laboratory within the effective band of each satellite to obtain the equivalent wideband spectrum of the laboratory corresponding to the effective band of each satellite. A laboratory equivalent wideband spectral matrix is ​​constructed based on the laboratory equivalent wideband spectra corresponding to the effective bands of each satellite.

3. The method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral reference as described in claim 2, characterized in that, The acquisition of the spectral response function includes: When the satellite hyperspectral sensor provides a measured spectral response function, the measured spectral response function is used; When the satellite hyperspectral sensor does not provide a measured spectral response function, a Gaussian approximate response function is constructed based on the center wavelength and bandwidth of the satellite hyperspectral sensor.

4. The method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral reference as described in claim 1, characterized in that, The expression corresponding to the band-by-band nonlinear spectral correction model is: ， In the formula, L is the laboratory equivalent wideband spectral matrix, X is the satellite hyperspectral matrix of the sampling point, n is the number of soil sampling points, G is the spectral correction nonlinear mapping function, and p is the number of effective satellite bands.

5. The method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral reference as described in claim 4, characterized in that, The spectral correction nonlinear mapping function is one or more of the following: random forest regression, gradient boosting tree regression, support vector regression, neural network regression, and equivalent regression models that can establish nonlinear mapping relationships.

6. The method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral reference as described in claim 1, wherein the extraction of satellite hyperspectral reflectance includes: The satellite hyperspectral reflectance of the pixel to be inverted where the soil sampling point is located is extracted from the original satellite hyperspectral image, and it is used as the satellite hyperspectral reflectance of the soil sampling point. Alternatively, extract the satellite hyperspectral reflectance of all pixels to be inverted within a neighborhood window centered on the soil sampling point from the original satellite hyperspectral image, calculate the mean of the satellite hyperspectral reflectance of all pixels to be inverted, and use it as the satellite hyperspectral reflectance of the soil sampling point.

7. The method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral reference as described in claim 1, characterized in that, The wavelength range of the laboratory spectral reflectance is 400-2500 nm; the satellite hyperspectral image includes multiple effective satellite bands in the 400-2500 nm range, and the multiple effective satellite bands correspond one-to-one with the laboratory equivalent wideband spectrum.

8. The method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral reference as described in claim 1, characterized in that, The soil sampling points are divided into model training points and independent validation points. The model training points are used for training the band-by-band nonlinear spectral correction model and the soil carbon inversion model, while the independent validation points are used to evaluate the generalization ability of the band-by-band nonlinear spectral correction model and the soil carbon inversion model.

9. The method for satellite hyperspectral pixel-by-pixel correction and soil carbon inversion constrained by laboratory spectral reference as described in claim 1, characterized in that, The preprocessing includes radiometric calibration, atmospheric correction, geometric correction, image mosaicking, masking, and effective band selection. The masking includes removing pixels with clouds, shadows, water bodies, strong vegetation cover, and anomalous reflectivity, while retaining pixels with bare soil or low vegetation cover.

10. A satellite hyperspectral pixel-by-pixel correction and soil carbon retrieval system constrained by laboratory spectral standards, characterized in that, include: The soil sampling data acquisition module for the study area is used to acquire the geographic coordinates, measured soil carbon values, and laboratory spectral reflectance of multiple soil sampling points within the study area. The satellite hyperspectral image acquisition module is used to acquire satellite hyperspectral images covering the study area and preprocess the satellite hyperspectral images to obtain raw satellite hyperspectral images. The spectral response matching module is used to perform narrow-band to wide-band conversion on the laboratory spectral reflectance based on the center wavelength, bandwidth and spectral response function of the satellite hyperspectral sensor, so as to obtain the laboratory equivalent wide-band spectral matrix corresponding to the effective band of the satellite. The sample point-level spectral matching data construction module is used to extract the satellite hyperspectral reflectance of the corresponding pixel to be inverted from the original satellite hyperspectral image based on the geographic coordinates of the soil sampling point, and construct the satellite hyperspectral matrix of the sampling point. The band-by-band nonlinear spectral correction model establishment module is used to establish a nonlinear mapping relationship for each target band by taking the satellite hyperspectral matrix of the sampling point as input and the laboratory equivalent wideband spectral matrix as output target, thereby forming a band-by-band nonlinear spectral correction model; wherein, each target band is jointly modeled by all effective satellite bands at the soil sampling point; The corrected satellite hyperspectral image generation module is used to apply the band-by-band nonlinear spectral correction model to each pixel to be inverted in the original satellite hyperspectral image according to the target band, and generate the corrected satellite hyperspectral image pixel by pixel. The soil carbon inversion and spatial evaluation module is used to establish a soil carbon inversion model based on the corrected satellite hyperspectral image and the measured soil carbon values, and to obtain the spatial distribution results of soil carbon in the study area.

Images