Soil organic carbon detection hyperspectral band optimization method based on deep learning

Abstract

The invention relates to the technical field of environment detection, and discloses a soil organic carbon detection hyperspectral band optimization method based on deep learning, which comprises the following steps: filtering and denoising original hyperspectral data by adopting an improved filtering mode; calculating the determination importance of the hyperspectral band to the determined organic carbon content, and dynamically selecting the hyperspectral band with the highest information gain as a candidate hyperspectral band; calculating correlation weights between the candidate hyperspectral data and the soil environment factors by adopting a canonical correlation analysis mode; and performing multi-factor fusion optimal hyperspectral band prediction by using the optimal hyperspectral band prediction model, and detecting the organic carbon content of the soil based on an optimal hyperspectral band prediction result. According to the method, the content of the organic carbon in the soil is efficiently predicted by collecting the hyperspectral data and the measured content of the organic carbon, screening the candidate hyperspectral wave band with high information gain, constructing the hyperspectral wave band prediction model and screening the optimal hyperspectral wave band in combination with environmental factors.

Description

Technical Field

[0001] The present invention relates to the field of environmental detection technology, and specifically to a hyperspectral band optimization method for soil organic carbon detection based on deep learning. Background Art

[0002] Soil organic carbon (SOC) is a key indicator of soil health, soil fertility, and the global carbon cycle. Soil organic carbon not only affects soil structure and fertility but is also directly linked to climate change, particularly greenhouse gas emissions and carbon storage. Effectively monitoring and accurately assessing the distribution and content of SOC is crucial for global climate change mitigation, land management, and sustainable agricultural development.

[0003] In recent years, the continuous development of spectral technology, particularly hyperspectral technology, has gradually gained recognition in soil testing. By measuring the spectral reflectance of surfaces with high precision and multiple wavelengths, hyperspectral technology can provide richer information about surface features, including the various components of the soil. Consequently, the quantitative prediction of soil organic carbon using hyperspectral data has become a popular research topic.

[0004] In existing research, patent publication CN113971989A discloses a hyperspectral modeling method for forest soil organic carbon content based on OPLS, which belongs to the field of soil spectral acquisition and analysis technology. The method includes the following technical steps: (1) soil sample collection and processing; (2) soil organic carbon determination; (3) spectral reflectance data determination; (4) spectral reflectance data preprocessing; (5) spectral reflectance data transformation; (6) initial model establishment and verification; and (7) hyperspectral prediction model establishment.

[0005] Although hyperspectral remote sensing technology has made significant progress in soil organic carbon monitoring, current research still faces challenges: First, hyperspectral remote sensing technology provides hundreds of bands, but different bands respond differently to soil organic carbon. How to screen out the optimal bands that are highly correlated with soil organic carbon content from a large number of spectral bands remains an urgent problem to be solved. In addition, the distribution of soil organic carbon is not only closely related to soil type, but is also affected by environmental factors such as climate, topography, and vegetation. Existing soil organic carbon prediction models are mostly based on a single factor (such as spectral data), but ignore the interaction of multiple factors such as soil type and geographical environment, resulting in limited generalization ability and prediction accuracy of the model. Summary of the Invention

[0006] In view of this, the present invention provides a hyperspectral band optimization method for soil organic carbon detection based on deep learning. The denoising ability of hyperspectral data is enhanced by improved Savitzky-Golay filtering, the optimal candidate bands are screened in combination with band importance, environmental factors such as soil type, vegetation index and terrain moisture index are introduced, and feature fusion modeling is performed using canonical correlation analysis. The optimal hyperspectral band is selected, and the calculation process of soil organic carbon prediction is simplified, which has better practical application value.

[0007] To achieve the above objectives, the present invention provides a method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning, comprising the following steps:

[0008] S1: Collect the original hyperspectral data of soil sampling points and the organic carbon content, and use the improved Savitzky-Golay filtering method to filter and denoise the original hyperspectral data to obtain filtered hyperspectral data;

[0009] S2: Calculate the importance of each hyperspectral band in the filtered hyperspectral data for the determination of organic carbon content. Based on the importance of the determination, dynamically select the hyperspectral band with the highest information gain as the candidate hyperspectral band to form a candidate hyperspectral band set.

[0010] S3: Collect candidate hyperspectral data of soil environmental factors and soil sampling points under the candidate hyperspectral band set, calculate the association weights between the candidate hyperspectral data and the soil environmental factors using canonical correlation analysis, and fuse the candidate hyperspectral data and soil environmental factors based on the association weights to obtain a fusion feature matrix;

[0011] S4: Construct an optimal hyperspectral band prediction model, use the optimal hyperspectral band prediction model to perform optimal hyperspectral band prediction with multi-factor fusion, and detect soil organic carbon content based on the optimal hyperspectral band prediction results.

[0012] As a further improvement method of the present invention:

[0013] Optionally, collect raw hyperspectral data from soil sampling points and determine organic carbon content, including:

[0014] Select N soil sampling points from the soil to be tested for organic carbon, use a soil drill to collect surface soil samples at a depth of 0–20 cm at the soil sampling points, remove impurities from the surface soil samples, and evenly spread the surface soil samples on a black background;

[0015] The hyperspectral fiber optic sensor head was fixed at a 45-degree angle so that the pointing position of the hyperspectral fiber optic sensor head was the center of the black background. The hyperspectral fiber optic sensor head was 10–15 cm away from the surface soil sample, and the hyperspectral band range was set to collect hyperspectral data of each surface soil sample within the set hyperspectral band range.

[0016] Specifically, the range of the hyperspectral band is 400 nanometers to 2500 nanometers, each nanometer is a hyperspectral band, and the hyperspectral data is the reflectivity of the surface soil sample in the hyperspectral band;

[0017] The original hyperspectral data is represented as follows: X=(X nm ) N×M The original hyperspectral data X is in the form of a matrix with N rows and M columns, where M represents the number of hyperspectral bands. nm Represents the hyperspectral data of the surface soil sample at the nth soil sampling point in the mth hyperspectral band, n∈[1,N], m∈[[1,M];

[0018] The organic carbon content of the surface soil sample is determined to obtain the organic carbon content of the soil sampling point associated with the surface soil sample.

[0019] Optionally, an improved Savitzky-Golay filtering method is used to filter and denoise the original hyperspectral data, including:

[0020] The improved Savitzky-Golay filtering process is as follows:

[0021] Set the filter window size to 2k+1, the polynomial order to d, and use the least squares fitting method to determine the filter coefficient at each filter window position. The range of the filter window position is [[-k, k]];

[0022] Extracting matrix elements of each row in the original hyperspectral data in sequence, wherein each row in the original hyperspectral data represents hyperspectral data of a surface soil sample associated with a soil sampling point within a set hyperspectral band range;

[0023] The matrix element in the nth row of the original hyperspectral data is: X(n)=((X n1 ,X n2 ,...,X nm ,...,X nM ), X n1 ,X n2 ,...,X nm ,...,X nM Represents the hyperspectral data of the surface soil sample at the nth soil sampling point in the 1st to Mth hyperspectral bands;

[0024] The filter coefficient is used to improve the filtering of each row of matrix elements extracted to obtain the improved filtering results of each row of matrix elements. The improved filtering results of N rows of matrix elements constitute the filtered hyperspectral data, where X in the nth row of matrix elements X(n) is nm The improved filtering formula is:

[0025]

[0026] Among them, X n(m+j) represents the m+jth hyperspectral data in the nth row matrix element, Represents hyperspectral data X nm The filtering result is Represents hyperspectral data X nm Normalized filter weight at the jth filter window position, g nm (j) represents the hyperspectral data X nm The filter weight at the jth filter window position, c j represents the filter coefficient of the jth filter window position, β represents the skewness adjustment coefficient, γ nm Represents the hyperspectral data X nm The skewness of the filter window sequence centered on nm Represents the hyperspectral data X nm The standard deviation of the hyperspectral data of the filter window sequence centered on Represents the hyperspectral data X nm is the mean of the hyperspectral data of the filtered window sequence centered at , j∈[-k,k], and sgn(·) represents the sign function.

[0027] Optionally, calculate the measured importance of each hyperspectral band in the filtered hyperspectral data for the measured organic carbon content, including:

[0028] The representation of the filtered hyperspectral data is: in Represents the hyperspectral data X of the surface soil sample at the nth soil sampling point in the mth hyperspectral band nm The filtering result of , n∈[[1,N], m∈[1,M], M represents the number of hyperspectral bands, and N represents the number of soil sampling points;

[0029] The organic carbon content of the nth soil sampling point is E n ;

[0030] Calculating the first-order derivative matrix of the filtered hyperspectral data and performing a standard normal transformation on the first-order derivative matrix to obtain an enhanced hyperspectral matrix;

[0031] Specifically, the filtered hyperspectral data The first-order derivative matrix of is X′:

[0032]

[0033] Among them, X′ nm represents the element in the nth row and mth column of the first-order derivative matrix X′;

[0034] The enhanced hyperspectral matrix is X″:

[0035]

[0036] Among them, X″ nm represents the element in the nth row and mth column of the enhanced hyperspectral matrix X″, μ((m)) represents the mean of the mth column in the first-order derivative matrix X′, and σ(m) represents the standard deviation of the mth column in the first-order derivative matrix X′;

[0037] Calculate the measurement importance between each hyperspectral band in the enhanced hyperspectral matrix and the measured organic carbon content. Specifically, the calculation formula for the measurement importance is a Pearson correlation coefficient calculation formula, where the measurement importance between the mth hyperspectral band and the measured organic carbon content is the Pearson correlation coefficient between the matrix element in the mth column of the enhanced hyperspectral matrix and the sequence of measured organic carbon contents at N soil sampling points.

[0038] Optionally, based on the measured importance, a hyperspectral band with the highest information gain is dynamically selected as a candidate hyperspectral band, and a candidate hyperspectral band set is formed, including:

[0039] Initialize the candidate hyperspectral band set;

[0040] Extracting the most important hyperspectral band and its spectral vector, and adding the extracted hyperspectral band as a candidate hyperspectral band to a candidate hyperspectral band set; the spectral vector of the hyperspectral band is a vector formed by the hyperspectral data of the hyperspectral band at N soil sampling points;

[0041] The orthogonal residual between the spectral vector of the current non-candidate hyperspectral band and the spectral vector matrix of all candidate hyperspectral bands in the candidate hyperspectral band set is calculated, and the orthogonal residual is processed by L2 norm as the residual of the current non-candidate hyperspectral band. The current non-candidate hyperspectral band with the largest residual is selected as the candidate hyperspectral band, and the candidate hyperspectral band is added to the candidate hyperspectral band set. The current non-candidate hyperspectral band is updated, and the orthogonal residual of the current non-candidate hyperspectral band is recalculated. The candidate hyperspectral band is selected until the maximum residual is lower than the preset residual threshold to obtain the candidate hyperspectral band set.

[0042] In the candidate hyperspectral selection process, by calculating the "orthogonality" residual of the non-candidate hyperspectral band relative to the existing candidate hyperspectral band set, the complementary spectral information contained in the current band and the candidate hyperspectral band set is identified, so as to maximize the retention of independent information related to the target variable but not yet captured by the existing bands, wherein the orthogonal residual represents the minimum projection error between the spectral response vector of the current non-candidate hyperspectral band and the candidate hyperspectral band set. The larger the orthogonal residual, the lower the information redundancy between the non-candidate hyperspectral band and the selected band, and the stronger the independent information supplementation capability. The L2 norm is used to quantize the residual vector into a scalar for easy sorting and selection, ensuring that the band selection process is stable and physically interpretable, that is, whether the introduction of a new hyperspectral band brings actual information gain.

[0043] Specifically, the spectral vector of the current non-candidate hyperspectral band λ is X λ =[[X λ ((1),X λ ((2),...,X λ ((N)] T , where T represents transpose, X λ ((1),X λ (2),...,X λ (N) represents the hyperspectral data of the current non-candidate hyperspectral band λ at N soil sampling points, and the spectral vector matrix of all candidate hyperspectral bands in the candidate hyperspectral band set is φ=[φ1,φ2,...,φ s ],φ1,φ2,...,φ s Represents the spectral vectors of S candidate hyperspectral bands in the candidate hyperspectral band set, where S represents the number of candidate hyperspectral bands in the current candidate hyperspectral band set;

[0044] The residual of the current non-candidate hyperspectral band λ is residual(λ):

[0045] residual(λ)=||I-φ(φ T φ) -1 φ T X λ ||2;

[0046] Among them, φ(φ T φ) -1 φ T X λ Represents the spectral vector X λ The orthogonal projection on the subspace φ, ||·||2 represents the L2 norm, and I represents the identity matrix of N rows and N columns.

[0047] Optionally, soil environmental factors and candidate hyperspectral data of soil sampling points under a candidate hyperspectral band set are collected, including:

[0048] Extract the candidate hyperspectral data of the soil sampling point in the candidate hyperspectral band set from step S1, and the representation of the candidate hyperspectral data is Y=(Y nb ) N×B , where the candidate hyperspectral data Y is in the form of a matrix with N rows and B columns, Y nb represents the hyperspectral data of the surface soil sample at the nth soil sampling point under the bth candidate hyperspectral band, b∈[1,B], n∈[1,N], N represents the number of soil sampling points, and B represents the number of candidate hyperspectral bands in the candidate hyperspectral band set;

[0049] The soil environmental factors include soil type vectors, NDVI vegetation index and terrain moisture index at different soil sampling points. The soil type vector of the nth soil sampling point is

[0050] in, Indicates the nth soil sampling point and the rth soil type Label r The encoding result between , r∈[1,R], R represents the total number of soil types.

[0051] Optionally, the association weights between the candidate hyperspectral data and the soil environmental factors are calculated using a canonical correlation analysis method, including:

[0052] The soil environmental factor is expressed as F=[F1, F2, ..., F n ,...,F N ] T , where F n The environmental vector representing the nth soil sampling point includes the soil type vector, NDVI vegetation index, and terrain moisture index of the nth soil sampling point;

[0053] Calculate the autocovariance matrix and cross-covariance matrix of the soil environmental factor F and the candidate hyperspectral data Y respectively. The autocovariance matrix of the soil environmental factor F is ∑ FF , the autocovariance matrix of the candidate hyperspectral data Y is ∑ YY , the cross covariance matrix of soil environmental factor F and candidate hyperspectral data Y is ∑ FY ,∑ YF ;

[0054] The autocovariance matrix and the cross covariance matrix are constructed as a joint covariance matrix ∑:

[0055]

[0056] Perform singular value decomposition on the joint covariance matrix ∑ to obtain the left singular vector U and the right singular vector V, and extract the first G columns of eigenvectors of the singular vectors to construct the spectral correlation weight w1 and the environmental correlation weight w2:

[0057]

[0058] in, represents the covariance matrix ∑ YY The inverse square root matrix of represents the covariance matrix ∑ FF The inverse square root matrix of , U(G) represents the first G columns of eigenvectors of the left singular vector U, V(G) represents the first G columns of eigenvectors of the right singular vector V, and G represents the singular vector projection coefficient;

[0059] Specifically, the autocovariance matrix ∑ of the soil environmental factor F is FF The correlation and variation range between the various dimensions of the environmental factors are reflected. The dimensions of the soil environmental factors include soil type vector, NDVI vegetation index and terrain moisture index; the autocovariance matrix ∑ YY , characterizes the linear correlation and collaborative change ability between candidate hyperspectral bands, reflects the redundancy and diversity characteristics of the bands themselves, and is conducive to subsequent dimensionality reduction and de-redundancy; the joint covariance matrix ∑ characterizes the cross-modal correlation between hyperspectral bands and soil environmental factors, and is the key basis for revealing the strength of the relationship between spectral bands and soil environmental factors. The canonical correlation analysis (CCA) method is used to extract the strongest correlation direction between hyperspectral bands and soil environmental factors, where the spectral correlation weight is the linear transformation matrix that projects the candidate hyperspectral data to the strongest correlation direction, reflecting the importance of each band's contribution to joint modeling, and the environmental correlation weight is the strongest collaborative transformation direction corresponding to the soil environmental factor. By using the correlation weight for fusion processing, the feature collaborative compression and information complementary enhancement between modalities are achieved;

[0060] The spectral association weight w1 and the environmental association weight w2 are used as association weights to fuse the candidate hyperspectral data Y and the soil environmental factor F to obtain the fusion feature matrix:

[0061] H=Yw1+Fw2;

[0062] Among them, H represents the fusion feature matrix.

[0063] Specifically, in hyperspectral band selection and multifactor modeling, spectral data and soil environmental factors often suffer from dimensional inconsistencies, high distribution heterogeneity, and severe data redundancy, limiting their collaborative modeling capabilities. This paper introduces a canonical correlation analysis method, mapping the high-dimensional spectral band matrix and the multi-source environmental factor matrix onto the same set of canonical variable spaces. Within this space, the directions with the strongest linear correlations between the two are extracted. By maximizing the correlations between pairs of canonical variables, the potential synergistic relationships between spectral data and environmental factors can be effectively explored, improving the discriminability and representativeness of hyperspectral data in a multifactor context.

[0064] Compared with traditional direct splicing or independent modeling methods, the fused features not only retain the main information of the original spectral data and environmental factors, but also significantly reduce the feature dimension, improving the stability and generalization ability of model training. This method solves the core problems of "structural mismatch", "unbalanced feature weights", and "weak variable redundancy coupling" in the fusion of hyperspectral and environmental factors, providing a feature expression with a reasonable structure and clear physical meaning for subsequent band prediction and soil organic carbon detection. It is particularly suitable for regional modeling tasks with complex soil properties and significant geographical variation.

[0065] Optionally, construct an optimal hyperspectral band prediction model, including:

[0066] The optimal hyperspectral band prediction model includes a lightweight convolution layer, a depth-separable convolution layer, a feature splicing layer, and a fully connected layer;

[0067] The lightweight convolution layer is used to receive the fusion feature matrix and perform lightweight convolution processing on the fusion feature matrix using a lightweight convolution kernel to obtain a fusion feature map;

[0068] The depth-wise separable convolution layer includes a DepthwiseConv2D module and a PointwiseConv2D module, which respectively perform channel-wise independent convolution and 1×1 convolution integration on the fused feature map, and connect BatchNorm+ReLU after each module to obtain the statistical separation feature map corresponding to the fused feature matrix;

[0069] The feature splicing layer is used to splice the fusion feature matrix and the statistical separation feature map to obtain a spliced feature map;

[0070] The fully connected layer is a multi-layer perceptron structure, which is used to receive splicing features and use the softmax function to output the environmental perception values of different candidate hyperspectral bands, and select the candidate hyperspectral band with the highest environmental perception value as the optimal hyperspectral band prediction result.

[0071] Optionally, the optimal hyperspectral band prediction model is used to perform multi-factor fusion optimal hyperspectral band prediction, and soil organic carbon content detection is performed based on the optimal hyperspectral band prediction result, including:

[0072] The optimal hyperspectral band prediction model is used to receive the fusion feature matrix and output the optimal hyperspectral band prediction result. The predicted spectral data of the soil sampling point in the optimal hyperspectral band prediction result and the environmental vector of the soil sampling point are extracted. The predicted spectral data and the soil environmental factor are spliced to obtain the encoding vector of the soil sampling point. The soil organic carbon content of the soil sampling point is predicted using a lightweight prediction model. The environmental vector of the soil sampling point includes the soil type vector, NDVI vegetation index and terrain moisture index of the soil sampling point.

[0073] In order to solve the above problem, the present invention provides an electronic device, comprising:

[0074] a memory storing at least one instruction;

[0075] Communication interfaces to enable electronic equipment to communicate; and

[0076] The processor executes the instructions stored in the memory to implement the above-mentioned optimal hyperspectral band prediction method for soil organic carbon environment detection.

[0077] In order to solve the above problems, the present invention also provides a computer-readable storage medium, which stores at least one instruction, and the at least one instruction is executed by a processor in an electronic device to implement the above-mentioned optimal hyperspectral band prediction method for soil organic carbon environment detection.

[0078] Compared with the existing technology, this paper proposes a hyperspectral band optimization method for soil organic carbon detection based on deep learning, which has the following advantages:

[0079] First, this application improves the traditional SG (Savitzky-Golay) filter so that the improved filter is specifically used to process the asymmetric disturbances and multi-scale change trends in soil hyperspectral data. The traditional SG filter uses a fixed window and symmetric weights, and is insensitive to changes in the local morphology of the signal, especially in the "shoulder peak", "slope" and "micro-fluctuation" areas commonly seen in soil spectra, which often have problems of over-filtering or feature loss. To this end, this application introduces a local skewness analysis mechanism, calculates the skewness value for the spectral distribution morphology of each central band within a set window, and dynamically adjusts the filter window length and weight distribution. The skewness value reflects the asymmetric characteristics of the spectral curve of the band. A positive skewness indicates that the right tail is elongated (upward trend), and a negative skewness indicates that the left tail is elongated (downward trend). By constructing a skewness-window mapping function, the smoothing direction and strength are adaptively adjusted. Specifically, an asymmetric window weighting strategy is adopted in high-skewness areas to enhance the ability to retain local anomalies; a standard symmetric window is used in low-skewness areas to improve the overall smoothing effect; this method significantly solves the problem that traditional filters cannot adapt to areas of spectral heterogeneity, and improves the window modeling stability of soil key bands. At the same time, the window mechanism under dynamic filtering weights can be automatically adjusted according to the actual spectral noise structure, without the need for manual repeated testing of window parameters, effectively reducing human dependence and window modeling uncertainty in the hyperspectral preprocessing process, and is suitable for universal application under complex soil types.

[0080] At the same time, this application proposes an algorithm for hyperspectral selection, which integrates partial correlation constraints and a dynamic iterative strategy driven by orthogonal residuals. Although the traditional continuous projection algorithm can effectively avoid multicollinearity, it has the problem of insufficient information in the selected bands when processing bands with weak correlation with target variables or when there are redundant features. To this end, this application first introduces a partial correlation coefficient pre-screening mechanism to eliminate bands with low correlation with organic carbon content and improve the quality of initial features; secondly, in the iterative stage, the orthogonal projection residuals of the candidate bands to the currently selected subspace are calculated, and the bands with the least redundancy and the largest information difference are given priority, thereby ensuring that each new band introduced can bring a new discriminant dimension to the model gain. At the same time, the algorithm adopts a norm-controlled termination strategy to avoid the introduction of too many redundant features. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 A flowchart of a method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning is provided in accordance with one embodiment of the present invention.

[0082] Figure 2 This is an example diagram of the spectral data of the soil sampling point in the red light band, near infrared band, and shortwave infrared band.

[0083] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0084] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0085] The embodiment of the present application provides a method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning. The execution subject of the method for predicting the optimal hyperspectral band for soil organic carbon environmental detection includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided by the embodiment of the present application. In other words, the method for predicting the optimal hyperspectral band for soil organic carbon environmental detection can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc.

[0086] Reference Figure 1 , embodiment 1 of the present invention is:

[0087] A deep learning-based hyperspectral band optimization method for soil organic carbon detection includes the following steps:

[0088] S1: Collect the original hyperspectral data of soil sampling points and the organic carbon content, and use the improved Savitzky-Golay filtering method to filter and denoise the original hyperspectral data to obtain filtered hyperspectral data.

[0089] Collect raw hyperspectral data and organic carbon content from soil sampling points, including:

[0090] Select N soil sampling points from the soil to be tested for organic carbon, use a soil drill to collect surface soil samples at a depth of 0–20 cm at the soil sampling points, remove impurities from the surface soil samples, and evenly spread the surface soil samples on a black background;

[0091] The hyperspectral fiber optic sensor head was fixed at a 45-degree angle so that the pointing position of the hyperspectral fiber optic sensor head was the center of the black background. The hyperspectral fiber optic sensor head was 10–15 cm away from the surface soil sample, and the hyperspectral band range was set to collect hyperspectral data of each surface soil sample within the set hyperspectral band range.

[0092] Specifically, the range of the hyperspectral band is 400 nanometers to 2500 nanometers, each nanometer is a hyperspectral band, and the hyperspectral data is the reflectivity of the surface soil sample in the hyperspectral band;

[0093] The original hyperspectral data is represented as follows: X=(X nm ) N×MThe original hyperspectral data X is in the form of a matrix with N rows and M columns, where M represents the number of hyperspectral bands. nm represents the hyperspectral data of the surface soil sample at the nth soil sampling point in the mth hyperspectral band, n∈[1,N], m∈[[1,M]; in the embodiment of the present invention, M=2101;

[0094] The organic carbon content of the surface soil sample is determined to obtain the organic carbon content of the soil sampling point associated with the surface soil sample.

[0095] As an embodiment of the present invention, the organic carbon determination method of the surface soil sample is potassium dichromate-external heating oxidation method, which uses potassium dichromate K2Cr2O7 as an oxidant, uses excess potassium dichromate K2Cr2O7 to oxidize the organic carbon in the surface soil sample, then heats with sulfuric acid to promote the reaction, and uses a titrant to remove the unreacted Cr2O7. 2- The titration is performed, and the amount of oxidant consumed by the organic carbon in the surface soil sample is inferred based on the volume of titrant consumed, and then the organic carbon content is calculated. The determination formula of the organic carbon content is:

[0096]

[0097] Wherein, SOC represents the measured organic carbon content, and the unit of the measured organic carbon content is (g / kg), which represents the mass of organic carbon in each kilogram of soil;

[0098] weight represents the mass of the surface soil sample, c represents the titrant concentration, τ=1 / 0.58 represents the conversion coefficient of organic matter to organic carbon; V0-V1 represents the amount of oxidant consumed by organic carbon, the V0 represents the control titrant volume, which means the volume of titrant consumed by titrating the mixed solution of oxidant and sulfuric acid without adding the surface soil sample, and the V1 represents the volume of titrant consumed by titrating the mixed solution of oxidant and sulfuric acid after adding the surface soil sample; the titrant is sodium thiosulfate solution.

[0099] The original hyperspectral data is filtered and denoised using an improved Savitzky-Golay filtering method, including:

[0100] The improved Savitzky-Golay filtering process is as follows:

[0101] Set the filter window size to 2k+1, the polynomial order to d, and use the least squares fitting method to determine the filter coefficient at each filter window position. The range of the filter window position is [[-k, k]];

[0102] Extracting matrix elements of each row in the original hyperspectral data in sequence, wherein each row in the original hyperspectral data represents hyperspectral data of a surface soil sample associated with a soil sampling point within a set hyperspectral band range;

[0103] The matrix element in the nth row of the original hyperspectral data is: X(n)=((X n1 ,X n2 ,...,X nm ,...,X nM ), X n1 ,X n2 ,...,X nm ,...,X nM Represents the hyperspectral data of the surface soil sample at the nth soil sampling point in the 1st to Mth hyperspectral bands;

[0104] The filter coefficient is used to improve the filtering of each row of matrix elements extracted to obtain the improved filtering results of each row of matrix elements. The improved filtering results of N rows of matrix elements constitute the filtered hyperspectral data, where X in the nth row of matrix elements X(n) is nm The improved filtering formula is:

[0105]

[0106] Among them, X n(m+j) represents the m+jth hyperspectral data in the nth row matrix element, Represents hyperspectral data X nm The filtering result is Represents hyperspectral data X nm Normalized filter weight at the jth filter window position, g nm (j) represents the hyperspectral data X nm The filter weight at the jth filter window position, c j represents the filter coefficient of the jth filter window position, β represents the skewness adjustment coefficient, γ nm Represents the hyperspectral data X nm The skewness of the filter window sequence centered on nm Represents the hyperspectral data X nm The standard deviation of the hyperspectral data of the filter window sequence centered on Represents the hyperspectral data X nm is the mean of the hyperspectral data of the filtered window sequence centered at , j∈[-k,k], and sgn(·) represents the sign function.

[0107] As an embodiment of the present invention, if j is greater than 0, sgn(j)=1; if j is less than 0, sgn(j)=-1; if j is equal to 0, sgn(j)=0.

[0108] S2: Calculate the importance of each hyperspectral band in the filtered hyperspectral data for the determination of organic carbon content. Based on the measurement importance, dynamically select the hyperspectral band with the highest information gain as the candidate hyperspectral band to form a set of candidate hyperspectral bands.

[0109] Calculate the importance of each hyperspectral band in the filtered hyperspectral data for the determination of organic carbon content, including:

[0110] The representation of the filtered hyperspectral data is: in Represents the hyperspectral data X of the surface soil sample at the nth soil sampling point in the mth hyperspectral band nm The filtering result of , n∈[[1,N], m∈[1,M]M represents the number of hyperspectral bands, and N represents the number of soil sampling points;

[0111] The organic carbon content of the nth soil sampling point is E n ;

[0112] Calculating the first-order derivative matrix of the filtered hyperspectral data and performing a standard normal transformation on the first-order derivative matrix to obtain an enhanced hyperspectral matrix;

[0113] Specifically, the filtered hyperspectral data The first-order derivative matrix of is X′:

[0114]

[0115] Among them, X′ nm represents the element in the nth row and mth column of the first-order derivative matrix X′;

[0116] The enhanced hyperspectral matrix is X″:

[0117]

[0118] Among them, X″ nm represents the element in the nth row and mth column of the enhanced hyperspectral matrix X″, μ(m) represents the mean of the mth column in the first-order derivative matrix X′, and σ(m) represents the standard deviation of the mth column in the first-order derivative matrix X′;

[0119] Calculate the measurement importance between each hyperspectral band in the enhanced hyperspectral matrix and the measured organic carbon content. Specifically, the calculation formula for the measurement importance is a Pearson correlation coefficient calculation formula, where the measurement importance between the mth hyperspectral band and the measured organic carbon content is the Pearson correlation coefficient between the matrix element in the mth column of the enhanced hyperspectral matrix and the sequence of measured organic carbon contents at N soil sampling points.

[0120] Based on the measurement importance, the hyperspectral band with the highest information gain is dynamically selected as the candidate hyperspectral band, and a candidate hyperspectral band set is formed, including:

[0121] Initialize the candidate hyperspectral band set;

[0122] Extracting the most important hyperspectral band and its spectral vector, and adding the extracted hyperspectral band as a candidate hyperspectral band to a candidate hyperspectral band set; the spectral vector of the hyperspectral band is a vector formed by the hyperspectral data of the hyperspectral band at N soil sampling points;

[0123] The orthogonal residual between the spectral vector of the current non-candidate hyperspectral band and the spectral vector matrix of all candidate hyperspectral bands in the candidate hyperspectral band set is calculated, and the orthogonal residual is processed by L2 norm as the residual of the current non-candidate hyperspectral band. The current non-candidate hyperspectral band with the largest residual is selected as the candidate hyperspectral band, and the candidate hyperspectral band is added to the candidate hyperspectral band set. The current non-candidate hyperspectral band is updated, and the orthogonal residual of the current non-candidate hyperspectral band is recalculated. The candidate hyperspectral band is selected until the maximum residual is lower than the preset residual threshold to obtain the candidate hyperspectral band set.

[0124] S3: Collect soil environmental factors and candidate hyperspectral data of soil sampling points under the candidate hyperspectral band set, use canonical correlation analysis to calculate the association weights between the candidate hyperspectral data and soil environmental factors, and fuse the candidate hyperspectral data and soil environmental factors based on the association weights to obtain a fusion feature matrix.

[0125] Collect soil environmental factors and candidate hyperspectral data of soil sampling points under the candidate hyperspectral band set, including:

[0126] Extract the candidate hyperspectral data of the soil sampling point in the candidate hyperspectral band set from step S1, and the representation of the candidate hyperspectral data is Y=(Y nb ) N×B , where the candidate hyperspectral data Y is in the form of a matrix with N rows and B columns, Y nb represents the hyperspectral data of the surface soil sample at the nth soil sampling point under the bth candidate hyperspectral band, b∈[1,B], n∈[1,N], N represents the number of soil sampling points, and B represents the number of candidate hyperspectral bands in the candidate hyperspectral band set;

[0127] The soil environmental factors include soil type vectors, NDVI vegetation index and terrain moisture index at different soil sampling points. The soil type vector of the nth soil sampling point is

[0128]

[0129] in, Indicates the nth soil sampling point and the rth soil type Label r The encoding result between , r∈[1,R], R represents the total number of soil types.

[0130] As an embodiment of the present invention, remote sensing images of soil sampling points are obtained, RED and NIR bands are extracted, and the NDVI vegetation index of the soil sampling points is generated using the NDVI formula; a high-precision digital elevation model of the soil sampling points is established, the slope and catchment area of the soil sampling points are calculated using GIS software, and the terrain moisture index of the soil sampling points is generated using the TWI formula.

[0131] The correlation weights between the candidate hyperspectral data and soil environmental factors are calculated using a canonical correlation analysis method, including:

[0132] The soil environmental factor is expressed as F=[F1, F2, ..., F n ,...,F N ] T , where F n The environmental vector representing the nth soil sampling point includes the soil type vector, NDVI vegetation index, and terrain moisture index of the nth soil sampling point;

[0133] Calculate the autocovariance matrix and cross-covariance matrix of the soil environmental factor F and the candidate hyperspectral data Y respectively. The autocovariance matrix of the soil environmental factor F is ∑ FF , the autocovariance matrix of the candidate hyperspectral data Y is ∑ YY , the cross covariance matrix of soil environmental factor F and candidate hyperspectral data Y is ∑ FY ,∑ YF As an embodiment of the present invention, ∑ FF It is a matrix with R+2 rows and R+2 columns, ∑ YY is a matrix with B rows and B columns, ∑ FY It is a matrix with R+2 rows and B columns, ∑ YF It is a matrix with B rows and R + 2 columns;

[0134] The autocovariance matrix and the cross covariance matrix are constructed as a joint covariance matrix ∑:

[0135]

[0136] Perform singular value decomposition on the joint covariance matrix ∑ to obtain the left singular vector U and the right singular vector V, and extract the first G columns of eigenvectors of the singular vectors to construct the spectral correlation weight w1 and the environmental correlation weight w2:

[0137]

[0138] in, represents the covariance matrix ∑ YY The inverse square root matrix of represents the covariance matrix ∑ FF The inverse square root matrix of , U(G) represents the first G columns of eigenvectors of the left singular vector U, V(G) represents the first G columns of eigenvectors of the right singular vector V, and G represents the singular vector projection coefficient; as an embodiment of the present invention, G≤min{B,R+2}

[0139] The spectral association weight w1 and the environmental association weight w2 are used as association weights to fuse the candidate hyperspectral data Y and the soil environmental factor F to obtain the fusion feature matrix:

[0140] H=Yw1+Fw2;

[0141] Among them, H represents the fusion feature matrix.

[0142] In an embodiment of the present invention, the spectral association weight w1 is in a matrix form of B rows and G columns, the environmental association weight w2 is in a matrix form of R+2 rows and G columns, F is in a matrix form of N rows and R+2 columns, Y is in a matrix form of N rows and B columns, and H is in a matrix form of N rows and G columns.

[0143] S4: Construct an optimal hyperspectral band prediction model, use the optimal hyperspectral band prediction model to perform optimal hyperspectral band prediction with multi-factor fusion, and detect soil organic carbon content based on the optimal hyperspectral band prediction results.

[0144] Construct an optimal hyperspectral band prediction model, including:

[0145] The optimal hyperspectral band prediction model includes a lightweight convolution layer, a depth-separable convolution layer, a feature splicing layer, and a fully connected layer;

[0146] The lightweight convolution layer is used to receive the fusion feature matrix and perform lightweight convolution processing on the fusion feature matrix using a lightweight convolution kernel to obtain a fusion feature map;

[0147] The depth-wise separable convolution layer includes a DepthwiseConv2D module and a PointwiseConv2D module, which respectively perform channel-wise independent convolution and 1×1 convolution integration on the fused feature map, and connect BatchNorm+ReLU after each module to obtain the statistical separation feature map corresponding to the fused feature matrix;

[0148] The feature splicing layer is used to splice the fusion feature matrix and the statistical separation feature map to obtain a spliced feature map;

[0149] The fully connected layer is a multi-layer perceptron structure, which is used to receive splicing features and use the softmax function to output the environmental perception values of different candidate hyperspectral bands, and select the candidate hyperspectral band with the highest environmental perception value as the optimal hyperspectral band prediction result.

[0150] The optimal hyperspectral band prediction model is used to perform multi-factor fusion optimal hyperspectral band prediction, and soil organic carbon content is detected based on the optimal hyperspectral band prediction result, including:

[0151] The optimal hyperspectral band prediction model is used to receive the fused feature matrix and output the optimal hyperspectral band prediction result. The predicted spectral data of the soil sampling point in the optimal hyperspectral band prediction result and the environmental vector of the soil sampling point are extracted. The predicted spectral data are spliced with the soil environmental factors to obtain the encoding vector of the soil sampling point. The soil organic carbon content of the soil sampling point is predicted using a lightweight prediction model. The environmental vector of the soil sampling point includes the soil type vector, NDVI vegetation index, and terrain moisture index of the soil sampling point. As one embodiment of the present invention, the lightweight prediction model is a SVM model.

[0152] Example 2:

[0153] Figure 2 This is an example diagram of the spectral data of the soil sampling point in the red light band, the near infrared band, and the shortwave infrared band, which represents the reflectivity of the soil sampling point in different hyperspectral bands, and calculates the average value of the reflectivity in the spectral data as the hyperspectral data of the soil sampling point in the red light band, the near infrared band, and the shortwave infrared band in Example 1, constituting the original hyperspectral data described in Example 1.

[0154] It should be understood that the embodiment is for illustration only and the scope of the patent application is not limited to this structure.

[0155] It should be noted that the serial numbers of the above-mentioned embodiments of the present invention are for descriptive purposes only and do not represent the advantages or disadvantages of the embodiments. In addition, the terms "including", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method comprising the element.

[0156] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0157] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A hyperspectral band optimization method for soil organic carbon detection based on deep learning, characterized in that: The method comprises: S1: Collect the original hyperspectral data of soil sampling points and the organic carbon content, and use the improved Savitzky-Golay filtering method to filter and denoise the original hyperspectral data to obtain filtered hyperspectral data; S2: calculating the measurement importance of each hyperspectral band in the filtered hyperspectral data for the determination of organic carbon content, and based on the measurement importance, dynamically selecting the hyperspectral band with the highest information gain as the candidate hyperspectral band to form a candidate hyperspectral band set; S3: Collect candidate hyperspectral data of soil environmental factors and soil sampling points under the candidate hyperspectral band set, calculate the association weights between the candidate hyperspectral data and the soil environmental factors using canonical correlation analysis, and fuse the candidate hyperspectral data and soil environmental factors based on the association weights to obtain a fusion feature matrix; S4: Construct an optimal hyperspectral band prediction model, use the optimal hyperspectral band prediction model to perform optimal hyperspectral band prediction of multi-factor fusion, and detect soil organic carbon content based on the optimal hyperspectral band prediction results. The optimal hyperspectral band prediction model takes the fusion feature matrix as input and the optimal hyperspectral band prediction results as output.

2. The method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning according to claim 1, characterized in that: Collect raw hyperspectral data and organic carbon content from soil sampling points, including: Select N soil sampling points from the soil to be tested for organic carbon, use a soil drill to collect surface soil samples at a depth of 0–20 cm at the soil sampling points, remove impurities from the surface soil samples, and evenly spread the surface soil samples on a black background; The optical fiber sensor head of the hyperspectrometer is fixed at a 45-degree angle so that the pointing position of the optical fiber sensor head of the hyperspectrometer is the center of the black background, and the hyperspectral band range is set to collect the hyperspectral data of each surface soil sample in the set hyperspectral band range; The original hyperspectral data is represented as follows: X=(X nm ) N×M The original hyperspectral data X is in the form of a matrix with N rows and M columns, where M represents the number of hyperspectral bands. nm Represents the hyperspectral data of the surface soil sample at the nth soil sampling point in the mth hyperspectral band, n∈[1,N], m∈[[1,M]; The organic carbon content of the surface soil sample is determined to obtain the organic carbon content of the soil sampling point associated with the surface soil sample.

3. The method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning according to claim 2, wherein: The original hyperspectral data is filtered and denoised using an improved Savitzky-Golay filtering method, including: The improved Savitzky-Golay filtering process is as follows: Set the filter window size to 2k+1, the polynomial order to d, and use the least squares fitting method to determine the filter coefficient at each filter window position. The range of the filter window position is [[-k, k]]; Extracting matrix elements of each row in the original hyperspectral data in sequence, wherein each row in the original hyperspectral data represents hyperspectral data of a surface soil sample associated with a soil sampling point within a set hyperspectral band range; The matrix element in the nth row of the original hyperspectral data is: X(n)=((X n1 ,X n2 ,...,X nm ,...,X nM ), X n1 ,X n2 ,...,X nm ,...,X nM Represents the hyperspectral data of the surface soil sample at the nth soil sampling point in the 1st to Mth hyperspectral bands; The filter coefficient is used to improve the filtering of each row of matrix elements extracted to obtain the improved filtering results of each row of matrix elements. The improved filtering results of N rows of matrix elements constitute the filtered hyperspectral data, where X in the nth row of matrix elements X(n) is nm The improved filtering formula is: Among them, X n(m+j) represents the m+jth hyperspectral data in the nth row matrix element, Represents hyperspectral data X nm The filtering result is Represents hyperspectral data X nm Normalized filter weight at the jth filter window position, g nm (j) represents the hyperspectral data X nm The filter weight at the jth filter window position, c j represents the filter coefficient of the jth filter window position, β represents the skewness adjustment coefficient, γ nm Represents the hyperspectral data X nm The skewness of the filter window sequence centered on nm Represents the hyperspectral data X nm The standard deviation of the hyperspectral data of the filter window sequence centered on Represents the hyperspectral data X nm is the mean of the hyperspectral data of the filtered window sequence centered at , j∈[-k,k], and sgn(·) represents the sign function.

4. The method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning according to claim 3, wherein: Calculate the importance of each hyperspectral band in the filtered hyperspectral data for the determination of organic carbon content, including: The representation of the filtered hyperspectral data is: in Represents the hyperspectral data X of the surface soil sample at the nth soil sampling point in the mth hyperspectral band nm The filtering result of , n∈[[1,N], m∈[1,M], M represents the number of hyperspectral bands, and N represents the number of soil sampling points; The organic carbon content of the nth soil sampling point is E n ; Calculating the first-order derivative matrix of the filtered hyperspectral data and performing a standard normal transformation on the first-order derivative matrix to obtain an enhanced hyperspectral matrix; The determination importance between each hyperspectral band in the enhanced hyperspectral matrix and the measured organic carbon content is calculated.

5. The method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning according to claim 4, characterized in that: Based on the measurement importance, the hyperspectral band with the highest information gain is dynamically selected as the candidate hyperspectral band, and a candidate hyperspectral band set is formed, including: Initialize the candidate hyperspectral band set; Extracting the most important hyperspectral band and its spectral vector, and adding the extracted hyperspectral band as a candidate hyperspectral band to a candidate hyperspectral band set; the spectral vector of the hyperspectral band is a vector formed by the hyperspectral data of the hyperspectral band at N soil sampling points; The orthogonal residual between the spectral vector of the current non-candidate hyperspectral band and the spectral vector matrix of all candidate hyperspectral bands in the candidate hyperspectral band set is calculated, and the orthogonal residual is processed by L2 norm as the residual of the current non-candidate hyperspectral band. The current non-candidate hyperspectral band with the largest residual is selected as the candidate hyperspectral band, and the candidate hyperspectral band is added to the candidate hyperspectral band set. The current non-candidate hyperspectral band is updated, and the orthogonal residual of the current non-candidate hyperspectral band is recalculated. The candidate hyperspectral band is selected until the maximum residual is lower than the preset residual threshold to obtain the candidate hyperspectral band set.

6. The method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning according to claim 1, wherein: Collect soil environmental factors and candidate hyperspectral data of soil sampling points under the candidate hyperspectral band set, including: Extract the candidate hyperspectral data of the soil sampling point in the candidate hyperspectral band set from step S1, and the representation of the candidate hyperspectral data is Y=(Y nb ) N×B , where the candidate hyperspectral data Y is in the form of a matrix with N rows and B columns, Y nb represents the hyperspectral data of the surface soil sample at the nth soil sampling point under the bth candidate hyperspectral band, b∈[1,B], n∈[1,N], N represents the number of soil sampling points, and B represents the number of candidate hyperspectral bands in the candidate hyperspectral band set; The soil environmental factors include soil type vectors, NDVI vegetation index and terrain moisture index at different soil sampling points. The soil type vector of the nth soil sampling point is in, Indicates the nth soil sampling point and the rth soil type Label r The encoding result between , r∈[1,R], R represents the total number of soil types.

7. The method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning according to claim 6, wherein: The correlation weights between the candidate hyperspectral data and soil environmental factors are calculated using a canonical correlation analysis method, including: The soil environmental factor is expressed as F=[F1, F2, ..., F n ,...,F N ] T , where F n The environmental vector representing the nth soil sampling point includes the soil type vector, NDVI vegetation index, and terrain moisture index of the nth soil sampling point; Calculate the autocovariance matrix and cross-covariance matrix of the soil environmental factor F and the candidate hyperspectral data Y respectively. The autocovariance matrix of the soil environmental factor F is ∑ FF , the autocovariance matrix of the candidate hyperspectral data Y is ∑ YY , the cross covariance matrix of soil environmental factor F and candidate hyperspectral data Y is ∑ FY ,∑ YF ; The autocovariance matrix and the cross covariance matrix are constructed as a joint covariance matrix ∑: Perform singular value decomposition on the joint covariance matrix ∑ to obtain the left singular vector U and the right singular vector V, and extract the first G columns of eigenvectors of the singular vectors to construct the spectral correlation weight w1 and the environmental correlation weight w2: in, represents the covariance matrix ∑ YY The inverse square root matrix of represents the covariance matrix ∑ FF The inverse square root matrix of , U(G) represents the first G columns of eigenvectors of the left singular vector U, V(G) represents the first G columns of eigenvectors of the right singular vector V, and G represents the singular vector projection coefficient; The spectral association weight w1 and the environmental association weight w2 are used as association weights, and the candidate hyperspectral data Y and the soil environmental factor F are fused to obtain a fusion feature matrix.

8. The method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning according to claim 1, wherein: Construct an optimal hyperspectral band prediction model, including: The optimal hyperspectral band prediction model includes a lightweight convolution layer, a depth-separable convolution layer, a feature splicing layer, and a fully connected layer; The lightweight convolution layer is used to receive the fusion feature matrix and perform lightweight convolution processing on the fusion feature matrix using a lightweight convolution kernel to obtain a fusion feature map; The depth-wise separable convolution layer includes a DepthwiseConv2D module and a PointwiseConv2D module, which respectively perform channel-wise independent convolution and 1×1 convolution integration on the fused feature map, and connect BatchNorm+ReLU after each module to obtain the statistical separation feature map corresponding to the fused feature matrix; The feature splicing layer is used to splice the fusion feature matrix and the statistical separation feature map to obtain a spliced feature map; The fully connected layer is a multi-layer perceptron structure, which is used to receive splicing features and use the softmax function to output the environmental perception values of different candidate hyperspectral bands, and select the candidate hyperspectral band with the highest environmental perception value as the optimal hyperspectral band prediction result.

9. The method for optimizing hyperspectral bands for soil organic carbon detection based on deep learning according to claim 8, characterized in that: The optimal hyperspectral band prediction model is used to perform multi-factor fusion optimal hyperspectral band prediction, and soil organic carbon content is detected based on the optimal hyperspectral band prediction result, including: The optimal hyperspectral band prediction model is used to receive the fusion feature matrix and output the optimal hyperspectral band prediction result. The predicted spectral data of the soil sampling point in the optimal hyperspectral band prediction result and the environmental vector of the soil sampling point are extracted. The predicted spectral data and the soil environmental factor are spliced to obtain the encoding vector of the soil sampling point. The soil organic carbon content of the soil sampling point is predicted using a lightweight prediction model. The environmental vector of the soil sampling point includes the soil type vector, NDVI vegetation index and terrain moisture index of the soil sampling point.

Images