Methods and apparatus for retrieving vegetation cover from hyperspectral remote sensing image data

Abstract

The application provides a kind of vegetation coverage inversion method and device for hyperspectral remote sensing image data, it is related to hyperspectral remote sensing image processing technical field, the method comprises: obtaining the hyperspectral remote sensing data and corresponding high-resolution remote sensing data of target region in the same period;The pure end element contained in hyperspectral remote sensing data is extracted and identified, at least one type of pure end element is determined;Combined with high-resolution remote sensing data, the pure end element category contained in mixed pixel of hyperspectral remote sensing data, pure end element area and pure end element area proportion are obtained;Further, the initial mixed pixel unmixing model is trained, and the target mixed pixel unmixing model is obtained to invert the vegetation coverage in the region to be monitored.The application realizes the unmixing of mixed pixel in hyperspectral remote sensing image, can improve the detection accuracy of end element composition ratio, thereby improving the inversion accuracy of vegetation coverage for hyperspectral remote sensing image.

Description

Technical Field

[0001] This application relates to the field of hyperspectral remote sensing image processing technology, and in particular to a method and apparatus for inverting vegetation cover in hyperspectral remote sensing image data. Background Technology

[0002] Vegetation cover is an important indicator reflecting vegetation growth and the most effective indicator for assessing the degree of land desertification. It is widely used in agriculture, land use, disaster risk monitoring, and drought monitoring. In arid regions, research on how to improve the accuracy of large-scale vegetation cover retrieval is crucial for maintaining ecosystem stability and protecting the ecological environment.

[0003] Currently, there are two main methods for estimating vegetation cover: ground-based measurement and remote sensing inversion. Among these methods, the remote sensing inversion method, when performing vegetation cover retrieval on medium-resolution satellite imagery, encounters numerous mixed pixels, which can easily lead to misclassification and misidentification of pixels. Therefore, its applicability in extracting vegetation information from vegetation cover is relatively poor. Summary of the Invention

[0004] The purpose of this application is to provide a method and apparatus for inverting vegetation cover in hyperspectral remote sensing image data, which can demix mixed pixels in hyperspectral remote sensing images, improve the detection accuracy of end-member composition ratio, and thus improve the inversion accuracy of vegetation cover in hyperspectral remote sensing images.

[0005] In a first aspect, the present invention provides a method for inverting vegetation cover from hyperspectral remote sensing image data, comprising: acquiring hyperspectral remote sensing data and corresponding high-resolution remote sensing data from the same period within a target area; extracting and identifying clean endmembers contained in the hyperspectral remote sensing data to determine at least one type of clean endmember; superimposing the hyperspectral remote sensing data and high-resolution remote sensing data including clean endmembers to obtain the clean endmember category, clean endmember area, and clean endmember area ratio contained in the mixed pixels in the hyperspectral remote sensing data; training an initial mixed pixel unmixing model based on the clean endmember category, clean endmember area, and clean endmember area ratio contained in the mixed pixels to obtain a target mixed pixel unmixing model, and using the target mixed pixel unmixing model to invert the hyperspectral remote sensing image data within the area to be monitored to obtain the vegetation cover within the area to be monitored.

[0006] In an optional implementation, the extraction and identification of clean endmembers contained in the hyperspectral remote sensing data, and the determination of at least one type of clean endmember, includes: performing dimensionality reduction processing on the matrix data corresponding to the hyperspectral remote sensing data to obtain dimensionality-reduced remote sensing data; synthesizing a first simplex from the dimensionality-reduced remote sensing dataset, and projecting the first simplex onto a hyperplane to obtain a second simplex; the vertices of the second simplex correspond one-to-one with the vertices of the first simplex; orthogonally projecting the observation data corresponding to the second simplex onto a subspace composed of known endmembers according to preset projection conditions, and determining the largest projection vector as the extracted target endmember, repeating this step until the extracted clean endmember dataset is obtained; and performing clustering identification on the clean endmember dataset to determine at least one type of clean endmember.

[0007] In an optional implementation, clustering identification is performed on the clean endmember dataset to determine at least one type of clean endmember, including: determining a first number of endmember types contained in the clean endmember dataset based on the land cover types in the target area; performing clustering identification on the clean endmember dataset to determine a second number of clusters; calculating the sample density in each of the second number of clusters respectively, and determining the clusters whose sample density meets a preset threshold as target clusters; matching the endmember spectra of the endmembers in the target clusters with standard spectra to determine the land cover types of the endmembers, thereby determining at least one type of clean endmember.

[0008] In an optional implementation, the endmember spectra of the endmembers in the target cluster are matched with the standard spectra to determine at least one type of pure endmembers. This includes: matching the endmember spectra of the endmembers in the target cluster with the standard spectra to determine the endmember land cover types, and performing matching and identification based on UAV-captured images and / or high-resolution remote sensing data to determine at least one type of pure endmembers.

[0009] In an optional implementation, hyperspectral remote sensing data including clean endmembers and high-resolution remote sensing data are overlaid to obtain the clean endmember categories, clean endmember areas, and clean endmember area proportions contained in the mixed pixels of the hyperspectral remote sensing data. This includes: determining the data size of the hyperspectral remote sensing data including clean endmembers; matching the number of targets in the corresponding region with high-resolution remote sensing data based on the data size; overlaying the hyperspectral remote sensing data including clean endmembers and a preset number of high-resolution remote sensing data to obtain the clean endmember categories and corresponding clean endmember areas contained in the mixed pixels of the hyperspectral remote sensing data; and determining the clean endmember area proportion of the corresponding category of clean endmembers in the mixed pixels based on the clean endmember area of ​​each category of clean endmembers in the mixed pixels of the hyperspectral remote sensing data and the area of ​​a single pixel in the hyperspectral remote sensing data.

[0010] In an optional implementation, the mixed-pixel demixing model includes a least-squares support vector machine; the prediction function corresponding to the least-squares support vector machine is:

[0011]

[0012] In the formula, a i It is a Lagrange multiplier, x i x represents the pure endmember category, pure endmember area, and pure endmember area percentage of the endmembers in the i-th pixel. j σ represents the pure endmember category, pure endmember area, and pure endmember area percentage of the endmembers in the j-th pixel, where σ is the radial basis parameter and b is the bias.

[0013] The target mixed pixel demixing model is to use a two-layer grid search and cross-validation method to demix the Lagrange multiplier a. i The target model is obtained after parameter optimization of the radial basis parameter σ and the bias.

[0014] In an optional implementation, the vegetation cover of the monitoring area is obtained by inverting the hyperspectral remote sensing image data of the monitoring area using a target mixed pixel unmixing model, including: inverting the hyperspectral remote sensing image data of the monitoring area using a target mixed pixel unmixing model to obtain the vegetation endmember abundance of the monitoring area; and calculating the vegetation cover of the monitoring area based on the vegetation endmember abundance and endmember type.

[0015] Secondly, the present invention provides an inversion device for vegetation cover of hyperspectral remote sensing image data, comprising: a data acquisition module for acquiring hyperspectral remote sensing data and corresponding high-resolution remote sensing data of the same period in a target area; an endmember extraction and identification module for extracting and identifying clean endmembers contained in the hyperspectral remote sensing data to determine at least one type of clean endmember; a data overlay and matching module for overlaying the hyperspectral remote sensing data including clean endmembers and the high-resolution remote sensing data to obtain the clean endmember category, clean endmember area, and clean endmember area ratio contained in the mixed pixels in the hyperspectral remote sensing data; and an inversion module for training an initial mixed pixel unmixing model based on the clean endmember category, clean endmember area, and clean endmember area ratio contained in the mixed pixels to obtain a target mixed pixel unmixing model, so as to invert the hyperspectral remote sensing image data of the area to be monitored through the target mixed pixel unmixing model to obtain the vegetation cover of the area to be monitored.

[0016] Thirdly, the present invention provides an electronic device including a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the vegetation cover inversion method for hyperspectral remote sensing image data according to any of the foregoing embodiments.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are invoked and executed by a processor, the computer-executable instructions cause the processor to implement the vegetation cover inversion method for hyperspectral remote sensing image data according to any of the foregoing embodiments.

[0018] The vegetation cover inversion method and apparatus for hyperspectral remote sensing image data provided in this application use hyperspectral image as the main data source. By combining high-resolution image to determine the information of clean endmembers (including type, area, and area ratio), and then by performing artificial intelligence training on the initial mixed pixel unmixing model, the relationship between hyperspectral image data and endmembers in pixels is determined, thereby realizing the unmixing of mixed pixels in hyperspectral remote sensing image. This can improve the detection accuracy of endmember composition ratio, thereby improving the inversion accuracy of vegetation cover for hyperspectral remote sensing image. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a method for retrieving vegetation cover from hyperspectral remote sensing image data, provided as an embodiment of this application;

[0021] Figure 2 A diagram illustrating an endmember extraction process provided in an embodiment of this application;

[0022] Figure 3 A model optimization flowchart provided for an embodiment of this application;

[0023] Figure 4 A structural diagram of a vegetation cover inversion device for hyperspectral remote sensing image data provided in an embodiment of this application;

[0024] Figure 5 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0027] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0028] Vegetation refers to the total number of plants covering the ground in a given area at a certain density, such as forests, grasslands, shrubs, and crops, and is an important component of ecosystems. Vegetation cover, the percentage of the vertically projected area of ​​aboveground vegetation (including leaves, stems, and branches) to the total area of ​​a statistically analyzed region, is a key parameter for monitoring ecosystems and their functions. Vegetation cover and aboveground biomass are important indicators reflecting vegetation growth and are among the most effective indicators for assessing the degree of land desertification, widely used in agriculture, land use, disaster risk monitoring, and drought monitoring. In arid regions, researching how to improve the accuracy of large-area vegetation cover retrieval is crucial for maintaining ecosystem stability and protecting the environment.

[0029] Currently, there are two main methods for estimating vegetation cover: ground-based measurement and remote sensing inversion. Ground-based measurement is limited by factors such as manpower, time, weather, and terrain, and requires sophisticated equipment, making it difficult to implement over large areas. However, it yields highly accurate results and is generally used as validation data for remote sensing inversion algorithms. Remote sensing technology can acquire surface information at various temporal and spatial scales, offering advantages such as large coverage areas, short revisit cycles, and relatively low acquisition costs, making it possible to quickly and accurately obtain vegetation cover over large areas. Remote sensing inversion methods for vegetation cover primarily establish the relationship between vegetation spectral information and vegetation cover, mainly including three types: 1) Regression model method: Based on mathematical statistics theory, different mathematical regression models are established between remote sensing data surface reflectance and measured vegetation cover to invert vegetation cover. This method is clear in principle and simple to operate, but it requires a large amount of ground-based measurement data, and the models are often only applicable to specific regions and specific vegetation types, resulting in relatively poor universality. 2) Nonparametric model methods use more flexible algorithms and do not require assumptions about the normal or non-normal distribution of variables. They are more effective in improving the estimation accuracy of vegetation cover in desertified areas. Examples include artificial neural networks, random forests, support vector machines, and k-nearest neighbors. 3) Mixed pixel decomposition methods determine the information of each component within a pixel by establishing various relationships. The percentage of vegetation in all components is considered as vegetation cover. The components within a pixel are called endmembers, and their percentage in the mixed pixels is called abundance. This method is not dependent on measured data, but its estimation accuracy is highly correlated with image data and varies with the spatial resolution of the image. Medium-resolution satellite imagery has good temporal resolution and is widely used in large-scale vegetation cover studies. However, medium-resolution remote sensing images contain a large number of mixed pixels, especially in arid regions where most desertified vegetation cover is sparse. Vegetation mixes with spectral information from soil, shadows, etc., producing mixed pixels, which can easily lead to misclassification and misidentification of pixels. Existing vegetation cover extraction methods have poor universality when extracting vegetation information.

[0030] Based on this, embodiments of this application provide a method and apparatus for inverting vegetation cover in hyperspectral remote sensing image data, which realizes the demixing of mixed pixels in hyperspectral remote sensing images, can improve the detection accuracy of end-member composition ratio, and thus improve the inversion accuracy of vegetation cover in hyperspectral remote sensing images.

[0031] This application provides a method for inverting vegetation cover from hyperspectral remote sensing image data. See [link to relevant documentation]. Figure 1 As shown, the method mainly includes the following steps:

[0032] Step S110: Obtain hyperspectral remote sensing data and corresponding high-resolution remote sensing data within the target area during the same period.

[0033] Step S120: Extract and identify clean endmembers contained in the hyperspectral remote sensing data to determine at least one type of clean endmember.

[0034] Step S130: Overlay the hyperspectral remote sensing data including clean endmembers and the high-resolution remote sensing data to obtain the clean endmember categories, clean endmember areas, and clean endmember area proportions contained in the mixed pixels in the hyperspectral remote sensing data.

[0035] Step S140: The initial mixed pixel unmixing model is trained based on the pure endmember categories, pure endmember areas, and pure endmember area ratios contained in the mixed pixels to obtain the target mixed pixel unmixing model. The target mixed pixel unmixing model is then used to invert the hyperspectral remote sensing image data of the area to be monitored to obtain the vegetation coverage of the area to be monitored.

[0036] The vegetation cover inversion method for hyperspectral remote sensing image data provided in this application uses hyperspectral image as the main data source. By combining high-resolution image to determine the information of clean endmembers (including type, area, and area ratio), and then by performing artificial intelligence training on the initial mixed pixel unmixing model, the relationship between hyperspectral image data and endmembers in pixels is determined, thereby achieving unmixing of mixed pixels in hyperspectral remote sensing image. This can improve the detection accuracy of endmember composition ratio, thereby improving the inversion accuracy of vegetation cover for hyperspectral remote sensing image.

[0037] In an optional implementation, the above-described extraction and identification of clean endmembers contained in hyperspectral remote sensing data to determine at least one type of clean endmember may include the following steps 1.1) to 1.4):

[0038] Step 1.1) Dimensionality reduction is performed on the matrix data corresponding to the hyperspectral remote sensing data to obtain dimensionality-reduced remote sensing data.

[0039] Hyperspectral remote sensing data using matrix M (n×m) This indicates that n is the total number of pixels and m is the total number of bands. Matrix element x ji Let x be the spectral measure of pixel j in the i-th band. It can also be expressed as x = (x... j (j = 1, ..., n), where j is the pixel index and n is the total number of pixels. Here, i is the pixel spectral metric vector, i is the band index, and m is the total number of bands. Let be the spectral measure of pixel j in the i-th band.

[0040] In one implementation, satellite remote sensing data is represented by matrix M. (n×m)This indicates that when performing dimensionality reduction on hyperspectral remote sensing data, the goal is to reduce the dimension m to K. (K＜m) Then it may include the following steps (1) to (7):

[0041] (1) Generate an m×2K Gaussian test matrix Ω;

[0042] (2) Y = (MM) * ) q MΩ;

[0043] (3) Find the orthogonal matrix Q of Y;

[0044] (4) N = Q * M;

[0045] (5) Calculate the SVD decomposition (singular value decomposition) of N.

[0046] (6)

[0047] (7) Then the data matrix M' after dimensionality reduction (n×k) : M` = MU.

[0048] Step 1.2) Synthesize the first simplex from the dimensionality-reduced remote sensing dataset, and project the first simplex onto the hyperplane to obtain the second simplex; the vertices of the second simplex correspond one-to-one with the vertices of the first simplex.

[0049] Pure pixels are extracted using vertex component analysis, and the data matrix M' is assembled into a simplex (i.e., the first simplex) S. x ={x∈R k :x=Es,1 T s=1,s≥0}, where E∈R k×n It is an endmember matrix, s∈R n×1 This is the abundance vector of the mixed pixel, representing the proportion of endmembers in the mixed pixel, where each element value is 1. Setting 1 to an n×1 dimensional vector with all element values ​​being 1, the vertices of this simplex correspond to the endmember vectors. Let the simplex S... x Projected onto hyperplane x R u=1 forms a simplex S y ={y∈R k y = x / (x T u),x∈S x}, whose vertices intersect the simplex S x The vertices correspond one-to-one, ensuring that no observation vector is orthogonal to u. Here, the observation vector refers to the vector composed of the band values ​​of each hyperspectral pixel.

[0050] Step 1.3): According to the preset projection conditions, the observation data corresponding to the second simplex is orthogonally projected onto the subspace composed of known endmembers, and the largest projection vector is determined as the extracted target endmember. This step is repeated until the pure endmember dataset is obtained.

[0051] After determining the second simplex S y Then, the observed data can be orthogonally projected onto the subspace composed of known endmembers according to preset projection conditions. The largest projection vector is the endmember. These preset projection conditions can be a specified number of projections, that is, by repeatedly orthogonally projecting the observed vector corresponding to the second simplex onto the subspace composed of known endmembers, the largest projection vector extracted in each projection is obtained. This projection vector refers to the vector after the observed vector is projected onto the hyperplane. The final clean endmember dataset M`` is determined through multiple projections. See [link to documentation]. Figure 2 The diagram shows an illustration. In the diagram, f1 is a randomly generated initial vector; f2 is a vector orthogonal to the first endmember.

[0052] Step 1.4) involves clustering the clean endmember dataset to identify at least one type of clean endmember. In practice, this step may further include steps 1.4.1) to 1.4.4):

[0053] Step 1.4.1) Based on the land cover types in the target area, determine the first number of end-member types contained in the clean end-member dataset.

[0054] In one implementation, different regions have different land cover types, so end-member types can be determined based on land cover types. For example, some regions have a rich variety of land cover types such as grassland, woodland, and bare soil, so more end-member types are set accordingly. In other regions, most of the land may be arid with only a small portion containing wild grass, so fewer end-member types are set accordingly. In practical applications, the end-member types can be set according to the actual situation of the target region. In this embodiment, the first quantity is described as m.

[0055] Step 1.4.2) Perform cluster identification on the pure endmember dataset to determine the second number of clusters.

[0056] To facilitate accurate determination of endmember types, the second quantity can be set to a number greater than the first quantity. This allows for the selection of endmember types that match vegetation cover inversion after determining the second quantity of clusters. For example, if the target area includes non-vegetated areas such as bare soil and concrete surfaces, setting the second quantity to a number greater than the first allows for clustering all categories within the target area during clustering, followed by filtering out the categories corresponding to vegetation. This approach avoids the situation where insufficient clusters result in missing endmember categories, thus affecting subsequent vegetation cover calculations.

[0057] In one implementation, assuming the second quantity is 2m, 2m cluster centers can be selected one by one, including the following steps (1) to (7):

[0058] (1) Randomly select a sample point from the pure endmember dataset M`` as the first initial cluster center c. p ;

[0059] (2) Calculate the shortest distance, D(x), from each sample point to the existing cluster center. For example, sample point x... j Currently, P cluster centers have been identified;

[0060]

[0061] (3) Calculate the probability P(x) that each sample point is selected as the next cluster center;

[0062]

[0063] (4) Repeat steps (2) and (3) until 2m cluster centers are selected;

[0064] (5) For each sample x in the dataset j Calculate its distance to the 2m cluster centers and assign it to the class corresponding to the cluster center with the smallest distance;

[0065] (6) For each category c i Recalculate its cluster centers (i.e., the centroid of all samples belonging to this class);

[0066] (7) Repeat steps 5) and 6) until the iteration count is 20 or the cluster center position no longer changes;

[0067] The result is 2m clusters.

[0068] Step 1.4.3) Calculate the sample density in each of the second number of clusters, and determine the clusters whose sample density meets the preset threshold as the target clusters.

[0069] Sample density is the average distance from each sample point in a cluster to the cluster center. The lower the density, the more concentrated the sample points within the cluster, the shorter the clusters between sample points, and in other words, the higher the similarity within the cluster.

[0070] The aforementioned preset threshold can be m. The value m can correspond to the land cover types in the corresponding area, sort them according to their density, and select the m clusters with the lowest density as the target clusters.

[0071] Step 1.4.4) Match the endmember spectra of the endmembers in the target cluster with the standard spectra to determine at least one type of pure endmember.

[0072] Samples are extracted from each cluster, and endmember spectra are compared with standard spectra using a cross-correlation spectral matching method to identify endmember land cover types (grassland, woodland, bare soil). Manual verification is then performed using UAV or satellite high-resolution data.

[0073] The degree of similarity between two spectra is determined by calculating the cross-correlation coefficients of the endmember spectra and the standard spectra at different matching positions. The cross-correlation coefficient between the endmember spectra and the standard spectra at each matching position is equal to the covariance between the two spectra divided by the product of their respective variances.

[0074] Furthermore, to improve the accuracy of pure endmember identification, the endmember spectra of the endmembers in the target cluster are matched with the standard spectra to determine at least one type of pure endmember. This can be achieved by matching the endmember spectra of the endmembers in the target cluster with the standard spectra, and by matching and identifying based on UAV images and / or high-resolution remote sensing data to determine at least one type of pure endmember.

[0075] Furthermore, the hyperspectral remote sensing data including clean endmembers and the high-resolution remote sensing data are overlaid to obtain the clean endmember categories, clean endmember areas, and clean endmember area proportions contained in the mixed pixels of the hyperspectral remote sensing data. This can include the following steps 2.1) to 2.4):

[0076] Step 2.1) Determine the data size of the hyperspectral remote sensing data, including the pure endmembers.

[0077] Considering that the resolution of actual hyperspectral remote sensing data is usually not very high, the vegetation information may be quite blurry. Therefore, to obtain accurate vegetation cover information, we can first determine the data size of the hyperspectral remote sensing data, which is also the resolution of the remote sensing data. For example, when the resolution of the hyperspectral remote sensing data used is 30m, 50m, etc., that is, one pixel of the hyperspectral remote sensing data represents an area of ​​30m×30m, 50m×50m, etc., the ground vegetation information is difficult to identify from the data, or the identification is inaccurate. Therefore, subsequent steps are used to improve accuracy.

[0078] Step 2.2) Match the number of targets in the corresponding area using high-resolution remote sensing data based on the data size.

[0079] For ease of explanation, we will use a 50m×50m dataset as an example. Assuming the data size is 50m×50m, to accurately identify the type of each endmember, we can match the corresponding high-resolution remote sensing data (i.e., high-resolution remote sensing data) for this 50m×50m dataset. If the matched high-resolution remote sensing data is 1m×1m, then 2500 1m×1m high-resolution remote sensing data points are needed for matching.

[0080] Step 2.3) Overlay the hyperspectral remote sensing data including clean endmembers with a preset number of high-resolution remote sensing data to obtain the clean endmember categories and corresponding clean endmember areas of the mixed pixels in the hyperspectral remote sensing data.

[0081] By overlaying hyperspectral remote sensing data including clean endmembers with a preset number of high-resolution remote sensing data, the originally unclear pixels in the hyperspectral remote sensing data can be made clearer, making it easier to obtain information on each endmember in the region, thereby determining the type of each clean endmember and the area of ​​the corresponding clean endmember, which facilitates the subsequent acquisition of vegetation cover information.

[0082] Pure end-member types are end-members that contain only one type of land feature information, such as end-members that contain only grassland, end-members that contain only woodland, end-members that contain only shrubs, end-members that contain only bare soil, and so on.

[0083] The area of ​​clean endmembers can be determined based on hyperspectral remote sensing data. When the resolution of the hyperspectral remote sensing data is determined, the area of ​​each clean endmember can also be identified and obtained.

[0084] Step 2.4) Determine the percentage of pure endmember area of ​​each category of pure endmember in the mixed pixel based on the pure endmember area of ​​each category in the hyperspectral remote sensing data and the area of ​​a single pixel in the hyperspectral remote sensing data.

[0085] In one implementation, after determining the area of ​​each type of pure endmember, the area ratio of pure units can be determined based on the area of ​​individual pixels in the hyperspectral remote sensing data and the area of ​​a certain type of pure unit. For example, the pure endmember area ratio = area of ​​a certain type of pure endmember / area of ​​the hyperspectral mixed pixel.

[0086] Furthermore, after obtaining the clean endmember categories, clean endmember areas, and clean endmember area proportions contained in the mixed pixels of the hyperspectral remote sensing data, these parameters can be used as training data to train the deep learning model until convergence. The converged, well-trained model is the target mixed pixel demixing model of this application. For example, the deep learning model can be a support vector machine, a least squares support vector machine, or a convolutional neural network.

[0087] In one implementation, the mixed-pixel unmixing model can be a least-squares support vector machine. Model training may include the following steps (1) to (3):

[0088] (1) Constructing the dataset

[0089] The hyperspectral data and the endmember dataset extracted and identified after processing in steps 2.1) to 2.4 above are integrated into a training set (y). j ,x j (j = 1, 2, ..., p), where p represents the number of training samples, y j Let x be the spectral value of the j-th pixel in the hyperspectral data of the training samples. j It represents the type, area, and proportion of endmembers in the j-th pixel.

[0090] (2) Establish a neural network LSSVM (Least Squares Support Vector Machine)

[0091] The linear function for constructing the Least Squares Support Vector Machine (LSSVM) is:

[0092]

[0093] In the formula, w is the weight vector, and b is the bias. Let f(x) be a nonlinear mapping function. x represents the dataset containing the types, areas, and proportions of endmembers, and f(x) represents the hyperspectral pixel value dataset. Based on the principle of minimizing structural risk, and considering both fitting error and algorithm complexity, the problem is transformed into an optimization problem with equality constraints:

[0094]

[0095] In the formula, d is the error variable, d is the adjustment parameter factor, p is the number of training samples, and the sample refers to the sample in LSSVM.

[0096] To solve the above optimization problem, we introduce the Lagrange equation, a i These are Lagrange multipliers, which transform constrained optimization problems into unconstrained optimization problems, as shown in the following equation:

[0097]

[0098] The model parameters a are obtained based on the conditions of nonlinear optimal programming (Karush-Kuhn-Tucker, KKT). i Substituting b into the linear function of the high-dimensional feature space of LSSVM, we obtain the output of the support vector machine:

[0099]

[0100] Finally, the radial basis function is chosen as the kernel function of LSSVM to obtain the prediction function of LSSVM:

[0101]

[0102] In the formula, σ is the radial basis parameter, which reflects the characteristics of the hyperspectral pixel dataset.

[0103] (3) Model Optimization

[0104] In model optimization, a two-layer grid search technique and cross-validation method are combined to optimize the parameters, and the training experiments are also optimized to finally obtain the inversion model parameters a. i b, radial basis parameter σ, and penalty factor See the model optimization flowchart. Figure 3 As shown, by judging whether the parameters meet the optimization requirements in the two-layer grid search, the RMSE of the parameter set in the two-layer grid can be minimized, thereby minimizing the error and making the prediction results more accurate.

[0105] 1) Parameter optimization

[0106] A two-layer grid search technique is employed. First, the approximate parameter range is roughly determined in the first layer, followed by a more detailed search in the second layer. RMSE is used as the model's prediction error, and 10-fold cross-validation is performed on each parameter pair to minimize the model's RMSE, as shown in the following equation:

[0107]

[0108] In the formula, p is the number of training samples, and y i For sample pixel values, For y i The predicted pixel value.

[0109] The optimal parameter combination (γ,σ) is determined with the goal of minimizing the RMSE of the model.

[0110] 2) Optimize training experiments

[0111] In LSSVM, the absolute value of a support value corresponds to its contribution to the regression. Based on this correspondence, training samples corresponding to support values ​​with small absolute values ​​can be removed from the support vectors. If the absolute value exceeds a threshold, the corresponding penalty coefficient is adjusted to limit its impact. The rules for optimizing training are shown in the following formula:

[0112]

[0113] Where b1 is the rejection threshold, and vectors α below this threshold are excluded. i′ b1 is removed; b2 and b3 are adjustment thresholds.

[0114] After model optimization, the final model parameters a are obtained. i b, radial basis parameters σ, and the penalty factor for support vectors

[0115] After the above training and optimization, the obtained target mixed pixel demixing model is obtained by using a two-layer grid search and cross-validation method to optimize the Lagrange multiplier a. i The target model is obtained after parameter optimization of the radial basis parameter σ and the bias.

[0116] Furthermore, the vegetation cover of the monitored area is obtained by inverting the hyperspectral remote sensing image data within the monitored area using a target mixed pixel unmixing model. In practical applications, only a hyperspectral remote sensing image needs to be input to obtain the vegetation cover of the area where the hyperspectral remote sensing image is located.

[0117] In summary, the vegetation cover retrieval method for hyperspectral remote sensing image data provided in this application uses hyperspectral imagery as the primary data source. By combining high-resolution imagery and UAV imagery, the number of endmembers and spectral values ​​are determined. Then, artificial intelligence training is used to determine the relationship model between hyperspectral imagery data and endmembers in pixels, achieving demixing of mixed pixels in medium-resolution satellite imagery. By comprehensively applying data from multiple data sources for endmember determination, spectral value determination, and training data calibration, accuracy is improved, and the labor costs and measurement errors of manual ground measurements are reduced.

[0118] This case utilizes dimensionality reduction processing of hyperspectral data to decrease the data volume and improve computational efficiency. High-resolution imagery and UAV imagery are used for training sample annotation, significantly reducing on-site manual measurement workload and errors while ensuring annotation accuracy, thus enabling the acquisition of more training samples.

[0119] By establishing an artificial intelligence model based on LSSVM, we can avoid the shortcomings of existing models and establish a more reasonable relationship model of the endmember ratio in the mixed pixels. This can improve the accuracy of the endmember composition ratio and thus improve the inversion accuracy of vegetation cover on a large scale.

[0120] The data acquisition and processing costs are low, and models can be trained for typical areas. In model applications, different models can be automatically called for inversion based on the area type, enabling the acquisition of high-precision vegetation cover products over a large area at low cost.

[0121] Based on the above method embodiments, this application also provides an inversion device for vegetation cover of hyperspectral remote sensing image data, see [link to relevant documentation]. Figure 4 As shown, the device mainly includes the following parts:

[0122] Data acquisition module 410 is used to acquire hyperspectral remote sensing data and corresponding high-resolution remote sensing data of the same period within the target area;

[0123] The endmember extraction and identification module 420 is used to extract and identify clean endmembers contained in hyperspectral remote sensing data, and to determine at least one type of clean endmember;

[0124] The data overlay matching module 430 is used to overlay hyperspectral remote sensing data including clean endmembers and high-resolution remote sensing data to obtain the clean endmember categories, clean endmember areas and clean endmember area proportions contained in the mixed pixels in the hyperspectral remote sensing data.

[0125] The inversion module 440 is used to train the initial mixed pixel unmixing model based on the pure endmember category, pure endmember area and pure endmember area ratio contained in the mixed pixel to obtain the target mixed pixel unmixing model, so as to invert the hyperspectral remote sensing image data in the area to be monitored through the target mixed pixel unmixing model to obtain the vegetation coverage in the area to be monitored.

[0126] The vegetation cover inversion device for hyperspectral remote sensing image data provided in this application uses hyperspectral image as the main data source, combines it with high-resolution image to determine the information of clean endmembers (including type, area and area ratio), and then uses artificial intelligence training on the initial mixed pixel unmixing model to determine the relationship between hyperspectral image data and endmembers in pixels, thereby achieving unmixing of mixed pixels in hyperspectral remote sensing image, improving the detection accuracy of endmember composition ratio, and thus improving the inversion accuracy of vegetation cover for hyperspectral remote sensing image.

[0127] In a feasible implementation, the aforementioned endmember extraction and identification module 420 is further configured to: perform dimensionality reduction processing on the matrix data corresponding to the hyperspectral remote sensing data to obtain dimensionality-reduced remote sensing data; synthesize the dimensionality-reduced remote sensing dataset into a first simplex, and project the first simplex onto a hyperplane to obtain a second simplex; the vertices of the second simplex correspond one-to-one with the vertices of the first simplex; orthogonally project the observation data corresponding to the second simplex onto a subspace composed of known endmembers according to preset projection conditions, and determine the largest projection vector as the extracted target endmember, repeating this step until a pure endmember dataset is obtained; perform clustering identification on the pure endmember dataset to determine at least one type of pure endmember.

[0128] In a feasible implementation, the endmember extraction and identification module 420 is further configured to: determine a first number of endmember types contained in the clean endmember dataset based on the land cover types in the target area; perform cluster identification on the clean endmember dataset to determine a second number of clusters; calculate the sample density in each of the second number of clusters respectively, and determine the clusters whose sample density meets a preset threshold as target clusters; and match the endmember spectra of the endmembers in the target clusters with the standard spectra to determine at least one type of clean endmember.

[0129] In one feasible implementation, the aforementioned endmember extraction and identification module 420 is further configured to: match the endmember spectra of the endmembers in the target cluster with the standard spectra to determine the endmember land cover type, and perform matching and identification based on UAV-captured images and / or high-resolution remote sensing data to determine at least one type of pure endmember.

[0130] In one feasible implementation, the data overlay matching module 430 is further configured to: determine the data size of the hyperspectral remote sensing data including clean endmembers; match the number of high-resolution remote sensing data of targets within the corresponding region based on the data size; overlay the hyperspectral remote sensing data including clean endmembers and a preset number of high-resolution remote sensing data to obtain the clean endmember categories and corresponding clean endmember areas contained in the mixed pixels of the hyperspectral remote sensing data; and determine the proportion of the clean endmember area of ​​the corresponding category of clean endmembers in the mixed pixels based on the clean endmember area of ​​each category of clean endmembers in the mixed pixels of the hyperspectral remote sensing data and the area of ​​a single pixel in the hyperspectral remote sensing data.

[0131] In one feasible implementation, the mixed-pixel demixing model includes a least-squares support vector machine; the prediction function corresponding to the least-squares support vector machine is:

[0132]

[0133] In the formula, a i It is a Lagrange multiplier, x i x represents the pure endmember category, pure endmember area, and pure endmember area percentage of the endmembers in the i-th pixel. j σ represents the pure endmember category, pure endmember area, and pure endmember area percentage of the endmembers in the j-th pixel, where σ is the radial basis parameter and b is the bias.

[0134] The target mixed pixel demixing model is to use a two-layer grid search and cross-validation method to demix the Lagrange multiplier a. i The target model is obtained after parameter optimization of the radial basis parameter σ and the bias.

[0135] In a feasible implementation, the aforementioned inversion module 440 is further configured to: invert hyperspectral remote sensing image data within the monitoring area using a target mixed pixel unmixing model to obtain vegetation endmember abundance within the monitoring area; and calculate vegetation coverage within the monitoring area based on vegetation endmember abundance and endmember type.

[0136] The vegetation cover inversion device for hyperspectral remote sensing image data provided in this application has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the embodiment of the vegetation cover inversion device for hyperspectral remote sensing image data can be referred to the corresponding content in the aforementioned embodiment of the vegetation cover inversion method for hyperspectral remote sensing image data.

[0137] This application also provides an electronic device, such as... Figure 5 The diagram shows the structure of the electronic device 100, which includes a processor 51 and a memory 50. The memory 50 stores computer-executable instructions that can be executed by the processor 51. The processor 51 executes the computer-executable instructions to implement any of the above-mentioned methods for inverting vegetation cover in hyperspectral remote sensing image data.

[0138] exist Figure 5 In the illustrated embodiment, the electronic device further includes a bus 52 and a communication interface 53, wherein the processor 51, the communication interface 53, and the memory 50 are connected via the bus 52.

[0139] The memory 50 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 53 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 52 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 52 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0140] The processor 51 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 51 or by instructions in software form. The processor 51 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory. The processor 51 reads the information in the memory and, in conjunction with its hardware, completes the steps of the vegetation cover inversion method for hyperspectral remote sensing image data in the aforementioned embodiment.

[0141] This application also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the above-described method for inverting vegetation cover in hyperspectral remote sensing image data. For specific implementation details, please refer to the foregoing method embodiments, which will not be repeated here.

[0142] The computer program product of the vegetation cover inversion method and apparatus for hyperspectral remote sensing image data provided in this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0143] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0144] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0145] In the description of this application, it should be noted that the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for retrieving vegetation coverage from hyperspectral remote sensing image data, characterized in that, The method comprises the following steps: acquiring hyperspectral remote sensing data and corresponding high-resolution remote sensing data in a target region at the same period; extracting and identifying pure end members contained in the hyperspectral remote sensing data, and determining at least one type of pure end member; superimposing the hyperspectral remote sensing data containing pure end members and the high-resolution remote sensing data to obtain the pure end member category, pure end member area and pure end member area proportion contained in the mixed pixels in the hyperspectral remote sensing data; training an initial mixed pixel unmixing model based on the pure end member category, pure end member area and pure end member area proportion contained in the mixed pixels to obtain a target mixed pixel unmixing model, so as to invert the hyperspectral remote sensing image data in the monitoring region by using the target mixed pixel unmixing model to obtain the vegetation coverage in the monitoring region; extracting and identifying pure end members contained in the hyperspectral remote sensing data, and determining at least one type of pure end member, which comprises the following steps: performing dimension reduction processing on the matrix data corresponding to the hyperspectral remote sensing data to obtain reduced dimension remote sensing data; 2. The method for retrieving vegetation fraction from hyperspectral remote sensing image data according to claim 1, wherein, grouping the reduced dimension remote sensing data into a first simplex, and projecting the first simplex to a hyperplane to obtain a second simplex; the vertices of the second simplex correspond to the vertices of the first simplex one by one; orthogonally projecting the observation data corresponding to the second simplex to a subspace formed by known end members according to a preset projection condition, and determining the maximum projection vector as the extracted target end member; repeating the step until the pure end member data set is extracted; clustering and identifying the pure end member data set to determine at least one type of pure end member. Clustering and identifying the pure end member data set to determine at least one type of pure end member, which comprises the following steps: determining the first number of end member types contained in the pure end member data set based on the ground object types in the target region; clustering and identifying the pure end member data set to determine the second number of clustering clusters; calculating the sample tightness in each clustering cluster in the second number of clustering clusters respectively, and determining the clustering cluster whose sample tightness meets a preset threshold as a target clustering cluster; 3. The method of claim 2, wherein, matching the end member spectrum of the end member in the target clustering cluster with the standard spectrum to determine at least one type of pure end member. Matching the end member spectrum of the end member in the target clustering cluster with the standard spectrum to determine at least one type of pure end member, which comprises the following steps:

4. The method of claim 1, wherein, matching the end member spectrum of the end member in the target clustering cluster with the standard spectrum, and performing matching and identification based on the unmanned aerial vehicle image and / or high-resolution remote sensing data to determine at least one type of pure end member. Superimposing the hyperspectral remote sensing data containing pure end members and the high-resolution remote sensing data to obtain the pure end member category, pure end member area and pure end member area proportion contained in the mixed pixels in the hyperspectral remote sensing data, which comprises the following steps: determining the data size of the hyperspectral remote sensing data containing pure end members; matching the high-resolution remote sensing data of the target number in the corresponding region based on the data size. superimpose the hyperspectral remote sensing data comprising the pure end members and the high-resolution remote sensing data of the preset number to obtain a pure end member category contained in the mixed pixel in the hyperspectral remote sensing data and a pure end member area of the corresponding category; determine a pure end member area proportion of the pure end member of the corresponding category in the mixed pixel based on the pure end member area of each category of pure end member in the hyperspectral remote sensing data mixed pixel and a single pixel area in the hyperspectral remote sensing data.

5. The method of claim 1, wherein, The mixed pixel unmixing model comprises a least squares support vector machine; a prediction function corresponding to the least squares support vector machine is: In the formula, is a Lagrange multiplier, is the first pure endmember category, pure endmember area and pure endmember area proportion in the first pure endmember category, pure endmember area and pure endmember area proportion in the first pure endmember category, pure endmember area and pure endmember area proportion in the first is a radial basis parameter, b is a bias amount; The target mixed pixel unmixing model is a target model obtained by parameter optimization of the Lagrange multipliers , radial basis parameters and bias amounts through two-layer grid search and cross-validation method.

6. The method of claim 1, wherein, inverse the hyperspectral remote sensing image data in the to-be-monitored region through the target mixed pixel unmixing model to obtain the vegetation coverage in the to-be-monitored region, comprising: inverse the hyperspectral remote sensing image data in the to-be-monitored region through the target mixed pixel unmixing model to obtain the vegetation end member abundance in the to-be-monitored region; calculate the vegetation coverage in the to-be-monitored region based on the vegetation end member abundance and the end member type.

7. An apparatus for retrieving vegetation coverage from hyperspectral remote sensing image data, characterized in that, comprising: a data acquisition module configured to acquire contemporaneous hyperspectral remote sensing data and corresponding high-resolution remote sensing data in a target region; an end member extraction and identification module configured to extract and identify pure end members contained in the hyperspectral remote sensing data to determine at least one type of pure end member; a data superimposition and matching module configured to superimpose the hyperspectral remote sensing data comprising the pure end members and the high-resolution remote sensing data to obtain a pure end member category contained in the mixed pixel in the hyperspectral remote sensing data, a pure end member area, and a pure end member area proportion; an inversion module configured to train an initial mixed pixel unmixing model based on the pure end member category contained in the mixed pixel, the pure end member area, and the pure end member area proportion to obtain a target mixed pixel unmixing model, so as to inverse the hyperspectral remote sensing image data in a to-be-monitored region through the target mixed pixel unmixing model to obtain the vegetation coverage in the to-be-monitored region; The end member extraction and identification module is specifically configured to perform dimension reduction processing on matrix data corresponding to the hyperspectral remote sensing data to obtain reduced dimension remote sensing data; group the reduced dimension remote sensing data into a first simplex, and project the first simplex to a hyperplane to obtain a second simplex; the vertices of the second simplex correspond one by one to the vertices of the first simplex; orthogonally project observation data corresponding to the second simplex to a subspace formed by known end members according to a preset projection condition, and determine the largest projection vector as an extracted target end member, and repeat the step until a pure end member data set is extracted; cluster and identify the pure end member data set to determine at least one type of pure end member.

8. An electronic device, comprising: comprise a processor and a memory, the memory stores computer executable instructions capable of being executed by the processor, and the processor executes the computer executable instructions to implement the inverse method for vegetation coverage of the hyperspectral remote sensing image data according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions, when the computer executable instructions are invoked and executed by the processor, the computer executable instructions cause the processor to implement the method for retrieving the vegetation coverage of the hyperspectral remote sensing image data according to any one of claims 1 to 6.

Images