Plant target identification method based on multi-source remote sensing image fusion

Abstract

The invention relates to the technical field of remote sensing image processing and target identification, in particular to a plant target identification method based on multi-source remote sensing image fusion, which comprises the following steps of: S1, acquiring a hyperspectral remote sensing image and a synthetic aperture radar (SAR) image of the same observation area, and preprocessing the hyperspectral remote sensing image and the SAR image; s2, extracting a spectral feature principal component graph; s3, extracting a scattering characteristic component graph; s4, performing pixel-level fusion to generate a multi-source feature fusion image; s5, performing superpixel segmentation to generate a plurality of homogeneous region blocks; and S6, matching is carried out, and the plant category of each homogeneous region block is identified. According to the method, the multi-source feature map fusing the spectrum and the scattering features is constructed, and superpixel segmentation and knowledge base matching are combined, so that high-precision region-level identification of the plant target is realized, and the boundary fitting degree of an identification result is improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing and target recognition technology, and in particular to a method for plant target recognition based on multi-source remote sensing image fusion. Background Technology

[0002] With the development of remote sensing technology, hyperspectral remote sensing images and synthetic aperture radar images have become important data sources for vegetation monitoring and classification. Hyperspectral images can provide detailed spectral information in continuous bands, which can be used to characterize the reflection characteristics of plants under different electromagnetic wave bands. SAR images can obtain the structural scattering characteristics of ground objects under all-weather conditions, and are particularly sensitive to the geometric structure and scattering mechanism of vegetation in multi-polarization mode.

[0003] However, existing technologies still have significant shortcomings in plant target identification: on the one hand, hyperspectral images have high dimensionality and large data redundancy, resulting in low information extraction efficiency and susceptibility to noise; on the other hand, although SAR images can provide structural information, their scattering mechanisms are complex, making it difficult to accurately distinguish different plant categories when used alone; simultaneously, there are radiometric differences and geometric shifts between multi-source images, and without effective registration and feature fusion methods, inaccurate information alignment can easily occur, leading to unstable identification results. Therefore, it is necessary to propose a plant target identification method based on multi-source remote sensing image fusion to address the shortcomings of existing technologies. Summary of the Invention

[0004] To achieve the above objectives, this invention provides a plant target recognition method based on multi-source remote sensing image fusion.

[0005] A plant target recognition method based on multi-source remote sensing image fusion includes the following steps: S1: Acquire hyperspectral remote sensing images and synthetic aperture radar (SAR) images of the same observation area, and perform radiometric calibration and geometric registration preprocessing on the hyperspectral remote sensing images and SAR images to obtain the registered hyperspectral images and registered SAR images. S2: Perform spectral feature dimensionality reduction processing on the registered hyperspectral image to extract the principal component map of spectral features that characterize the core spectral information of ground features; S3: Perform polarimetric target decomposition on the registered SAR image to extract scattering feature component maps that characterize the scattering properties of ground structures; S4: Perform pixel-level fusion of the spectral feature principal component map and the scattering feature component map to generate a multi-source feature fusion map that simultaneously contains spectral and scattering information; S5: A linear iterative clustering algorithm is used to perform superpixel segmentation on the multi-source feature fusion map, generating multiple homogeneous region blocks with boundaries that fit the distribution of ground features. Each homogeneous region block contains spectral-scattering integrated features. S6: Match the spectral-scattering integrated features of each homogeneous region block with the pre-established plant target spectral-scattering feature knowledge base, identify the plant category of each homogeneous region block based on the matching results, and output the final plant target category map.

[0006] Optionally, S1 specifically includes: S11: Retrieve hyperspectral remote sensing satellite images and synthetic aperture radar images covering the same observation area through the remote sensing satellite image acquisition interface; S12: Perform radiometric calibration on the acquired hyperspectral image to convert the original digital quantization values ​​into apparent reflectance values ​​of ground features; S13: Perform radiometric calibration on the SAR image, convert the square of the echo amplitude value into the backscattering coefficient, and uniformly convert it into a logarithmic scale amplitude map according to the polarization channel type; S14: Using the hyperspectral image as the main control reference image, based on the ground control points in the two types of images, perform affine transformation and interpolation resampling of the SAR image to complete the geometric registration process, and output the registered hyperspectral image and the registered SAR image.

[0007] Optionally, S14 specifically includes: S141: Identify common targets with clear spatial locations in hyperspectral and SAR images as ground control points, extract their pixel coordinates in the two images, and construct a one-to-one set of control points. S142: Based on the set of control points, fit the affine transformation model from SAR image to hyperspectral image using the least squares method; S143: Perform coordinate transformation on all pixels in the SAR image according to the affine model, and calculate the gray value of the pixels that fall at non-integer coordinate positions after transformation using the bilinear interpolation method. S144: The SAR image after transformation and interpolation is resampled to the same spatial size and coordinate system as the hyperspectral image, thus obtaining the registered SAR image.

[0008] Optionally, S2 specifically includes: S21: Expand the registered hyperspectral image into a two-dimensional spectral matrix by pixel. ; S22: For two-dimensional spectral data matrix Mean centering is performed to eliminate mean deviations between bands. The mean vector is calculated, and the mean of the corresponding band is subtracted from each sample to obtain the centered matrix. ; S23: For the centralized matrix Perform principal component analysis to construct the covariance matrix and solve for its eigenvalues ​​and eigenvectors; S24: Sort by feature value from largest to smallest, and select the top performers whose cumulative contribution rate exceeds a set threshold. Using eigenvectors, construct a transformation matrix, and then convert the original two-dimensional spectral matrix... Projecting onto the principal component space yields the dimensionality-reduced principal component map. .

[0009] Optionally, S3 specifically includes: S31: Construct a complex polarization scattering matrix from the polarization channel data in the registered SAR image; S32: Construct the Pauli vector representation based on the scattering matrix, and calculate the coherence matrix within a sliding window. ; S33: For the coherence matrix Eigenvalue decomposition is performed to extract the eigenvalues ​​corresponding to the main polarization scattering mechanisms, and the power proportion of each scattering mechanism is calculated. S34: Encode the power values ​​of the three scattering mechanisms corresponding to each pixel by channel to form a three-channel scattering feature component map, with each channel representing the surface scattering map, volume scattering map, and double reflection scattering map, respectively.

[0010] Optionally, S4 specifically includes: S41: Normalize each channel of the spectral feature principal component map and the scattering feature component map to uniformly map the feature values ​​from different sources in the image to the standard range. S42: While maintaining the pixel spatial alignment of the original image, a pixel-by-pixel stitching method is used to fuse and stitch the normalized spectral feature vector and the scattering feature vector to construct a fused multi-source feature vector image.

[0011] Optionally, S42 specifically includes: S421: Obtain the normalized principal component map of spectral features and the scattering feature map, and represent them as two-dimensional matrix forms, where the spectral feature map is denoted as a matrix. , dimension The scattering feature map is denoted as a matrix. , dimension ,in Image size, , These are the spectral feature dimension and the scattering feature dimension, respectively. S422: For each pixel location in the image, extract the spectral feature vector corresponding to that location. With scattering eigenvectors They are then concatenated along the vector dimension to form a multi-source fusion feature vector. The expression is: ,in, This represents a vector concatenation operation; S423: Recombines the fused feature vectors at all pixel locations into a three-dimensional matrix, with an output dimension of... The fused image.

[0012] Optionally, S5 specifically includes: S51: Construct each cell in the multi-source feature fusion map into a five-dimensional feature vector containing spatial coordinates and fused feature vectors; S52: Set the initial number of cluster centers according to the image size, and uniformly initialize the positions of cluster centers on the image; S53: A linear iterative clustering algorithm is used to calculate the weighted distance between the cell and the neighboring cluster centers in each iteration, and the cell is assigned to the class corresponding to the cluster center with the smallest distance. S54: Update the spatial location and feature mean of each cluster center, repeat the iteration until the cluster centers converge, and finally generate multiple homogeneous region blocks with boundaries that fit the distribution of ground features.

[0013] Optionally, S53 specifically includes: S531: For each cluster center in the current round, select a predetermined range of pixels within its spatial neighborhood as a candidate set; S532: Calculate the spatial distance and feature distance between the candidate pixel and the current cluster center, and then combine the two with a weighted sum to form the overall distance metric. ; S533: For each cell, iterate through its nearest cluster centers and select the one with the smallest weighted distance. Using the cluster centers as the target, pixels are labeled with the category corresponding to the cluster centers.

[0014] Optionally, S6 specifically includes: S61: For each homogeneous region block, extract the spectral-scattering fusion feature vector of all pixels within it, and calculate the average vector of the fusion features within the region; S62: Call the pre-established plant target spectral-scattering feature knowledge base. The knowledge base stores a set of standard fused feature vectors corresponding to several plant categories. Each category corresponds to one or more normalized reference feature vectors. S63: Feature vector for the current region block Calculate the similarity score between it and the reference feature vector of each category; S64: Among all category similarity scores, select the category label with the highest score as the recognition result of the current region block, and assign the label to all pixels in the corresponding region to complete the category labeling of homogeneous regions. Finally, output the plant target category map of the entire image.

[0015] The beneficial effects of this invention are: This invention performs radiometric calibration and geometric registration on hyperspectral and SAR images. While ensuring the consistency of physical quantities and spatial alignment accuracy of multi-source images, it uses principal component analysis to extract principal components of spectral features and polarization decomposition to extract scattering feature components, achieving complementary use of spectral and scattering information and constructing a unified multi-source feature fusion map. Furthermore, it combines linear iterative clustering to perform superpixel segmentation, enabling the generated region blocks to fit the boundaries of ground features and maintain the consistency of internal features.

[0016] This invention achieves accurate identification of regional-level plant targets by matching the spectral-scattering fusion features of regional blocks with the pre-defined standard features of plant categories using a knowledge base matching method. This method not only improves the accuracy and stability of plant classification results but also reduces pixel-level noise interference. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the plant target identification method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a superpixel segmentation method according to an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0020] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0021] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0022] like Figures 1-2 As shown, the plant target recognition method based on multi-source remote sensing image fusion includes the following steps: S1: Acquire hyperspectral remote sensing images and synthetic aperture radar (SAR) images of the same observation area, and perform radiometric calibration and geometric registration preprocessing on the hyperspectral remote sensing images and SAR images to obtain the registered hyperspectral images and registered SAR images. S1 specifically includes: S11: Through the remote sensing satellite image acquisition interface, retrieve hyperspectral remote sensing satellite images and synthetic aperture radar (SAR) images covering the same observation area respectively. Hyperspectral images have continuous narrow band spectral information, and SAR images have structural information under multi-polarization mode. Both types of images are limited to multi-source remote sensing image data with the same time window and spatial coverage. S12: Perform radiometric calibration on the acquired hyperspectral image, converting the original digital quantization (DN) values ​​into apparent reflectance values ​​of ground features. The conversion formula is as follows: ,in, Indicates the reflectivity value. Indicates radiance. Indicates the Earth-Sun distance. Indicates solar irradiance, Indicates the solar zenith angle; S13: Perform radiometric calibration on the SAR image, converting the squared echo amplitude values ​​into backscattering coefficients, and then uniformly converting them into logarithmic scale amplitude maps according to the polarization channel type; the corresponding conversion and calculation formulas are as follows: ; ;in, Represents the backscattering coefficient. This represents the original echo amplitude value. As the scaling factor, The backscattering value is on a logarithmic scale; S14: Using the hyperspectral image as the main control reference image, based on the ground control points in the two types of images, perform affine transformation and interpolation resampling of the SAR image to complete the geometric registration process, and output the registered hyperspectral image and the registered SAR image. Through the above steps, it can be ensured that the acquired multi-source remote sensing data has a unified spatial resolution and physical quantity meaning, laying a highly consistent input foundation for the subsequent fusion of spectral and scattering features, thereby improving the accuracy of multi-source image fusion and target recognition effect.

[0023] S14 specifically includes: S141: Identify common targets with clear spatial locations in hyperspectral and SAR images as ground control points, extract their pixel coordinates in the two images, and construct a one-to-one set of control points. S142: Based on the control point set, the least squares method is used to fit the affine transformation model from the SAR image to the hyperspectral image, establishing the transformation relationship between pixel coordinates. The specific transformation form is as follows: ; in, These are the pixel coordinates in the original SAR image; Represents the pixel coordinates of the target after transformation; These are the affine transformation parameters, representing the rotation, scaling, and translation coefficients, obtained by fitting control points. S143: Perform coordinate transformation on all pixels in the SAR image according to the affine model, and calculate the gray value of the pixels that fall at non-integer coordinate positions after transformation using bilinear interpolation. The interpolation process is as follows: ,in, Floating-point coordinates after transformation The corresponding interpolated pixel value; The grayscale values ​​at four adjacent integer pixel positions in the floating-point coordinate system; This represents the horizontal distance between the floating-point coordinates and the left neighboring pixel. This represents the distance between the floating-point coordinates and the upper neighbor pixel in the vertical direction; S144: The SAR image after transformation and interpolation is resampled to have the same spatial dimensions and coordinate system as the hyperspectral image, thus obtaining the registered SAR image; the hyperspectral image serves as the registration reference image and is directly output as the registration result without transformation; the above steps, through the control point-driven affine model registration and bilinear interpolation calculation method, can effectively achieve high-precision spatial alignment of the SAR image with the hyperspectral image, ensuring pixel spatial consistency during subsequent feature fusion, thereby improving the quality of multi-source image fusion and the accuracy of plant identification.

[0024] S2: Perform spectral feature dimensionality reduction processing on the registered hyperspectral image to extract the principal component map of spectral features that characterize the core spectral information of ground features; S2 specifically includes: S21: Expand the registered hyperspectral image into a two-dimensional spectral matrix by pixel. Let the image have a total of Each of the pixels contains 1000 pixels, and each pixel contains 10000 pixels. Continuous spectral bands, constructed with dimensions of Two-dimensional spectral data matrix ; S22: For two-dimensional spectral data matrix Mean centering is performed to eliminate mean deviations between bands. The mean vector is calculated, and the mean of the corresponding band is subtracted from each sample to obtain the centered matrix. The calculation formula is as follows: ,in, This represents the band mean vector calculated by column, with dimension 1. ; S23: For the centralized matrix Principal component analysis (PCA) is performed to construct the covariance matrix, and its eigenvalues ​​and eigenvectors are solved to determine the principal component directions that best represent spectral variation information. The expression for the covariance matrix is ​​as follows: ,in, The inter-band covariance matrix has dimensions of . ; S24: Sort the features from largest to smallest, and select those whose cumulative contribution rate exceeds a set threshold (e.g., ...). (before) Using eigenvectors, construct a transformation matrix, and then convert the original two-dimensional spectral matrix... Projecting onto the principal component space yields the dimensionality-reduced principal component map. The calculation is as follows: ,in, Indicates from the previous The transformation matrix consists of eigenvectors, with dimension . The above steps, by performing principal component analysis to reduce the dimensionality of the registered hyperspectral image, can effectively compress redundant band information, highlight the main spectral differences between ground features, and thus extract the principal component map that characterizes the spectral properties of plant targets, providing a highly discriminative spectral representation basis for subsequent effective fusion with SAR features.

[0025] S3: Perform polarimetric target decomposition on the registered SAR image to extract scattering feature component maps that characterize the scattering properties of ground structures; S3 specifically includes: S31: Construct a complex polarization scattering matrix from the polarization channel data in the registered SAR image to characterize the scattering response of pixels under different polarization modes. The scattering matrix is ​​represented as follows: ;in, and These are the complex scattering coefficients under horizontal-horizontal polarization and vertical-vertical polarization, respectively; and These are the scattering coefficients of the cross-polarization channels, satisfying... ; S32: Construct the Pauli vector representation based on the scattering matrix, and calculate the coherence matrix within a sliding window. To quantify the scattering characteristics of a local region, the specific expression is as follows: ; ; in, For the first Pauli decomposition vector of 1 pixel; Indicates its conjugate transpose; This is the coherence matrix, used to statistically describe the correlation between different polarization components; This represents the number of pixels within the sliding window. S33: For the coherence matrix Eigenvalue decomposition is performed to extract the eigenvalues ​​corresponding to the main polarization scattering mechanisms, and the power proportion of each scattering mechanism is calculated, as shown in the following formula: ,in, Coherence matrix The descending eigenvalues ​​correspond to the dominant energies of surface scattering, volume scattering, and double reflection scattering, respectively. For the first The normalized power component of the scattering mechanism reflects its energy proportion in the current pixel; S34: The power values ​​of the three scattering mechanisms corresponding to each pixel are encoded by channel to form a three-channel scattering feature component map. Each channel represents the surface scattering map, volume scattering map, and double reflection scattering map, respectively, as the structural feature output of the SAR image. Through the above polarimetric target decomposition steps, the geometric structural information contained in the SAR image is transformed into a physically meaningful scattering component expression, which can effectively enhance the ability to characterize the structural differences between different ground features and provide a more robust structural feature foundation for subsequent spectral information fusion and plant target identification.

[0026] S4: Perform pixel-level fusion of the spectral feature principal component map and the scattering feature component map to generate a multi-source feature fusion map that simultaneously contains spectral and scattering information; S4 specifically includes: S41: Normalize each channel of the spectral feature principal component map and the scattering feature component map to uniformly map the feature values ​​from different sources in the image to the standard range, so as to eliminate the amplitude difference between cross-source data and improve the feature balance of subsequent fusion. S42: While maintaining the pixel spatial alignment of the original image, a pixel-by-pixel stitching method is used to fuse and stitch the normalized spectral feature vector and the scattering feature vector to construct a fused multi-source feature vector image, so that each pixel simultaneously contains composite feature information from the spectral principal component and the scattering component. Through the above steps, complementary spectral and scattering information from different remote sensing source images can be efficiently integrated to form a multi-source feature map with stronger fusion expression capabilities, providing a high-information, high-alignment input foundation for subsequent target region segmentation and plant category identification.

[0027] S42 specifically includes: S421: Obtain the normalized principal component map of spectral features and the scattering feature map, and represent them as two-dimensional matrix forms, where the spectral feature map is denoted as a matrix. , dimension The scattering feature map is denoted as a matrix. , dimension ,in Image size, , These are the spectral feature dimension and the scattering feature dimension, respectively. S422: For each pixel location in the image, extract the spectral feature vector corresponding to that location. With scattering eigenvectors They are then concatenated along the vector dimension to form a multi-source fusion feature vector. The expression is: ,in, This represents a vector concatenation operation; S423: Recombines the fused feature vectors at all pixel locations into a three-dimensional matrix, with an output dimension of... The fused image serves as the unified input data for subsequent processing stages. By employing a pixel-by-pixel feature vector stitching method, the above steps can fully preserve the independent structure and joint representation capabilities of both spectral and scattering information. This allows the fused image to maintain spatial consistency while enhancing the expressive power of feature dimensions, thereby improving the accuracy and discriminative power of subsequent plant region identification.

[0028] S5: A linear iterative clustering algorithm is used to perform superpixel segmentation on the multi-source feature fusion map, generating multiple homogeneous region blocks with boundaries that fit the distribution of ground features. Each homogeneous region block contains spectral-scattering integrated features. S5 specifically includes: S51: Construct each cell in the multi-source feature fusion map into a five-dimensional feature vector containing spatial coordinates and fused feature vectors, which is used to simultaneously represent spatial location and land cover attributes; S52: Set the initial number of cluster centers according to the image size, and uniformly initialize the cluster center positions on the image as the initial seeds for subsequent superpixel segmentation; S53: A linear iterative clustering algorithm is used to calculate the weighted distance between the cell and the neighboring cluster centers in each iteration, and the cell is assigned to the class corresponding to the cluster center with the smallest distance. S54: Update the spatial location and feature mean of each cluster center, repeat the iteration until the cluster centers converge, and finally generate multiple homogeneous region blocks with boundaries that fit the distribution of ground features; the above steps, by using linear iterative clustering in the fused image for superpixel segmentation, can effectively fit the spatial boundaries of ground features while ensuring feature similarity, enhance the integrity and edge accuracy of plant target region extraction, and provide structurally clear and feature-consistent segmentation basic units for subsequent identification.

[0029] S53 specifically includes: S531: For each cluster center in the current round, select a predetermined range of pixels within its spatial neighborhood as a candidate set; S532: Calculate the spatial distance and feature distance between the candidate pixel and the current cluster center, and then combine the two with a weighted sum to form the overall distance metric. The calculation formula is as follows: ;in, For pixels With the The weighted distance between the cluster centers; For pixels fused feature vectors; Cluster center The mean of the eigenvectors; For pixels The coordinates of the location; Cluster center Spatial coordinates; The normalized scaling factor for the feature components; This is the normalized scale factor for the spatial components; S533: For each cell, iterate through its nearest cluster centers and select the one with the smallest weighted distance. The cluster centers are used as the target of assignment. The pixels are marked as the category corresponding to the cluster centers, and the update results are recorded for the next round of center update. The above steps introduce a weighted distance calculation method that combines fusion features and spatial location. This can enhance the feature similarity aggregation ability while maintaining spatial coherence, thereby generating region division results with natural boundaries and consistent interiors in the superpixel segmentation process, providing a clearer target structure basis for the subsequent recognition stage.

[0030] S6: Match the spectral-scattering integrated features of each homogeneous region block with the pre-established plant target spectral-scattering feature knowledge base, identify the plant category of each homogeneous region block based on the matching results, and output the final plant target category map; S6 specifically includes: S61: For each homogeneous region block, extract the spectral-scattering fusion feature vector of all pixels within it, and calculate the average vector of the fusion features within the region to represent the overall feature expression of the region block. S62: Call the pre-established plant target spectral-scattering feature knowledge base. The knowledge base stores a set of standard fused feature vectors corresponding to several plant categories. Each category corresponds to one or more normalized reference feature vectors. S63: Feature vector for the current region block Calculate the similarity score between each feature vector and the reference feature vector of each category, using cosine similarity as the matching degree measure. The calculation formula is as follows: ,in, This represents the average fused feature vector of the current region block; _i_refs the reference feature vector of a certain plant category in the knowledge base; _t_ represents the vector dot product operation; The L2 norm of a vector; This represents the similarity score, with a value range of [value range missing]. ; S64: Among all category similarity scores, the category label with the highest score is selected as the recognition result of the current region block, and the label is assigned to all pixels in the corresponding region to complete the category labeling of homogeneous regions. Finally, the plant target category map of the entire image is output. The above steps achieve efficient identification and classification of regional plant targets by matching the similarity between the region fusion features and the standard features of plant categories in the knowledge base, avoiding pixel-level noise interference, improving the overall recognition accuracy and image semantic consistency, and providing support for remote sensing applications such as plant resource surveys, classification and monitoring.

[0031] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0032] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A plant target recognition method based on multi-source remote sensing image fusion, characterized in that, Includes the following steps: S1: Acquire hyperspectral remote sensing images and synthetic aperture radar (SAR) images of the same observation area, and perform radiometric calibration and geometric registration preprocessing on the hyperspectral remote sensing images and SAR images to obtain the registered hyperspectral images and registered SAR images. S2: Perform spectral feature dimensionality reduction processing on the registered hyperspectral image to extract the principal component map of spectral features that characterize the core spectral information of ground features; S3: Perform polarimetric target decomposition on the registered SAR image to extract scattering feature component maps that characterize the scattering properties of ground structures; S4: Perform pixel-level fusion of the spectral feature principal component map and the scattering feature component map to generate a multi-source feature fusion map that simultaneously contains spectral and scattering information; S5: A linear iterative clustering algorithm is used to perform superpixel segmentation on the multi-source feature fusion map, generating multiple homogeneous region blocks with boundaries that fit the distribution of ground features. Each homogeneous region block contains spectral-scattering integrated features. S6: Match the spectral-scattering integrated features of each homogeneous region block with the pre-established plant target spectral-scattering feature knowledge base, identify the plant category of each homogeneous region block based on the matching results, and output the final plant target category map.

2. The plant target recognition method based on multi-source remote sensing image fusion according to claim 1, characterized in that, S1 specifically includes: S11: Retrieve hyperspectral remote sensing satellite images and synthetic aperture radar images covering the same observation area through the remote sensing satellite image acquisition interface; S12: Perform radiometric calibration on the acquired hyperspectral image to convert the original digital quantization values ​​into apparent reflectance values ​​of ground features; S13: Perform radiometric calibration on the SAR image, convert the square of the echo amplitude value into the backscattering coefficient, and uniformly convert it into a logarithmic scale amplitude map according to the polarization channel type; S14: Using the hyperspectral image as the main control reference image, based on the ground control points in the two types of images, perform affine transformation and interpolation resampling of the SAR image to complete the geometric registration process, and output the registered hyperspectral image and the registered SAR image.

3. The plant target recognition method based on multi-source remote sensing image fusion according to claim 2, characterized in that, S14 specifically includes: S141: Identify common targets with clear spatial locations in hyperspectral and SAR images as ground control points, extract their pixel coordinates in the two images, and construct a one-to-one set of control points. S142: Based on the set of control points, fit the affine transformation model from SAR image to hyperspectral image using the least squares method; S143: Perform coordinate transformation on all pixels in the SAR image according to the affine model, and calculate the gray value of the pixels that fall at non-integer coordinate positions after transformation using the bilinear interpolation method. S144: The SAR image after transformation and interpolation is resampled to the same spatial size and coordinate system as the hyperspectral image, thus obtaining the registered SAR image.

4. The plant target recognition method based on multi-source remote sensing image fusion according to claim 1, characterized in that, S2 specifically includes: S21: Expand the registered hyperspectral image into a two-dimensional spectral matrix by pixel. ; S22: For two-dimensional spectral data matrix Mean centering is performed to eliminate mean deviations between bands. The mean vector is calculated, and the mean of the corresponding band is subtracted from each sample to obtain the centered matrix. ; S23: For the centralized matrix Perform principal component analysis to construct the covariance matrix and solve for its eigenvalues ​​and eigenvectors; S24: Sort by feature value from largest to smallest, and select the top performers whose cumulative contribution rate exceeds a set threshold. Using eigenvectors, construct a transformation matrix, and then convert the original two-dimensional spectral matrix... Projecting onto the principal component space yields the dimensionality-reduced principal component map. .

5. The plant target recognition method based on multi-source remote sensing image fusion according to claim 1, characterized in that, S3 specifically includes: S31: Construct a complex polarization scattering matrix from the polarization channel data in the registered SAR image; S32: Construct the Pauli vector representation based on the scattering matrix, and calculate the coherence matrix within a sliding window. ; S33: For the coherence matrix Eigenvalue decomposition is performed to extract the eigenvalues ​​corresponding to the main polarization scattering mechanisms, and the power proportion of each scattering mechanism is calculated. S34: Encode the power values ​​of the three scattering mechanisms corresponding to each pixel by channel to form a three-channel scattering feature component map, with each channel representing the surface scattering map, volume scattering map, and double reflection scattering map, respectively.

6. The plant target recognition method based on multi-source remote sensing image fusion according to claim 1, characterized in that, S4 specifically includes: S41: Normalize each channel of the spectral feature principal component map and the scattering feature component map to uniformly map the feature values ​​from different sources in the image to the standard range. S42: While maintaining the pixel spatial alignment of the original image, a pixel-by-pixel stitching method is used to fuse and stitch the normalized spectral feature vector and the scattering feature vector to construct a fused multi-source feature vector image.

7. The plant target recognition method based on multi-source remote sensing image fusion according to claim 6, characterized in that, S42 specifically includes: S421: Obtain the normalized principal component map of spectral features and the scattering feature map, and represent them as two-dimensional matrix forms, where the spectral feature map is denoted as a matrix. , dimension The scattering feature map is denoted as a matrix. , dimension ,in Image size, , These are the spectral feature dimension and the scattering feature dimension, respectively. S422: For each pixel location in the image, extract the spectral feature vector corresponding to that location. With scattering eigenvectors They are then concatenated along the vector dimension to form a multi-source fusion feature vector. The expression is: ,in, This represents a vector concatenation operation; S423: Recombines the fused feature vectors at all pixel locations into a three-dimensional matrix, with an output dimension of... The fused image.

8. The plant target recognition method based on multi-source remote sensing image fusion according to claim 1, characterized in that, S5 specifically includes: S51: Construct each cell in the multi-source feature fusion map into a five-dimensional feature vector containing spatial coordinates and fused feature vectors; S52: Set the initial number of cluster centers according to the image size, and uniformly initialize the positions of cluster centers on the image; S53: A linear iterative clustering algorithm is used to calculate the weighted distance between the cell and the neighboring cluster centers in each iteration, and the cell is assigned to the class corresponding to the cluster center with the smallest distance. S54: Update the spatial location and feature mean of each cluster center, repeat the iteration until the cluster centers converge, and finally generate multiple homogeneous region blocks with boundaries that fit the distribution of ground features.

9. The plant target recognition method based on multi-source remote sensing image fusion according to claim 8, characterized in that, S53 specifically includes: S531: For each cluster center in the current round, select a predetermined range of pixels within its spatial neighborhood as a candidate set; S532: Calculate the spatial distance and feature distance between the candidate pixel and the current cluster center, and then combine the two with a weighted sum to form the overall distance metric. ; S533: For each cell, iterate through its nearest cluster centers and select the one with the smallest weighted distance. Using the cluster centers as the target, pixels are labeled with the category corresponding to the cluster centers.

10. The plant target recognition method based on multi-source remote sensing image fusion according to claim 1, characterized in that, S6 specifically includes: S61: For each homogeneous region block, extract the spectral-scattering fusion feature vector of all pixels within it, and calculate the average vector of the fusion features within the region; S62: Call the pre-established plant target spectral-scattering feature knowledge base. The knowledge base stores a set of standard fused feature vectors corresponding to several plant categories. Each category corresponds to one or more normalized reference feature vectors. S63: Feature vector for the current region block Calculate the similarity score between it and the reference feature vector of each category; S64: Among all category similarity scores, select the category label with the highest score as the recognition result of the current region block, and assign the label to all pixels in the corresponding region to complete the category labeling of homogeneous regions. Finally, output the plant target category map of the entire image.

Images