A multi-source remote sensing image classification method based on pattern recognition

Abstract

The application discloses a kind of based on pattern recognition's multi-source remote sensing image classification method, comprising the following steps: S1, obtains source domain remote sensing image, target domain remote sensing image;S2, using singular spectrum analysis extracts source domain spectral feature and target domain spectral feature, using spatial filtering extracts source domain spatial feature and target domain spatial feature;S3, introduce spectral weight matrix mechanism;S4, introduce edge weight mapping mechanism;S5, input improved MDANet model, introduce geodesic mutual attention mechanism, output source domain fusion feature and target domain fusion feature;S6, construct segmentation loss;S7, by fusion generates joint loss function, and updates parameter, and outputs target domain classification chart.The application effectively alleviates the domain shift problem caused by imaging condition, sensor difference and ground object distribution change of multi-source remote sensing image, realizes high-precision, strong robustness automatic classification of ground object class under cross-domain scene.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a classification method for multi-source remote sensing images based on pattern recognition. Background Technology

[0002] With the rapid development of high-resolution remote sensing sensors, multi-platform Earth observation systems, and multi-source remote sensing data sharing mechanisms, intelligent interpretation and fine classification technologies for multi-source remote sensing images in complex surface scenarios have attracted widespread attention. Existing remote sensing image classification methods mostly rely on single-source spectral information or simple spatial texture features for supervised learning modeling, and commonly suffer from the following problems in cross-sensor, cross-imaging condition, and cross-regional applications: Images acquired from different remote sensing platforms exhibit significant differences in radiometric response, spectral band settings, and spatial resolution, directly leading to inconsistent data distributions between the source and target domains. Existing preprocessing and feature alignment methods based on statistical normalization or linear transformations struggle to characterize complex nonlinear domain shift relationships, resulting in a significant decrease in the model's generalization ability in the target region. In multi-source remote sensing images, spectral and spatial structural information are tightly coupled. Existing methods often treat spectral feature extraction and spatial feature extraction separately, lacking systematic modeling of cross-modal correlations and geometric consistency, leading to blurred ground feature boundaries and severe confusion between heterogeneous features. For cross-domain remote sensing classification tasks, commonly used maximum mean difference or adversarial domain adaptation methods primarily constrain at the overall distribution level, ignoring differences in class conditional structure and geometric subspace relationships, easily resulting in feature collapse and class drift problems. When labeled samples are lacking in the target domain, existing pseudo-labeling strategies are sensitive to samples with high uncertainty, exhibiting significant error accumulation and further amplifying cross-domain error propagation.

[0003] Therefore, how to provide a classification method for multi-source remote sensing images based on pattern recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a classification method for multi-source remote sensing images based on pattern recognition. This invention integrates singular spectral analysis and spatial filtering for collaborative extraction of spectral and spatial features. It introduces Grassmann manifold mapping combined with spectral weighting and edge weighting mechanisms to achieve cross-domain spectral and spatial alignment. Furthermore, it uses an improved MDANet model to embed a geodesic mutual attention mechanism in the fusion module to achieve unified subspace representation and efficient fusion of multimodal features. Simultaneously, it constructs a joint loss function composed of segmentation loss, weighted maximum mean difference loss, and class conditional maximum mean difference loss to drive cross-domain adaptive learning, achieving high-precision classification output of remote sensing images in the source and target domains. This method has the advantages of strong cross-domain robustness, high feature alignment accuracy, good classification consistency, and strong sample adaptability.

[0005] A method for classifying multi-source remote sensing images based on pattern recognition according to an embodiment of the present invention includes the following steps: S1. Obtain the source domain remote sensing image, the target domain remote sensing image, and the source domain annotation map corresponding to the source domain remote sensing image, and perform preprocessing to obtain the source domain preprocessed remote sensing image and the target domain preprocessed remote sensing image; S2. Based on the preprocessed remote sensing images of the source domain and the preprocessed remote sensing images of the target domain, singular spectral analysis is used to extract the spectral features of the source domain and the target domain, and spatial filtering is used to extract the spatial features of the source domain and the target domain. S3. Based on the source domain spectral features and the target domain spectral features, a spectral weighting mechanism is introduced through Grassmann manifold mapping to perform spectral alignment and obtain aligned spectral features and spectral principal angle vectors. S4. Based on the spatial features of the source domain and the spatial features of the target domain, the edge weight mapping mechanism is introduced through Grassmann manifold mapping to perform spatial alignment and obtain the aligned spatial features and spatial principal angle vector. S5. Input the aligned spectral features, aligned spatial features, spectral principal angle vector and spatial principal angle vector into the improved MDANet model. Introduce the geodesic mutual attention mechanism into the fusion module of the improved MDANet model and output the source domain fusion features and the target domain fusion features. S6. Input the source domain fusion feature and the target domain fusion feature into the segmentation operator to generate the source domain segmentation probability map and the target domain segmentation probability map, and construct the segmentation loss. Based on the target domain segmentation probability map, generate the target domain pseudo-label map and confidence mask. S7. Construct the maximum mean difference loss based on the source domain fusion features and the target domain fusion features, and introduce the corner kernel mechanism to calculate the weighted maximum mean difference loss; calculate the class conditional maximum mean difference loss based on the target domain pseudo-label map and confidence mask; generate a joint loss function through fusion, and update the parameters of the improved MDANet model and the segmentation operator parameters to output the target domain classification map.

[0006] Optionally, S1 specifically includes: Acquire source domain remote sensing images, target domain remote sensing images, and source domain annotation maps corresponding to the source domain remote sensing images; Radiometric calibration is performed on the source domain remote sensing image and the target domain remote sensing image respectively to obtain the source domain calibrated remote sensing image and the target domain calibrated remote sensing image; Geometric registration processing is performed on the source domain calibrated remote sensing image and the target domain calibrated remote sensing image respectively to obtain the source domain registered remote sensing image and the target domain registered remote sensing image; Perform spectral band unification processing on the source domain registered remote sensing image and the target domain registered remote sensing image respectively to obtain the source domain unified remote sensing image and the target domain unified remote sensing image; Spatial resolution uniformity processing is performed on the source domain uniform remote sensing image and the target domain uniform remote sensing image respectively to obtain the source domain preprocessed remote sensing image and the target domain preprocessed remote sensing image.

[0007] Optionally, S2 specifically includes: The source domain preprocessed remote sensing images and the target domain preprocessed remote sensing images are respectively constructed into spectral data matrices according to the band dimension. The rows of the spectral data matrix correspond to the pixel positions, and the columns correspond to the spectral bands. Trajectory matrices are constructed from the spectral data matrices. The trajectory matrices are formed by shifting and splicing continuous band spectral vectors according to a preset window length. Singular value decomposition is performed on the trajectory matrix to obtain the singular value sequence and the corresponding feature subspace basis vectors; Based on the proportion of the sum of singular values ​​to the total sum of singular values, the feature subspace basis vectors corresponding to the main spectral components are selected to form the spectral subspace representation. The spectral subspace representation is then mapped back to the original pixel space to obtain the source domain spectral features and the target domain spectral features. Spatial convolution kernel groups are constructed for the source domain preprocessed remote sensing images and the target domain preprocessed remote sensing images, respectively. The spatial convolution kernel groups contain multi-scale directional response kernels. The aforementioned spatial convolution kernel group is used to perform convolution operations on the source domain preprocessed remote sensing image and the target domain preprocessed remote sensing image respectively to obtain a multi-scale spatial response map; Multi-scale spatial response maps are stitched together along the channel dimension to form a spatial feature tensor, thus obtaining the source domain spatial features and the target domain spatial features.

[0008] Optionally, the spectral weighting mechanism introduced through Grassmann manifold mapping is specifically as follows: The band weight vector is calculated based on the source domain spectral features and the target domain spectral features, and each dimension of the band weight vector corresponds to a spectral band. The source domain spectral features and the target domain spectral features are weighted based on the band weight vector to obtain the weighted source domain spectral features and the weighted target domain spectral features. Based on the weighted source domain spectral features and the weighted target domain spectral features, spectral covariance matrices are constructed respectively. The source domain spectral subspace basis and the target domain spectral subspace basis are obtained by eigenvalue decomposition of the spectral covariance matrices. The principal angle matrix is ​​calculated based on the source domain spectral subspace basis and the target domain spectral subspace basis, and the spectral principal angle vector is obtained by performing an inverse cosine operation on the principal angle matrix. Based on the source domain spectral subspace basis, the target domain spectral subspace basis, and the spectral principal angle vector, a geodesic transport mapping is constructed to map the source domain spectral features and the target domain spectral features to a unified subspace coordinate system, thereby obtaining aligned spectral features.

[0009] Optionally, the edge-weight mapping mechanism introduced through Grassmann manifold mapping is specifically as follows: An edge weight map is constructed based on the spatial features of the source domain and the spatial features of the target domain. The edge weight map is a weight raster map that corresponds one-to-one with the spatial features of the source domain and the spatial features of the target domain in spatial coordinates. The weight of each pixel position in the edge weight map is determined by the gradient magnitude of the corresponding spatial feature. The gradient magnitude is obtained by taking the square root of the sum of the squares of the horizontal spatial response components, the vertical spatial response components, and the diagonal spatial response components. The source domain spatial features and target domain spatial features are weighted based on the edge weight map to obtain weighted source domain spatial features and weighted target domain spatial features; Based on the weighted source domain spatial features and the weighted target domain spatial features, spatial covariance matrices are constructed respectively, and the source domain spatial subspace basis and the target domain spatial subspace basis are obtained by eigenvalue decomposition of the spatial covariance matrices. The principal angle matrix is ​​calculated based on the source domain space subspace basis and the target domain space subspace basis, and the spatial principal angle vector is obtained by performing an inverse cosine operation on the principal angle matrix. Based on the source domain spatial subspace basis, the target domain spatial subspace basis, and the spatial principal angle vector, a geodesic transmission mapping is constructed to map the source domain spatial features and the target domain spatial features to a unified subspace coordinate system, thereby obtaining aligned spatial features.

[0010] Optionally, the improved MDANet model specifically includes a spectral coding module, a null coding module, and a fusion module; The spectral coding module performs multi-channel convolution operations on the aligned spectral features. The pixel value of any channel is equal to the sum of the element-wise multiplication of the convolution kernel weight matrix and the corresponding spectral neighborhood pixel matrix, plus a bias term to form a linear response value. The linear response value is then subjected to a nonlinear activation mapping to obtain the spectral coding features. The nonlinear activation mapping uses a piecewise linear function or an exponential normalization function. The empty coding module performs multi-channel convolution operation on the aligned spatial features. The pixel value of any channel is equal to the sum of the element-wise multiplication of the convolution kernel weight matrix and the corresponding spatial neighborhood pixel matrix, plus a bias term to form a linear response value. A nonlinear activation mapping is applied to the linear response value to obtain the spatial coding feature. The improved MDANet model introduces a geodesic mutual attention mechanism in its fusion module. A geodesic transport mapping matrix is ​​constructed based on the spectral principal angle vector and the spatial principal angle vector. Any element in the geodesic transport mapping matrix is ​​equal to the mapping coefficient formed by combining the cosine and sine functions of the corresponding principal angle value according to weighted coefficients. The spectral coding features and spatial coding features are respectively multiplied with the geodesic transport mapping matrix to obtain the spectral transport features and spatial transport features. The spectral coding features and spatial coding features are then projected onto a unified subspace coordinate system. A cross-modal attention weight matrix is ​​constructed based on spectral transmission features and spatial transmission features. Any weight in the cross-modal attention weight matrix is ​​equal to the ratio of the inner product of the corresponding spectral transmission feature vector and spatial transmission feature vector after normalization by an exponential function. The spectral transmission features and spatial transmission features are weighted by element-wise multiplication according to their corresponding weights, and then summed after being concatenated in order along the channel dimension to form spectral enhancement features and spatial enhancement features. Angle-gated weight vectors are generated based on the spectral principal angle vector and the spatial principal angle vector. Any weight in the angle-gated weight vector is equal to the result of the corresponding principal angle component after exponential decay mapping and normalization. The spectral enhancement features and spatial enhancement features are weighted by channel-by-channel product in channel order and summed in the pixel dimension to form fusion features. The source domain fusion features are fusion features formed based on the source domain preprocessed remote sensing images. The target domain fusion features are fusion features formed based on the target domain preprocessed remote sensing images.

[0011] Optionally, the step of inputting the source domain fusion features and the target domain fusion features into the segmentation operator to generate a source domain segmentation probability map and a target domain segmentation probability map, and constructing a segmentation loss, specifically involves: The source domain fusion features are input into the segmentation operator and convolution operation is performed to obtain the source domain pixel score tensor. The source domain pixel score tensor gives the corresponding score value in the pixel position and category channel. The source domain pixel score tensor is subjected to exponential normalization operation in the category channel to obtain the source domain segmentation probability map. The target domain fusion feature is input into the segmentation operator and convolution operation is performed to obtain the target domain pixel score tensor. The target domain pixel score tensor gives the corresponding score value in the pixel position and category channel. The target domain pixel score tensor is subjected to exponential normalization operation in the category channel to obtain the target domain segmentation probability map. The segmentation loss is constructed based on the source domain segmentation probability map and the source domain annotation map. The source domain annotation map provides a class indicator value at the pixel position. The segmentation loss is equal to the sum of the logarithmic products of the class indicator values ​​in the source domain annotation map and the corresponding class probabilities in the source domain segmentation probability map, and then the negative value is taken. The loss is normalized according to the number of pixels in the source domain annotation map.

[0012] Optionally, the step of generating the target domain pseudo-label map and confidence mask based on the target domain segmentation probability map specifically involves: The maximum value selection operation is performed on the target domain segmentation probability map in the category channel to obtain the target domain maximum probability map and the target domain category index map. The target domain maximum probability map gives the maximum category probability value at the cell position, and the target domain category index map gives the category index corresponding to the maximum probability value at the cell position. Use the target domain category index map as the target domain pseudo-annotation map; The maximum probability image of the target domain is compared with a set confidence threshold. Positions where the maximum probability image value of the target domain is greater than or equal to the set confidence threshold are assigned a value of 1, and positions where the maximum probability image value of the target domain is less than the set confidence threshold are assigned a value of 0, thus forming a confidence mask.

[0013] Optionally, the maximum mean difference loss is constructed based on the source domain fusion features and the target domain fusion features, and a corner kernel mechanism is introduced to calculate the weighted maximum mean difference loss; the class-conditional maximum mean difference loss is calculated based on the target domain pseudo-labeled map and the confidence mask, specifically as follows: The source domain fusion features and the target domain fusion features are expanded according to the pixel position to form the source domain feature set and the target domain feature set; Generating angular kernel weighting coefficients based on spectral principal angle vectors and spatial principal angle vectors; A weighted maximum mean difference loss is constructed based on the source domain feature set and the target domain feature set, combined with the corner kernel weight coefficients. The target domain pseudo-annotated map and confidence mask are used to filter the target domain feature set to form the effective feature set of the target domain; Based on the target domain pseudo-annotated map and the source domain annotated map, the effective feature set of the target domain and the feature set of the source domain are classified into target domain category feature subsets and source domain category feature subsets; The maximum mean difference loss of class conditions is constructed based on the subsets of category features in the source domain and the subsets of category features in the target domain, and the maximum mean difference loss of class conditions for all categories is summed to obtain the total maximum mean difference loss of class conditions.

[0014] Optionally, the step of generating a joint loss function through fusion, updating the parameters of the improved MDANet model and the segmentation operator, and outputting a target domain classification map specifically involves: A joint loss function is constructed based on the segmentation loss, the weighted maximum mean difference loss, and the total conditional maximum mean difference loss. The joint loss function is equal to the sum of the segmentation loss multiplied by the first weight coefficient, the weighted maximum mean difference loss multiplied by the second weight coefficient, and the total conditional maximum mean difference loss. The gradients of the parameters of the improved MDANet model and the parameters of the segmentation operator are calculated based on the joint loss function, and the parameters are updated. The parameter update amount is equal to the product of the learning rate and the corresponding gradient. The updated parameter value is equal to the parameter value before the update minus the parameter update amount. The updated improved MDANet model parameters are used to preprocess the remote sensing image of the target domain to generate target domain fusion features. The updated segmentation operator parameters are used to generate the target domain pixel score tensor and generate the target domain segmentation probability map from the target domain fusion features. The target domain segmentation probability map is subjected to a maximum value selection operation in the category channel to obtain the target domain category index map, which serves as the target domain classification map.

[0015] The beneficial effects of this invention are: By combining singular spectral analysis and spatial filtering, source domain spectral features and source domain spatial features, as well as target domain spectral features and target domain spatial features, and by combining Grassmann manifold mapping to introduce spectral weighting and edge weighting mechanisms to complete spectral alignment and spatial alignment, the feature drift caused by cross-sensor radiation differences and spatial structure mismatch is effectively suppressed, and the consistency and discrimination stability of cross-domain features of multi-source remote sensing images are improved. In the fusion module of the improved MDANet model, a geodesic mutual attention mechanism is introduced. Based on the spectral principal angle vector and the spatial principal angle vector, a geodesic transport map and cross-modal attention weights are constructed to achieve adaptive fusion of aligned spectral features and aligned spatial features in a unified subspace coordinate system. This significantly enhances the ability to express multimodal complementary information and the accuracy of distinguishing complex land cover categories. By combining segmentation operators to generate source domain segmentation probability maps and target domain segmentation probability maps, and using target domain pseudo-labeled maps and confidence masks to select high-confidence target domain samples for training, the target domain classification boundary is continuously optimized under the condition of no manual labeling of the target domain, effectively alleviating the classification performance degradation problem caused by domain offset; We construct a weighted maximum mean difference loss and a class-conditional maximum mean difference loss and fuse them to form a joint loss function. Through the corner kernel mechanism, we guide the dual alignment of source domain fusion features and target domain fusion features at the overall distribution and class structure levels, achieving adaptive convergence of cross-domain feature distribution and significantly improving the overall accuracy and robustness of the target domain classification map. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a multi-source remote sensing image classification method based on pattern recognition proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved MDANet model proposed in this invention; Figure 3 This is a data flow diagram of a multi-source remote sensing image classification method based on pattern recognition proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3A classification method for multi-source remote sensing images based on pattern recognition includes the following steps: S1. Acquire source domain remote sensing images, target domain remote sensing images, and source domain annotation maps corresponding to the source domain remote sensing images. Perform radiometric calibration and geometric registration on the source domain remote sensing images and the target domain remote sensing images, and perform spectral band consistency and spatial resolution consistency processing to obtain source domain preprocessed remote sensing images and target domain preprocessed remote sensing images. S2. Based on the preprocessed remote sensing images of the source domain and the preprocessed remote sensing images of the target domain, singular spectral analysis is used to extract the spectral features of the source domain and the target domain, and spatial filtering is used to extract the spatial features of the source domain and the target domain. S3. Based on the source domain spectral features and the target domain spectral features, a spectral weighting mechanism is introduced through Grassmann manifold mapping to perform spectral alignment and obtain aligned spectral features and spectral principal angle vectors. S4. Based on the spatial features of the source domain and the spatial features of the target domain, the edge weight mapping mechanism is introduced through Grassmann manifold mapping to perform spatial alignment and obtain the aligned spatial features and spatial principal angle vector. S5. Input the aligned spectral features, aligned spatial features, spectral principal angle vector and spatial principal angle vector into the improved MDANet model. Introduce the geodesic mutual attention mechanism in the fusion module of the improved MDANet model and output the source domain fusion features and the target domain fusion features. S6. Input the source domain fusion feature and the target domain fusion feature into the segmentation operator to generate the source domain segmentation probability map and the target domain segmentation probability map, and construct the segmentation loss. Based on the target domain segmentation probability map, generate the target domain pseudo-label map and confidence mask. S7. Construct the maximum mean difference loss based on the source domain fusion features and the target domain fusion features, and introduce the corner kernel mechanism to calculate the weighted maximum mean difference loss; calculate the class conditional maximum mean difference loss based on the target domain pseudo-label map and confidence mask; generate a joint loss function through fusion, and update the parameters of the improved MDANet model and the segmentation operator parameters to output the target domain classification map.

[0019] In this embodiment, S1 specifically refers to: The process involves acquiring source domain remote sensing images, target domain remote sensing images, and source domain annotation maps corresponding to the source domain remote sensing images. The source domain annotation map is a pixel-level raster map that corresponds one-to-one with the source domain remote sensing images in spatial coordinates. Each pixel position in the source domain annotation map stores the land cover category number to which the corresponding remote sensing pixel belongs. The land cover category number uses discrete integer encoding to represent different land cover types. The spatial resolution of the source domain annotation map is consistent with the spatial resolution of the source domain preprocessed remote sensing images and is fully aligned with the source domain preprocessed remote sensing images in the row and column dimensions. Radiometric calibration is performed on the source domain remote sensing image and the target domain remote sensing image respectively. The radiometric calibration process includes converting the digital quantization value of each pixel into the radiance value received by the sensor. The radiance value is equal to the gain coefficient multiplied by the digital quantization value and the offset coefficient added. The gain coefficient and the offset coefficient are determined by the calibration parameters of the corresponding remote sensing sensor. Based on the radiance value, the surface reflectance value is further calculated. The surface reflectance value is equal to the radiance value multiplied by the distance correction factor and divided by the product of the incident solar irradiance and the cosine of the solar zenith angle, to obtain the source domain calibrated remote sensing image and the target domain calibrated remote sensing image. Geometric registration processing is performed on the source domain calibrated remote sensing image and the target domain calibrated remote sensing image respectively. Geometric registration processing includes selecting control point pairs and constructing a spatial transformation model. The spatial transformation model consists of translation parameters, rotation parameters and scale parameters. The transformation parameters that minimize the control point error are solved by the least squares method. The pixel coordinates of the source domain calibrated remote sensing image and the target domain calibrated remote sensing image are mapped to a unified spatial reference coordinate system. Bilinear interpolation is used to resample the mapped pixel values ​​to obtain the source domain registered remote sensing image and the target domain registered remote sensing image. The source domain registered remote sensing images and the target domain registered remote sensing images are subjected to spectral band unification processing. The spectral band unification processing includes constructing a band response matrix. The elements of the band response matrix are determined by the center wavelength and bandwidth of the corresponding bands of different sensors. The high-dimensional spectral response is mapped to a unified band space through matrix multiplication. The pixel spectral value in the unified band space is equal to the product of the band response matrix and the original spectral vector, thereby achieving the unification of the band dimension and band position of different sensors, and obtaining the source domain unified remote sensing image and the target domain unified remote sensing image. Spatial resolution unification processing is performed on the source domain unified remote sensing image and the target domain unified remote sensing image respectively. Spatial resolution unification processing includes constructing a resampling grid based on the target resolution, downsampling the high-resolution image using a region-weighted average method, and upsampling the low-resolution image using a bicubic interpolation method, so that the pixel size of the two domain remote sensing images is unified to a preset spatial scale, thus obtaining the source domain preprocessed remote sensing image and the target domain preprocessed remote sensing image.

[0020] In this embodiment, S2 specifically refers to: The source domain preprocessed remote sensing image and the target domain preprocessed remote sensing image are expanded in the spectral band dimension to form a spectral data matrix. Each row in the spectral data matrix corresponds to the reflectance vector of a uniform pixel in all spectral bands, and each column corresponds to the reflectance value of the uniform band at all pixel positions. Based on the preset window length, the spectral data matrix is ​​slid-sampled along the band direction. The spectral vectors of adjacent continuous bands are shifted and spliced ​​in column vector order to form a trajectory matrix. Each column of the trajectory matrix is ​​composed of the end-to-end connection of multiple continuous band spectral values ​​within the window. Singular value decomposition is performed on the trajectory matrix to obtain a left singular vector matrix, a singular value diagonal matrix and a right singular vector matrix. Each singular value in the singular value diagonal matrix is ​​arranged in descending order of its numerical value, and the corresponding left singular vectors form the basis vector set of the spectral subspace. The total sum of squares is obtained by summing the squares of the singular values. The sum of squares of the first few singular values ​​is then compared with the total sum of squares. When the ratio reaches a preset ratio threshold, the corresponding number of subspace basis vectors are selected as the main spectral component subspace. The basis vectors of the main spectral component subspace are multiplied with the original spectral data matrix to restore the pixel dimension space, forming a spectral subspace representation, and the source domain spectral features and target domain spectral features are obtained. Multi-scale directional spatial convolution kernel groups are constructed on the source domain preprocessed remote sensing images and the target domain preprocessed remote sensing images. The spatial convolution kernel group contains horizontal response kernels, vertical response kernels and diagonal response kernels of different scales. The weights of each convolution kernel satisfy the relationship of central symmetry or directional gradient distribution. Spatial convolution kernel groups are used to perform pixel-by-pixel convolution operations on the source domain preprocessed remote sensing image and the target domain preprocessed remote sensing image in two-dimensional space to generate convolution results. The convolution results are equal to the sum of the corresponding kernel weights and the reflectance values ​​of the covered pixels, resulting in spatial response maps of different scales and directions. Spatial response maps of different scales and directions are stacked side by side according to the channel dimension to form a spatial feature tensor. The channel arrangement order corresponds one-to-one with the convolution kernel group arrangement order, thus obtaining the source domain spatial features and the target domain spatial features.

[0021] In this embodiment, a spectral weighting mechanism is introduced through Grassmann manifold mapping, specifically as follows: Based on the source domain spectral features and the target domain spectral features, a band statistical matrix is ​​constructed in the pixel dimension. Each row of the band statistical matrix corresponds to the set of values ​​of the spectral band at all pixel positions. The mean vector and standard deviation vector are calculated along the pixel dimension of the band statistical matrix. The weight of each dimension of the band weight vector is equal to the ratio of the corresponding standard deviation to the mean and is normalized. The normalization process satisfies that the sum of the weights of all dimensions of the band weight vector is equal to 1. Based on the band weight vector, the source domain spectral features and the target domain spectral features are multiplied dimension by dimension in the band dimension to form weighted spectral features. In the weighted source domain spectral features, the spectral vector of each pixel is equal to the element-wise product of the original spectral vector and the band weight vector. In the weighted target domain spectral features, the spectral vector of each pixel is equal to the element-wise product of the original spectral vector and the band weight vector. The spectral covariance matrix is ​​calculated in the pixel dimension based on the weighted source domain spectral features and the weighted target domain spectral features respectively. The spectral covariance matrix is ​​equal to the product of the transpose of the weighted spectral feature matrix and the weighted spectral feature matrix, divided by the number of pixels minus 1. The spectral covariance matrix is ​​decomposed into eigenvalues ​​to obtain the set of eigenvalues ​​and the eigenvector matrix. The first k columns of eigenvectors are selected in descending order of eigenvalues ​​to form the basis of the spectral subspace. The subspace correlation matrix is ​​formed by performing matrix multiplication on the source domain spectral subspace basis and the target domain spectral subspace basis. The subspace correlation matrix is ​​equal to the product of the transpose of the source domain spectral subspace basis and the target domain spectral subspace basis. The subspace correlation matrix is ​​then subjected to singular value decomposition to obtain a set of singular values. Each dimension of the spectral principal angle vector is equal to the inverse cosine of the corresponding singular value. An angle diagonal matrix is ​​constructed based on the principal angle vector of the spectrum, and an orthogonal complementary basis matrix is ​​constructed based on the source domain spectral subspace basis. The source domain spectral subspace basis, the orthogonal complementary basis matrix, and the angle diagonal matrix are combined in block matrix form to form a geodesic transport mapping matrix. The geodesic transport mapping matrix is ​​multiplied by the source domain spectral feature matrix and the target domain spectral feature matrix, and the spectral features of the two domains are projected onto a unified subspace coordinate system to obtain aligned spectral features.

[0022] In this implementation, an edge-weighted mapping mechanism is introduced through Grassmann manifold mapping, specifically as follows: Based on the spatial features of the source domain and the spatial features of the target domain, a spatial gradient response matrix is ​​constructed in the spatial direction channel dimension. The spatial gradient response matrix contains horizontal response components, vertical response components and diagonal response components. Each channel of the spatial gradient response matrix corresponds one-to-one with the spatial filter output direction. The edge weight map is calculated based on the spatial gradient response matrix in the spatial location dimension. The weight of each pixel in the edge weight map is equal to the square root of the sum of the square of the horizontal response component, the square of the vertical response component, and the square of the diagonal response component. The weights of all pixel positions are normalized so that the weight range falls within the interval of 0 to 1. Based on the edge weight map, the source domain spatial features and the target domain spatial features are multiplied pixel by pixel in the spatial coordinate dimension to form a weighted spatial feature. The weighted source domain spatial features and the weighted target domain spatial features maintain the original spatial response structure unchanged in the channel dimension. Based on the weighted source domain spatial features and the weighted target domain spatial features, spatial covariance matrices are constructed in the pixel dimension. The spatial covariance matrix is ​​equal to the product of the transpose of the weighted spatial feature matrix and the weighted spatial feature matrix, divided by the number of pixels minus 1. The spatial covariance matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalue set and the eigenvector matrix. The first few columns of eigenvectors are selected in descending order of eigenvalue size to form the spatial subspace basis. A spatial correlation matrix is ​​formed by performing matrix multiplication based on the source domain spatial subspace basis and the target domain spatial subspace basis. The spatial correlation matrix is ​​then subjected to singular value decomposition to obtain a set of singular values. The angle value of each dimension in the spatial principal angle vector is equal to the inverse cosine value of the corresponding singular value. An angle diagonal matrix is ​​constructed based on the principal angle vector of the space, and an orthogonal complementary basis matrix is ​​constructed based on the source domain space subspace basis. The source domain space subspace basis, the orthogonal complementary basis matrix and the angle diagonal matrix are combined in block matrix form to form a geodesic transmission mapping matrix. The geodesic transmission mapping matrix is ​​multiplied by the source domain space feature matrix and the target domain space feature matrix. The spatial features of the two domains are projected onto a unified subspace coordinate system to obtain aligned spatial features.

[0023] In this embodiment, the improved MDANet model specifically includes a spectral coding module, a null coding module, and a fusion module; The spectral coding module constructs a multi-channel convolutional mapping structure for aligned spectral features. The pixel value of any channel in the convolution output is equal to the sum of the sum of the element-wise multiplication of each weight in the convolution kernel weight matrix and the pixel value at the same position in the corresponding spectral neighborhood pixel matrix, plus the bias term, forming a linear response matrix. A piecewise linear activation function or an exponentially normalized activation function is applied to each pixel in the linear response matrix to generate a spectral coding feature matrix. The empty coding module constructs a multi-channel convolutional mapping structure for the aligned spatial features. The pixel value of any channel in the convolution output is equal to the sum of the sum of the element-wise multiplication of each weight in the convolution kernel weight matrix and the pixel value at the same position in the corresponding spatial neighborhood pixel matrix, plus the bias term, forming a linear response matrix. A nonlinear activation mapping is applied to each pixel in the linear response matrix to generate a spatial coding feature matrix. The fusion module constructs a geodesic transmission mapping matrix based on the spectral principal angle vector and the spatial principal angle vector. The mapping coefficient in the i-th row and j-th column of the geodesic transmission mapping matrix is ​​equal to the sum of the cosine function value of the corresponding principal angle component multiplied by the first weight coefficient and the sine function value multiplied by the second weight coefficient. The spectral coding feature matrix and the spatial coding feature matrix are respectively multiplied by the geodesic transmission mapping matrix to form the spectral transmission feature matrix and the spatial transmission feature matrix, so that the spectral coding feature matrix and the spatial coding feature matrix are projected onto the unified subspace coordinate system in the column space direction. A cross-modal attention weight matrix is ​​constructed based on the spectral transmission feature matrix and the spatial transmission feature matrix. The weight in the m-th row and n-th column of the cross-modal attention weight matrix is ​​equal to the ratio of the sum of the inner product of the m-th eigenvector in the spectral transmission feature matrix and the n-th eigenvector in the spatial transmission feature matrix after exponential mapping to the sum of all inner product exponential mapping results. The spectral transmission feature matrix and the spatial transmission feature matrix are weighted element-wise by their corresponding weights, and then concatenated in the channel dimension in the order of spectral channel first and spatial channel last to form a weighted feature group. The weighted feature group is then summed channel-wise in the concatenation channel direction to form the spectral enhancement feature matrix and the spatial enhancement feature matrix. An angle-gated weight vector is constructed based on the spectral principal angle vector and the spatial principal angle vector. The kth weight in the angle-gated weight vector is equal to the proportion of the corresponding principal angle component after being mapped by a negative exponential function and normalized by all weights. The spectral enhancement feature matrix and the spatial enhancement feature matrix are weighted by channel-by-channel multiplication in channel order. The weighted result is summed element-by-element in the pixel dimension to form a fusion feature matrix. The source domain fusion features are output based on the fusion feature matrix formed by preprocessing the remote sensing images in the source domain, and the target domain fusion features are output based on the fusion feature matrix formed by preprocessing the remote sensing images in the target domain.

[0024] In this embodiment, the improved MDANet model introduces spectral manifold alignment results and spatial manifold alignment results as unified geometric constraint inputs on the basis of a multi-domain feature adaptive network structure. This extends the traditional multi-domain fusion structure based on statistical distribution alignment to a geometrically consistent fusion structure based on the Grassmann subspace principal angle evolution relationship. By constructing a geodesic transport mapping matrix jointly controlled by the spectral principal angle vector and the spatial principal angle vector within the fusion module, spectral and spatial encoded features achieve continuous geometric transport alignment within a unified subspace coordinate system, breaking through the existing MDANet model's flat space fusion method that relies solely on feature concatenation or weighted summation. Furthermore, a cross-modal attention weight matrix is ​​introduced within the unified subspace coordinate system to establish explicit... The model establishes a cross-correlation relationship to achieve dynamic synergistic enhancement of spectral and spatial structural information. Based on the principal angle evolution relationship, an angle-gated weight vector is constructed to implement channel-level geometric constraint fusion of spectral and spatial enhancement features, enabling the fusion process to adaptively adjust according to the degree of subspace offset between the source and target domains. By integrating manifold geometric transmission, cross-modal mutual attention modeling, and angle-gated fusion mechanism into the MDANet model fusion module, a geometric perception fusion network structure for multi-source remote sensing cross-domain scenarios is formed. This breaks through the domain adaptation paradigm of the existing MDANet model, which is based on statistical consistency, at the structural level. It constructs a multimodal deep fusion path based on subspace geometric evolution constraints, and has synergistic innovative features such as continuous subspace alignment, cross-modal correlation enhancement, and adaptive adjustment of geometric constraints.

[0025] In this embodiment, the source domain fusion features and the target domain fusion features are input into the segmentation operator to generate a source domain segmentation probability map and a target domain segmentation probability map, and a segmentation loss is constructed, specifically as follows: The source domain fusion feature and the convolution kernel weight tensor in the segmentation operator are convolved and mapped channel by channel. The pixel score of any class channel in the convolution output is equal to the sum of the element-wise multiplication of the corresponding convolution kernel weight matrix and the neighborhood pixel matrix of the fusion feature, and a bias term is added to form the source domain pixel score tensor. The ratio of the c-th component of the source domain pixel score tensor in the class channel direction after being mapped by the exponential function to the sum of the exponential mapping results of all classes forms the corresponding class probability value in the source domain segmentation probability map. The target domain fusion features and the convolution kernel weight tensor in the segmentation operator are subjected to channel-wise convolution mapping. The pixel score of any class channel in the convolution output is equal to the sum of the element-wise multiplication of the corresponding convolution kernel weight matrix and the neighborhood pixel matrix of the fusion features, and a bias term is added to form the target domain pixel score tensor. The ratio of the c-th component of the target domain pixel score tensor in the class channel direction after being mapped by an exponential function to the sum of the exponential mapping results of all classes forms the corresponding class probability value in the target domain segmentation probability map. The source domain annotation map represents the category indication information in the form of one-hot vectors at the pixel positions. The one-hot vectors have a value of 1 for the position corresponding to the true category and a value of 0 for the other category positions. The category indication vectors in the source domain annotation map are multiplied pixel by pixel with the corresponding category probability values ​​in the source domain segmentation probability map. The natural logarithm of the product is taken, and the negative value is obtained after summing the category channel and spatial pixel dimensions to obtain the segmentation loss. This loss is then normalized according to the total number of pixels in the source domain annotation map. The normalization process is to divide the sum of the segmentation loss in the spatial pixel dimension and the category channel dimension by the total number of pixel positions N in the source domain annotation map, where N is equal to the product of the number of rows and columns in the source domain annotation map. The normalized segmentation loss is equal to the ratio of the original segmentation loss sum to N, so that remote sensing image samples of different sizes have consistent scale weights in the segmentation loss calculation process.

[0026] In this embodiment, a pseudo-label map and a confidence mask for the target domain are generated based on the target domain segmentation probability map, specifically as follows: The target domain segmentation probability map forms a set of probability vectors in the category channel direction. The probability vector at any pixel position contains the probability component of the corresponding category. The magnitude of each component in the probability vector is compared, and the value of the largest probability component is selected as the probability value of the corresponding pixel position in the target domain maximum probability map. At the same time, the channel index where the largest probability component is located is recorded as the category index value of the corresponding pixel position in the target domain category index map. The category index value of each pixel in the target domain category index map is used as the category identifier value of the target domain pseudo-label map, forming a target domain pseudo-label map that corresponds one-to-one with the target domain preprocessed remote sensing image in terms of spatial size. The probability value of any pixel in the maximum probability map of the target domain is compared with a set confidence threshold pixel by pixel. When the result of the maximum probability value of the target domain minus the confidence threshold is greater than or equal to 0, the corresponding pixel is assigned a value of 1. When the result of the maximum probability value of the target domain minus the confidence threshold is less than 0, the corresponding pixel is assigned a value of 0. This forms a confidence mask that corresponds one-to-one with the preprocessed remote sensing image of the target domain in terms of spatial size. Pixels with a value of 1 in the confidence mask are designated as high-confidence pseudo-labeled areas, and pixels with a value of 0 are designated as low-confidence exclusion areas.

[0027] In this implementation, a maximum mean difference loss is constructed based on source domain fusion features and target domain fusion features, and a corner kernel mechanism is introduced to calculate the weighted maximum mean difference loss; the class-conditional maximum mean difference loss is calculated based on the target domain pseudo-label map and the confidence mask, specifically as follows: The source domain fusion features are flattened according to pixel position to form a source domain feature set, and the target domain fusion features are flattened according to pixel position to form a target domain feature set. Each feature vector in the source domain feature set is aligned with each feature vector in the target domain feature set in the feature dimension. The angular kernel weighting coefficient is constructed based on the spectral principal angle vector and the spatial principal angle vector. The angular kernel weighting coefficient is equal to the scalar value obtained by mapping the result of the summation of the spectral principal angle vector and the spatial principal angle vector by components through a negative exponential function. The weighted maximum mean difference loss is constructed based on the source domain feature set and the target domain feature set. The weighted maximum mean difference loss is equal to the sum of the kernel function values ​​of any two feature vectors in the source domain feature set multiplied by the angle kernel weight coefficient, and the sum of the kernel function values ​​of any two feature vectors in the target domain feature set multiplied by the angle kernel weight coefficient, minus twice the sum of the kernel function values ​​of any two feature vectors between the source domain feature set and the target domain feature set multiplied by the angle kernel weight coefficient. The kernel function value is equal to the similarity value obtained by mapping the square norm of the difference between the two feature vectors through the negative exponential function. The target domain feature set is filtered based on the target domain pseudo-annotated map and confidence mask. The corresponding target domain feature vector is retained at the pixel position with a value of 1 in the confidence mask, and the corresponding target domain feature vector is removed at the pixel position with a value of 0 in the confidence mask, so as to obtain the effective feature set of the target domain. Based on the pseudo-annotated map of the target domain, the effective feature set of the target domain is grouped by category. The effective feature vectors of the target domain corresponding to the same category index in the pseudo-annotated map of the target domain constitute the category feature subset of the target domain. Based on the source domain annotation map, the source domain feature set is grouped by category. The source domain feature vectors corresponding to the same category index in the source domain annotation map constitute the source domain category feature subset. For any category index, construct the class-conditional maximum mean difference loss based on the source domain category feature subset and the target domain category feature subset. The class-conditional maximum mean difference loss is equal to the sum of the kernel function values ​​of any two feature vectors in the source domain category feature subset and the sum of the kernel function values ​​of any two feature vectors in the target domain category feature subset, minus twice the sum of the kernel function values ​​of any two feature vectors between the source domain category feature subset and the target domain category feature subset. The total maximum mean difference loss of class conditions is obtained by summing the maximum mean difference loss of class conditions for all category indexes.

[0028] In this embodiment, a joint loss function is generated by fusion, and the parameters of the improved MDANet model and the segmentation operator are updated to output a target domain classification map, specifically: A joint loss function is constructed based on the segmentation loss, the weighted maximum mean difference loss, and the total condition maximum mean difference loss. The joint loss function is equal to the algebraic sum of the segmentation loss and the weight coefficient 1 multiplied by the weighted maximum mean difference loss and the weight coefficient 2 multiplied by the total condition maximum mean difference loss, where the weight coefficient 1 and the weight coefficient 2 are positive real numbers and satisfy that the sum of the weight coefficient 1 and the weight coefficient 2 is less than or equal to 1. The gradients of the parameters of the spectral coding module, the empty coding module, the fusion module, the convolution kernel weight matrix and the bias term in the segmentation operator in the improved MDANet model are calculated based on the joint loss function. The gradient of any parameter is equal to the partial derivative of the joint loss function with respect to that parameter. The parameter update amount is equal to the product of the learning rate and the corresponding gradient. The updated parameter value is equal to the parameter value before the update minus the parameter update amount. The target domain preprocessed remote sensing images are sequentially input into the improved MDANet model after parameter updates to obtain target domain fusion features. The target domain fusion features are then input into the segmentation operator after parameter updates to obtain the target domain pixel score tensor. The ratio of the c-th component of the target domain pixel score tensor in the category channel direction after being mapped by an exponential function to the sum of the exponential mapping results of all categories forms the corresponding category probability in the target domain segmentation probability map. The target domain segmentation probability map is subjected to a pixel-by-pixel maximum value selection operation in the category channel. The category index corresponding to the maximum probability constitutes the target domain category index map, which is then output as the target domain classification map.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to land cover classification in a coastal plain and suburban area. The source domain remote sensing image came from a high-resolution satellite with a spatial resolution of 1.0m, and a source domain annotation map corresponding to each pixel of the source domain remote sensing image was provided. The target domain remote sensing image came from another multispectral platform with a spatial resolution of 10.0m. The imaging conditions included fog and differences in solar altitude angle, resulting in significant inconsistencies in radiometric response, spectral band settings, and texture details. Traditional direct training resulted in problems such as confusion between water bodies and shadows, fragmented boundaries between cultivated land and bare land, and outward expansion of road boundaries in the target domain.

[0030] In this scenario, radiometric calibration, geometric registration, spectral band consistency, and spatial resolution consistency are first performed on the source and target domain remote sensing images to obtain preprocessed source and target domain remote sensing images. Then, singular spectral analysis is used to extract source and target domain spectral features from these images, and spatial filtering is used to extract source and target domain spatial features. Finally, based on Grassmann manifold mapping, spectral weighting and edge weighting mechanisms are introduced to complete spectral and spatial alignment, resulting in aligned spectral features, aligned spatial features, spectral principal angle vectors, and spatial principal angle vectors. The above input is then fed into the improved MDANet model. A geodesic mutual attention mechanism is introduced in the fusion module to output source domain fusion features and target domain fusion features. The source domain segmentation probability map and target domain segmentation probability map are obtained through the segmentation operator. The target domain segmentation probability map is used to generate a target domain pseudo-label map and a confidence mask. The parameters of the improved MDANet model and the segmentation operator are updated by fusing the segmentation loss, weighted maximum mean difference loss and class conditional maximum mean difference loss to form a joint loss function. The target domain classification map is then output.

[0031] Table 1. Comparison of Accuracy and Efficiency of Target Domain Classification Map

[0032] Analysis of Table 1 shows that with only preprocessing and segmentation operators, the overall accuracy of the target domain is 72.8%, while the boundary F1 is only 41.3%, indicating that domain offset leads to significant boundary expansion and class confusion. After adding weighted maximum mean difference loss and class conditional maximum mean difference loss, the overall distribution and class structure are simultaneously constrained, and the overall accuracy of the target domain improves to 80.2%. When the spectral weight moment mechanism and edge weight mapping mechanism are removed, the overall accuracy of the target domain remains at 82.7%, indicating that the lack of spectral alignment and spatial alignment under Grassmann manifold mapping limits the upper limit. The entire process of this invention achieves 88.9%, the boundary F1 is improved to 61.8%, the road and water body boundaries converge more tightly, and the inference time only increases by 9ms, meeting the requirements of near real-time mapping.

[0033] Table 2 Comparison of pseudo-labeling and loss convergence process data

[0034] Analysis of Table 2 shows that the confidence mask coverage was 34.2% in the early training stage, and the accuracy of the target domain pseudo-annotated map was 83.1%. While some uncertain pixels were still excluded, error accumulation was suppressed. As the joint loss function drove the updates of the improved MDANet model parameters and segmentation operator parameters, the weighted maximum mean difference loss decreased from 0.92 to 0.49, and the class-conditional maximum mean difference loss decreased from 0.68 to 0.35, reflecting the synchronous convergence of the source domain fusion features and target domain fusion features in terms of overall distribution and class-conditional structure. The confidence mask coverage gradually increased to 61.3%, and the accuracy of the target domain pseudo-annotated map improved to 91.6%, indicating that the high-confidence region continued to expand and become more reliable. Finally, the overall accuracy of the target domain stabilized at 88.9%, verifying that the present invention can still output a highly consistent target domain classification map even without manual target domain annotation.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A classification method for multi-source remote sensing images based on pattern recognition, characterized in that, Includes the following steps: S1. Obtain the source domain remote sensing image, the target domain remote sensing image, and the source domain annotation map corresponding to the source domain remote sensing image, and perform preprocessing to obtain the source domain preprocessed remote sensing image and the target domain preprocessed remote sensing image; S2. Based on the preprocessed remote sensing images of the source domain and the preprocessed remote sensing images of the target domain, singular spectral analysis is used to extract the spectral features of the source domain and the target domain, and spatial filtering is used to extract the spatial features of the source domain and the target domain. S3. Based on the source domain spectral features and the target domain spectral features, a spectral weighting mechanism is introduced through Grassmann manifold mapping to perform spectral alignment and obtain aligned spectral features and spectral principal angle vectors. S4. Based on the spatial features of the source domain and the spatial features of the target domain, the edge weight mapping mechanism is introduced through Grassmann manifold mapping to perform spatial alignment and obtain the aligned spatial features and spatial principal angle vector. S5. Input the aligned spectral features, aligned spatial features, spectral principal angle vector and spatial principal angle vector into the improved MDANet model. Introduce the geodesic mutual attention mechanism into the fusion module of the improved MDANet model and output the source domain fusion features and the target domain fusion features. S6. Input the source domain fusion feature and the target domain fusion feature into the segmentation operator to generate the source domain segmentation probability map and the target domain segmentation probability map, and construct the segmentation loss. Based on the target domain segmentation probability map, generate the target domain pseudo-label map and confidence mask. S7. Construct the maximum mean difference loss based on the source domain fusion features and the target domain fusion features, and introduce the corner kernel mechanism to calculate the weighted maximum mean difference loss; calculate the class conditional maximum mean difference loss based on the target domain pseudo-label map and confidence mask; generate a joint loss function through fusion, and update the parameters of the improved MDANet model and the segmentation operator parameters to output the target domain classification map.

2. The classification method for multi-source remote sensing images based on pattern recognition according to claim 1, characterized in that, Specifically, S1 is: Acquire source domain remote sensing images, target domain remote sensing images, and source domain annotation maps corresponding to the source domain remote sensing images; Radiometric calibration is performed on the source domain remote sensing image and the target domain remote sensing image respectively to obtain the source domain calibrated remote sensing image and the target domain calibrated remote sensing image; Geometric registration processing is performed on the source domain calibrated remote sensing image and the target domain calibrated remote sensing image respectively to obtain the source domain registered remote sensing image and the target domain registered remote sensing image; Perform spectral band unification processing on the source domain registered remote sensing image and the target domain registered remote sensing image respectively to obtain the source domain unified remote sensing image and the target domain unified remote sensing image; Spatial resolution uniformity processing is performed on the source domain uniform remote sensing image and the target domain uniform remote sensing image respectively to obtain the source domain preprocessed remote sensing image and the target domain preprocessed remote sensing image.

3. The classification method for multi-source remote sensing images based on pattern recognition according to claim 1, characterized in that, Specifically, S2 is: The source domain preprocessed remote sensing images and the target domain preprocessed remote sensing images are respectively constructed into spectral data matrices according to the band dimension. The rows of the spectral data matrix correspond to the pixel positions, and the columns correspond to the spectral bands. Trajectory matrices are constructed from the spectral data matrices. The trajectory matrices are formed by shifting and splicing continuous band spectral vectors according to a preset window length. Singular value decomposition is performed on the trajectory matrix to obtain the singular value sequence and the corresponding feature subspace basis vectors; Based on the proportion of the sum of singular values ​​to the total sum of singular values, the feature subspace basis vectors corresponding to the main spectral components are selected to form the spectral subspace representation. The spectral subspace representation is then mapped back to the original pixel space to obtain the source domain spectral features and the target domain spectral features. Spatial convolution kernel groups are constructed for the source domain preprocessed remote sensing images and the target domain preprocessed remote sensing images, respectively. The spatial convolution kernel groups contain multi-scale directional response kernels. The aforementioned spatial convolution kernel group is used to perform convolution operations on the source domain preprocessed remote sensing image and the target domain preprocessed remote sensing image respectively to obtain a multi-scale spatial response map; Multi-scale spatial response maps are stitched together along the channel dimension to form a spatial feature tensor, thus obtaining the source domain spatial features and the target domain spatial features.

4. The classification method for multi-source remote sensing images based on pattern recognition according to claim 1, characterized in that, The spectral weighting mechanism introduced through Grassmann manifold mapping is specifically as follows: The band weight vector is calculated based on the source domain spectral features and the target domain spectral features, and each dimension of the band weight vector corresponds to a spectral band. The source domain spectral features and the target domain spectral features are weighted based on the band weight vector to obtain the weighted source domain spectral features and the weighted target domain spectral features. Based on the weighted source domain spectral features and the weighted target domain spectral features, spectral covariance matrices are constructed respectively. The source domain spectral subspace basis and the target domain spectral subspace basis are obtained by eigenvalue decomposition of the spectral covariance matrices. The principal angle matrix is ​​calculated based on the source domain spectral subspace basis and the target domain spectral subspace basis, and the spectral principal angle vector is obtained by performing an inverse cosine operation on the principal angle matrix. Based on the source domain spectral subspace basis, the target domain spectral subspace basis, and the spectral principal angle vector, a geodesic transport mapping is constructed to map the source domain spectral features and the target domain spectral features to a unified subspace coordinate system, thereby obtaining aligned spectral features.

5. The classification method for multi-source remote sensing images based on pattern recognition according to claim 1, characterized in that, The edge-weight mapping mechanism introduced through Grassmann manifold mapping is specifically as follows: An edge weight map is constructed based on the spatial features of the source domain and the spatial features of the target domain. The edge weight map is a weight raster map that corresponds one-to-one with the spatial features of the source domain and the spatial features of the target domain in spatial coordinates. The weight of each pixel position in the edge weight map is determined by the gradient magnitude of the corresponding spatial feature. The gradient magnitude is obtained by taking the square root of the sum of the squares of the horizontal spatial response components, the vertical spatial response components, and the diagonal spatial response components. The source domain spatial features and target domain spatial features are weighted based on the edge weight map to obtain weighted source domain spatial features and weighted target domain spatial features; Based on the weighted source domain spatial features and the weighted target domain spatial features, spatial covariance matrices are constructed respectively, and the source domain spatial subspace basis and the target domain spatial subspace basis are obtained by eigenvalue decomposition of the spatial covariance matrices. The principal angle matrix is ​​calculated based on the source domain space subspace basis and the target domain space subspace basis, and the spatial principal angle vector is obtained by performing an inverse cosine operation on the principal angle matrix. Based on the source domain spatial subspace basis, the target domain spatial subspace basis, and the spatial principal angle vector, a geodesic transmission mapping is constructed to map the source domain spatial features and the target domain spatial features to a unified subspace coordinate system, thereby obtaining aligned spatial features.

6. The classification method for multi-source remote sensing images based on pattern recognition according to claim 1, characterized in that, The improved MDANet model specifically includes a spectral coding module, a null coding module, and a fusion module; The spectral coding module performs multi-channel convolution operations on the aligned spectral features. The pixel value of any channel is equal to the sum of the element-wise multiplication of the convolution kernel weight matrix and the corresponding spectral neighborhood pixel matrix, plus a bias term to form a linear response value. The linear response value is then subjected to a nonlinear activation mapping to obtain the spectral coding features. The nonlinear activation mapping uses a piecewise linear function or an exponential normalization function. The empty coding module performs multi-channel convolution operation on the aligned spatial features. The pixel value of any channel is equal to the sum of the element-wise multiplication of the convolution kernel weight matrix and the corresponding spatial neighborhood pixel matrix, plus a bias term to form a linear response value. A nonlinear activation mapping is applied to the linear response value to obtain the spatial coding feature. The improved MDANet model introduces a geodesic mutual attention mechanism in its fusion module. A geodesic transport mapping matrix is ​​constructed based on the spectral principal angle vector and the spatial principal angle vector. Any element in the geodesic transport mapping matrix is ​​equal to the mapping coefficient formed by combining the cosine and sine functions of the corresponding principal angle value according to weighted coefficients. The spectral coding features and spatial coding features are respectively multiplied with the geodesic transport mapping matrix to obtain the spectral transport features and spatial transport features. The spectral coding features and spatial coding features are then projected onto a unified subspace coordinate system. A cross-modal attention weight matrix is ​​constructed based on spectral transmission features and spatial transmission features. Any weight in the cross-modal attention weight matrix is ​​equal to the ratio of the inner product of the corresponding spectral transmission feature vector and spatial transmission feature vector after normalization by an exponential function. The spectral transmission features and spatial transmission features are weighted by element-wise multiplication according to their corresponding weights, and then summed after being concatenated in order along the channel dimension to form spectral enhancement features and spatial enhancement features. Angle-gated weight vectors are generated based on the spectral principal angle vector and the spatial principal angle vector. Any weight in the angle-gated weight vector is equal to the result of the corresponding principal angle component after exponential decay mapping and normalization. The spectral enhancement features and spatial enhancement features are weighted by channel-by-channel product in channel order and summed in the pixel dimension to form fusion features. The source domain fusion features are fusion features formed based on the source domain preprocessed remote sensing images. The target domain fusion features are fusion features formed based on the target domain preprocessed remote sensing images.

7. The classification method for multi-source remote sensing images based on pattern recognition according to claim 1, characterized in that, The process involves inputting the source domain fusion features and the target domain fusion features into the segmentation operator to generate source domain segmentation probability maps and target domain segmentation probability maps, and constructing a segmentation loss, specifically as follows: The source domain fusion features are input into the segmentation operator and convolution operation is performed to obtain the source domain pixel score tensor. The source domain pixel score tensor gives the corresponding score value in the pixel position and category channel. The source domain pixel score tensor is subjected to exponential normalization operation in the category channel to obtain the source domain segmentation probability map. The target domain fusion feature is input into the segmentation operator and convolution operation is performed to obtain the target domain pixel score tensor. The target domain pixel score tensor gives the corresponding score value in the pixel position and category channel. The target domain pixel score tensor is subjected to exponential normalization operation in the category channel to obtain the target domain segmentation probability map. The segmentation loss is constructed based on the source domain segmentation probability map and the source domain annotation map. The source domain annotation map provides a class indicator value at the pixel position. The segmentation loss is equal to the sum of the logarithmic products of the class indicator values ​​in the source domain annotation map and the corresponding class probabilities in the source domain segmentation probability map, and then the negative value is taken. The loss is normalized according to the number of pixels in the source domain annotation map.

8. The classification method for multi-source remote sensing images based on pattern recognition according to claim 1, characterized in that, The process of generating a pseudo-labeled map and a confidence mask for the target domain based on the target domain segmentation probability map is as follows: The maximum value selection operation is performed on the target domain segmentation probability map in the category channel to obtain the target domain maximum probability map and the target domain category index map. The target domain maximum probability map gives the maximum category probability value at the cell position, and the target domain category index map gives the category index corresponding to the maximum probability value at the cell position. Use the target domain category index map as the target domain pseudo-annotation map; The maximum probability image of the target domain is compared with a set confidence threshold. Positions where the maximum probability image value of the target domain is greater than or equal to the set confidence threshold are assigned a value of 1, and positions where the maximum probability image value of the target domain is less than the set confidence threshold are assigned a value of 0, thus forming a confidence mask.

9. A classification method for multi-source remote sensing images based on pattern recognition according to claim 1, characterized in that, The maximum mean difference loss is constructed based on the source domain fusion features and the target domain fusion features, and a corner kernel mechanism is introduced to calculate the weighted maximum mean difference loss; the class-conditional maximum mean difference loss is calculated based on the target domain pseudo-label map and the confidence mask, specifically as follows: The source domain fusion features and the target domain fusion features are expanded according to the pixel position to form the source domain feature set and the target domain feature set; Generating angular kernel weighting coefficients based on spectral principal angle vectors and spatial principal angle vectors; A weighted maximum mean difference loss is constructed based on the source domain feature set and the target domain feature set, combined with the corner kernel weight coefficients. The target domain pseudo-annotated map and confidence mask are used to filter the target domain feature set to form the effective feature set of the target domain; Based on the target domain pseudo-annotated map and the source domain annotated map, the effective feature set of the target domain and the feature set of the source domain are classified into target domain category feature subsets and source domain category feature subsets; The maximum mean difference loss of class conditions is constructed based on the subsets of category features in the source domain and the subsets of category features in the target domain, and the maximum mean difference loss of class conditions for all categories is summed to obtain the total maximum mean difference loss of class conditions.

10. A classification method for multi-source remote sensing images based on pattern recognition according to claim 1, characterized in that, The process involves fusing and generating a joint loss function, updating the parameters of the improved MDANet model and the segmentation operator, and outputting a target domain classification map. Specifically: A joint loss function is constructed based on the segmentation loss, the weighted maximum mean difference loss, and the total conditional maximum mean difference loss. The joint loss function is equal to the sum of the segmentation loss multiplied by the first weight coefficient, the weighted maximum mean difference loss multiplied by the second weight coefficient, and the total conditional maximum mean difference loss. The gradients of the parameters of the improved MDANet model and the parameters of the segmentation operator are calculated based on the joint loss function, and the parameters are updated. The parameter update amount is equal to the product of the learning rate and the corresponding gradient. The updated parameter value is equal to the parameter value before the update minus the parameter update amount. The updated improved MDANet model parameters are used to preprocess the remote sensing image of the target domain to generate target domain fusion features. The updated segmentation operator parameters are used to generate the target domain pixel score tensor and generate the target domain segmentation probability map from the target domain fusion features. The target domain segmentation probability map is subjected to a maximum value selection operation in the category channel to obtain the target domain category index map, which serves as the target domain classification map.

Images