Remote sensing image classification method based on multilevel differential aggregation gating discrimination network

Abstract

The invention discloses a remote sensing image classification method based on a multi-stage differential aggregation gating discrimination network. The method comprises the following steps: step 1, constructing the multi-stage differential aggregation gating discrimination network; step 2, data preprocessing; step 3, feature extraction based on ML-DAM; step 4, feature screening based on ADM; step 5, edge feature extraction based on Morpo-Sconv is carried out; step 6, feature fusion based on an adaptive weight coupling module; and step 7, outputting a result based on the classification module. According to the invention, the characteristic channel response is dynamically adjusted by using the gating mechanism of the adaptive discrimination module, the redundant information interference is suppressed, and the recognition precision of a specific category is improved; through a hierarchical differential aggregation strategy of the multi-stage differential aggregation module, the discrimination of the features is remarkably enhanced; the morphological strip convolution module is combined with morphological expansion and erosion operation, so that the sensing ability of the model to irregular boundaries and geometric structures is remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of ground object cover classification, and in particular to a remote sensing image classification method based on a multi-level differential aggregation gated discriminant network. Background Art

[0002] In recent years, with the continuous advancement of remote sensing observation technology, complex object classification based on hyperspectral imagery (HSI) has gradually become a research hotspot in remote sensing. Each pixel in HSI contains a complete spectral curve, accurately representing the reflectance characteristics of an object in different bands. This provides strong support for refined classification and identification, and has broad application prospects in fields such as urban planning and precision agriculture. By combining pixel-level spectral information with spatial context, HSI pixels can be accurately labeled. However, due to its relatively low spatial resolution and susceptibility to interference from factors such as the atmosphere and illumination, HSI still has shortcomings in representing spatial structure. In particular, in complex scenes with blurred object boundaries and similar inter-class structures, HSI alone cannot accurately distinguish object categories. In contrast, LiDAR, as an active remote sensing technology, offers high spatial sampling accuracy and three-dimensional modeling capabilities, can obtain precise elevation structure information, and has excellent anti-interference performance. With the rapid development of multimodal fusion technology in the field of artificial intelligence, the combination of HSI and LiDAR data has become a research hotspot in remote sensing image classification. Making full use of the complementary characteristics of the two heterogeneous data can effectively alleviate the occurrence of "same spectrum, different objects" and "same object, different spectrum" phenomena caused by a single HSI in remote sensing image classification.

[0003] Although convolutional neural networks and Transformer-based methods have achieved promising results in remote sensing image classification, they still suffer from the following shortcomings: 1. When processing multi-source remote sensing images with complex spatial structures, existing methods retain a large amount of redundant information during feature extraction and fusion. This redundant information interferes with each other, weakening the network's ability to distinguish specific categories and making it difficult to accurately model complex geometric structures and spatial dependencies, resulting in a decrease in the network's ability to perceive subtle spatial features. 2. Existing methods are often limited to extracting and fusing shallow, single-scale features, neglecting the extraction and fusion of multi-level features in complex landform scenes. 3. Although the attention mechanism can focus on important feature information, the weighted summation-based calculation method assigns weights to all input features. Even if the weights are low, they retain worthless information, making the model susceptible to noise interference and affecting classification performance. 4. Due to the irregular distribution of multi-source remote sensing images, grid-based methods have difficulty effectively modeling complex spatial geometric structures, resulting in insufficient ability to capture spatial details such as target boundaries and texture directionality. Classification confusion is prone to occur in scenes with variable morphologies of similar landforms or interlaced distribution of different landforms. Summary of the Invention

[0004] The purpose of the present invention is to provide a remote sensing image classification method based on a multi-level differential aggregation gated discriminant network to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a remote sensing image classification method based on a multi-level differential aggregation gated discriminant network, comprising the following steps: step 1, constructing a multi-level differential aggregation gated discriminant network; step 2, data preprocessing; step 3, feature extraction based on ML-DAM; step 4, feature screening based on ADM; step 5, edge feature extraction based on Morpho-Sconv; step 6, feature fusion based on an adaptive weight coupling module; step 7, result output based on a classification module;

[0006] In the above step 1, the constructed multi-level differential aggregation gated discriminant network includes a multi-level differential aggregation module ML-DAM, an adaptive discriminant module ADM, a morphological strip convolution module Morpho-Sconv, an adaptive weight coupling module and a classification module;

[0007] The above step 2 specifically includes the following steps:

[0008] 2.1 Input data: Input HSI data and LiDAR data. For each pixel, extract the adjacent cube from the HSI data, denoted as , extract adjacent patches from LiDAR data, denoted as ,in, Indicates the size of the patch. For edge pixels, use The width of the pixel is padded, and the category of each pixel can be represented by a one-hot encoding vector;

[0009] 2.2 Dataset Division: Collect all To form a collection , collect all To form a collection , divide the collected data into training set, validation set and test set;

[0010] The above step 3 specifically includes the following steps:

[0011] 3.1 Spectral feature extraction: For HSI data, input the multi-level differential aggregation module ML-DAM After a 50@1 1 1's 3D convolution increases the number of channels and generates feature maps , and then along the channel dimension Divided into five subgroups ,in, , , five-scale pseudo-3D convolution kernels are used to extract features layer by layer, and spectral features are obtained through differential operation and cascade;

[0012] 3.2 Elevation spatial feature extraction: For LiDAR data, input the multi-level differential aggregation module ML-DAM After a 50@1 1 2D convolution increases the number of channels and generates feature maps , and then along the channel dimension Divided into five subgroups ,in, , ,Use 2D convolution kernels of five scales to extract features layer by layer, and obtain the elevation spatial features through differential operation and cascade;

[0013] The above step 4 specifically includes the following steps:

[0014] 4.1 Multi-scale shallow feature extraction: Adaptive discriminant module ADM After the reconstruction operation, shallow features are extracted through three 2D convolution blocks of different scales. , and ,right Perform average pooling and max pooling To smooth the feature response, combine the features after the pooling operation, and activate the Sigmod function The features after the pooling operation The value of each pixel in is unified to between 0 and 1, and then and All are fed into the first gated discriminator;

[0015] 4.2 Gated Discriminator Screening Key Features: Introducing All-Zero Tensors in the Gated Discriminator , threshold and the discriminant factor Ω, through the gated discriminator to screen key features according to pixel ratio, dynamically adjust the feature channel response, suppress redundant information, and output the filtered spatial features;

[0016] The above step 5 specifically includes the following steps:

[0017] 5.1 Padding and morphological operations: Morphological strip convolution module Morpho-Sconv After the reconstruction operation, the feature map is obtained , for the feature map Filling operation is performed to retain edge information. Erosion and dilation operations are combined with two strip convolution kernels to extract four irregular morphological spatial features of HSI. The four irregular morphological spatial features are subjected to matrix multiplication and layer normalization to obtain erosion features. and expansion features ;

[0018] 5.2 Feature Fusion: Fusion of Erosion Features by Principal Element Addition and expansion features , output irregular spatial features ;

[0019] The above step 6 specifically includes the following steps:

[0020] 6.1 Weight Coupling: HSI Spectral Features Extracted by ML-DAM and HSI spatial features extracted by ADM , weight coupling is performed through the convolution and BN layers of the adaptive weight coupling module, and the scaling factor is used Assess redundancy;

[0021] 6.2 Dynamic Fusion: The adaptive weight coupling module adopts an adaptive weighted fusion strategy to output the spectral features after weight coupling. and spatial characteristics , and the LiDAR elevation spatial features extracted by ML-DAM and irregular spatial features of HSI extracted by Morpho-Sconv Perform dynamic feature fusion;

[0022] In the above step 7, the features after dynamic fusion in step 6.2 are Input into the classification module, and the classification module generates the final classification result.

[0023] Preferably, in step 2.2, the training set generation formula is as shown in formula (1):

[0024] (1)

[0025] The verification set generation formula is shown in formula (2):

[0026] (2)

[0027] The test set generation formula is shown in formula (3):

[0028] (3)

[0029] in, and They are HSI cube and LiDAR patch selected arbitrarily from the set, , and are the number of samples in the training set, validation set, and test set, respectively. is the corresponding one-hot encoded category vector.

[0030] Preferably, in step 3.1, the depth sequence of the five convolution kernels is expressed by formula (4):

[0031] (4)

[0032] in, is the depth of the jth pseudo-3D convolution kernel, and j is the sequence number of the current convolution layer;

[0033] The specific method for obtaining spectral features is: using the first pseudo-3D convolution , that is, the convolution kernel size is 10@1 1 The first set of feature maps after 11 convolution pair difference Perform feature extraction and then use the extracted features Divide it into two sub-feature maps again along the channel dimension and ,Will With the next set of feature maps After the cascade operation, it is input into the second set of corresponding pseudo-3D convolutions In this way, the first sub-feature map of each group is split The final output feature is obtained after cascading operations along the channel dimension , the above process can be expressed by equations (5), (6) and (7):

[0034] (5)

[0035] , (6)

[0036] (7)

[0037] in, and are the two sub-feature maps output by the j-th layer, Indicates the difference operation along the channel dimension, is the indicator function, when j 1 is 1, otherwise it is 0. This is the final output of the module.

[0038] Preferably, in step 3.2, the elevation spatial feature extraction process can be expressed by equations (8), (9) and (10):

[0039] (8)

[0040] (9)

[0041] (10)

[0042] in, and are the two sub-feature maps output by the j-th layer, Indicates the difference operation along the channel dimension, is the indicator function, when j 1 is 1, otherwise it is 0. This is the final output of the module.

[0043] Preferably, in step 4.1, the shallow feature extraction process can be expressed by equations (11) and (12):

[0044] (11)

[0045] (12)

[0046] in, represents a 2D convolutional block, Represents the LeakyReLU activation function, * represents the convolution operation, represents the weight of the convolution kernel, b represents bias, , and They represent the outputs of 2D convolutional blocks of different scales.

[0047] Preferably, in step 4.2, specifically: and At the same time, it is input into the discriminator, and the discriminator will automatically recognize The pixel information greater than ξ in the image is judged to be correct, that is, when the ratio of the number of pixels greater than ξ to the number of all pixels is greater than hour( ), the discriminator will think that the input features contain a large number of key features. The above process is expressed by equations (13), (14) and (15):

[0048] (13)

[0049] (14)

[0050] (15)

[0051] in, is the symbol prediction, is an indicator function used to screen key features. is the number of pixels in the discrete state, is the feature map area, is a feature map set, It is a sparse feature map after threshold judgment;

[0052] Afterwards, Input to a convolution kernel size of 2D convolutional blocks Further feature calibration is performed in Add together, and and Before adding conduct The decay of the ratio is because the more pixels with a value higher than ξ, the greater their contribution and the greater their influence on the features of different scales. When the judgment is wrong, it indicates that there is a lot of redundant information in this feature, and the discriminator will automatically give up. , direct output , the above process can be expressed by equations (16) and (17):

[0053] (16)

[0054] (17)

[0055] in, is the feature fusion function, is to find the optimal output, It is a distance metric whose purpose is to maintain the topological structure of the feature outline;

[0056] If the judgment is correct, the output of the first discriminator is and Perform the same operation as above to get the final output If the outputs of the two gated discriminators are both wrong, the features output by the two discriminators are and After adding, Multiply to get the final output , the above process is expressed by formula (19):

[0057] (19).

[0058] Preferably, in step 5.1, the filling operation can be expressed by equations (20) and (21):

[0059] (20)

[0060] (twenty one)

[0061] Among them, k represents the original convolution kernel size, represents the effective size after expansion, d represents the expansion rate, Represents a filling operation;

[0062] The erosion operation process can be expressed by formula (22):

[0063] (twenty two)

[0064] in, Represents the elements of the convolution kernel, Represents the structural element used for erosion operations;

[0065] The expansion operation process can be expressed by formula (23):

[0066] (twenty three)

[0067] in, Represents the structural element used for dilation operation;

[0068] Obtaining Erosion Characteristics and expansion features The process is as follows: Input into four morphological convolutions respectively , , and The four irregular morphological spatial features of HSI are obtained respectively , , and ,in, and are two strip convolution kernels, , then, respectively and , and Matrix multiplication is performed to represent the correlation between features. The above process can be expressed by Equations (24) and (25):

[0069] (twenty four)

[0070] (25)

[0071] Among them, i and j represent the position of each pixel, ⊙ represents matrix multiplication, is layer normalization, and Normalized summary of the correlation weights of irregular morphological features at all locations for erosion and dilation operations, respectively.

[0072] Preferably, in step 6.1, for and The weight coupling can be expressed by formula (26):

[0073] (26)

[0074] in, is the scaling factor of the feature, and are the mean and variance of the feature channels, is a non-zero constant, is the global offset term, and the redundant information is given by Determined, when If it is small enough, it means that the input features are redundant features. In the process of weight coupling, it is through evaluation that The size of the value is used to identify and eliminate redundant features, thereby deciding whether to retain Or use be replaced.

[0075] Preferably, in step 6.2, the dynamic feature fusion process is expressed by formula (27):

[0076] (27).

[0077] Preferably, in step seven, the features after dynamic fusion are: Input into the classification module and pass through the average pooling layer in turn , Batch Norm layer , Mish activation function , reshape layer , dropout layer and linear layers , generate the final classification result, and the classification process is expressed by formula (28):

[0078] (28).

[0079] Compared with the existing technology, the beneficial effects of the present invention are: the present invention uses the gating mechanism of the adaptive discriminant module to dynamically adjust the feature channel response, suppresses redundant information interference, and improves the recognition accuracy of specific categories; the hierarchical differential aggregation strategy of the multi-level differential aggregation module significantly enhances the discriminability of features; the morphological strip convolution module combines morphological dilation and erosion operations, which significantly enhances the model's perception of irregular boundaries and geometric structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 This is a structural diagram of the multi-level differential aggregation gated discrimination network of the present invention;

[0081] Figure 2 It is a structural diagram of the ML-DAM of the present invention;

[0082] Figure 3 It is the ADM structure diagram of the present invention;

[0083] Figure 4 It is a structural diagram of the Morpho-Sconv of the present invention;

[0084] Figure 5 is another structural diagram of the Morpho-Sconv of the present invention;

[0085] Figure 6 is the Trento dataset; where:

[0086] (a) Pseudo-color composite image (31, 14, and 2 bands); (b) LiDAR image; (c) ground truth map;

[0087] Figure 7 It is the Houston2013 dataset; among them:

[0088] (a) Pseudo-color composite image (64, 30, and 20 bands); (b) LiDAR image; (c) ground truth map;

[0089] Figure 8 is the MUUFL dataset; where:

[0090] (a) Pseudo-color composite image (31, 16, and 6 bands); (b) LiDAR image; (c) ground truth map;

[0091] Figure 9 It is the full pixel classification map of the Trento dataset; among them: (a) DBCTNet; (b) DBMST; (c) S 2ENet; (d) HCTNet; (e) MHST; (f) M2FNet; (g) MS2CANet; (h) S2ATNet; (i) MDAGNet; (j) MS2CANet; (k) S2ATNet; (l) MDAGNet;

[0092] Figure 10 It is the full pixel classification map of Houston2013 dataset; among them: (a) DBCTNet; (b) DBMST; (c) S 2 ENet; (d) HCTNet; (e) MHST; (f) M2FNet; (g) MS2CANet; (h) S2ATNet; (i) MDAGNet; (j) MS2CANet; (k) S2ATNet; (l) MDAGNet;

[0093] Figure 11 is the full pixel classification map of the MUUFL dataset; where: (a) DBCTNet; (b) DBMST; (c) S 2 ENet; (d) HCTNet; (e) MHST; (f) M2FNet; (g) MS2CANet; (h) S2ATNet; (i) MDAGNet; (j) MS2CANet; (k) S2ATNet; (l) MDAGNet;

[0094] Figure 12 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0095] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0096] Please see the attached Figure 1 -Attached Figure 12 The present invention provides an embodiment: a remote sensing image classification method based on a multi-level differential aggregation gated discriminant network, comprising the following steps: step 1, constructing a multi-level differential aggregation gated discriminant network; step 2, data preprocessing; step 3, feature extraction based on ML-DAM; step 4, feature screening based on ADM; step 5, edge feature extraction based on Morpho-Sconv; step 6, feature fusion based on an adaptive weight coupling module; step 7, result output based on a classification module;

[0097] In the above step 1, the constructed multi-level differential aggregation gated discriminant network includes a multi-level differential aggregation module ML-DAM, an adaptive discriminant module ADM, a morphological strip convolution module Morpho-Sconv, an adaptive weight coupling module and a classification module;

[0098] The above step 2 specifically includes the following steps:

[0099] 2.1 Input data: Input HSI data and LiDAR data. For each pixel, extract the adjacent cube from the HSI data, denoted as , extract adjacent patches from LiDAR data, denoted as ,in, Indicates the size of the patch. For edge pixels, use The width of the pixel is padded, and the category of each pixel can be represented by a one-hot encoding vector;

[0100] 2.2 Dataset Division: Collect all To form a collection , collect all To form a collection , the collected data is divided into training set, validation set and test set; the training set generation formula is shown in formula (1):

[0101] (1)

[0102] The verification set generation formula is shown in formula (2):

[0103] (2)

[0104] The test set generation formula is shown in formula (3):

[0105] (3)

[0106] in, and They are HSI cube and LiDAR patch selected arbitrarily from the set, , and are the number of samples in the training set, validation set, and test set, respectively. is the corresponding one-hot encoded category vector;

[0107] The above step 3 specifically includes the following steps:

[0108] 3.1 Spectral feature extraction: For HSI data, input the multi-level differential aggregation module ML-DAM After a 50@1 1 1's 3D convolution increases the number of channels and generates feature maps , and then along the channel dimension Divided into five subgroups ,in, , , using pseudo-3D convolution kernels of five scales to extract features layer by layer, and obtaining spectral features through differential operation and cascade; where the depth sequence of the five convolution kernels is expressed by formula (4):

[0109] (4)

[0110] in, is the depth of the jth pseudo-3D convolution kernel, and j is the sequence number of the current convolution layer;

[0111] The specific method for obtaining spectral features is: using the first pseudo-3D convolution , that is, the convolution kernel size is 10@1 1 The first set of feature maps after 11 convolution pair difference Perform feature extraction and then use the extracted features Divide it into two sub-feature maps again along the channel dimension and ,Will With the next set of feature maps After the cascade operation, it is input into the second set of corresponding pseudo-3D convolutions In this way, the first sub-feature map of each group is split The final output feature is obtained after cascading operations along the channel dimension , the above process can be expressed by equations (5), (6) and (7):

[0112] (5)

[0113] , (6)

[0114] (7)

[0115] in, and are the two sub-feature maps output by the j-th layer, Indicates the difference operation along the channel dimension, is the indicator function, when j 1 is 1, otherwise it is 0. The final output of this module;

[0116] 3.2 Elevation spatial feature extraction: For LiDAR data, input the multi-level differential aggregation module ML-DAM After a 50@1 1 2D convolution increases the number of channels and generates feature maps , and then along the channel dimension Divided into five subgroups ,in, , , five scales of 2D convolution kernels are used to extract features layer by layer, and the elevation spatial features are obtained through differential operation and cascade. The elevation spatial feature extraction process can be expressed by equations (8), (9) and (10):

[0117] (8)

[0118] (9)

[0119] (10)

[0120] in, and are the two sub-feature maps output by the j-th layer, Indicates the difference operation along the channel dimension, is the indicator function, when j 1 is 1, otherwise it is 0. The final output of this module;

[0121] The above step 4 specifically includes the following steps:

[0122] 4.1 Multi-scale shallow feature extraction: Adaptive discriminant module ADM After the reconstruction operation, shallow features are extracted through three 2D convolution blocks of different scales. , and ,right Perform average pooling and max pooling To smooth the feature response, combine the features after the pooling operation, and activate the Sigmod function The features after the pooling operation The value of each pixel in is unified to between 0 and 1, and then and are fed into the first gated discriminator; the shallow feature extraction process can be expressed by equations (11) and (12):

[0123] (11)

[0124] (12)

[0125] in, represents a 2D convolutional block, Represents the LeakyReLU activation function, * represents the convolution operation, represents the weight of the convolution kernel, b represents bias, , and Represent the output of 2D convolution blocks of different scales;

[0126] 4.2 Gated Discriminator Screening Key Features: Introducing All-Zero Tensors in the Gated Discriminator , threshold And the discriminant factor Ω, through the gated discriminator according to the pixel ratio to filter the key features, dynamically adjust the feature channel response, suppress redundant information, and output the filtered spatial features; specifically: and At the same time, it is input into the discriminator, and the discriminator will automatically recognize The pixel information greater than ξ in the image is judged to be correct, that is, when the ratio of the number of pixels greater than ξ to the number of all pixels is greater than hour( ), the discriminator will think that the input features contain a large number of key features. The above process is expressed by equations (13), (14) and (15):

[0127] (13)

[0128] (14)

[0129] (15)

[0130] in, is the symbol prediction, is an indicator function used to screen key features. is the number of pixels in the discrete state, is the feature map area, is a feature map set, It is a sparse feature map after threshold judgment;

[0131] Afterwards, Input to a convolution kernel size of 2D convolutional blocks Further feature calibration is performed in Add together, and and Before adding conduct The decay of the ratio is because the more pixels with a value higher than ξ, the greater their contribution and the greater their influence on the features of different scales. When the judgment is wrong, it indicates that there is a lot of redundant information in this feature, and the discriminator will automatically give up. , direct output , the above process can be expressed by equations (16) and (17):

[0132] (16)

[0133] (17)

[0134] in, is the feature fusion function, is to find the optimal output, It is a distance metric whose purpose is to maintain the topological structure of the feature outline;

[0135] If the judgment is correct, the output of the first discriminator is and Perform the same operation as above to get the final output If the outputs of the two gated discriminators are both wrong, the features output by the two discriminators are and After adding, Multiply to get the final output , the above process is expressed by formula (19):

[0136] (19);

[0137] The above step 5 specifically includes the following steps:

[0138] 5.1 Padding and morphological operations: Morphological strip convolution module Morpho-Sconv After the reconstruction operation, the feature map is obtained , for the feature map Filling operation is performed to retain edge information. Erosion and dilation operations are combined with two strip convolution kernels to extract four irregular morphological spatial features of HSI. The four irregular morphological spatial features are subjected to matrix multiplication and layer normalization to obtain erosion features. and expansion features ; Among them, the filling operation can be expressed by equations (20) and (21):

[0139] (20)

[0140] (twenty one)

[0141] Among them, k represents the original convolution kernel size, represents the effective size after expansion, d represents the expansion rate, Represents a filling operation;

[0142] The erosion operation process can be expressed by formula (22):

[0143] (twenty two)

[0144] in, Represents the elements of the convolution kernel, Represents the structural element used for erosion operations;

[0145] The expansion operation process can be expressed by formula (23):

[0146] (twenty three)

[0147] in, Represents the structural element used for dilation operation;

[0148] Obtaining Erosion Characteristics and expansion features The process is as follows: Input into four morphological convolutions respectively , , and The four irregular morphological spatial features of HSI are obtained respectively , , and ,in, and are two strip convolution kernels, , then, respectively and , and Matrix multiplication is performed to represent the correlation between features. The above process can be expressed by Equations (24) and (25):

[0149] (twenty four)

[0150] (25)

[0151] Among them, i and j represent the position of each pixel, ⊙ represents matrix multiplication, is layer normalization, and is the normalized summary of the correlation weights of irregular morphological features at all locations for erosion and dilation operations, respectively;

[0152] 5.2 Feature Fusion: Fusion of Erosion Features by Principal Element Addition and expansion features , output irregular spatial features ;

[0153] The above step 6 specifically includes the following steps:

[0154] 6.1 Weight Coupling: HSI Spectral Features Extracted by ML-DAM and HSI spatial features extracted by ADM , weight coupling is performed through the convolution and BN layers of the adaptive weight coupling module, and the scaling factor is used Evaluate redundancy; where, for and The weight coupling can be expressed by formula (26):

[0155] (26)

[0156] in, is the scaling factor of the feature, and are the mean and variance of the feature channels, is a non-zero constant, is the global offset term, and the redundant information is given by Determined, when If it is small enough, it means that the input features are redundant features. In the process of weight coupling, it is through evaluation that The size of the value is used to identify and eliminate redundant features, thereby deciding whether to retain Or use be replaced;

[0157] 6.2 Dynamic Fusion: The adaptive weight coupling module adopts an adaptive weighted fusion strategy to output the spectral features after weight coupling. and spatial characteristics , and the LiDAR elevation spatial features extracted by ML-DAM and irregular spatial features of HSI extracted by Morpho-Sconv Perform dynamic feature fusion; the dynamic feature fusion process is expressed by formula (27):

[0158] (27);

[0159] In the above step 7, the features after dynamic fusion in step 6.2 are Input into the classification module and pass through the average pooling layer in turn , Batch Norm layer , Mish activation function , reshape layer , dropout layer and linear layers , generate the final classification result, and the classification process is expressed by formula (28):

[0160] (28).

[0161] Experimental Example 1:

[0162] In order to evaluate the effectiveness of the proposed Multi-level Differential Aggregation Gated Discriminant Network (MDAGNet), eight deep learning-based comparison algorithms were selected. Extensive experiments were conducted on the Trento dataset, Houston2013 dataset, and MUUFL dataset to verify the superiority and generalization of the proposed MDAGNet in the multi-source remote sensing image classification task. The detailed descriptions of these three datasets are as follows:

[0163] (1) Trento dataset: The Trento dataset is one of the most widely used datasets in the field of remote sensing. This dataset was taken in the countryside of Trento, Italy, and contains a variety of ground object types. The HSI was collected by the ASIA Eagle sensor, and the LiDAR DSM was collected by the Optech ALTM 3100EA sensor. The size of each band in the dataset is 600×166 pixels, with a spatial resolution of 1m. The hyperspectral image contains 63 bands with a wavelength range of 400-990nm. The dataset includes 6 ground object categories and 29,602 ground object real samples. The dataset contains 6 ground object coverage categories. For detailed information, see Table I-1 and Appendix. Figure 6 ;

[0164] (2) Houston2013 dataset: The Houston2013 dataset is from the campus of the University of Houston in Texas, USA and its surrounding urban areas. It is provided by the IEEE Geoscience and Remote Sensing Society (GRSS). The hyperspectral data is obtained by the CASI-1500 sensor. The size of each band in the dataset is 349×1905 pixels, the spatial resolution is 2.5m, and the hyperspectral image contains 144 bands with a wavelength range of 364-1046nm. The dataset includes 15 ground object categories and 14,713 ground object real samples. Due to the influence of weather during collection, there is a large cloud shadow on the right side of the HSI image, which poses a severe challenge to ground object classification. For detailed information, see Table I-2 and Appendix. Figure 7 ;

[0165] (3) MUUFL dataset: The MUUFL dataset was collected at the University of Mississippi, Gulf Park, and published by a research team at the University of Mississippi. The hyperspectral data was obtained by a CASI-1500 sensor. Each band of the dataset contains 325×220 pixels with a spatial resolution of 1m. Since the original hyperspectral data was severely affected by noise and there were invalid data bands, 8 spectral bands were removed and 64 bands were retained. The wavelength range is 375-1050nm. The dataset includes 11 ground object categories and 52,785 ground object real samples. The imbalance in the number of samples in different categories has a great impact on the classification. For detailed information, see Table I-3 and Appendix. Figure 8 ;

[0166] To verify the classification performance of MDAGNet, classification accuracy experiments were conducted on 3 public datasets for 8 comparison methods and MDAGNet. The training set, validation set, and test set all used independent data, and the training samples were determined by random sampling. The results of each method were averaged based on 5 rounds of experiments. Table II - Table IV and Appendix Figure 9 -Attached Figure 11 The classification results and full-pixel classification maps of different algorithms on three datasets are shown respectively. In Tables II-IV, the best classification results are marked in bold. The experimental results and analysis are as follows:

[0167] (1) Experimental results and analysis of the Trento dataset: In order to fully verify the effectiveness of the proposed method in the task of ground object classification, a comparative experiment was conducted on the Trento dataset with eight advanced algorithms. The detailed classification results and full pixel classification maps are shown in Table II and Appendix. Figure 9As shown in the table, the best classification accuracy of each category as well as OA, AA and Kappa are highlighted in bold. Through multiple experiments, the hyperparameters are set as follows on this dataset: the learning rate is set to 1e-3, the patch size is set to 7×7, and the threshold is set to 0.7. These settings enable the network to show the best classification performance on this dataset. Since the Trento dataset has sufficient training samples and fewer ground object categories, most methods have shown good classification performance on this dataset. Overall, the joint method based on HSI+LiDAR data as the dataset source The classification performance of the combined classification method is better than that of the classification method using only a single HSI as the data source. As can be seen from Table II, the OA, AA and Kappa coefficients of the proposed MDAGNet are all the best, which are 99.21%, 98.78% and 98.95 respectively. Among them, compared with the latest multi-source fusion classification methods MHST, M2FNet, MS2CANet and S2ATNet, they are significantly improved. This may be because the proposed method has stronger feature expression and discrimination capabilities in dealing with the spatial structure of complex objects. In specific categories, it is better than that of several In categories with complex structures or fuzzy boundaries, the proposed method shows stronger advantages. For example, in the C02 (Buildings) category, the classification performance of most methods is limited. The classification accuracy of DBMST in this category is only 90.53%. Although S2ATNet has achieved the best classification accuracy in this category, reaching 96.68%, its standard deviation is high. In contrast, the proposed MDAGNet achieves the second highest classification accuracy, reaching 96.65%. While approaching the classification performance of S2ATNet in this category, it shows stronger stability. This shows that the proposed The proposed method has stronger robustness in distinguishing transition areas such as building edges. In addition, in the extremely challenging category C06 (Roads), most methods generally face the problem of structural discontinuity and boundary confusion. Among them, the classification accuracy of HCTNet and MHST in this category is only 95.91% and 95.47%, respectively, which is much lower than the 98.09% classification accuracy obtained by the proposed network. This may be because the proposed method effectively alleviates the problem of edge information loss through morphological structure modeling, thereby improving the effectiveness of the network in modeling complex spatial structures. In terms of visual contrast, the attached Figure 9 The full pixel classification map in the figure further proves the above quantitative results. The full pixel classification map generated by the proposed method (see Appendix Figure 9 (l)) The boundaries of adjacent categories such as C06, C02 and C05 show clearer and more complete boundaries;

[0168] (2) Experimental results and analysis of the Houston2013 dataset: To further verify the classification performance of the proposed method in complex urban scenes, extensive comparative experiments were conducted with eight other comparison algorithms on the Houston2013 dataset, which contains 15 land object coverage categories. The dataset has rich land object types, high spectral similarity between categories, and complex spatial structure, which greatly tests the model's ability to discriminate different categories in a complex land object environment. The best classification accuracy of each category, as well as OA, AA, and Kappa are highlighted in bold in the table. Through multiple experiments, the hyperparameters were set as follows on this dataset: the learning rate was set to 1e-3, the patch size was set to 7×7, and the threshold was set to 0.1. These settings enabled the network to show the best classification performance on this dataset. Table III shows the classification results of the proposed method and the eight comparison algorithms on the Houston2013 dataset. Figure 10 The classification visualization results of different methods on the dataset are further shown. In the Houston2013 dataset, the proposed model achieved the best classification results in the three core evaluation indicators OA, AA, and Kappa coefficient, which are 90.22%, 92.04% and 89.42% respectively, which are comparable to S2ATNet (OA: 89.69%, AA: 91.06, Kappa: 88.50%) and S 2 Compared with ENet (OA: 89.47%, AA: 91.27%, Kappa: 88.62%), the proposed MDAGNet has significantly improved in classification accuracy. This result fully demonstrates that the proposed method has stronger robustness and generalization ability in dealing with multi-category complex urban remote sensing scene classification tasks with dense distribution of objects and severe spectral aliasing. In terms of the classification accuracy of specific categories, the proposed method can still show significant advantages in several key categories. In C03 (SyntheticGrass), the proposed method obtained a classification result of 100.00%, which is higher than other methods. The classification results for this category demonstrate the strong modeling ability of the proposed method for regular texture targets. In the C08 (Commercial) category, the classification accuracy of all methods is lower than 87%, and the proposed method achieves a classification accuracy of 89.85%, which is significantly higher than other methods. For the C10 (Highway) category with complex structure and severe spectral aliasing, MHST and MS2CANet achieve classification accuracy of 83.11% and 81.71% respectively. However, the proposed method achieves an excellent performance of 91.05% in this category, verifying its strong discrimination ability for complex boundary targets. Figure 10It can be observed from the full pixel classification diagram that the classification results of the proposed method on the whole image are more continuous, especially in areas such as urban roads, rail transit areas and densely built areas, the classification boundaries are compact and smooth, compared with HCTNet (Appendix Figure 10 (g)) and MS2CANet (with Figure 10 (k)) The proposed method is more stable in the segmentation of detail areas, which further verifies its good perception ability of complex spatial structures.

[0169] (3) Experimental results and analysis of the MUUFL dataset: The MUUFL dataset contains 11 ground object coverage categories. The spectral similarity between the ground objects is high, the category distribution is extremely unbalanced, and the sample size of some categories is extremely small, which significantly increases the difficulty of the model in multi-category classification tasks. According to the characteristics of the above-mentioned MUUFL dataset, higher requirements are placed on the classification method in terms of category discrimination ability and small sample information utilization ability. The best classification accuracy of each category as well as OA, AA and Kappa are highlighted in bold in the table. Through multiple experiments, the hyperparameters are set as follows on this dataset: the learning rate is set to 5e-4, the patch size is set to 9×9, and the threshold is set to 0.2. These settings enable the network to show the best classification performance on this dataset. Table IV and Appendix Figure 11 The classification results and classification visualization performance of all methods on the dataset are shown respectively. From the overall performance point of view, the proposed method achieved the best classification results in OA, AA and Kappa coefficient, which are 89.18%, 76.58 and 85.7 respectively. This shows that the proposed method still has robust classification ability and spatial structure perception ability when facing complex land object situations such as unbalanced data samples and blurred land object boundaries. In terms of the comparison of specific category accuracy, the proposed method shows obvious advantages in multiple challenging categories, especially in categories with unbalanced sample size or high confusion. In the C01 (Trees) category, the proposed method achieved an accuracy of 95.75%, which is higher than all the classification methods. This shows that it has shown high stability in texture modeling of large-area vegetation classes, and for C06 (Water) and C10 (Yellow) with small sample sizes and greater difficulty in judgment, the proposed method has a relatively high accuracy of 95.75%. Curb), the proposed method achieved classification accuracy of 96.49% and 22.44% respectively. In C06, although the classification accuracy of the proposed method is lower than that of MHST, the standard deviation of the proposed method is much smaller than that of MHST, showing good stability. In addition, in this category, only the proposed MDAGNet and MHST obtained classification results above 96%, while the classification accuracy of most other classification methods in this category is less than 80%. By observing the full pixel classification map (attached Figure 11) can also be found that the classification results of the proposed model (Appendix Figure 11 (l)) The boundaries are clear and smooth, especially in the intersection area of ​​buildings, roads and ground objects, with stronger category distinction ability.

[0170] Experimental Example 2:

[0171] In order to systematically evaluate the contribution and synergistic effect of each module in the proposed MDAGNet in land object classification, a set of ablation experiments with gradually increasing structural complexity were designed and tested on three typical multi-source remote sensing datasets: Trento, Houston2013 and MUUFL. Each set of experiments introduced key modules in the network structure in turn and constructed multiple sub-networks (Net_0 to Net_7) from basic feature extraction to complete fusion strategy to verify the influence of each component on the final classification accuracy. The experimental results are shown in Table V. Specifically, Net_0 contains two modules, HSI_ML-DAM and LiDAR_ML-DAM, which process HSI and LiDAR data respectively; Net_1 contains two modules, ADM and Morpho-Sconv; compared with Net_0, Net_2 adds a Morpho-Sconv module to enhance the perception of irregular boundaries and geometric structures; Net_3 adds ADM to suppress redundant information on the basis of Net_1 to improve the pertinence of feature selection; Net_4 is based on the proposed MDAGNet. HSI_ML-DAM was removed. Net_5 added ADM and a weight coupling strategy based on Net_1 to enhance the information fusion between multimodal features. Net_6 removed the weight coupling strategy based on MDAGNet to verify the classification performance of the model without the fusion strategy. Net_7 integrated all modules, namely the proposed MDAGNet. From the experimental results, Net_0, which integrates the HSI_ML-DAM and LiDAR_ML-DAM modules, achieved OA of 97.65%, 86.09%, and 86.07% on the Trento, Houston2013, and MUUFL datasets, respectively. This demonstrates the effectiveness of ML-DAM for multi-scale feature extraction on HSI and LiDAR data. Net_1 integrates the two core structures of ADM and Morpho-Sconv. However, compared with Net_0, Net_1's classification performance decreased on all datasets, especially on the Houston2013 and MUUFL datasets, where it decreased by 2.53% and 2.66%. Further observation shows that Net_0 and Net_1 have lower OA on the three datasets compared with Net_6, which integrates four core components: HSI_ML-DAM, LiDAR_ML-DAM, ADM and Morpho-Sconv. This shows that the four components achieve multi-level dynamic interaction of multimodal features and suppression of redundant information through synergy, significantly improving the richness and discrimination of feature expression, thereby enhancing the adaptability of the model to complex scenes and heterogeneous data. Compared with Net_0, Net_2 adds a Morpho-Sconv module, and its OA is There are improvements on all three datasets, which further proves that the Morpho-Sconv module has strong capabilities in modeling the edge structure of objects and extracting irregular spatial information, and is particularly suitable for scenes with complex terrain or blurred target contours. Net_3 adds a LiDAR_ML-DAM on the basis of Net_1, and its OA is significantly improved on all three datasets, which shows that by introducing the three-dimensional structure information of LiDAR data, the model's ability to represent the spatial distribution and geometric morphology of complex objects is improved. Net_4 removes LiDAR_ML-DAM from the proposed MDAGNet. On the Trento dataset, OA dropped to 98.44%, which is 0.77% lower than that of Net_7. This shows that although HSI plays a dominant role in this dataset, the elevation spatial information provided by LiDAR still has an impact on the overall classification performance. In Houston2013 and MUUFL, two datasets containing a large number of complex spatial structures such as buildings and roads, OA dropped to 88.75% and 87.85% respectively. This further verifies the necessity of elevation information in structural scenes, especially for modeling irregular boundaries and local spatial relationships. Net_5 plays an important role in the proposed MDAGN. et al., removing the weight coupling strategy to examine its role in multimodal feature interaction. Table V shows that the OA on the Trento, Houston 2013, and MUUFL datasets dropped to 98.56%, 89.74%, and 88.39%, respectively. This demonstrates that the weight coupling strategy plays a key role in the dynamic regulation of multimodal features. Finally, Net_7, which integrates all modules, achieved optimal classification performance on all three datasets. This further demonstrates that the proposed modules form a synergistic and complementary mechanism in functional design, greatly enhancing the network's classification performance and generalization capabilities.

[0172] Table I-1 Categories and sample numbers in the Trento dataset

[0173] Table I-2 Categories and sample numbers in the Housotn2013 dataset

[0174] Table I-3 Categories and sample numbers in the MUUFL dataset

[0175] Table II Classification results of the Trento dataset, standard deviation, OA (%), AA (%) and Kappa (the best results are highlighted in bold)

[0176]

[0177] Table III Classification results of Houston2013 dataset, standard deviation, OA (%), AA (%) and Kappa (the best results are highlighted in bold)

[0178]

[0179] Table IV Classification results of the MUUFL dataset, standard deviation, OA (%), AA (%) and Kappa (the best results are highlighted in bold)

[0180]

[0181] Table V Ablation experiments and analysis of different components within MDAGNet on three datasets (the best results are highlighted in bold)

[0182]

[0183] Based on the above, the advantage of the present invention is that when the invention is used, the multi-level differential aggregation module ML-DAM uses a hierarchical differential aggregation strategy to perform cross-level modeling on the spectral features of HSI and the elevation spatial features of LiDAR, generating a more discriminative multi-level feature representation, which significantly improves the modeling ability of the model for complex geometric structures and spatial dependency relationships. The adaptive discrimination module ADM introduces a gating mechanism and dynamically adjusts the response intensity of the feature channel to adaptively suppress redundant information according to the feature importance, enhances the expression of key features related to the target category, reduces the interference of redundant information on classification, and improves the recognition accuracy of specific categories. The morphological strip convolution module Morpho-Sconv combines morphological dilation and erosion operations to effectively capture the non-uniformity in remote sensing images. The uniform and irregular spatial distribution structure alleviates the feature deviation problem of existing methods when dealing with irregular boundaries and enhances the perception of target boundaries and geometric structures. The adaptive weight coupling module realizes the adaptive fusion of HSI spectral features and LiDAR elevation features through convolution layers, BN layers and dynamic weighting strategies, giving full play to the complementarity of heterogeneous data and alleviating the limitations of single modality data. The classification module maps the fused features to the final classification results. Experiments on the Trento, Houston2013 and MUUFL public datasets show that the overall classification accuracy, average precision and Kappa coefficient of the proposed method are better than those of the comparison algorithms, especially in the categories of objects with fuzzy boundaries and complex structures, showing higher classification accuracy and stability.

[0184] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A remote sensing image classification method based on a multi-level differential aggregation gated discriminant network includes the following steps: Step 1: Construct a multi-level differential aggregation gated discriminant network; Step 2: Data preprocessing; Step 3: Feature extraction based on ML-DAM; Step Step 4: feature screening based on ADM; Step 5: edge feature extraction based on Morpho-Sconv; Step 6: feature fusion based on adaptive weight coupling module; Step 7: result output based on classification module; It is characterized by: In the above step 1, the constructed multi-level differential aggregation gated discriminant network includes a multi-level differential aggregation module ML-DAM, an adaptive discriminant module ADM, a morphological strip convolution module Morpho-Sconv, an adaptive weight coupling module and a classification module; The above step 2 specifically includes the following steps: 2.1 Input data: Input HSI data and LiDAR data. For each pixel, extract the adjacent cube from the HSI data, denoted as , extract adjacent patches from LiDAR data, denoted as ,in, Indicates the size of the patch. For edge pixels, use The width of the pixel is padded, and the category of each pixel can be represented by a one-hot encoding vector; 2.2 Dataset Division: Collect all To form a collection , collect all To form a collection , divide the collected data into training set, validation set and test set; The above step 3 specifically includes the following steps: 3.1 Spectral feature extraction: For HSI data, input the multi-level differential aggregation module ML-DAM After a 50@1 1 1's 3D convolution increases the number of channels and generates feature maps , and then along the channel dimension Divided into five subgroups ,in, , , five-scale pseudo-3D convolution kernels are used to extract features layer by layer, and spectral features are obtained through differential operation and cascade; 3.2 Elevation spatial feature extraction: For LiDAR data, input the multi-level differential aggregation module ML-DAM After a 50@1 1 2D convolution increases the number of channels and generates feature maps , and then along the channel dimension Divided into five subgroups ,in, , ,Use 2D convolution kernels of five scales to extract features layer by layer, and obtain the elevation spatial features through differential operation and cascade; The above step 4 specifically includes the following steps: 4.1 Multi-scale shallow feature extraction: Adaptive discriminant module ADM After the reconstruction operation, shallow features are extracted through three 2D convolution blocks of different scales. , and ,right Perform average pooling and max pooling To smooth the feature response, combine the features after the pooling operation, and activate the Sigmod function The features after the pooling operation The value of each pixel in is unified to between 0 and 1, and then and All are fed into the first gated discriminator; 4.2 Gated Discriminator Screening Key Features: Introducing All-Zero Tensors in the Gated Discriminator , threshold and the discriminant factor Ω, through the gated discriminator to screen key features according to pixel ratio, dynamically adjust the feature channel response, suppress redundant information, and output the filtered spatial features; The above step 5 specifically includes the following steps: 5.1 Padding and morphological operations: Morphological strip convolution module Morpho-Sconv After the reconstruction operation, the feature map is obtained , for the feature map Filling operation is performed to retain edge information. Erosion and dilation operations are combined with two strip convolution kernels to extract four irregular morphological spatial features of HSI. The four irregular morphological spatial features are subjected to matrix multiplication and layer normalization to obtain erosion features. and expansion features ; 5.2 Feature Fusion: Fusion of Erosion Features by Principal Element Addition and expansion features , output irregular spatial features ; The above step 6 specifically includes the following steps: 6.1 Weight Coupling: HSI Spectral Features Extracted by ML-DAM and HSI spatial features extracted by ADM , weight coupling is performed through the convolution and BN layers of the adaptive weight coupling module, and the scaling factor is used Assess redundancy; 6.2 Dynamic Fusion: The adaptive weight coupling module adopts an adaptive weighted fusion strategy to output the spectral features after weight coupling. and spatial characteristics , and the LiDAR elevation spatial features extracted by ML-DAM and irregular spatial features of HSI extracted by Morpho-Sconv Perform dynamic feature fusion; In the above step 7, the features after dynamic fusion in step 6.2 are Input into the classification module, and the classification module generates the final classification result.

2. The remote sensing image classification method based on a multi-level differential aggregation gated discriminant network according to claim 1, characterized in that: In step 2.2, the training set generation formula is shown in formula (1): (1) The verification set generation formula is shown in formula (2): (2) The test set generation formula is shown in formula (3): (3) in, and They are HSI cube and LiDAR patch selected arbitrarily from the set, , and are the number of samples in the training set, validation set, and test set, respectively. is the corresponding one-hot encoded category vector.

3. The remote sensing image classification method based on a multi-level differential aggregation gated discriminant network according to claim 1, characterized in that: In step 3.1, the depth sequence of the five convolution kernels is expressed by formula (4): (4) in, is the depth of the jth pseudo-3D convolution kernel, and j is the sequence number of the current convolution layer; The specific method for obtaining spectral features is: using the first pseudo-3D convolution , that is, the convolution kernel size is 10@1 1 The first set of feature maps after 11 convolution pair difference Perform feature extraction and then use the extracted features Divide it into two sub-feature maps again along the channel dimension and ,Will With the next set of feature maps After the cascade operation, it is input into the second set of corresponding pseudo-3D convolutions In this way, the first sub-feature map of each group is split The final output feature is obtained after cascading operations along the channel dimension , the above process can be expressed by equations (5), (6) and (7): (5) , (6) (7) in, and are the two sub-feature maps output by the j-th layer, Indicates the difference operation along the channel dimension, is the indicator function, when j 1 is 1, otherwise it is 0. This is the final output of the module.

4. The remote sensing image classification method based on a multi-level differential aggregation gated discriminant network according to claim 1, characterized in that: In step 3.2, the elevation spatial feature extraction process can be expressed by equations (8), (9) and (10): (8) (9) (10) in, and are the two sub-feature maps output by the j-th layer, Indicates the difference operation along the channel dimension, is the indicator function, when j 1 is 1, otherwise it is 0. This is the final output of the module.

5. The remote sensing image classification method based on a multi-level differential aggregation gated discriminant network according to claim 1, characterized in that: In step 4.1, the shallow feature extraction process can be expressed by equations (11) and (12): (11) (12) in, represents a 2D convolutional block, Represents the LeakyReLU activation function, * represents the convolution operation, represents the weight of the convolution kernel, b represents bias, , and They represent the outputs of 2D convolutional blocks of different scales.

6. The remote sensing image classification method based on a multi-level differential aggregation gated discriminant network according to claim 1, characterized in that: In the step 4.2, specifically: and At the same time, it is input into the discriminator, and the discriminator will automatically recognize The pixel information greater than ξ in the image is judged to be correct, that is, when the ratio of the number of pixels greater than ξ to the number of all pixels is greater than hour( ), the discriminator will think that the input features contain a large number of key features. The above process is expressed by equations (13), (14) and (15): (13) (14) (15) in, is the symbol prediction, is an indicator function used to screen key features. is the number of pixels in the discrete state, is the feature map area, is a feature map set, It is a sparse feature map after threshold judgment; Afterwards, Input to a convolution kernel size of 2D convolutional blocks Further feature calibration is performed in Add together, and and Before adding conduct The decay of the ratio is because the more pixels with a value higher than ξ, the greater their contribution and the greater their influence on the features of different scales. When the judgment is wrong, it indicates that there is a lot of redundant information in this feature, and the discriminator will automatically give up. , direct output , the above process can be expressed by equations (16) and (17): (16) (17) in, is the feature fusion function, is to find the optimal output, It is a distance metric whose purpose is to maintain the topological structure of the feature outline; If the judgment is correct, the output of the first discriminator is and Perform the same operation as above to get the final output If the outputs of the two gated discriminators are both wrong, the features output by the two discriminators are and After adding, Multiply to get the final output , the above process is expressed by formula (19): (19)。 7. The remote sensing image classification method based on a multi-level differential aggregation gated discriminant network according to claim 1, characterized in that: In step 5.1, the filling operation can be expressed by equations (20) and (21): (20) (21) Among them, k represents the original convolution kernel size, represents the effective size after expansion, d represents the expansion rate, Represents a filling operation; The erosion operation process can be expressed by formula (22): (22) in, Represents the elements of the convolution kernel, Represents the structural element used for erosion operations; The expansion operation process can be expressed by formula (23): (23) in, Represents the structural element used for dilation operation; Obtaining Erosion Characteristics and expansion features The process is as follows: Input into four morphological convolutions respectively , , and The four irregular morphological spatial features of HSI are obtained respectively , , and ,in, and are two strip convolution kernels, , then, respectively and , and Matrix multiplication is performed to represent the correlation between features. The above process can be expressed by Equations (24) and (25): (24) (25) Among them, i and j represent the position of each pixel, ⊙ represents matrix multiplication, is layer normalization, and Normalized summary of the correlation weights of irregular morphological features at all locations for erosion and dilation operations, respectively.

8. The remote sensing image classification method based on a multi-level differential aggregation gated discriminant network according to claim 1, characterized in that: In step 6.1, for and The weight coupling can be expressed by formula (26): (26) in, is the scaling factor of the feature, and are the mean and variance of the feature channels, is a non-zero constant, is the global offset term, and the redundant information is given by Determined, when If it is small enough, it means that the input features are redundant features. In the process of weight coupling, it is through evaluation that The size of the value is used to identify and eliminate redundant features, thereby deciding whether to retain Or use be replaced.

9. The remote sensing image classification method based on a multi-level differential aggregation gated discriminant network according to claim 1, characterized in that: In step 6.2, the dynamic feature fusion process is expressed by formula (27): (27)。 10. The remote sensing image classification method based on a multi-level differential aggregation gated discriminant network according to claim 1, characterized in that: In the step 7, specifically: the dynamic fusion feature Input into the classification module and pass through the average pooling layer in turn , Batch Norm layer , Mish activation function , reshape layer , dropout layer and linear layers , generate the final classification result, and the classification process is expressed by formula (28): (28)。

Images