Hyperspectral cross-domain wetland mapping method and system based on fractional fourier rwkv network

Abstract

The application provides a hyperspectral cross-domain wetland mapping method and system of a fractional Fourier RWKV network, relates to the technical field of remote sensing intelligent processing, and comprises the following steps: acquiring a source domain hyperspectral image with labels and a target domain hyperspectral image without labels; performing feature conversion according to the source domain hyperspectral image and the target domain hyperspectral image to obtain a feature vector; performing space-spectrum interaction according to the feature vector to obtain a space-spectrum joint feature map with a large receptive field; performing network hierarchical design according to the space-spectrum joint feature map to obtain hierarchical space-spectrum features; performing generation optimization processing according to the hierarchical space-spectrum features to obtain space-spectrum features; and performing cross-domain alignment according to the space-spectrum features to obtain a cross-domain wetland classification map. The application realizes the joint mining of spatial, spectral, amplitude and phase information of the hyperspectral image, thereby enhancing the learning ability of the texture, details and structural features of the wetland.

Description

Technical Field

[0001] This invention relates to the field of remote sensing intelligent processing technology, and more specifically, to a hyperspectral trans-domain wetland mapping method and system using fractional Fourier RWKV networks. Background Technology

[0002] Coastal wetlands, as ecosystems of immense research value, provide habitats for flora and fauna, food for humans, and raw materials for industry. They play a crucial role in climate and hydrological regulation, promoting carbon sequestration and emission reduction, and maintaining ecological balance. Reliable monitoring methods are essential for wetland restoration, protection, and sustainable management. As an Earth observation technology integrating imaging and spectral techniques, hyperspectral remote sensing acquires images with rich spatial details and spectral information, offering unique advantages in identifying physical structure and composition. It has now become an important means of refined monitoring of coastal wetlands.

[0003] Deep learning technology, with its powerful feature learning capabilities and data generalization advantages, has been widely applied to tasks such as hyperspectral remote sensing image classification and coastal wetland mapping. However, the dynamic characteristics of coastal wetlands (such as tidal fluctuations and phenological changes) easily lead to significant spectral variations over time and in the environment. Furthermore, wetlands typically have multiple land cover types and mixed pixels, and due to their remote and complex environments, sample collection is costly and time-consuming. Hyperspectral images are usually collected from different geographical areas, at different times, or using different sensors. Differences in wetland types and spectral characteristics result in severe domain shifts between different scenes, posing challenges to the practical application of existing hyperspectral image classification and wetland mapping methods.

[0004] Existing hyperspectral wetland mapping algorithms based on deep learning and domain adaptation have the following shortcomings: First, due to the dynamic changes in wetlands, existing cross-domain classification and mapping methods often neglect the frequency and phase information of hyperspectral images, making them ill-equipped to handle domain shifts caused by variations in sensors, illumination, and phase. Second, Transformer-based methods typically suffer from quadratic complexity, and existing frequency domain analysis methods often neglect the effective utilization of phase information that characterizes the texture and structural features of hyperspectral images, leading to the loss of important details during data generation and cross-domain transfer.

[0005] Based on the shortcomings of the existing technologies, there is an urgent need for a hyperspectral trans-domain wetland mapping method and system using fractional Fourier RWKV networks. Summary of the Invention

[0006] The purpose of this invention is to provide a hyperspectral trans-domain wetland mapping method and system using fractional Fourier transform RWKV networks to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0007] In a first aspect, this application provides a hyperspectral trans-domain wetland mapping method using fractional Fourier RWKV networks, including:

[0008] Acquire labeled source domain hyperspectral images and unlabeled target domain hyperspectral images, wherein both the source domain hyperspectral images and the target domain hyperspectral images are hyperspectral remote sensing data containing spatial and spectral dimensions;

[0009] Based on the source domain hyperspectral image and the target domain hyperspectral image, feature transformation is performed by dividing pixels into token sequences and expanding them into channel representations, while applying rotation transformation in the fractional Fourier domain to obtain a feature vector that integrates spatial spectral information and fractional Fourier domain amplitude-phase information.

[0010] Based on the feature vector, spatial-spectral interaction is performed. By shifting and recurrently convolving the token sequence along the spatial dimension, long-range dependencies between positions are established. The token information is interactively fused along the channel dimension to obtain a spatial-spectral joint feature map with a large receptive field.

[0011] Based on the aforementioned spatial-spectral joint feature map, a hierarchical network design is performed. By stacking the feature representation and spatial-spectral interaction module layer by layer and alternately performing spatial and channel dimension mixing operations between layers, the amplitude-phase modulated hierarchical spatial-spectral features are obtained.

[0012] Based on the hierarchical spatial spectral features, the generation optimization process is performed. By constraining the amplitude component, the energy distribution of the generated features is consistent with the real features to maintain spectral fidelity. The structural correlation between the two is measured in the phase component to obtain the optimized spatial spectral features.

[0013] Cross-domain alignment is performed based on the optimized spatial spectral features. By inputting the spatial spectral features of the source domain and the target domain into a dual classifier and applying adversarial constraints, while using source domain labels for supervised training, the model is made capable of class recognition. Furthermore, catastrophic forgetting during the training process is mitigated by predictive consistency constraints at each stage, resulting in a cross-domain wetland classification map.

[0014] Secondly, this application also provides a hyperspectral trans-domain wetland mapping system using a fractional Fourier RWKV network, comprising:

[0015] The acquisition module is used to acquire a source domain hyperspectral image with labels and a target domain hyperspectral image without labels. Both the source domain hyperspectral image and the target domain hyperspectral image are hyperspectral remote sensing data containing spatial and spectral dimensions.

[0016] The conversion module is used to perform feature conversion based on the source domain hyperspectral image and the target domain hyperspectral image. It divides pixels into token sequences and expands them into channel representations, while applying a rotation transformation in the fractional Fourier domain to obtain a feature vector that integrates spatial spectral information and fractional Fourier domain amplitude-phase information.

[0017] The interaction module is used to perform spatial-spectral interaction based on the feature vector. It establishes long-range dependencies between positions by shifting and cyclically convolving the token sequence along the spatial dimension, and performs interactive fusion of token information along the channel dimension to obtain a spatial-spectral joint feature map with a large receptive field.

[0018] The layered module is used to design a network layered structure based on the joint spatial-spectral feature map. By stacking the feature representation and the spatial-spectral interaction module layer by layer and alternately performing spatial and channel dimension mixing operations between layers, the layered spatial-spectral features with amplitude-phase modulation are obtained.

[0019] The optimization module is used to generate optimization processes based on the hierarchical spatial spectral features. It maintains spectral fidelity by generating energy distributions consistent with the true features through amplitude component constraints and measures the structural correlation between the two in the phase component to obtain optimized spatial spectral features.

[0020] The output module is used to perform cross-domain alignment based on the optimized spatial spectral features. It inputs the spatial spectral features of the source domain and the target domain into a dual classifier and applies adversarial constraints. At the same time, it uses source domain labels for supervised training to enable the model to have class recognition capabilities. It also alleviates catastrophic forgetting during the training process by using prediction consistency constraints at each stage, thereby obtaining a cross-domain wetland classification map.

[0021] The beneficial effects of this invention are as follows:

[0022] This invention combines the amplitude-phase analysis capability of fractional Fourier transform with the efficient sequence modeling capability of RWKV network to achieve joint mining of spatial, spectral, amplitude, and phase information of hyperspectral images, thereby enhancing the learning ability of wetland texture, detail, and structural features. Furthermore, it utilizes a hierarchical network structure to expand the receptive field to adapt to complex land cover distributions. Addressing the domain offset problem in cross-domain mapping, the invention designs a dual-classifier adversarial network that achieves feature-level domain alignment and decision-level class boundary enhancement, effectively extracting domain-invariant features. Simultaneously, it incorporates a replay-based self-distillation learning method to alleviate instability and catastrophic forgetting issues during model training. Ultimately, the synergistic effect of these technologies results in a method that outperforms existing methods in overall classification accuracy, average accuracy, and Kappa coefficient, improving the performance of hyperspectral remote sensing images in cross-domain mapping tasks of coastal wetlands. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a hyperspectral trans-domain wetland mapping method using a fractional Fourier RWKV network as described in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of a hyperspectral trans-domain wetland mapping system using a fractional Fourier RWKV network as described in an embodiment of the present invention.

[0026] Figure 3 The overall framework diagram of the fractional Fourier transform RWKV network model;

[0027] Figure 4 This is a schematic diagram comparing the heterogeneous shift mechanism with existing shift mechanisms;

[0028] Figure 5 This is a schematic diagram of the fractional Fourier WKV attention mechanism.

[0029] The diagram is labeled as follows: 901, Acquisition Module; 902, Conversion Module; 903, Interaction Module; 904, Layering Module; 905, Optimization Module; 906, Output Module. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0031] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0032] In dynamic monitoring of coastal wetlands, such as when using data from different years, seasons, or sensors to detect changes or classify and map the same wetland area, a typical "cross-temporal / cross-sensor" domain offset problem is encountered. Specifically, even for the same geographical area, phenological changes, tidal fluctuations, and differences in illumination conditions caused by different data collection times can result in the same type of land cover exhibiting drastically different spectral reflectance curves and spatial texture features in images from different periods. Furthermore, wetland features typically exhibit a highly scattered, interwoven, and blurred spatial distribution pattern. This spatial heterogeneity and complexity require feature extraction models to have a large receptive field and strong resolution of detailed textures. However, many existing methods, when directly processing such cross-domain data, often only perform shallow spatial-spectral alignment, failing to capture, at the more fundamental frequency domain level, features that are relatively invariant to illumination and sensor differences but sensitive to the physical structure of land cover. This leads to models training well in the source domain but experiencing a significant performance drop when transferred to the target domain, making it difficult to produce accurate wetland classification maps with consistent spatiotemporal consistency. This invention addresses the core problem of "cross-domain feature representation and alignment" by proposing a systematic solution encompassing feature extraction, optimization, and domain adaptation. The invention will now be described in detail through specific embodiments.

[0033] Example 1:

[0034] This embodiment provides a hyperspectral transdomain wetland mapping method using fractional Fourier RWKV networks.

[0035] See Figure 1 The figure shows that the method includes steps S100 to S600.

[0036] Step S100: Obtain a labeled source domain hyperspectral image and an unlabeled target domain hyperspectral image. Both the source domain hyperspectral image and the target domain hyperspectral image are hyperspectral remote sensing data containing spatial and spectral dimensions.

[0037] Understandably, the purpose of step S100 is to acquire the basic data necessary for model training and transfer. In the scenario of continuous monitoring of coastal wetlands, this involves collecting hyperspectral remote sensing images from the same area but taken at different times or by different types of sensors. Some of these images have been manually and accurately labeled with the categories of various land cover with the assistance of field surveys or historical data; this part constitutes "labeled source domain hyperspectral images." The other part, newly acquired images to be classified, do not have existing labels and constitute "unlabeled target domain hyperspectral images."

[0038] Step S200: Perform feature transformation based on the source domain hyperspectral image and the target domain hyperspectral image. Divide the pixels into token sequences and expand them into channel representations. At the same time, apply a rotation transformation in the fractional Fourier domain to obtain a feature vector that integrates spatial spectral information and fractional Fourier domain amplitude-phase information.

[0039] It should be noted that this step treats pixels as discrete sequence units (tokens) and considers their spectral channels, which is fundamental to transforming image data into something that can be processed by a sequence network. By introducing a fractional Fourier transform, the image information is projected into a fractional domain between pure space and pure frequency. In this domain, the spatial structure information and spectral vibration characteristics of ground features are fused and represented as phase and amplitude components, respectively, thus providing the model with a joint representation starting point that utilizes both spatial context and frequency response.

[0040] Step S300: Perform spatial-spectral interaction based on feature vectors. By shifting and recurrently convolving the token sequence along the spatial dimension, long-range dependencies between positions are established. The token information is then interactively fused along the channel dimension to obtain a spatial-spectral joint feature map with a large receptive field.

[0041] Understandably, the spatial-spectral interaction processing in this step is designed as a mechanism to enable sufficient information mixing and integration within the feature vector in both spatial and spectral dimensions. Specific shifting and convolution operations on the token sequence along the spatial dimension aim to break the limitations of the local receptive field, simulating and establishing correlations between distant pixels, which is crucial for capturing widely distributed or structurally continuous land features in wetlands. Simultaneously, the interactive fusion along the channel dimension focuses on coordinating and reorganizing the information carried by hundreds of spectral bands, enhancing the ability to discriminate subtle spectral differences in land features, and ultimately outputting a joint spatial-spectral feature map that integrates global spatial context and deep spectral features.

[0042] Step S400: Based on the joint spatial-spectral feature map, a hierarchical network design is performed. By stacking the feature representation and spatial-spectral interaction module layer by layer and alternately performing spatial and channel dimension mixing operations between layers, the hierarchical spatial-spectral features of amplitude-phase modulation are obtained.

[0043] It should be noted that this step uses hierarchical spatial-spectral features as the basic building blocks. By stacking these units layer by layer and inserting downsampling feature maps and expanding the receptive field between layers, a deep hierarchical network is constructed. Shallow networks focus more on local details and textures, while as the network deepens and the receptive field increases, higher-level features tend to capture the overall semantics and macroscopic layout of the region. This hierarchical design enables the network to adaptively extract multi-scale spatial-spectral features, from subtle to macroscopic, modulated by amplitude and phase.

[0044] Step S500: Generate and optimize the spatial spectrum features based on the hierarchical spatial spectrum features. Maintain spectral fidelity by generating energy distribution consistent with the real features through amplitude component constraints, and measure the structural correlation between the two in the phase component to obtain the optimized spatial spectrum features.

[0045] Understandably, step S500 compares the network reconstruction result with the original input in the fractional Fourier domain and separately constrains the consistency of its amplitude components. Essentially, this requires the network to retain the spectral characteristics of the energy reflected by the ground objects. Measuring the structural correlation of the phase components guides the network to focus on and maintain the spatial morphology, edge contours, and other geometric structural information of the ground objects, preventing the loss of key shape and texture details during feature extraction, thereby obtaining feature representations that are optimized in both spectral and spatial structure.

[0046] Step S600: Perform cross-domain alignment based on the optimized spatial spectral features. By inputting the spatial spectral features of the source domain and the target domain into a dual classifier and applying adversarial constraints, while using source domain labels for supervised training, the model is made capable of class recognition. Furthermore, the catastrophic forgetting during the training process is mitigated by the prediction consistency constraints at each stage, resulting in a cross-domain wetland classification map.

[0047] It should be noted that step S600, which utilizes dual classifiers and imposes adversarial constraints on their predictions, is a strategy that encourages the network to learn more generalizable essential features stripped of domain-specific differences, aiming to achieve domain alignment at the feature level. Simultaneously, supervised training using source domain labels ensures the model masters basic classification capabilities. The introduction of replay-based consistency constraints stabilizes this complex adversarial training process. By allowing the model to maintain relatively stable predictions on the same data at different training stages, it mitigates the rapid forgetting of learned source domain knowledge that may occur when learning new domain knowledge, ultimately achieving the generation of a reliable wetland classification map on the unlabeled target domain.

[0048] Further, step S200 includes steps S210 to S230.

[0049] Step S210: Perform token construction processing based on the source domain hyperspectral image and the target domain hyperspectral image. By converting the hyperspectral image into a feature sequence representation, an input sequence containing spatial location encoding and spectral channel representation is obtained.

[0050] Step S220: Perform fractional Fourier domain transformation processing on the input sequence. By selecting a transformation order that matches the spectral characteristics of hyperspectral wetland land cover and applying a rotation transformation according to the transformation order, the token sequence is expanded in the fractional Fourier domain between the spatial domain and the frequency domain to simultaneously preserve the spatial domain energy of the spatial texture and the frequency domain energy of the spectral absorption features, thus obtaining a fractional domain characterization sequence that integrates spatial texture and spectral absorption features.

[0051] Step S230: Perform amplitude and phase extraction processing based on the fractional domain representation sequence. Obtain the energy distribution intensity information of the ground target by calculating the amplitude information of each token in the fractional Fourier domain, and extract the phase information to obtain the feature vector that integrates the spatial spectrum information and the amplitude-phase information of the fractional Fourier domain.

[0052] Specifically, assuming that the hyperspectral data of the source domain and the target domain in this embodiment are respectively and , and These are the source domain dataset and the target domain dataset, respectively. and Representing the source domain and the target domain respectively. and They are source domain image data and target domain image data, respectively. , The symbol for the set of real numbers, Represents each pixel Spatial neighborhood size, Indicates the number of spectral bands. and They represent The class has a known source domain label space and an unknown target domain label space. The above process performs feature transformation based on a novel RWKV module using fractional Fourier transform. For example... Figure 3 As shown, the Fractional Fourier Transform RWKV (FrFT-RWKV) module comprises a spatial mixing submodule and a channel mixing submodule. The core components of the spatial mixing submodule are the Het-Shift mechanism and the FrFT-RWKV attention mechanism, used for spatial dimension token interaction. The channel mixing submodule is used for channel dimension feature fusion. The Het-Shift mechanism aims to improve the effective utilization of multi-scale spatial information in hyperspectral wetland images under large receptive field conditions; the FrFT-RWKV attention mechanism introduces amplitude-phase feature extraction from the fractional Fourier transform, extracting joint spatial-spectral features from the hyperspectral wetland image through the optimal domain between the spatial and frequency domains. Assuming the sequence transformation of the input spatial-spectral features of the entire FrFT-RWKV module is represented as... ,in, and These represent the number of feature tokens and the channel dimension, respectively. In step S210, each pixel of the hyperspectral image and its spatial neighborhood are converted into a feature sequence representation. Each token contains spectral information of the corresponding spatial location and spatial neighborhood context information, thus obtaining an input sequence containing spatial location encoding and spectral channel representation. In step S220, by selecting a transform order that matches the spectral characteristics of hyperspectral wetland features and applying a rotation transform, the token sequence is expanded in the fractional Fourier domain between the spatial and frequency domains, while preserving the spatial domain energy of the spatial texture and the frequency domain energy of the spectral absorption features, enabling the features to represent wetland feature information in the optimal transform domain between the spatial and frequency domains. In step S230, the amplitude spectrum and phase spectrum are extracted from the complex representation in the fractional Fourier domain, respectively. The amplitude spectrum reflects the energy distribution intensity of the feature target, and the phase spectrum reflects the spatial structure information of the feature target. Together, they constitute a feature vector that integrates spatial spectral information and fractional Fourier domain amplitude-phase information.

[0053] Further, step S300 includes steps S310 to S330.

[0054] Step S310: Perform spatial mixing processing based on feature vectors. By cyclically shifting the token sequence along the spatial dimension according to a preset shift ratio, and applying linear projection and activation based on the weight matrix to the shifted token sequence to establish long-range dependencies between different spatial locations, so that information can be transferred between different spatial tokens within the same wetland land cover area, and a spatial aggregation feature map is obtained.

[0055] Step S320: Perform channel mixing processing based on the spatial aggregation feature map. By applying gated linear projection along the channel dimension to the token sequence, the nonlinear combination relationship between different spectral bands is learned, and a spatial-spectral joint feature map that integrates spatial-spectral information and spectral information is obtained.

[0056] Step S330: Perform receptive field expansion processing based on the spatial-spectral joint feature map. By using a heterogeneous shift mechanism, the token sequence is input into multiple group convolutional layers with different hole rates. After being averaged along the channel dimension, the layers are sequentially mixed and cascaded. A global attention map is obtained by bidirectional scanning from the horizontal and vertical directions. The current attention result is combined with the previous attention results from different scanning directions to obtain a spatial-spectral joint feature map with a large receptive field.

[0057] Specifically, this process performs a hybrid interaction operation on the feature vector in both spatial and channel dimensions, achieved through three sub-steps: spatial mixing, channel mixing, and receptive field expansion. In step S310, the feature vector is input into the Het-Shift heterogeneous shift mechanism. The input features are fed into four group convolutional layers with different dilation rates, and their outputs are divided into four groups along the channel dimension. These features are sequentially mixed to obtain four new features, each containing the sum of the outputs of the group convolutions with different dilation rates. These new features are then concatenated along the channels to obtain the final output. This process is defined as Het-Shift. Multi-scale spatial information extraction is achieved through group convolutions with different dilation rates, and sequential mixing and concatenation operations enable full interaction between features of different scales, establishing long-range dependencies between spatial locations. The mathematical expression of the Het-Shift heterogeneous shift mechanism is:

[0058] ;

[0059] ;

[0060] ;

[0061] In the formula, For input features; For the first The output features of the grouped convolutional layers, where These correspond to four groups of convolutional layers with different porosity. Presentation layer normalization operation; Indicates the first Each 3×3 set of convolutional layer weights; Indicates equal division Group, ; This represents the final output characteristic of the Het-Shift heterogeneous shift mechanism; , , and These represent the four new feature vectors obtained after sequential mixing; , ... , representing the characteristics of each position from group 1 to group 4; This indicates a cascading operation; for ease of subsequent analysis, the above process is defined as... .

[0062] See Figure 3 The figure shows the internal structure of the designed fractional Fourier transform RWKV module; see also... Figure 4The figure shows a comparison between the heterogeneous shift mechanism proposed in this invention and existing shift mechanisms. In step S330, local features are input into a fractional Fourier WKV attention layer and then passed through a linear projection layer. Obtain the receiver matrix The system iteratively executes horizontal H-scan and vertical V-scan directions to acquire a global attention map. This design models global dependencies, combining the current attention result with previous attention results obtained from different scan directions. The reception matrix is ​​input into a sigmoid activation function, and a linear projection layer is used to generate the output projection of the spatial mixing submodule. Through bidirectional scanning in both horizontal and vertical directions, spatial contextual information can be captured from different directions, enabling the attention mechanism to maintain linear complexity while possessing a global receptive field. The output of the fractional Fourier WKV attention mechanism is expressed as follows:

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] In the formula, This represents the output characteristics of the Het-Shift heterogeneous shift mechanism; This is the attention feature map after fractional Fourier transform; and These represent the 1-dimensional fractional Fourier transforms along the sequence direction and the hidden dimension, respectively. , and For linear projection layers; For the first The key vector in the next iteration; Indicates a change in direction operation; For the first The value vector in the next iteration; For the first The outer key vector in the next iteration; For the first The WKV attention output of the next iteration; For the first The WKV attention output of the next iteration; This represents a bidirectional WKV attention mechanism. This design models global dependencies, combining the current attention result with previous attention results obtained from different scanning directions to achieve a more accurate spatial-spectral-frequency characterization. For ease of subsequent analysis, the above process is defined as... ,in, For the first WKV attention output of the next iteration Indicates the first WKV attention computation using fractional Fourier transform. Indicates the iteration number. Indicates the total number of iterations. This represents the input features. Finally, the reception matrix is ​​input into the Sigmoid function to adjust the reception probability of wkv, and then passed through a linear projection layer. The output projection of the generated spatial mixing submodule :

[0068] ;

[0069] ;

[0070] ;

[0071] In the formula, This represents the output characteristics of the Het-Shift heterogeneous shift mechanism; This is the receiver matrix for the spatial mixing submodule; This is the output of the FrFT-WKV attention mechanism; For the output projection of the spatial mixing submodule; For linear projection layers; This represents element-wise multiplication. implement The output of the FrFT-WKV attention mechanism in the next iteration; This represents the Sigmoid function.

[0072] See Figure 5 , Figure 5 This is the structure of the fractional Fourier WKV attention mechanism proposed in this invention. In the channel mixing process of step S320, in the channel mixing submodule, firstly, ... Processing is performed through LN and Het-Shift layers to generate the receive matrix, key matrix, and value matrix. Subsequently, the receive matrix is ​​used to modulate the value matrix, yielding the output projection of this submodule. , represented as:

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] In the formula, For the output projection of the spatial mixing submodule; For receiving matrix; The key matrix; It is a value matrix; This indicates that the output of the spatial hybrid submodule is subjected to layer normalization. For the output projection of the channel mixing submodule; , and These are the linear projection layer weight matrices of the receiver matrix, key matrix, and value matrix in the channel mixing submodule, respectively. It is a linear rectification function. Represents the Sigmoid function; This represents element-wise multiplication. Indicates a heterogeneous shift operation; Output the weight matrix of the projection layer for the channel mixing submodule.

[0078] Further, step S400 includes steps S410 to S430.

[0079] Step S410: Perform local feature extraction processing based on the spatial-spectral joint feature map. By aligning the channel dimension using a two-dimensional convolutional layer in the network stem module and performing a spatial-channel dimension mixing operation of a multi-layer fractional Fourier transform RWKV module in the shallow network, the local texture and spectral details of wetland vegetation, water bodies and bare land are captured to obtain shallow local spatial-spectral features.

[0080] Step S420: Perform receptive field increment processing based on shallow local spatial spectrum features. By using a deep convolutional layer with a stride of a set value in the downsampling module to halve the feature space dimension and using a convolutional layer to double the channel dimension, a global spatial spectrum feature with a large receptive field is obtained.

[0081] Step S430: Perform cross-layer fusion processing based on shallow local spatial spectral features and global spatial spectral features. By using bilinear interpolation upsampling layer and convolutional layer in the upsampling module, the spatial dimension of global spatial spectral features is doubled and the channel dimension is halved. Then, the upsampled global spatial spectral features and shallow local spatial spectral features are stitched together along the channel dimension by the fusion projection head to simultaneously preserve the local details and global semantics of ground features, thereby obtaining amplitude-phase modulated hierarchical spatial spectral features.

[0082] Specifically, this process uses FrFT-RWKV modules as basic units to build a hierarchical feature extraction network. For example... Figure 3As shown in "I. Fractional Fourier Transform RWKV Backbone Network", this network is designed based on an encoder-decoder structure, including a network stem module based on a two-dimensional convolutional layer, a multi-layer fractional Fourier Transform RWKV (FrFT-RWKV) module, and multiple downsampling modules. In step S410, the source domain data and target domain data are first subjected to preliminary feature extraction through the network stem module, converting the input hyperspectral data into a feature representation suitable for subsequent processing. The stem module aligns the channel dimensions through a two-dimensional convolutional layer, and effectively captures the local texture and spectral detail features of wetland vegetation, water bodies, and bare land through spatial and channel dimension mixing operations of the multi-layer FrFT-RWKV module in the shallow network. In step S420, The fractional Fourier transform (RWKV) module and a downsampling module are used alternately. The downsampling module uses a deep convolutional layer with a stride of 2 to halve the feature space dimension, while using convolutional layers to double the channel dimension, gradually expanding the receptive field to obtain more global semantic information. In step S430, the decoder uses another fractional Fourier transform (RWKV) module, a deep convolutional layer, multiple upsampling layers based on bilinear interpolation, and a fusion projection head to reconstruct the input hyperspectral data. The upsampling layer doubles the spatial dimension of the input features while reducing the channel dimension by half. The fusion projection head includes a channel cascade operation and a convolutional layer to achieve fine-grained feature alignment. Through skip connections between the encoder and decoder, the multi-scale features of each level of the encoder are fused with the corresponding level of the decoder, while preserving the local details and global semantic information of the ground objects, resulting in amplitude-phase modulated spatial-spectral feature extraction, which is represented as source domain spatial-spectral features and target domain spatial-spectral features, respectively.

[0083] Further, step S500 includes steps S510 to S530.

[0084] Step S510: Perform fractional Fourier domain decomposition based on the layered spatial spectrum features. By mapping the layered spatial spectrum features to the fractional Fourier domain, extract the amplitude component of the reflected energy intensity of the corresponding ground object and the phase component of the spatial structure information of the corresponding ground object, and obtain the amplitude component and phase component.

[0085] Step S520: Perform energy constraint processing based on the amplitude components. By calculating the mean square error between the amplitude components of the generated features and the amplitude components of the real hyperspectral wetland image, the amplitude-constrained features are obtained.

[0086] Step S530: Perform structure-guided processing based on the phase component and amplitude-constrained features. Extract the spatial structure correlation between the phase component of the generated features and the phase component of the real hyperspectral wetland image by calculating the complex cross-correlation measure between the two. Guide the network to learn the structural information of the hyperspectral image and obtain the optimized spatial-spectral features.

[0087] Specifically, the above process optimizes the hierarchical RWKV network by designing a fractional Fourier domain amplitude-phase loss function. For example... Figure 3 Section II. Amplitude-Phase Loss Function for Data Generation: To enable the designed hierarchical fractional Fourier transform RWKV network to more effectively capture the frequency and structural information of hyperspectral images, a fractional Fourier domain amplitude-phase loss function was further designed to replace the traditional root mean square error (MSE). This function aims to regularize the amplitude and phase components of the prediction results to extract more representative amplitude-phase features. In step S510, the complex representations in the fractional Fourier domain of the hierarchical spatial spectral features obtained in step S400 and the real hyperspectral wetland image are calculated, and their amplitude components (corresponding to the reflected energy intensity of ground objects) and phase components (corresponding to the spatial structural information of ground objects) are extracted. In step S520, MSE loss is applied to these two amplitude components, and the difference between their amplitude spectra is calculated to constrain the generated features to maintain consistency with the real features in terms of energy distribution, thereby improving the clarity and sharpness of the generated data. In step S530, it is found that data generation cannot be accurately achieved using only amplitude information. An intuitive idea is to minimize phase differences; however, phase discontinuities prevent this from being achieved. An additional phase learning loss is designed to effectively learn the structural information of the hyperspectral image by calculating their correlation. The spatial structural correlation between generated and ground-based features is extracted using a complex cross-correlation metric, guiding the network to learn the fine structural information of the hyperspectral image while maintaining a consistent energy distribution. The entire fractional Fourier domain amplitude-phase loss function can be expressed as follows:

[0088] ;

[0089] In the formula, This represents the fractional Fourier domain amplitude-phase loss function; and These represent the amplitude spectrum and real part of the data, respectively. Represents the cross-correlation operation of complex numbers; As a balance factor; Input data; For the generated reconstructed hyperspectral image data.

[0090] Further, step S600 includes steps S610 to S630.

[0091] Step S610: Perform adversarial distribution alignment processing based on the optimized spatial spectral features. By inputting the source domain spatial spectral features and the target domain spatial spectral features into two independent classifiers and making the outputs of the two classifiers tend to be consistent, the spatial distribution distance between the source domain and the target domain features is reduced. At the same time, a minimum class confusion constraint is added to the classifier to establish the decision boundary of the task for different wetland land cover categories to reduce inter-category confusion, and the domain-aligned spatial spectral features are obtained.

[0092] Step S620: Perform supervised classification processing based on the spatial spectral features and the labels of the source domain hyperspectral images. Apply cross-entropy constraints to the source domain prediction output of the model using the labels so that the model can accurately distinguish different wetland land cover categories, and obtain a prediction model with source domain category recognition capability.

[0093] Step S630: Perform adaptive optimization processing based on the prediction model. By storing the prediction results of the model for the source domain samples in each training stage and using the prediction results as the target constraints for the corresponding samples in the subsequent training stages, the memory of the previously learned wetland category knowledge is maintained, and a cross-domain wetland classification map is obtained.

[0094] Specifically, this invention designs a dual-classifier adversarial network and a replay-based self-distillation learning method to establish task-specific decision margins between different categories, mitigating catastrophic forgetting and instability in adversarial training, and effectively improving the accuracy of hyperspectral cross-domain wetland mapping using fractional Fourier RWKV networks. A significant problem is that when neural networks receive new data for training, they quickly forget what they learned during previous training. A crucial mechanism for protecting memory in the human brain is reactivating the activity patterns of neurons representing these memories. To address this, a dual-classifier adversarial network and a replay-based self-distillation learning strategy are further designed to achieve cross-domain distributed adversarial alignment, while mitigating the instability of adversarial training. This extracts more accurate, domain-invariant spatial-spectral features from hyperspectral images, thereby achieving cross-domain coastal wetland mapping. In step S610, as... Figure 3 As shown in "III. Trans-domain Wetland Mapping", for a dual-classifier adversarial network, the extracted source domain spatial-spectral features are... and target domain spatial spectral features Input into the dual classifier The corresponding prediction results were obtained respectively. The entire network is trained under supervision using known source domain labels to enable it to identify wetland samples in the source domain. To reduce confusion between the two classifiers for different categories in the target domain, a minimum class confusion loss is introduced. This not only determines whether a feature comes from the source or target domain but also sets a decision boundary for a specific task for different categories. Through the adversarial training mechanism of the dual classifiers, the spatial spectral features of the source and target domains are brought closer in the feature space, achieving feature-level domain alignment. In step S620, the entire network is trained under supervision using source domain labels to enable it to identify wetland samples in the source domain. Cross-entropy loss constraints enable the model to accurately distinguish different wetland land cover categories in the source domain. In step S630, as... Figure 3 As shown in "IV. Replay-based Self-Distillation Learning", a data repository is defined. and prediction repository Store input data and its classifier In the The predicted value corresponding to the next iteration process During training, the data corresponding to a certain iteration process is randomly read. and and obtain using the updated model New forecast values Use with temperature coefficient The softmax function transforms the old and new predictions into a probability distribution, thus obtaining their soft labels. and In the formula, For the generated soft tags, For the first The original output of each sample, For temperature coefficient, For the softmax function, As a historical soft label, sample The self-distillation learning strategy stores the model's prediction results for source domain samples at each training stage, using these prediction results as target constraints for subsequent training stages. This dynamically constrains the consistency of the model's prediction results at different training times, thereby maintaining the model's memory of previously learned wetland category knowledge and eliminating the problem of catastrophic forgetting.

[0095] Furthermore, the Kullback-Leibler divergence loss is used to ensure that the new predictions remain consistent with the old predictions in different iterations, thus improving the overall model training stability. The designed self-distillation learning loss function can be expressed as:

[0096] ;

[0097] In the formula, The self-distillation learning loss function; Kullback-Leibler divergence loss; For generated soft tags; This is a historical soft tag.

[0098] Example 2:

[0099] To verify the effectiveness of the proposed method, this embodiment selects a publicly available hyperspectral wetland dataset (Yellow River Estuary dataset) for simulation experiments. This dataset consists of two sets of Yellow River Estuary data collected by the domestic ZY1-02D-AHSI satellite on June 28, 2020, and September 29, 2021, respectively. The former is the source domain, with an image size of 1147×1600 pixels; the latter is the target domain, with a size of 1050×1219 pixels, containing 108 bands and 8 land cover types for cross-domain analysis. Furthermore, this embodiment compares three cross-domain classification algorithms from the past five years, and the experimental results are shown in Table 1 below.

[0100] Table 1. Classification results (%) of different classification algorithms on the Yellow River Estuary dataset

[0101]

[0102] Table 1 lists the classification accuracy, overall accuracy (OA), average accuracy (AA), and Kappa coefficient of the algorithm proposed in this invention compared with different comparison methods. According to the experimental results shown in Table 1, compared with other methods, the algorithm of this invention improves the OA, AA, and Kappa coefficient by at least 1.05%, 2.33%, and 1.86%, respectively, demonstrating its advantage in improving the performance of cross-domain wetland mapping using hyperspectral images.

[0103] Example 3:

[0104] like Figure 2 As shown, this embodiment provides a hyperspectral trans-domain wetland mapping system using fractional Fourier RWKV networks. The system includes:

[0105] The acquisition module 901 is used to acquire labeled source domain hyperspectral images and unlabeled target domain hyperspectral images. Both the source domain hyperspectral images and the target domain hyperspectral images are hyperspectral remote sensing data containing spatial and spectral dimensions.

[0106] The conversion module 902 is used to perform feature conversion based on the source domain hyperspectral image and the target domain hyperspectral image. It divides the pixels into token sequences and expands them into channel representations. At the same time, it applies a rotation transformation in the fractional Fourier domain to obtain a feature vector that integrates spatial spectral information and fractional Fourier domain amplitude-phase information.

[0107] The interaction module 903 is used to perform spatial-spectral interaction based on the feature vector. It establishes long-range dependencies between positions by shifting and recurrent convolution of the token sequence along the spatial dimension, and interacts and fuses the token information along the channel dimension to obtain a spatial-spectral joint feature map with a large receptive field.

[0108] The layered module 904 is used to design a network layered structure based on the joint spatial-spectral feature map. By stacking the feature representation and the spatial-spectral interaction module layer by layer and alternately performing spatial and channel dimension mixing operations between layers, the layered spatial-spectral features with amplitude-phase modulation are obtained.

[0109] The optimization module 905 is used to generate optimization processing based on the hierarchical spatial spectral features. It maintains spectral fidelity by generating energy distribution consistent with the real features through amplitude component constraints, and measures the structural correlation between the two in the phase component to obtain the optimized spatial spectral features.

[0110] Output module 906 is used to perform cross-domain alignment based on the optimized spatial spectral features. By inputting the spatial spectral features of the source domain and the target domain into a dual classifier and applying adversarial constraints, while using source domain labels for supervised training, the model is made capable of class recognition. Furthermore, the catastrophic forgetting during the training process is mitigated by the prediction consistency constraints at each stage, resulting in a cross-domain wetland classification map.

[0111] In one specific embodiment of this application, the conversion module 902 includes:

[0112] The first conversion unit is used to perform token construction processing based on the source domain hyperspectral image and the target domain hyperspectral image. By converting the hyperspectral image into a feature sequence representation, an input sequence containing spatial location encoding and spectral channel representation is obtained.

[0113] The second transformation unit is used to perform fractional Fourier domain transformation processing on the input sequence. By selecting a transformation order that matches the spectral characteristics of hyperspectral wetland land cover and applying a rotation transformation according to the transformation order, the token sequence is expanded in the fractional Fourier domain between the spatial domain and the frequency domain to simultaneously preserve the spatial domain energy of the spatial texture and the frequency domain energy of the spectral absorption features, thus obtaining a fractional domain characterization sequence that integrates spatial texture and spectral absorption features.

[0114] The third transformation unit is used to perform amplitude and phase extraction processing based on the fractional domain representation sequence. It obtains the energy distribution intensity information of the ground target by calculating the amplitude information of each token in the fractional Fourier domain and extracting the phase information to obtain a feature vector that integrates the spatial spectrum information and the amplitude-phase information of the fractional Fourier domain.

[0115] In one specific embodiment of this application, the interaction module 903 includes:

[0116] The first interactive unit is used to perform spatial mixing processing based on feature vectors. It cyclically shifts the token sequence along the spatial dimension according to a preset shift ratio, and applies linear projection and activation based on the weight matrix to the shifted token sequence to establish long-range dependencies between different spatial locations, so that information can be transferred between different spatial tokens within the same wetland land cover area, and a spatial aggregation feature map is obtained.

[0117] The second interactive unit is used to perform channel mixing processing based on the spatial aggregation feature map. By applying gated linear projection along the channel dimension to the token sequence, it learns the nonlinear combination relationship between different spectral bands and obtains a spatial-spectral joint feature map that integrates spatial and spectral information.

[0118] The third interaction unit is used to expand the receptive field based on the spatial-spectral joint feature map. By using a heterogeneous shift mechanism, the token sequence is input into multiple group convolutional layers with different hole rates. After being averaged along the channel dimension, the layers are sequentially mixed and cascaded. The global attention map is obtained by bidirectional scanning from the horizontal and vertical directions. The current attention result is combined with the previous attention results from different scanning directions to obtain a spatial-spectral joint feature map with a large receptive field.

[0119] In one specific embodiment of this application, the layering module 904 includes:

[0120] The first layer unit is used to perform local feature extraction processing based on the joint spatial-spectral feature map. By aligning the channel dimension with a two-dimensional convolutional layer in the network stem module, and performing a spatial-channel dimension mixing operation of the multi-layer fractional Fourier transform RWKV module in the shallow network, local texture and spectral details of wetland vegetation, water bodies and bare land are captured to obtain shallow local spatial-spectral features.

[0121] The second layer unit is used to perform receptive field increment processing based on shallow local spatial spectrum features. By using a deep convolutional layer with a set stride in the downsampling module to halve the feature space dimension and using a convolutional layer to double the channel dimension, a global spatial spectrum feature with a large receptive field is obtained.

[0122] The third layer unit is used to perform cross-layer fusion processing based on shallow local spatial spectral features and global spatial spectral features. By using bilinear interpolation upsampling layers and convolutional layers in the upsampling module, the spatial dimension of the global spatial spectral features is doubled and the channel dimension is halved. Then, the upsampled global spatial spectral features and shallow local spatial spectral features are stitched together along the channel dimension by the fusion projection head to simultaneously preserve the local details and global semantics of the ground features, resulting in amplitude-phase modulated layered spatial spectral features.

[0123] In one specific embodiment of this application, the optimization module 905 includes:

[0124] The first optimization unit is used to perform fractional Fourier domain decomposition based on the layered spatial spectrum features. By mapping the layered spatial spectrum features to the fractional Fourier domain, the amplitude component of the reflected energy intensity of the corresponding ground object and the phase component of the spatial structure information of the corresponding ground object are extracted to obtain the amplitude component and the phase component.

[0125] The second optimization unit is used to perform energy constraint processing based on the amplitude component. By calculating the mean square error between the amplitude component of the generated feature and the amplitude component of the real hyperspectral wetland image, the amplitude-constrained feature is obtained.

[0126] The third optimization unit is used to perform structure-guided processing based on the phase component and amplitude-constrained features. It extracts the spatial structure correlation between the phase component of the generated features and the phase component of the real hyperspectral wetland image by calculating the complex cross-correlation measure between the two, guides the network to learn the structural information of the hyperspectral image, and obtains the optimized spatial-spectral features.

[0127] In one specific embodiment of this application, the output module 906 includes:

[0128] The first output unit is used to perform adversarial distribution alignment processing based on the optimized spatial spectral features. By inputting the source domain spatial spectral features and the target domain spatial spectral features into two independent classifiers and making the outputs of the two classifiers tend to be consistent, the spatial distribution distance between the source domain and the target domain features is reduced. At the same time, a minimum class confusion constraint is added to the classifier to establish the decision boundary of the task for different wetland land cover categories to reduce inter-category confusion, and the domain-aligned spatial spectral features are obtained.

[0129] The second output unit is used to perform supervised classification processing based on the spatial spectral features and the labels of the source domain hyperspectral image. The labels are used to apply cross-entropy constraints to the source domain prediction output of the model so that the model can accurately distinguish different wetland land cover categories, thus obtaining a prediction model with source domain category recognition capability.

[0130] The third output unit is used to perform adaptive optimization processing based on the prediction model. It maintains the memory of previously learned wetland category knowledge by storing the prediction results of the model for the source domain samples in each training stage and using the prediction results as the target constraints for the corresponding samples in the subsequent training stages. This results in a cross-domain wetland classification map.

[0131] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A hyperspectral trans-domain wetland mapping method using fractional Fourier RWKV networks, characterized in that, include: Acquire labeled source domain hyperspectral images and unlabeled target domain hyperspectral images, wherein both the source domain hyperspectral images and the target domain hyperspectral images are hyperspectral wetland datasets containing spatial and spectral dimensions; Based on the source domain hyperspectral image and the target domain hyperspectral image, feature transformation is performed by dividing pixels into token sequences and expanding them into channel representations. Simultaneously, feature learning is performed in the fractional Fourier domain to obtain a feature vector that integrates spatial spectral information and fractional Fourier domain amplitude-phase information. Based on the feature vector, spatial-spectral interaction is performed. By shifting and recurrently convolving the token sequence along the spatial dimension, long-range dependencies between positions are established. The token information is interactively fused along the channel dimension to obtain a spatial-spectral joint feature map with a large receptive field. Based on the aforementioned spatial-spectral joint feature map, a hierarchical network design is performed. By stacking the feature representation and spatial-spectral interaction module layer by layer and alternately performing spatial and channel dimension mixing operations between layers, the amplitude-phase modulated hierarchical spatial-spectral features are obtained. Based on the hierarchical spatial spectral features, the generation optimization process is performed. By constraining the amplitude component, the energy distribution of the generated features is consistent with the real features to maintain spectral fidelity. The structural correlation between the two is measured in the phase component to obtain the optimized spatial spectral features. Cross-domain alignment is performed based on the optimized spatial spectral features. By inputting the spatial spectral features of the source domain and the target domain into a dual classifier and applying adversarial constraints, while using source domain labels for supervised training, the model is made capable of class recognition. Furthermore, the catastrophic forgetting during the training process is mitigated by the prediction consistency constraints at each stage, resulting in a cross-domain wetland classification map. The feature transformation based on the source domain hyperspectral image and the target domain hyperspectral image includes: Token construction processing is performed on the source domain hyperspectral image and the target domain hyperspectral image. By converting the hyperspectral image into a feature sequence representation, an input sequence containing spatial location encoding and spectral channel representation is obtained. The input sequence is processed by fractional Fourier domain transformation. By selecting a transformation order that matches the spectral characteristics of hyperspectral wetland land cover and applying a rotation transformation according to the transformation order, the token sequence is expanded in the fractional Fourier domain between the spatial domain and the frequency domain to simultaneously preserve the spatial domain energy of the spatial texture and the frequency domain energy of the spectral absorption features, thus obtaining a fractional domain characterization sequence that integrates spatial texture and spectral absorption features. Amplitude and phase extraction processing is performed based on the fractional domain representation sequence. The energy distribution intensity information of the ground target is obtained by calculating the amplitude information of each token in the fractional Fourier domain, and the phase information is extracted to obtain a feature vector that integrates the spatial spectrum information and the amplitude-phase information of the fractional Fourier domain. The network layering design based on the joint spatial-spectral feature map includes: Local feature extraction is performed based on the spatial-spectral joint feature map. By aligning the channel dimension using a two-dimensional convolutional layer in the network stem module and performing a spatial-channel dimension mixing operation of a multi-layer fractional Fourier transform (RWKV) module in the shallow network, local texture and spectral details of wetland vegetation, water bodies and bare land are captured to obtain shallow local spatial-spectral features. Based on the shallow local spatial spectrum features, the receptive field is increased by using a deep convolutional layer with a set stride in the downsampling module to halve the feature space dimension and use a convolutional layer to double the channel dimension, thereby obtaining a global spatial spectrum feature with a large receptive field. Cross-layer fusion processing is performed based on the shallow local spatial spectral features and the global spatial spectral features. By using a bilinear interpolation upsampling layer and a convolutional layer in the upsampling module, the spatial dimension of the global spatial spectral features is doubled and the channel dimension is halved. Then, the upsampled global spatial spectral features and the shallow local spatial spectral features are stitched together along the channel dimension by a fusion projection head to simultaneously preserve the local details and global semantics of the ground features, resulting in amplitude-phase modulated hierarchical spatial spectral features.

2. The hyperspectral trans-domain wetland mapping method using fractional Fourier RWKV networks according to claim 1, characterized in that, Spatial-spectral interaction based on the eigenvectors includes: Spatial mixing processing is performed based on the feature vector. The token sequence is cyclically shifted along the spatial dimension according to a preset shift ratio. Linear projection and activation based on the weight matrix are applied to the shifted token sequence to establish long-range dependencies between different spatial locations. This enables information transfer between different spatial tokens within the same wetland land cover area, resulting in a spatial aggregation feature map. Based on the spatial aggregation feature map, channel mixing processing is performed. By applying gated linear projection along the channel dimension to the token sequence, the nonlinear combination relationship between different spectral bands is learned, resulting in a spatial-spectral joint feature map that integrates spatial and spectral information. The receptive field is expanded based on the spatial-spectral joint feature map. The token sequence is input into multiple group convolutional layers with different porosities by using a heterogeneous shift mechanism. After being averaged along the channel dimension, the layers are sequentially mixed and cascaded. A global attention map is obtained by scanning bidirectionally from the horizontal and vertical directions. The current attention result is combined with the previous attention results from different scanning directions to obtain a spatial-spectral joint feature map with a large receptive field.

3. The hyperspectral trans-domain wetland mapping method using fractional Fourier RWKV networks according to claim 1, characterized in that, The generation optimization is performed based on the hierarchical spatial spectral features, including: Based on the layered spatial spectrum features, fractional Fourier domain decomposition is performed. By mapping the layered spatial spectrum features to the fractional Fourier domain, the amplitude component of the reflected energy intensity of the corresponding ground object and the phase component of the spatial structure information of the corresponding ground object are extracted to obtain the amplitude component and the phase component. Energy constraint processing is performed based on the amplitude components, and the amplitude-constrained features are obtained by calculating the mean square error between the amplitude components of the generated features and the amplitude components of the real hyperspectral wetland image. Based on the phase components and the amplitude-constrained features, structural guidance processing is performed. By calculating the complex cross-correlation metric between the phase components of the generated features and the phase components of the real hyperspectral wetland image, the spatial structural correlation between the two is extracted. This guides the network to learn the structural information of the hyperspectral image and obtains the optimized spatial-spectral features.

4. A hyperspectral trans-domain wetland mapping system using a fractional Fourier RWKV network, characterized in that, include: The acquisition module is used to acquire labeled source domain hyperspectral images and unlabeled target domain hyperspectral images, wherein both the source domain hyperspectral images and the target domain hyperspectral images are hyperspectral wetland datasets containing spatial and spectral dimensions. The conversion module is used to perform feature conversion based on the source domain hyperspectral image and the target domain hyperspectral image. It divides pixels into token sequences and expands them into channel representations. At the same time, it performs feature learning in the fractional Fourier domain to obtain a feature vector that integrates spatial spectral information and fractional Fourier domain amplitude-phase information. The interaction module is used to perform spatial-spectral interaction based on the feature vector. It establishes long-range dependencies between positions by shifting and cyclically convolving the token sequence along the spatial dimension, and performs interactive fusion of token information along the channel dimension to obtain a spatial-spectral joint feature map with a large receptive field. The layered module is used to design a network layered structure based on the joint spatial-spectral feature map. By stacking the feature representation and the spatial-spectral interaction module layer by layer and alternately performing spatial and channel dimension mixing operations between layers, the layered spatial-spectral features with amplitude-phase modulation are obtained. The optimization module is used to generate optimization processes based on the hierarchical spatial spectral features. It maintains spectral fidelity by generating energy distributions consistent with the true features through amplitude component constraints and measures the structural correlation between the two in the phase component to obtain optimized spatial spectral features. The output module is used to perform cross-domain alignment based on the optimized spatial spectral features. By inputting the spatial spectral features of the source domain and the target domain into a dual classifier and applying adversarial constraints, while using source domain labels for supervised training, the model is made capable of class recognition. Furthermore, the catastrophic forgetting during the training process is mitigated by the prediction consistency constraints at each stage, resulting in a cross-domain wetland classification map. The conversion module includes: The first conversion unit is used to perform token construction processing based on the source domain hyperspectral image and the target domain hyperspectral image, and obtain an input sequence containing spatial location encoding and spectral channel representation by converting the hyperspectral image into a feature sequence representation; The second conversion unit is used to perform fractional Fourier domain transformation processing on the input sequence. By selecting a transformation order that matches the spectral characteristics of hyperspectral wetland land cover and applying a rotation transformation according to the transformation order, the token sequence is expanded in the fractional Fourier domain between the spatial domain and the frequency domain to simultaneously retain the spatial domain energy of the spatial texture and the frequency domain energy of the spectral absorption features, thereby obtaining a fractional domain characterization sequence that integrates spatial texture and spectral absorption features. The third conversion unit is used to perform amplitude and phase extraction processing based on the fractional domain representation sequence. It obtains the energy distribution intensity information of the ground target by calculating the amplitude information of each token in the fractional Fourier domain and extracting the phase information to obtain a feature vector that integrates the spatial spectrum information and the amplitude-phase information of the fractional Fourier domain. The hierarchical module includes: The first layer unit is used to perform local feature extraction processing based on the spatial-spectral joint feature map. By aligning the channel dimension with a two-dimensional convolutional layer in the network stem module and performing a spatial-channel dimension mixing operation of a multi-layer fractional Fourier transform (RWKV) module in the shallow network, local texture and spectral details of wetland vegetation, water bodies and bare land are captured to obtain shallow local spatial-spectral features. The second layer unit is used to perform receptive field increment processing based on the shallow local spatial spectrum features. By using a deep convolutional layer with a set stride in the downsampling module to halve the feature space dimension and using a convolutional layer to double the channel dimension, a global spatial spectrum feature with a large receptive field is obtained. The third layer unit is used to perform cross-layer fusion processing based on the shallow local spatial spectral features and the global spatial spectral features. By using a bilinear interpolation upsampling layer and a convolutional layer in the upsampling module, the spatial dimension of the global spatial spectral features is doubled and the channel dimension is halved. Then, the upsampled global spatial spectral features and the shallow local spatial spectral features are stitched together along the channel dimension by a fusion projection head to simultaneously preserve the local details and global semantics of the ground features, thereby obtaining the amplitude-phase modulated layered spatial spectral features.

5. The hyperspectral trans-domain wetland mapping system using a fractional Fourier RWKV network according to claim 4, characterized in that, The interaction module includes: The first interaction unit is used to perform spatial mixing processing based on the feature vector. By cyclically shifting the token sequence along the spatial dimension according to a preset shift ratio, and applying linear projection and activation based on the weight matrix to the shifted token sequence, a long-range dependency relationship between different spatial locations is established, so that information can be transmitted between different spatial tokens within the same wetland land cover area, and a spatial aggregation feature map is obtained. The second interaction unit is used to perform channel mixing processing based on the spatial aggregation feature map. By applying gated linear projection along the channel dimension to the token sequence, the nonlinear combination relationship between different spectral bands is learned, and a spatial-spectral joint feature map that integrates spatial and spectral information is obtained. The third interaction unit is used to perform receptive field expansion processing based on the spatial-spectral joint feature map. By using a heterogeneous shift mechanism, the token sequence is input into multiple group convolutional layers with different hole rates. After being averaged along the channel dimension, the layers are sequentially mixed and cascaded. A global attention map is obtained by bidirectional scanning from the horizontal and vertical directions. The current attention result is combined with the previous attention results from different scanning directions to obtain a spatial-spectral joint feature map with a large receptive field.

6. The hyperspectral trans-domain wetland mapping system using a fractional Fourier RWKV network according to claim 4, characterized in that, The optimization module includes: The first optimization unit is used to perform fractional Fourier domain decomposition processing based on the layered spatial spectrum features. By mapping the layered spatial spectrum features to the fractional Fourier domain, the amplitude component of the reflected energy intensity of the corresponding ground object and the phase component of the spatial structure information of the corresponding ground object are extracted respectively to obtain the amplitude component and the phase component. The second optimization unit is used to perform energy constraint processing based on the amplitude components, and obtain the amplitude-constrained features by calculating the mean square error between the amplitude components of the generated features and the amplitude components of the real hyperspectral wetland image. The third optimization unit is used to perform structure-guided processing based on the phase components and the amplitude-constrained features. It extracts the spatial structure correlation between the phase components of the generated features and the phase components of the real hyperspectral wetland image by calculating the complex cross-correlation measure between them, guides the network to learn the structural information of the hyperspectral image, and obtains the optimized spatial-spectral features.

Images