CN-122024074-A - Hyperspectral cross-domain wetland mapping method and system for fractional Fourier RWKV network

CN122024074ACN 122024074 ACN122024074 ACN 122024074ACN-122024074-A

Abstract

The invention provides a hyperspectral cross-domain wetland mapping method and system of a fractional Fourier RWKV network, which relate to the technical field of remote sensing intelligent processing and comprise the steps of obtaining a source domain hyperspectral image with a label and a target domain hyperspectral image without a label; performing feature conversion according to a source domain hyperspectral image and a target domain hyperspectral image to obtain feature vectors, performing space spectrum interaction according to the feature vectors to obtain a space spectrum combined feature map with a large receptive field, performing network hierarchical design according to the space spectrum combined 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 method realizes the joint mining of hyperspectral image space, spectrum, amplitude and phase information, thereby enhancing the learning ability of wetland texture, detail and structural characteristics.

Inventors

LI HENGCHAO
REN GUOLIANG
HU WENSHUAI
ZHAO XUDONG

Assignees

西南交通大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. A method for mapping a hyperspectral cross-domain wetland for a fractional fourier RWKV network, comprising: Acquiring a source domain hyperspectral image with a label and a target domain hyperspectral image without a label, wherein the source domain hyperspectral image and the target domain hyperspectral image are hyperspectral remote sensing data containing space dimension and spectrum dimension; performing feature conversion according to the source domain hyperspectral image and the target domain hyperspectral image, dividing pixels into token sequences, expanding the token sequences into channel representations, and performing feature learning in a fractional Fourier domain at the same time to obtain feature vectors fusing spatial spectrum information and fractional Fourier domain amplitude-phase information; performing spatial spectrum interaction according to the feature vector, establishing a long-range dependency relationship between positions by performing shift and cyclic convolution on the token sequence along a spatial dimension, and performing interaction fusion on token information along a channel dimension to obtain a spatial spectrum combined feature map with a large receptive field; Carrying out network layering design according to the spatial spectrum combined feature map, and obtaining layered spatial spectrum features of amplitude-phase modulation by stacking feature characterization and spatial spectrum interaction modules layer by layer and alternately executing spatial dimension and channel dimension mixing operation between layers; Generating and optimizing according to the layered spatial spectrum characteristics, generating energy distribution consistency with real characteristics through amplitude component constraint to keep spectrum fidelity, and measuring structural correlation of the two in phase component to obtain optimized spatial spectrum characteristics; And performing cross-domain alignment according to the optimized spatial spectrum features, inputting the spatial spectrum features of the source domain and the target domain into a double classifier, applying countermeasure constraint, performing supervision training by using a source domain label to enable the model to have class recognition capability, and relieving catastrophic forgetting in the training process by predicting consistency constraint at each stage to obtain a cross-domain wetland classification map.
2. The method of mapping a hyperspectral cross-domain wetland for a fractional fourier RWKV network according to claim 1, wherein performing feature transformation from the source domain hyperspectral image and the target domain hyperspectral image comprises: Carrying out token construction processing according to the source domain hyperspectral image and the target domain hyperspectral image, and obtaining an input sequence containing space position coding and spectrum channel representation by converting the hyperspectral image into a characteristic sequence representation; Performing fractional Fourier domain transformation processing according to the input sequence, and obtaining a fractional domain characterization sequence fusing the spatial texture and the spectral absorption characteristic by selecting a transformation order matched with the spectral characteristics of the hyperspectral wetland object and applying rotary transformation according to the transformation order to enable the token sequence to be unfolded in the fractional Fourier domain between the spatial domain and the frequency domain so as to simultaneously reserve the spatial domain energy of the spatial texture and the frequency domain energy of the spectral absorption characteristic; And performing amplitude-phase extraction processing according to the fractional Fourier domain characterization sequence, respectively calculating amplitude information of each token in the fractional Fourier domain to obtain energy distribution intensity information of the ground object target, and extracting phase information to obtain a feature vector fusing the spatial spectrum information and the fractional Fourier domain amplitude-phase information.
3. The method for hyperspectral cross-domain wetland mapping in fractional fourier RWKV network according to claim 1, wherein performing spatial-spectral interactions according to the feature vector comprises: performing space mixing processing according to the feature vectors, performing cyclic shift on the token sequence along the space dimension according to a preset shift proportion, and applying linear projection and activation based on a weight matrix to the shifted token sequence to establish a long-range dependency relationship between different space positions, so that information transfer between different space tokens in the same wetland object coverage area is realized, and a space aggregation feature map is obtained; Carrying out channel mixing processing according to the space aggregation feature map, and obtaining a space spectrum combination feature map fusing space and spectrum information by applying gating linear projection to the token sequence along the channel dimension so as to learn nonlinear combination relations among different spectrum bands; And performing receptive field expansion processing according to the spatial spectrum combined characteristic map, inputting a token sequence into a plurality of group convolution layers with different void ratios by adopting a heterogeneous shift mechanism, mixing and cascading sequentially after being evenly grouped along a channel dimension, acquiring global attention force map by bidirectional scanning from a horizontal direction and a vertical direction, and combining a current attention result with previous attention results in different scanning directions to obtain the spatial spectrum combined characteristic map with a large receptive field.
4. The method for hyperspectral cross-domain wetland mapping of fractional fourier RWKV network according to claim 1, wherein the network hierarchical design is performed according to the spatial spectrum joint feature map, comprising: Carrying out local feature extraction processing according to the spatial spectrum combined feature map, aligning channel dimensions by adopting a two-dimensional convolution layer in a network stem module, and executing spatial dimension and channel dimension mixing operation of a multi-layer fractional Fourier transform RWKV module in a shallow network to capture local textures and spectrum details of wetland vegetation, water and bare land so as to obtain shallow local spatial spectrum features; Performing receptive field increment processing according to the shallow local spatial spectrum characteristics, halving the characteristic space dimension by using a depth convolution layer with the step length as a set value in a downsampling module, and doubling the channel dimension by using the convolution layer to obtain global spatial spectrum characteristics with large receptive field; And performing cross-layer fusion processing according to the shallow local spatial spectrum features and the global spatial spectrum features, doubling the spatial dimension of the global spatial spectrum features and reducing the channel dimension by half by using a bilinear interpolation upsampling layer and a convolution layer in an upsampling module, and splicing the upsampled global spatial spectrum features and the shallow local spatial spectrum features along the channel dimension by using a fusion projection head so as to simultaneously reserve local details and global semantics of ground features, thereby obtaining the layered spatial spectrum features of amplitude-phase modulation.
5. The method for hyperspectral cross-domain wetland mapping in fractional fourier RWKV network according to claim 1, wherein generating optimization is performed according to the hierarchical spatial spectrum features, comprising: Carrying out fractional Fourier domain decomposition processing according to the layered spatial spectrum characteristics, and respectively extracting an amplitude component corresponding to the ground object target reflection energy intensity and a phase component corresponding to the ground object target space structure information by mapping the layered spatial spectrum characteristics to the fractional Fourier domain to obtain the amplitude component and the phase component; performing energy constraint processing according to the amplitude components, and obtaining the characteristics after the amplitude constraint by calculating the mean square error between the amplitude components of the generated characteristics and the amplitude components of the real hyperspectral wetland image; and carrying out structure guiding processing according to the phase component and the amplitude-constrained characteristic, and extracting the spatial structure correlation of the phase component of the generated characteristic and the phase component of the real hyperspectral wetland image by calculating the complex cross-correlation metric between the phase component and the phase component of the real hyperspectral wetland image, so as to guide a network to learn the structural information of the hyperspectral image and obtain the optimized spatial spectrum characteristic.
6. A hyperspectral cross-domain wetland mapping system for a fractional fourier RWKV network, comprising: the system comprises an acquisition module, a storage module and a display module, wherein the acquisition module is used for acquiring a source domain hyperspectral image with a label and a target domain hyperspectral image without a label, and the source domain hyperspectral image and the target domain hyperspectral image are hyperspectral remote sensing data containing space dimension and spectrum dimension; the conversion module is used for carrying out feature conversion according to the source domain hyperspectral image and the target domain hyperspectral image, dividing pixels into token sequences and expanding the token sequences into channel representations, and simultaneously carrying out feature learning in a fractional Fourier domain to obtain feature vectors fusing spatial spectrum information and fractional Fourier domain amplitude-phase information; the interaction module is used for carrying out space spectrum interaction according to the feature vector, establishing a long-range dependence relationship between positions by carrying out shift and cyclic convolution on the token sequence along a space dimension, and carrying out interaction fusion on the token information along a channel dimension to obtain a space spectrum combined feature map with a large receptive field; the layering module is used for carrying out network layering design according to the spatial spectrum combined feature map, and obtaining the layering spatial spectrum features of amplitude-phase modulation by stacking the feature characterization and spatial spectrum interaction module layer by layer and alternately executing space dimension and channel dimension mixing operation between layers; The optimization module is used for generating and optimizing according to the layered spatial spectrum characteristics, generating energy distribution consistency with real characteristics through amplitude component constraint to keep spectrum fidelity, and measuring structural correlation of the two in phase components to obtain optimized spatial spectrum characteristics; The output module is used for performing cross-domain alignment according to the optimized spatial spectrum characteristics, inputting the spatial spectrum characteristics of the source domain and the target domain into the double classifier, applying countermeasure constraint, performing supervision training by utilizing the source domain label to enable the model to have class identification capability, and relieving catastrophic forgetting in the training process by predicting consistency constraint at each stage to obtain a cross-domain wetland classification map.
7. The hyperspectral cross-domain wetland cartography system of a fractional order fourier RWKV network as recited in claim 6, wherein the transformation module comprises: the first conversion unit is used for carrying out token construction processing according to the source domain hyperspectral image and the target domain hyperspectral image, and obtaining an input sequence containing space position coding and spectrum channel representation by converting the hyperspectral image into a characteristic sequence representation; The second conversion unit is used for carrying out fractional Fourier domain transformation processing according to the input sequence, and obtaining a fractional domain representation sequence integrating the spatial texture and the spectral absorption characteristic by selecting a transformation order matched with the spectral characteristics of the hyperspectral wetland object and applying rotary transformation according to the transformation order to enable the token sequence to be unfolded in the fractional Fourier domain between the spatial domain and the frequency domain so as to simultaneously reserve the spatial domain energy of the spatial texture and the frequency domain energy of the spectral absorption characteristic; and the third conversion unit is used for carrying out amplitude-phase extraction processing according to the fractional domain characterization sequence, obtaining energy distribution intensity information of the ground object target by respectively calculating the amplitude information of each token in the fractional Fourier domain, and extracting phase information to obtain a feature vector fusing the spatial spectrum information and the fractional Fourier domain amplitude-phase information.
8. The hyperspectral cross-domain wetland cartography system of a fractional order fourier RWKV network as recited in claim 6, wherein the interaction module comprises: the first interaction unit is used for carrying out space mixing processing according to the feature vectors, circularly shifting the token sequence along the space dimension according to a preset shifting proportion, and applying linear projection and activation based on a weight matrix to the shifted token sequence to establish a long-range dependency relationship between different space positions so as to realize information transfer between different space tokens in the same wetland object coverage area and obtain a space aggregation feature map; The second interaction unit is used for carrying out channel mixing processing according to the space aggregation feature map, and obtaining a space spectrum combination feature map fusing space and spectrum information by applying gating linear projection to the token sequence along the channel dimension so as to learn nonlinear combination relations among different spectrum bands; And the third interaction unit is used for performing receptive field expansion processing according to the spatial spectrum joint feature map, inputting the token sequence into a plurality of group convolution layers with different void ratios by adopting a heterogeneous shift mechanism, mixing and cascading sequentially after being evenly grouped along the channel dimension, acquiring a global attention map by bidirectional scanning from the horizontal direction and the vertical direction, and combining the current attention result with the previous attention results in different scanning directions to obtain the spatial spectrum joint feature map with a large receptive field.
9. A hyperspectral cross-domain wetland cartography system of a fractional order fourier RWKV network as recited in claim 6, wherein said layering module comprises: the first layering unit is used for carrying out local feature extraction processing according to the spatial spectrum combined feature map, aligning channel dimensions by adopting a two-dimensional convolution layer in a network stem module, and executing space dimension and channel dimension mixing operation of a multi-layer fractional Fourier transform RWKV module in a shallow network to capture local textures and spectrum details of wetland vegetation, water and bare land so as to obtain shallow local spatial spectrum features; the second layering unit is used for performing receptive field increment processing according to the shallow local spatial spectrum characteristics, halving the characteristic space dimension by using a depth convolution layer with the step length as a set value in a downsampling module and doubling the channel dimension by using the convolution layer to obtain global spatial spectrum characteristics with large receptive fields; and the third layering unit is used for performing cross-layer fusion processing according to the shallow local spatial spectrum characteristics and the global spatial spectrum characteristics, doubling the spatial dimension of the global spatial spectrum characteristics and reducing the channel dimension by half by using a bilinear interpolation upsampling layer and a convolution layer in the upsampling module, and splicing the upsampled global spatial spectrum characteristics and the shallow local spatial spectrum characteristics along the channel dimension by using a fusion projection head so as to simultaneously reserve the local details and global semantics of the ground object, thereby obtaining the amplitude-phase modulated layered spatial spectrum characteristics.
10. The hyperspectral cross-domain wetland cartography system of a fractional order fourier RWKV network as recited in claim 6, wherein the optimization module comprises: The first optimizing unit is used for carrying out fractional Fourier domain decomposition processing according to the layered spatial spectrum characteristics, and respectively extracting an amplitude component corresponding to the ground object target reflection energy intensity and a phase component corresponding to the ground object target space structure information by mapping the layered spatial spectrum characteristics to the fractional Fourier domain to obtain the amplitude component and the phase component; The second optimizing unit is used for carrying out energy constraint processing according to the amplitude components, and obtaining the amplitude-constrained characteristics by calculating the mean square error between the amplitude components of the generated characteristics and the amplitude components of the real hyperspectral wetland image; And the third optimizing unit is used for conducting structure guiding processing according to the phase component and the amplitude-constrained characteristic, and extracting the spatial structure correlation of the phase component of the generated characteristic and the phase component of the real hyperspectral wetland image by calculating the complex cross-correlation metric between the phase component and the phase component of the real hyperspectral wetland image, so as to guide a network to learn the structural information of the hyperspectral image and obtain the optimized spatial spectrum characteristic.

Description

Hyperspectral cross-domain wetland mapping method and system for fractional Fourier RWKV network Technical Field The invention relates to the technical field of remote sensing intelligent processing, in particular to a hyperspectral cross-domain wetland mapping method and system of a fractional Fourier RWKV network. Background The coastal wetland is used as an ecological system with great research value, can provide animal and plant habitats, human foods, industrial raw materials and the like, and plays a key role in climate and hydrologic regulation, promotion of carbon sequestration, emission reduction, ecological balance maintenance and the like. Reliable monitoring methods are critical for wetland restoration, protection and sustainable management. As a ground observation technology integrating imaging and spectrum technology, the image obtained by the hyperspectral remote sensing technology has rich space details and spectrum information, has unique advantages in the aspects of physical structure and component identification, and becomes an important means for finely monitoring the coastal wetland. The deep learning technology has been widely applied to the tasks of hyperspectral remote sensing image classification, coastal wetland cartography and the like by the strong characteristic learning capability and data generalization advantages. However, the dynamic characteristics of the coastal wetland (such as tidal fluctuation, climatic change and the like) are easy to cause the spectrum to change significantly along with time and environment, meanwhile, the wetland is usually provided with various land coverage types and mixed pixels, the collection of samples is high in cost and time-consuming due to remote and complex environment, hyperspectral images are usually collected in different geographic areas, different times or different sensors, serious field shifts exist among different scenes due to the difference of the types and the spectral characteristics of the wetland, and challenges are presented to the practical application effects of the existing hyperspectral image classification and wetland mapping methods. The existing hyperspectral wetland mapping algorithm based on deep learning and domain self-adaption has the following defects that firstly, due to dynamic change of the wetland, the existing cross-domain classification and mapping method usually ignores frequency and phase information of hyperspectral images, so that the problem of domain offset caused by changes of sensors, illumination, phase and the like is difficult to solve. 2. The transform-based method generally has the problem of secondary complexity, and meanwhile, the existing frequency domain analysis method often neglects the effective utilization of phase information for characterizing the texture and structural characteristics of the hyperspectral image, so that important detail information is easy to lose in the processes of data generation and cross-domain migration. Based on the shortcomings of the prior art, a method and a system for mapping hyperspectral cross-domain wetland of a fractional Fourier RWKV network are needed. Disclosure of Invention The invention aims to provide a hyperspectral cross-domain wetland mapping method and system for a fractional Fourier RWKV network, so as to solve the problems. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: in a first aspect, the application provides a method for mapping a hyperspectral cross-domain wetland in a fractional Fourier RWKV network, which comprises the following steps: Acquiring a source domain hyperspectral image with a label and a target domain hyperspectral image without a label, wherein the source domain hyperspectral image and the target domain hyperspectral image are hyperspectral remote sensing data containing space dimension and spectrum dimension; Performing feature conversion according to the source domain hyperspectral image and the target domain hyperspectral image, dividing pixels into token sequences, expanding the token sequences into channel representations, and applying rotation conversion in a fractional Fourier domain to obtain feature vectors fusing spatial spectrum information and fractional Fourier domain amplitude-phase information; performing spatial spectrum interaction according to the feature vector, establishing a long-range dependency relationship between positions by performing shift and cyclic convolution on the token sequence along a spatial dimension, and performing interaction fusion on token information along a channel dimension to obtain a spatial spectrum combined feature map with a large receptive field; Carrying out network layering design according to the spatial spectrum combined feature map, and obtaining layered spatial spectrum features of amplitude-phase modulation by stacking feature characterization and spatial spectrum interaction modules layer by layer and alternate