A hyperspectral image category incremental classification method based on frequency domain prototype reconstruction

Abstract

The application provides a hyperspectral image category incremental classification method based on a frequency domain prototype reconstruction, and belongs to the technical field of remote sensing image processing and intelligent classification. First, a previous stage hyperspectral image classification model, a compact history memory and current incremental task samples are inputted to initialize history category to-be-reconstructed samples without saving history original image blocks. Second, a category space frequency domain template is used to constrain low-frequency layout, high-frequency details and a frequency domain ratio, so that the spatial representation of the history category is maintained. Then, a category prototype and a relationship memory bank are introduced to guide the to-be-reconstructed samples to be close to the target history category and far away from non-target categories. Finally, history category auxiliary samples are obtained by jointly optimizing multiple constraints, and the classification model is trained together with the current task samples. Experimental results show that the method maintains high classification accuracy on the WHU-Hi-Longkou dataset, and provides a new idea for hyperspectral image category incremental classification under the condition of no original samples.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image classification, and more specifically to an incremental classification method for hyperspectral images based on frequency domain prototype reconstruction. Background Technology

[0002] Hyperspectral images possess continuous, fine-grained spectral and spatial texture information, enabling precise identification of targets such as crops, urban features, vegetation cover, mineral distribution, and ecological environments. Therefore, they are widely used in precision agriculture, urban planning, environmental monitoring, disaster assessment, and resource exploration. As remote sensing observation areas, acquisition times, and application tasks constantly change, hyperspectral image classification models often need to incrementally learn newly emerging land cover categories while maintaining the ability to recognize previously learned historical categories.

[0003] In incremental hyperspectral image classification tasks, if the model is trained using only samples from the new category of the current task, the model parameters will rapidly shift towards the current category distribution, leading to a significant decrease in the ability to distinguish historical categories, i.e., catastrophic forgetting. Traditional sample replay methods mitigate forgetting by saving the original samples of historical categories and reusing these samples in subsequent training phases. However, these methods require long-term storage of historical hyperspectral original image patches, pixel-level spectral spatial tensors, or spatial location information, which may be constrained by factors such as storage costs, data sharing limitations, privacy protection, and data management regulations in practical remote sensing scenarios.

[0004] To avoid preserving historical original samples, existing incremental learning methods without original samples typically reconstruct the model using the previous stage model, generating historical class auxiliary samples, which are then used for subsequent training. Existing model reconstruction methods largely rely on classification confidence, statistical information from the model's normalization layer, spectral channel mean and variance, or local smoothing constraints. For hyperspectral images, simply maintaining low-order spectral statistics is insufficient to guarantee the reliability of the reconstructed samples in terms of spatial structure, edge texture, and class discrimination regions. This can easily lead to problems such as inaccurate spatial frequency domain representations of the reconstructed samples, unclear class boundaries, and confusion between different historical classes in the feature space.

[0005] The inventors discovered that when performing incremental classification of hyperspectral images without original samples, the key to preserving historical categories lies not only in providing training evidence for the old categories, but also in whether this training evidence can maintain the spatial frequency domain representation and class discriminative relationship of the target historical categories. Therefore, a hyperspectral image incremental classification method is needed that can obtain auxiliary samples for historical categories through frequency domain prototype reconstruction without preserving the original historical hyperspectral image patches, and then use these samples for current-stage classification training. Summary of the Invention

[0006] This invention proposes an incremental classification method for hyperspectral images based on frequency domain prototype reconstruction. It combines the spatial representation preservation capability of the category space frequency domain template, the target category constraint capability of the category prototype, and the non-target category discrimination capability of the relation memory. The method first inputs the previous stage hyperspectral image classification model, compact historical memory, and the current incremental task samples, initializing the historical category samples to be reconstructed without preserving the original historical hyperspectral image patches. Then, the category space frequency domain template constrains the low-frequency layout, high-frequency details, and frequency domain proportions of the historical category samples to be reconstructed, preserving the spatial representation of the target historical category. Subsequently, prototype relation guidance constraints are constructed by combining the category prototype and the relation memory, causing the samples to be reconstructed to be closer to the target historical category and farther from non-target historical categories in the feature space. Finally, multiple constraints are jointly optimized to obtain historical category auxiliary samples, which are then used to train the classification model together with the current incremental task samples to obtain the incremental classification results for hyperspectral images.

[0007] The objective of this invention is achieved as follows:

[0008] An incremental classification method for hyperspectral images based on frequency domain prototype reconstruction includes the following steps:

[0009] Step a: Input the hyperspectral image classification model from the previous stage, the compact historical memory, and the current incremental task samples. The compact historical memory includes category spectral statistics, category spatial frequency domain templates, category prototypes, and relation memory.

[0010] Step b: Initialize the historical category samples to be reconstructed based on the historical category labels, and input the historical category samples to be reconstructed into the frozen previous stage hyperspectral image classification model to obtain the classification response and feature representation;

[0011] Step c: Use the category space frequency domain template to constrain the low-frequency layout, high-frequency details and frequency domain ratio of the historical category samples to be reconstructed, so as to maintain the spatial representation of the target historical category.

[0012] Step d: Construct prototype relation guidance constraints using category prototypes and relation memory banks to make the feature representations of the samples to be reconstructed from historical categories close to the target historical category and far away from non-target historical categories;

[0013] Step e involves iteratively optimizing the historical category samples to be reconstructed using joint classification constraints, model statistical constraints, spectral statistical constraints, frequency domain template constraints, prototype relationship guidance constraints, and smoothing constraints to obtain historical category auxiliary samples.

[0014] Step f: Train the current stage hyperspectral image classification model together with the historical category auxiliary samples and the current incremental task samples, and use the trained current stage hyperspectral image classification model to output the category incremental classification result.

[0015] Beneficial effects:

[0016] This invention proposes a hyperspectral image category incremental classification method based on frequency domain prototype reconstruction, which can provide historical category auxiliary samples for the current stage classification model without preserving historical original hyperspectral image patches. By constraining low-frequency layout, high-frequency details, and frequency domain proportions using a category space frequency domain template, the spatial representation reliability of historical category auxiliary samples can be improved. Prototype relation guidance through category prototypes and relation memory enhances the category discriminativeness of historical category auxiliary samples. By co-training the classification model with historical category auxiliary samples and current incremental task samples, the problem of old class forgetting during the category incremental classification process can be alleviated, improving the stability of hyperspectral image category incremental classification under conditions without original samples. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 This is an overall flowchart of the hyperspectral image category incremental classification method based on frequency domain prototype reconstruction in the present invention.

[0019] Figure 2 This is a schematic diagram of the frequency domain prototype reconstruction module in the method of the present invention.

[0020] Figure 3 The images show the classification results of the method of the present invention on the WHU-Hi-LongKou dataset, where (a) is the classification result of the comparison method and (b) is the classification result of the method of the present invention. Detailed Implementation

[0021] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments of the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, or combinations thereof.

[0023] A specific embodiment of the present invention provides a hyperspectral image category incremental classification method based on frequency domain prototype reconstruction, the flowchart of which is shown below. Figure 1As shown, it includes the following steps:

[0024] Step a: Obtain the hyperspectral image classification model from the previous stage, the compact historical memory, and the hyperspectral image samples for the current incremental task.

[0025] In this step, the previous stage hyperspectral image classification model refers to the classification model that has completed historical category training before the current incremental task begins. To avoid the current reconstruction process changing the discriminative ability of the historical model, this embodiment sets the previous stage hyperspectral image classification model to a frozen state, that is, the network parameters of the previous stage hyperspectral image classification model are not updated during the reconstruction process.

[0026] Compact historical memory is used to provide necessary statistical information and feature references for reconstructing historical category samples without preserving the original historical hyperspectral image patches. The compact historical memory includes category spectral statistics, category spatial frequency domain templates, category prototypes, and a relational memory. Category spectral statistics may include the mean and variance of historical categories across each spectral channel; the category spatial frequency domain template may include low-frequency energy statistics, high-frequency energy statistics, and frequency domain proportion statistics of historical categories in the spatial frequency domain; the category prototype can be the central feature representation of historical categories in the feature space; and the relational memory can be a fixed-capacity feature reference storage structure that stores normalized feature representations and corresponding category labels.

[0027] Among them, represents the compact historical memory retained from the previous stage, represents the category spectral statistics, represents the category spatial frequency domain template, represents the category prototype, and represents the relational memory bank.

[0028] It should be noted that the compact historical memory does not include historical hyperspectral original image patches, pixel-level spectral space tensors, or spatial coordinate information. The feature representations in the relational memory are not used as input-level training samples, are not concatenated with current task samples for playback, are not assigned supervised classification loss, and are only used as a feature space reference during the reconstruction of historical category samples.

[0029] Step b: Initialize the historical category samples to be reconstructed based on the historical category labels, and input the frozen previous stage classification model.

[0030] In this step, the target historical category label is first determined from the learned historical category set. Then, the historical category reconstruction samples are initialized based on the target historical category label. The historical category reconstruction samples can be initialized with random noise, and their dimensions are consistent with the input dimensions of the hyperspectral image classification model, including the number of spectral channels and the spatial window size.

[0031] The initialized historical category samples to be reconstructed are input into the frozen previous-stage hyperspectral image classification model to obtain the corresponding classification response and feature representation. The classification response is used to construct classification constraints, so that the historical category samples to be reconstructed are identified as the target historical category by the previous-stage hyperspectral image classification model; the feature representation is used to construct prototype relation guiding constraints, so that the historical category samples to be reconstructed have the discriminative attribute of the target historical category in the feature space.

[0032] Step c: Construct frequency domain template constraints based on category space frequency domain templates to maintain low-frequency layout, high-frequency details, and frequency domain proportion features.

[0033] In this step, a two-dimensional frequency domain transformation is performed along the spatial dimension for each spectral channel of the historical category sample to be reconstructed, resulting in a frequency domain amplitude representation. To reduce the impact of the dynamic range of the frequency domain amplitude on the optimization process, a logarithmic transformation or normalization can be applied to the frequency domain amplitude representation.

[0034] After obtaining the frequency domain amplitude representation, low-frequency energy of the historical category samples to be reconstructed is extracted based on a low-frequency mask, and high-frequency energy is extracted based on a high-frequency mask. The frequency domain ratio is then calculated based on the low-frequency and high-frequency energies. Low-frequency energy is mainly used to characterize the overall spatial layout and smooth region structure of the historical category samples to be reconstructed, while high-frequency energy is mainly used to characterize edge, texture, and local abrupt change information. The frequency domain ratio is used to characterize the relative relationship between low-frequency structure and high-frequency details.

[0035] For a target historical category, the corresponding category space frequency domain template is retrieved from the compact historical memory. The category space frequency domain template may include the mean and variance of the low-frequency energy, the mean and variance of the high-frequency energy, and the mean and variance of the frequency domain proportion of the target historical category. Frequency domain template constraints are constructed by comparing the differences between the low-frequency energy, high-frequency energy, and frequency domain proportion of the historical category sample to be reconstructed and the category space frequency domain template of the target historical category.

[0036] In one alternative implementation, the frequency domain template constraint can be expressed by the following formula:

[0037]

[0038]

[0039] in, Indicates samples to be reconstructed in the historical category The One spectral channel, Indicates the number of spectral channels. and These represent the low-frequency region and the high-frequency region, respectively. , and These represent low-frequency energy, high-frequency energy, and frequency domain ratio, respectively. , , , , and For target history category Category spatial frequency domain template parameters, This represents a normalized measure of difference. To prevent constants with a denominator of zero.

[0040] Specifically, the frequency domain template constraints include low-frequency structural consistency constraints, high-frequency detail consistency constraints, and frequency domain scale consistency constraints. Low-frequency structural consistency constraints make the historical category samples to be reconstructed approximate the target historical category in terms of overall spatial structure; high-frequency detail consistency constraints make the historical category samples to be reconstructed approximate the target historical category in terms of local texture and edge variations; and frequency domain scale consistency constraints make the historical category samples to be reconstructed approximate the target historical category in terms of the relative energy distribution between low and high frequencies.

[0041] Step d: Construct prototype relation guidance constraints based on category prototypes and relation memory, and determine positive reference features and negative reference features.

[0042] In this step, the historical category samples to be reconstructed are input into the frozen previous stage hyperspectral image classification model to extract the feature representation of the historical category samples to be reconstructed, and the feature representation is normalized to obtain the normalized feature representation.

[0043] For a target historical category, the corresponding category prototype is read from the compact historical memory and determined as the positive reference feature. The category prototype is used to represent the central position of the target historical category in the feature space, and the normalized feature representation of the historical category sample to be reconstructed should be close to the positive reference feature.

[0044] Simultaneously, feature representations with class labels different from the target historical class are selected from the relational memory as negative reference features. These negative reference features represent the reference positions of other historical classes in the feature space, and the normalized feature representations of the historical class samples to be reconstructed should be far from these negative reference features. Through the combined effect of positive and negative reference features, the class discriminative power of the historical class samples to be reconstructed in the feature space can be enhanced.

[0045] In at least one embodiment, the prototype relation guiding constraint can be constructed based on similarity. The similarity can be the inner product similarity between normalized feature representations. The prototype relation guiding constraint increases the similarity between the historical category sample to be reconstructed and the positive reference feature, while decreasing the similarity between the historical category sample to be reconstructed and the negative reference feature.

[0046] In one alternative implementation, the prototype relation guiding constraint can be expressed by the following formula:

[0047]

[0048] in, This represents the feature extraction mapping from the previous stage of the classification model that is now frozen. The normalized feature representation of the historical category sample to be reconstructed. Indicates the target history category Category prototype, Represents a relational memory. and These represent the normalized feature representations and their category labels in the relational memory, respectively. This represents the selected set of negative reference features. This represents the temperature coefficient.

[0049] Step e: Adopt the course negative sample selection strategy and combine multiple constraints to iteratively optimize the historical category samples to be reconstructed.

[0050] In this step, to avoid introducing overly difficult negative reference features directly in the early stages of reconstruction, which could lead to optimization instability, this embodiment employs a course negative sample selection strategy. Specifically, the similarity between the normalized feature representation of the historical category sample to be reconstructed and the feature representation of the non-target historical category in the relational memory is calculated, and the non-target historical category feature representations are sorted according to their similarity. In the early stages of iterative optimization, non-target historical category feature representations with lower similarity are selected as negative reference features; as the iterative optimization progresses, non-target historical category feature representations with higher similarity are gradually selected as negative reference features.

[0051] Through the negative sample selection strategy described above, the historical category samples to be reconstructed can first form a stable target category identity in the early stage of optimization, and then gradually face the more confusing historical category feature references, thereby strengthening the feature boundary between the target historical category and the non-target historical category.

[0052] During the iterative optimization process, the historical category samples to be reconstructed are optimized by combining classification constraints, model statistical distribution constraints, spectral statistical constraints, frequency domain template constraints, prototype relationship guidance constraints, and smoothing constraints.

[0053] In one alternative implementation, the formula for the joint optimization objective is as follows:

[0054]

[0055] in, Represents classification constraints, This indicates the statistical distribution constraints of the model. Indicates spectral statistical constraints, Represents frequency domain template constraints. Represents prototype-based relationships and guiding constraints. Indicates a smoothing constraint; , , , , and These represent the weights of each constraint term. Indicates the iteration round, This indicates the step size for optimization.

[0056] Among them, classification constraints are used to ensure that the frozen previous-stage hyperspectral image classification model classifies the historical category to be reconstructed samples into the corresponding target historical category; model statistical distribution constraints are used to ensure that the historical category to be reconstructed samples are close to the historical training distribution in the normalized layer statistical distribution of the previous-stage hyperspectral image classification model; spectral statistical constraints are used to ensure that the spectral channel mean and variance of the historical category to be reconstructed samples are close to the category spectral statistical information of the target historical category; frequency domain template constraints are used to ensure that the historical category to be reconstructed samples maintain the low-frequency structure, high-frequency details and frequency domain proportion features of the target historical category; prototype relation guidance constraints are used to ensure that the feature representation of the historical category to be reconstructed samples is close to the prototype of the target historical category and far away from the reference features of non-target historical categories; and smoothing constraints are used to ensure that the historical category to be reconstructed samples maintain continuity in the local spatial neighborhood.

[0057] Through the above joint optimization, the historical category samples to be reconstructed can not only conform to the target historical category at the classification output level, but also approach the historical category distribution at the three levels of spectral statistics, spatial frequency domain representation and feature relationship.

[0058] Step f: Train the classification model based on historical category auxiliary samples and current incremental task samples, and output the category incremental classification results.

[0059] When the iterative optimization reaches a preset number of iterations, or when the joint optimization objective meets a preset convergence condition, the optimized historical category samples to be reconstructed are determined as historical category auxiliary samples. Further, the historical category auxiliary samples and the real hyperspectral samples in the current incremental task are used to construct the current stage classification training data, where the historical category auxiliary samples provide old class supervision labels, and the current task samples provide current class supervision labels. Based on the classification training data, the current stage hyperspectral image classification model is trained to maintain the historical category discrimination boundary while learning the current category.

[0060] After training, the trained hyperspectral image classification model for the current stage is used to predict the category of the hyperspectral image to be classified, and the classification result is output. It should be noted that the historical category auxiliary samples are only used temporarily during the current incremental task training phase and are discarded after training ends. They are not stored as a long-term sample library, nor are they used as original sample replay data for subsequent tasks. When subsequent tasks require old class evidence, historical category auxiliary samples can be regenerated based on the frozen previous stage model and the updated compact historical memory.

[0061] Furthermore, such as Figure 2 As shown, the frequency domain prototype reconstruction module in this invention includes a frozen previous-stage hyperspectral image classification model, compact historical memory, historical category samples to be reconstructed, category space frequency domain template constraint branch, prototype relationship guiding branch, and joint optimization part.

[0062] Figure 2 The input components include a frozen previous-stage hyperspectral image classification model, a compact history memory, and randomly initialized historical class samples to be reconstructed. The frozen previous-stage hyperspectral image classification model provides the classification response, model statistical distribution, and feature representation. The compact history memory provides class spectral statistics, class spatial frequency domain templates, class prototypes, and relational memory.

[0063] Figure 2 The frequency domain template consistency branch is used to constrain the historical class samples to be reconstructed in the input space. This branch obtains the spatial frequency representation of the historical class samples to be reconstructed through a two-dimensional frequency domain transformation, and constrains the low-frequency structure, high-frequency details, and frequency domain scale based on the class space frequency domain template. This branch can avoid the problem that the reconstructed samples only satisfy the spectral mean and variance and lack a stable spatial structure.

[0064] Figure 2 The prototype relation-guided branch is used to constrain the historical category samples to be reconstructed in the feature space. This branch uses the category prototype of the target historical category as a positive reference feature, uses the feature representation of non-target historical categories in the relation memory as a negative reference feature, and guides the feature representation of the historical category samples to be reconstructed to move closer to the target category region and further away from the non-target category region through similarity relations.

[0065] Figure 2 The joint optimization part of the algorithm combines classification constraints, model statistical distribution constraints, spectral statistical constraints, frequency domain template constraints, prototype relation guidance constraints, and smoothing constraints in a weighted manner. It updates the historical category samples to be reconstructed themselves through reverse optimization, without updating the frozen parameters of the previous stage hyperspectral image classification model. After multiple iterations, historical category auxiliary samples are obtained.

[0066] This invention is validated using the WHU-Hi-LongKou dataset.

[0067] The WHU-Hi-LongKou dataset is a hyperspectral remote sensing dataset of farmland scenes from unmanned aerial vehicles (UAVs), containing land cover categories such as corn, cotton, sesame, soybeans, rice, water bodies, roads and houses, and weeds. In the experiments, the nine categories of this dataset were divided into three incremental tasks, each containing three categories. Incremental classification was validated without preserving historical original hyperspectral image samples.

[0068] In a specific embodiment of the present invention, historical category auxiliary samples are obtained through steps a to e, and these historical category auxiliary samples are used together with the current incremental task hyperspectral image samples for training the hyperspectral image classification model in the current stage. After the current stage of training is completed, the historical category auxiliary samples are discarded. This training process is used to verify the application effect of the present invention in performing incremental classification of hyperspectral image categories without saving the original historical samples, and does not change the core technical solution of frequency domain prototype reconstruction to generate historical category auxiliary samples and serve classification training.

[0069] The experimental results are shown in Table 1. After completing the third incremental task, the method of this invention achieved an overall classification accuracy of 95.21%, an average classification accuracy of 95.42%, and a Kappa coefficient of 93.75 on the WHU-Hi-LongKou dataset. Combining the classification results for each category, it can be seen that categories such as corn, cotton, water bodies, roads and houses, and mixed weeds maintain high classification accuracy after incremental category learning. This indicates that the historical category auxiliary samples reconstructed by this invention can provide effective old class evidence for the current stage of model training even without original samples.

[0070] like Figure 3 As shown, the classification results using the historical category auxiliary samples of this invention exhibit good continuity and discriminative power in farmland boundaries, road and house areas, and adjacent crop category areas. This result demonstrates that the historical category auxiliary samples obtained through frequency domain template constraints and prototype relationships possess good structural reliability and feature discrimination assistance capabilities, and can mitigate the forgetting of old categories during incremental classification without original samples.

[0071] Table 1. Experimental results of incremental classification on the WHU-Hi-LongKou dataset.

[0072]

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A hyperspectral image category incremental classification method based on frequency domain prototype reconstruction, characterized in that, Includes the following steps: Step a: Input the hyperspectral image classification model from the previous stage, the compact historical memory, and the current incremental task samples, wherein the compact historical memory is used to provide historical category statistics and feature references without saving the original historical hyperspectral image patches; Step b: Initialize the historical category to be reconstructed samples according to the historical category labels, and input the historical category to be reconstructed samples into the frozen previous stage hyperspectral image classification model to obtain the classification response and feature representation; Step c: Using the category space frequency domain template in the compact historical memory, constrain the low-frequency layout, high-frequency details and frequency domain ratio of the historical category to be reconstructed sample, so as to maintain the spatial representation of the target historical category. Step d: Construct prototype relation guidance constraints using category prototypes and relation memory banks to make the feature representations of the samples to be reconstructed from historical categories close to the target historical category and far away from non-target historical categories; Step e involves iteratively optimizing the historical category samples to be reconstructed using joint classification constraints, model statistical constraints, spectral statistical constraints, frequency domain template constraints, prototype relationship guidance constraints, and smoothing constraints to obtain historical category auxiliary samples. Step f: Train the current stage hyperspectral image classification model together with the historical category auxiliary samples and the current incremental task samples, and use the trained current stage hyperspectral image classification model to output the category incremental classification result.

2. The hyperspectral image category incremental classification method based on frequency domain prototype reconstruction according to claim 1, characterized in that, The compact historical memory in step a includes category spectral statistics, category spatial frequency domain template, category prototype, and relation memory bank; wherein, category spectral statistics are used to characterize the statistical distribution of historical category samples in the spectral channels, category spatial frequency domain template is used to characterize the low frequency, high frequency, and frequency domain ratio information of historical category samples in the spatial frequency domain, category prototype is used to characterize the central representation of historical categories in the feature space, and relation memory bank is used to store normalized feature representations and their category labels.

3. The hyperspectral image category incremental classification method based on frequency domain prototype reconstruction according to claim 1, characterized in that, The category space frequency domain template in step c is constructed in the following way: performing a two-dimensional frequency domain transformation along the spatial dimension on each spectral channel of the historical category training samples to obtain the frequency domain amplitude representation; extracting low-frequency energy and high-frequency energy respectively, and calculating the frequency domain ratio based on the low-frequency energy and high-frequency energy; and statistically calculating the mean and variance of low-frequency energy, high-frequency energy and frequency domain ratio according to the category to obtain the category space frequency domain template.

4. The hyperspectral image category incremental classification method based on frequency domain prototype reconstruction according to claim 1, characterized in that, Step c specifically includes: converting the historical category samples to be reconstructed to the spatial frequency domain, calculating their low-frequency energy, high-frequency energy and frequency domain ratio respectively, and matching them with the category spatial frequency domain template of the target historical category to maintain the consistency of the historical category samples to be reconstructed in terms of overall layout, local details and frequency domain ratio.

5. The hyperspectral image category incremental classification method based on frequency domain prototype reconstruction according to claim 1, characterized in that, In step d, the category prototype serves as the positive reference feature of the target historical category, and the feature representation of the non-target historical category in the relation memory serves as the negative reference feature. By increasing the similarity between the historical category sample to be reconstructed and the positive reference feature and decreasing its similarity with the negative reference feature, the class discriminativeness of the historical category auxiliary sample is improved.

6. The hyperspectral image category incremental classification method based on frequency domain prototype reconstruction according to claim 5, characterized in that, The negative reference features are determined using a course selection strategy: in the early stage of iterative optimization, non-target historical category features with low similarity to the historical category samples to be reconstructed are selected as negative reference features, and in the later stage of iterative optimization, non-target historical category features with high similarity are gradually selected as negative reference features.

7. The method for incremental classification of hyperspectral images based on frequency domain prototype reconstruction according to claim 1, characterized in that, The classification constraints in step e are used to make the frozen previous stage hyperspectral image classification model identify the historical category to be reconstructed sample as the target historical category; the model statistical constraints are used to make the historical category to be reconstructed sample match the statistical distribution of the previous stage hyperspectral image classification model; the spectral statistical constraints are used to make the spectral channel statistical features of the historical category to be reconstructed sample close to the target historical category; and the smoothing constraints are used to maintain the local spatial continuity of the historical category to be reconstructed sample.

8. The hyperspectral image category incremental classification method based on frequency domain prototype reconstruction according to claim 1, characterized in that, The historical category auxiliary samples obtained in step e are only used temporarily during the current incremental task training phase and are discarded after training ends. They are not saved as a long-term sample library, nor are they used as historical original sample replay data for subsequent tasks.

9. The hyperspectral image category incremental classification method based on frequency domain prototype reconstruction according to claim 1, characterized in that, In step f, historical category auxiliary samples and their historical category labels are used as old category training evidence, and current incremental task samples and their category labels are used as current category training evidence. Together, they form the current stage training set to train the current stage hyperspectral image classification model.

10. The hyperspectral image category incremental classification method based on frequency domain prototype reconstruction according to claim 1, characterized in that, In step f, the trained hyperspectral image classification model for the current stage is used to predict the category of the hyperspectral image to be classified, and obtains an incremental classification result that includes the historical category and the current category.

Images