CN-121982564-A - Universal domain adaptive hyperspectral remote sensing image classification method based on graph

CN121982564ACN 121982564 ACN121982564 ACN 121982564ACN-121982564-A

Abstract

The invention relates to the technical field of intelligent processing and analysis of remote sensing images, and discloses a graph-based general domain adaptive hyperspectral remote sensing image classification method. Weights are calculated for the target domain and the source domain samples respectively through a reconstruction network and an auxiliary domain classifier so as to distinguish public classes from private classes and inhibit interference of the private classes on characteristic alignment. On the basis, a weighted circular consistency alignment strategy is adopted, and the conditional distribution alignment of the two-domain common class features is realized at the sample level. And finally, training an extension classifier by using the weighted samples to finish accurate classification of public classes and effective identification of private classes in the target domain. The method can remarkably improve the classification precision and the robustness of the hyperspectral image under the scenes of cross-domain and inconsistent categories.

Inventors

AI HAO
MA LI

Assignees

中国地质大学（武汉）

Dates

Publication Date: 20260505
Application Date: 20260116

Claims (10)

1. A graph-based general domain adaptive hyperspectral remote sensing image classification method is characterized by comprising the following steps: s1, acquiring a hyperspectral remote sensing image source domain data set and a target domain data set, wherein the source domain data set comprises a label sample, the target domain data set comprises a non-label sample, C categories shared by two domains are defined as public categories, and the categories unique to the source domain and the target domain are private categories; s2, constructing a graph-based feature extraction network, extracting features from a source domain sample and a target domain sample respectively, and dynamically constructing and updating a graph structure for representing the relationship between the samples; s3, calculating weights for the source domain sample and the target domain sample respectively based on the features extracted in the step S2; S4, based on the features extracted in the step S2 and the weights calculated in the step S3, carrying out sample-level conditional distribution alignment on the common class features of the source domain and the target domain by adopting a weighted cyclic consistency alignment method; S5, training an extended classifier by using the weighted source domain labeled samples and the weighted target domain samples, classifying the target domain samples, and dividing the samples into public classes or target domain private classes.
2. The method for classifying the hyperspectral remote sensing images based on the universal domain adaptation of the graph as claimed in claim 1, wherein the step S2 specifically comprises the following steps: s21, obtaining node feature vectors of each sample including a source domain sample and a target domain sample through a full connection layer of a feature extraction network; S22, calculating KL divergence between any two node feature vectors based on node feature vectors of all samples, constructing an edge weight matrix representing similarity among nodes according to the KL divergence, and normalizing the edge weight matrix to obtain a graph adjacency matrix; s23, aggregating the characteristics of adjacent nodes according to the graph adjacent matrix, and carrying out characteristic transformation through a learnable full-connection layer to update the characteristics of the nodes; s24, based on the updated node characteristics, steps S22 to S23 are circularly executed, and the graph structure and the node characteristics are dynamically updated.
3. The method for classifying the universal domain adaptive hyperspectral remote sensing images based on the graph of claim 1, wherein the step S3 comprises the following steps: S31, inputting the features extracted in the step S2 into a reconstruction network to calculate a reconstruction error of the target domain sample, and distributing weight for the target domain sample according to the reconstruction error, wherein the weight is larger as the error is smaller; S32, inputting the features extracted in the step S2 into an auxiliary domain classifier to calculate the probability that the source domain sample belongs to the source domain, and distributing weights for the source domain sample according to the probability, wherein the weights are larger when the probability is lower.
4. The method for classifying the image of the universal domain adaptive hyperspectral remote sensing image based on the graph of claim 3, wherein the step S31 specifically comprises the following steps: S311, training a reconstruction network by using the source domain data, so that the reconstruction network learns the characteristic reconstruction capability of the source domain data; s312, inputting the characteristics of the target domain sample into a trained reconstruction network to obtain reconstructed target domain data; S313, calculating a reconstruction error between the original sample characteristic of each target domain and the reconstruction characteristic of each target domain; s314, converting the reconstruction errors into weights according to the reconstruction errors of all the target domain samples, wherein the weights obtained by the target domain samples with smaller reconstruction errors are larger.
5. The method for classifying the image of the universal domain adaptive hyperspectral remote sensing image based on the graph of claim 3, wherein the step S32 specifically comprises the following steps: S321, constructing an auxiliary domain classifier, distributing a pseudo tag 1 for a source domain sample, and distributing a pseudo tag 0 for a target domain sample; s322, inputting the source domain sample characteristics into the auxiliary domain classifier to obtain a probability value which is predicted to be from the source domain; s323, calculating the weight for the source domain samples according to the probability value, wherein the source domain samples with lower probability value have larger weight.
6. The method for classifying the hyperspectral remote sensing images based on the universal domain adaptation as claimed in claim 1, wherein the weighted cyclic consistency alignment method in the step S4 specifically comprises the following steps: s41, calculating a connection probability matrix from the source domain to the target domain based on weighted sample characteristics of the source domain and the target domain, and fusing the connection probability matrix from the target domain to the source domain of the target domain prediction probability; S42, calculating the cyclic connection probability of the source domain sample after the target domain sample is connected back to the source domain sample based on the two connection probability matrixes; S43, constructing a cyclic consistency loss function, wherein the loss function comprises a first loss term for restraining cyclic connection probability by using a source domain label and a second loss term for restraining uniform distribution of the connection probability from a source domain to a target domain; S44, driving the feature extraction network learning inter-domain invariant feature by minimizing the loop consistency loss function.
7. The method for classifying the hyperspectral remote sensing images based on the universal domain adaptation of the graph as claimed in claim 3, wherein the step S5 specifically comprises the following steps: S51, calculating source domain classification loss by using a labeled source domain sample and a real label thereof, wherein the source domain classification loss is used for training a feature extraction network and an expansion classifier; s52, calculating the weighted classification loss of the target domain by using the target domain sample and the weight thereof obtained in the step S31, wherein the smaller the weight is, the larger the contribution of the sample to the loss is, so as to enhance the recognition capability of the classifier to the private class of the target domain; and S53, combining the source domain classification loss and the target domain weighted classification loss, and optimizing the extended classifier.
8. The method for classifying the universal domain adaptive hyperspectral remote sensing image based on the graph of claim 1, further comprising the step of model offline training: a. initializing parameters of a feature extraction network, a reconstruction network, an auxiliary domain classifier and an extension classifier; b. In each round of training iteration, a reconstruction loss L r , a source domain classification loss L s , a target domain weighted classification loss L t and a loop consistency loss L cycle are sequentially calculated; c. All network parameters are updated by the back propagation algorithm according to the joint loss function L total : L total = L s + λ t * L t + λ r * L r + λ cycle * L cycle Wherein lambda t 、λ r 、λ cycle is the weight super parameter of each loss; d. And c, repeating the steps b to c until the model converges to obtain a trained general domain adaptive hyperspectral remote sensing image classification model.
9. A storage medium is characterized in that the storage medium stores instructions and data for realizing the graph-based general domain adaptive hyperspectral remote sensing image classification method according to any one of claims 1 to 8.
10. The universal domain adaptive hyperspectral remote sensing image classification device based on the graph is characterized by comprising a processor and a storage medium, wherein the processor loads and executes instructions and data in the storage medium to realize the universal domain adaptive hyperspectral remote sensing image classification method based on the graph according to any one of claims 1-8.

Description

Universal domain adaptive hyperspectral remote sensing image classification method based on graph Technical Field The invention relates to the technical field of intelligent processing and analysis of remote sensing images, in particular to a general domain adaptive hyperspectral remote sensing image classification method based on a graph. Background The spectrum image has hundreds of spectrum bands, has rich spatial information and spectrum characteristics, is favorable for classifying the image at the pixel level, and has wide application space in the fields of precise agriculture, environment monitoring, urban research and the like. The deep learning is used as a powerful feature extraction tool, and can effectively solve the problem of image classification. Many conventional deep learning-based hyperspectral image classification methods assume that the data follow independent co-distribution principles and are not effectively addressed when cross-domain problems are encountered. Classification methods, such as Convolutional Neural Networks (CNNs), are difficult to generalize efficiently in cross-domain tasks. In practical applications, acquiring tagged data is often time consuming and labor intensive, and even impractical. Unsupervised domain adaptation (Unsupervised Domain Adaptation) migrates knowledge from tagged source domain data to untagged target domain data by narrowing the distribution gap between the source domain and the target domain. The existing hyperspectral remote sensing image unsupervised domain adaptation algorithm mainly adopts a distribution alignment strategy. In recent years, domain adaptation methods based on countermeasure learning have been attracting attention, and such methods learn a common feature space through countermeasure learning by a feature extractor and a domain discriminator. The traditional unsupervised domain adaptation algorithm follows the condition that the source domain and the target domain must share the same tag set, and the domain adaptation in this scenario is called closed-set domain adaptation. However, in a real environment, because the hyperspectral remote sensing image has wide coverage area, the types of ground features are complex and continuously changed, it is difficult to ensure that the types of ground features in a source domain and a target domain are kept completely consistent. How to apply the domain adaptation method to a wider range of ground object conditions is an urgent problem to be solved. Busto et al added the concept of the open-set scene to the domain adaptation method for the first time in 2017, put forward the open-set domain adaptation method OSDA, and the target domain contains not only the existing class of the source domain but also the unknown class specific to the target domain. In recent years, general domain adaptation (UniDA) has been proposed based on open domain adaptation. You et al in 2019 proposed the concept of universal domain adaptation and proposed a Universal Adaptive Network (UAN). The general domain adaptation defines a scenario in which the source domain and the target domain have a common class. Meanwhile, the two domains respectively have unique categories, the unique categories of the source domain are called source domain private categories, and the unique categories of the target domain are called target domain private categories. The tag set of the source domain data is known and the tag set of the target domain data is unknown. The generic domain adaptation aims at correctly classifying public classes in the target domain while separating target domain private class samples from the data. The existing hyperspectral remote sensing image classification method based on universal domain adaptation faces a plurality of challenges. On the one hand, for input data, the public class data of the target domain and the public class data of the source domain have certain similarity. If a target domain data is very similar to the source domain, it is likely to belong to the public class, otherwise it is likely to belong to the private class of the target domain. During the training process of the model, the features of the two-domain common class extracted by the feature extractor have stronger and stronger similarity. The private class data of the target domain does not participate in the domain adaptation process, and the extracted private class features and public class features have strong differences. Because the target domain has no condition of any prior information, in order to avoid the generation of negative migration, the model only carries out domain adaptation on the common class of the two domains. On the other hand, compared with the traditional unsupervised domain adaptation scene, the source domain data in the general scene has the private class, and the addition of the private class of the source domain can have a certain influence on the characteristic alignment of the public class during doma