Cross-Scene Hyperspectral Image Classification Based On Domain Adaptation

Hyperspectral image classification is a technique that uses the spectral and spatial information of hyperspectral images for recognition and classification.It is one of the important techniques in remote sensing.Traditional classification methods rely heavily on labeled samples for model training,and it is difficult to achieve accurate classification when labeled samples are unavailable or when the data distribution of labeled and unlabeled samples is inconsistent.Data distribution inconsistency is a common problem due to differences in collection sensors,collection areas,and time between source and target scenes.To address these issues,this thesis combines deep learning and domain adaptation technology to study how to improve the accuracy of cross-scene hyperspectral image classification,and the main research content and innovation points can be summarized as follows:1.To address the problem of data distribution bias between cross-scene hyperspectral images,a cross-scene hyperspectral image classification method based on discriminative feature norm network is proposed.First,in order to fully utilize the spatial-spectral information of hyperspectral images,a 3D residual convolutional neural network is used to extract deep spatial-spectral features.Then,starting from the perspective that the smaller feature norm value of the target domain than that of the source domain will cause an unstable discriminative angle,the feature norms of the two domains are gradually expanded to a constrained value to obtain domain-invariant features.Then,the discriminative features are obtained by reducing the distance between features and class centers and reducing the intra-class distance.Finally,by separately reducing the distances between the class centroids and the overall centroid in the two domains,class-level distribution adaptation is carried out to further reduce domain discrepancy.2.To address the problem that existing methods based on convolutional neural networks cannot fully exploit the sequence attributes of spectral features,we propose a hyperspectral image classification method based on adversarial transfer Transformer.First,the data of the two domains are linearly projected and positionally encoded from the spectral dimension into groups,and the data after position encoding is sent to the Transformer encoder to obtain discriminative features.Then,to make the model focus on transferable features,the features are fed into an adversarial transferability adaptation module to increase the weight of transferable features.Then,the target domain features are pushed away from the decision boundary and pushed toward the source domain using a dual classifier adversarial strategy to adapt the features of the two domains.At the same time,the maximum batch Nuclear-norm is used on the output matrix to ensure the diversity of the model predictions.Finally,the model is used to classify the target domain data.Experiments and analysis on the Houston,Hy RANK,and Botswana datasets validate the effectiveness of the two methods proposed in this thesis for cross-scene hyperspectral image classification tasks.This thesis has 25 figures,14 tables and 107 references.