Hyperspectral Image Cross-Domain Few-Shot Classification Based On Meta-Learning

Hyperspectral image contains rich spectral information and wide spatial coverage,which plays a great role in urban development,environmental monitoring,agricultural management and mineral exploration.Hyperspectral image classification has also become a popular research direction in the field of remote sensing.Deep learning can obtain more satisfactory classification accuracy with sufficient training annotation samples.However,obtaining high-quality labeled samples is a time-consuming and resource-intensive task.Therefore,hyperspectral image with sufficient labeled samples(source domain)can be used to classify hyperspectral image with fewer labeled samples(target domain).Few-Shot learning based on meta-learning is able to learn general general meta-knowledge from the source domain and classify new classes in the target domain.In this thesis,we use meta-learning as the basis for cross-domain few-shot classification combined with domain adaptation,and the main research work is summarized as follows:1.A cross-domain few-shot classification method based on contrastive learning is proposed for hyperspectral image with high inter-class similarity and large intra-class variance due to the differences in domain distribution and hyperspectral image.Firstly,the spatial-spectral features of hyperspectral samples are extracted by using a deep residual three-dimensional convolutional network for the hyperspectral “combining image with spectrum” feature.Secondly,to solve the problem of the degradation in meta-learning generalization performance due to inter-domain distribution differences,a kernel parametrization is introduced in the category classifier to align two-domain features.Meanwhile,for the characteristics of high inter-class similarity and large intraclass variance of hyperspectral image,contrastive learning is used to map similar hyperspectral samples to adjacent spatial regions,enabling the method to better distinguish between different classes of data.2.A cross-domain few-shot classification method based on disentangle confidence prototype network to solve the problem of the difference in domain distribution and few labeled samples of hyperspectral image.Firstly,the spatial-spectral features of hyperspectral samples are extracted using a deep residual 3D convolutional network.Secondly,domain-invariant features are aligned as much as possible using a domain classifier so that the method is not influenced by domain-specific features.Meanwhile,the domain invariant features are fed into the confidence prototype network to filter the samples with high confidence and assign weights to alleviate the unreliability of the category prototype due to the few labeled samples.Experiments are conducted on the Chikusei,University of Pavia,Pavia Center,Salinas,and Indian Pines datasets,and the experimental results show that the proposed methods in this thesis can achieve good classification results with the few labeled samples,and the classification accuracy is improved in all cases.There are 30 figures,14 tables,and 86 references in this thesis.