A Study Of Few-shot Hyperspectral Image Classification Based On Deep Learning

Hyperspectral images are rich in spectral and spatial information and have the potential to provide more accurate identification of feature types of interest than traditional color images;they have become an important tool for Earth observation and even for space exploration.As a very important and core research topic in the field of hyperspectral research,hyperspectral image classification has been receiving great attention and research,and has been widely used in the related fields of national economy.At present,the task of hyperspectral remote sensing image classification still faces some difficulties and challenges: how to further explore the inherent deep features of hyperspectral images;how to solve the problem of lack of training samples and unsatisfactory performance of high-dimensional classification in small samples;how to face the problem that in practical applications,for the "closed set",the unknown class is forced to be assigned to the In practical applications,the unknown classes are forced to be assigned to the existing labeled values,resulting in the overestimation of the labeled values of the known classes.In order to solve these problems,this paper analyzes and discusses the characteristics of hyperspectral data itself,and rethinks the current hyperspectral image processing techniques and methods for dealing with the traditional image small sample problem.1.To explore more accurate classification of hyperspectral images with a small number of labeled samples,this paper proposes a graph neural network-based smallsample classification algorithm to solve the problem of hyperspectral image classification.The feature encoder is used to extract the spatial-spectral characteristics of hyperspectral,and the hyperspectral feature extraction network structure is designed according to the characteristics of small sample data;while the metric space is used to calculate the correlation between the support set and the query set,and the graph neural network is introduced in the metric space to calculate the correlation,which extends the traditional analysis in the Euclidean space to the non-Euclidean space to obtain a better correlation judgment.2.In order to solve the problem of "unknown classes",a new method is proposed to classify unknown classes in open sets and by reconstruction recognition at the same time.A feature extraction network is designed for the classification of unknown classes with small hyperspectral samples and a reconstruction network that fuses the reconstructed features from different reconstruction stages after a three-dimensional densely connected network,and finally two methods are used to classify the two cases of few samples and enough samples,namely,global mode and category mode.The results show that compared with other small sample methods based on Euclidean space,the algorithm proposed in this paper can achieve better classification results with better generalization ability for the same small number of labeled samples;in the classification recognition of unknown classes,it can identify the unknown classes of hyperspectral images more efficiently,and the robustness and generalization ability are also improved.