Deep Learning Based Hyperspectral Image Classification In Few-shot Scenes

Hyperspectral image classification is an important problem to be solved when hyperspectral technology is applied to geological mapping,environmental monitoring,vegetation analysis,atmospheric characterization,biochemical detection and other fields.Because the pixel level annotation of hyperspectral images consumes high labor cost and time cost,it is significant to study how to classify hyperspectral images in few-shot scenes for the promotion of hyperspectral technology.At present,the hyperspectral image classification based on deep learning in few-shot scenes has some problems,such as low utilization of sample information and poor generalization of feature extraction model,which leads to the hyperspectral image classification in few-shot scenes far from reaching the standards that can be applied to practice.To solve the above problems,this thesis studies hyperspectral image classification based on deep learning in few-shot scenes from the perspective of optimizing feature extraction model and labeled samples.The main research work is summarized as follows:(1)To solve the problem of poor performance when the number of labels in the deep learning method is limited,a few-shot hyperspectral image classification algorithm based on three-dimensional convolution siamese network is presented,which combines few-shot learning with semi-supervised learning.The algorithm uses siamese network structure to combine contrastive information with label information to further mine information from hyperspectral image data in samples-limited scenarios.Relevant experiments show that the algorithm can be jointly trained and achieve higher classification accuracy for hyperspectral images in few-shot scenes.It only focuses on the information carried by the labeled samples,and lacks the ability to extract the information of unlabeled samples,so it is suitable for scenarios with high quality labeled samples.(2)To overcome the shortage of information sources for hyperspectral image classification in few-shot scenes,a few-shot hyperspectral image classification algorithm based on threedimensional convolution is presented,which combines deep learning classification method with auto-encoder structure.The algorithm combines convolution structures with an autoencoder to perform semi-supervised learning using both labeled and unlabeled data,and to classify hyperspectral images in ferw-shot scenes.The experimental results show that the classification algorithm can make use of unlabeled samples and labeled samples,being less dependent on labeled samples.(3)To solve the problem that quality of labeled samples seriously affects the performance of deep learning model in few-shot scenes,a few-shot hyperspectral image classification algorithm based on active learning model is proposed from the perspective of optimizing the quality of labeled samples.The algorithm combines semi-supervised clustering with active learning,gathers representative features from a small number of samples,searches for the samples worth labeling,and establishes a high-quality labeled samples pool.The experimental results show that the accuracy and robustness of the algorithm for hyperspectral image classification in fewshot scenes are significantly better than those of the baseline methods because it can find highquality samples in complex scenes.(4)Aiming at the problems that unsupervised contrastive learning lacks suitable data augmentation and high computational resource requirements in the field of hyperspectral images,a hyperspectral feature learning algorithm based on auto-encoder and prototypical contrastive learning is presented,which combines generative learning with discriminative learning.The algorithm uses two auto-encoders to preprocess hyperspectral image data,and uses two set of transformed features to perform prototypical contrastive learning,which can extract information from all samples without supervision.Experiments show that the feature learning algorithm can extract useful hyperspectral image features for classification in an unsupervised way.