Research On Key Algorithms Of Small Sample Hyperspectral Image Classification Based On Deep Learning

Optical remote sensing images acquire surface feature information of the Earth by recording changes in electromagnetic radiation wavelength in the visible and infrared bands.Due to the differences in the reflection and absorption of electromagnetic waves by surface materials,hyperspectral images containing rich spectral spatial information can be acquired using imaging spectrometers.Hyperspectral remote sensing is widely used in various fields,including land cover mapping,detection,and agriculture,as the spectral-spatial information in hyperspectral images reflects the properties of ground materials.Hyperspectral image classification is a crucial step in remote sensing image analysis,where each pixel in the acquired image is assigned a class label.Labeling pixels in hyperspectral images is a time-consuming and laborious task.Therefore,achieving hyperspectral image classification with a small number of labeled samples has become a hot research topic.Existing methods for addressing the problem of small-sample hyperspectral image classification suffer from issues such as insufficient feature extraction from labeled samples,inadequate use of unlabeled samples,and poor generalization.As a result,there is a significant gap between experimental accuracy and practical demand.This thesis focuses on optimizing feature extraction methods and fully utilizing unlabeled samples to address these issues.The main research contents are summarized as follows:(1)To address the problem of poor classification accuracy of hyperspectral images in small sample scenes,a hyperspectral image classification method based on self-supervised learning and adaptive distillation is proposed by combining soft labels with self-supervised learning.The proposed soft label generation algorithm is used to generate soft labels for unlabeled samples of the dataset,and the full utilization of unlabeled samples is achieved.The discriminative information contained in the training samples is fully extracted by a self-supervised auxiliary task based on spectral space transformation.In addition,the concept of soft label quality is proposed for the first time in hyperspectral image classification,and a simple metric is proposed.The experimental results show that the proposed method achieves better classification results in small sample scenes compared with traditional classification methods and some advanced deep learning classification methods.(2)To address the issues of weak transferable information extraction from the source domain dataset and insufficient consideration of unlabeled samples in the target domain in cross-domain hyperspectral image classification,we propose a cross-domain hyperspectral image classification method based on few-shot learning and self-supervised learning.Specifically,we create a joint task of self-supervised and few-shot learning in the source domain to extract transferable knowledge.In the target domain,we propose a new adaptive soft label generation method to fully utilize the discriminative information contained in the unlabeled samples given the limited number of labeled samples.To enhance the network representation ability,we design a residual contraction network as a feature extractor and incorporate unsupervised domain adaptation methods to alleviate the domain shift problem during training.Experimental results demonstrate that the proposed classification method effectively solves the problem of insufficient transferable knowledge extraction in existing methods and achieves excellent classification accuracy with extremely limited samples by combining the unlabeled samples in the target domain.(3)To address the problems of poor metric space discrimination and poor generalization of existing soft label generation algorithms in the process of small-sample hyperspectral image classification,a cross-domain hyperspectral image classification method based on knowledge distillation and few-shot learning is proposed.Specifically,pre-training is first performed on the source domain dataset,and then the pre-training model is used to train the initialization parameters for the subsequent model.In each iteration of training,soft labels are generated for unlabeled samples in the target domain using the trained model,and a joint loss of representation learning and few-shot learning is constructed in the target domain.The method of generating soft labels for unlabeled samples through the previous iteration of network generation overcomes the limitations of hand-designed algorithms.In addition,the inclusion of representation learning enhances the ability of the metric space to identify the variability among different samples.Experimental results show that the proposed method effectively improves the discriminative ability of the metric space and has an accuracy advantage over the current state-of-the-art classification methods.