Model Pretraining And Network Regularization For Hyperspectral Image Classification With Small-Sized Labeled Samples

Hyperspectral image(HSI)classification is an important branch in the field of remote sensing for earth observation,and its application is wide and thus highly valuable to study.The problem of small samples is a pressing problem in the field of hyperspectral image classification because of the difficulty of obtaining actual label samples and the poor performance of classification algorithms.Hyperspectral image elements are rich in spectral information due to their hundreds of dimensions,and the spatial information of the surrounding neighborhoods is crucial to improve the classification performance because of the strong correlation between adjacent image elements.At the same time,although there are few HSI labeled samples,there is a large amount of unlabeled data,and how to make full use of the information of these data will become an important element to improve the classification performance of hyperspectral images under small sample conditions.In addition,due to the high information content and complex data of hyperspectral data,only large-capacity deep models are capable of their classification tasks,but in the case of small training samples,the large-capacity models cannot be adequately trained and lead to their poor classification performance.Therefore,this paper proposes a series of hyperspectral image classification methods to solve the existing small-sized labeled samples problem of hyperspectral image classification from the perspectives of pre-training unlabeled samples and reducing model capacity,respectively,as follows.Firstly,to address the problem of small-sized labeled samples,a hyperspectral image classification method based on pre-training of pseudo-labeled samples is proposed to generate reliable pseudo-labels by unsupervised clustering of unlabeled samples,pretraining with pseudo-labeled samples,and then fine-tuning the two-branch fusion network with a small number of labeled samples.In this method,hyperspectral image elements are first spatially clustered,followed by spectral clustering to further reduce the number of clusters,which fully considers the spatial correlation and spectral similarity of the image elements and generates reliable pseudolabels.The pre-training method makes full use of a large amount of unlabeled sample information to enhance the feature extraction capability of the model.Finally,the designed two-branch fusion network makes full use of the pretrained model while compensating for the loss of original information.Experimental analysis shows that this method obtains consistent performance improvement on three hyperspectral datasets,and the classification accuracy is much higher than other algorithms.Secondly,for the problem of small-sized labeled samples,considering that the pre-trained model should try to extract discriminative information for distinguishing different samples,a hyperspectral image classification method based on contrast representation learning pretrained network is proposed to use contrast representation learning to the field of spectral classification.Specifically,the spectral and spatial samples are augmented to form suitable spectral-space sample pairs,respectively,to maximize the feature consistency of positive samples and minimize the consistency of negative sample features using the contrast representation learning method,followed by multi-level fusion of the learned representations from the spectral and spatial branches.The pre-trained network used in this method makes full use of unlabeled samples to improve the performance of the learning system with small samples.The sample pair generation method and the comparison learning paradigm are designed to specify the network to learn more invariant or covariant attributes that are beneficial for category determination.The designed multilevel fusion network is able to take full advantage of the representation of each layer obtained by contrast learning.Comparative and data sensitivity experiments validate that the method has better classification performance and is more robust to sample size on three classical datasets.Finally,to address the problem of small-sized labeled samples,a hyperspectral image classification algorithm based on regular network of spectral-spatial characteristics relative localization is proposed from the perspective of reducing the model capacity,and the subtasks of relative positioning of spectral features and relative positioning of spatial features are designed to improve the network’s ability to extract spatial information.The spatial relationship of features in the domain and the position relationship between spectral band information greatly reduces the model capacity,thus improving the classification accuracy of hyperspectral images under small sample training.In addition,the algorithm performance can be enhanced by sub-tasking with appropriate pre-training using unlabeled sample information.Experimental results on three hyperspectral image datasets demonstrate the superiority of the proposed algorithm over other advanced classification algorithms.