In recent years, deep-learning-based hyperspectral image (HSI) processing and analysis have made significant progress. However, models with high performance require sufficient training samples because scarce labeled samples limit their generalization ability. To solve this problem, we adopt a self-supervised learning strategy and conduct self-training for a neural network model by obtaining different views of the same sample (positive pairs). As a result, the network can learn representative features for classification from unlabeled samples. In addition, to increase the spatial receptive field compared with the use of conventional convolutions, we use the transformer to capture long-distance dependencies for feature enhancement and adequately combine their advantages. Experimental results on two publicly available HSI datasets demonstrate that the proposed method can extract robust features through self-training on unlabeled samples and can be adapted to HSI classification tasks under the small sample conditions.
Recently, convolutional neural networks have greatly improved the performance of hyperspectral image (HSI) classification. However, these methods mainly use local spatial–spectral information for the HSI classification and require a large number of labeled samples to ensure high classification accuracy. In our study, we propose a multiscale nested U-Net (MsNU-Net) to capture the global context information and improve the HSI classification accuracy with a small number of labeled samples. We took an HSI as the input and constructed a nested U-Net to complete the classification. Because scale is very important for image recognition, we propose a simple but effective multiscale loss function. Apart from introducing multiscale features into the network, this method uses Gaussian filters to construct multiscale data, inputs the multiscale data into the nested U-Net with shared parameters, and calculates the sum of loss functions of different scales as the final loss function. Furthermore, it introduces different scales of global context information, thus improving the classification accuracy. To demonstrate its effectiveness, we carried out classification experiments on four widely used HSIs. The results show that this method could achieve a higher classification accuracy than the compared methods when only a small number of labeled samples is available. Furthermore, the codes of the proposed method will be made available freely at https://github.com/liubing220524/MsNU-Net.
Recently, deep learning models based on convolutional neural networks (CNN) remain dominant in hyperspectral image (HSI) classification. However, there are some problems in CNN models, such as not good at modeling the long-distance dependencies and obtaining global context information. Different from the existing CNN-based models, an innovative classification method based on the transformer model is proposed to further improve the classification accuracy of HSI. Specifically, the proposed method first extracts the extended morphological profile (EMP) features of HSI to make full use of the spatial and spectral information while effectively reducing the number of bands. Next, a deep network model is constructed by introducing the transformer-iN-transformer (TNT) modules to carry out end-to-end classification. The outer and inner transformer models in the TNT module can extract the patch-level and pixel-level features, respectively, to make full use of the global and local information in the input EMP cubes. Experimental results on three public HSI data sets show that the proposed method can achieve better classification performance than the existing CNN-based models. In addition, using the transformer-based deep model without convolution to classify HSI provides a new idea for related research.
Deep learning has been widely used in hyperspectral image (HSI) classification. However, a deep learning model is a data-driven machine learning method, and collecting labeled data is quite time-consuming for an HSI classification task, which means that a deep learning model needs a lot of labeled data and cannot deal with the small sample problem. We explore the small sample classification problem of HSI with graph convolutional network (GCN). First, HSI with a small number of labeled samples are treated as a graph. Then, the GCN (an efficient variant of convolutional neural networks) operates directly on the graph constructed from the HSI. GCN utilizes the adjacency nodes in graph to approximate the convolution. In other words, graph convolution can use both labeled and unlabeled nodes. Therefore, our method is a semisupervised method. Three HSI are used to assess the performance of the proposed method. The experimental results show that the proposed method outperforms the traditional semisupervised methods and advanced deep learning methods.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.