Recent research has shown that spatial-spectral information can help to improve the classification of hyperspectral images (HSIs). Therefore, three-dimensional convolutional neural networks (3D-CNNs) have been applied to HSI classification. However, a lack of HSI training samples restricts the performance of 3D-CNNs. To solve this problem and improve the classification, an improved method based on 3D-CNNs combined with parameter optimization, transfer learning, and virtual samples is proposed in this paper. Firstly, to optimize the network performance, the parameters of the 3D-CNN of the HSI to be classified (target data) are adjusted according to the single variable principle. Secondly, in order to relieve the problem caused by insufficient samples, the weights in the bottom layers of the parameter-optimized 3D-CNN of the target data can be transferred from another well trained 3D-CNN by a HSI (source data) with enough samples and the same feature space as the target data. Then, some virtual samples can be generated from the original samples of the target data to further alleviate the lack of HSI training samples. Finally, the parameter-optimized 3D-CNN with transfer learning can be trained by the training samples consisting of the virtual and the original samples. Experimental results on real-world hyperspectral satellite images have shown that the proposed method has great potential prospects in HSI classification.

Target detection is one of the most important applications of hyperspectral technology. However, due to spectral variations caused by noise or environment, the within-class variation is enlarged which degrades the performance of detectors, especially when the target size is small. Therefore, improving the detection performance of small targets and noisy targets is a key task. Considering the great feature extraction and representation ability of deep learning models, denoising autoencoder (DAE) is introduced to reconstruct spectrums and exploit the invariant information for target detection. To fully extract the features from the original spectrums, a multiscale denoising autoencoder (MSDAE) model is designed to incorporate complementary informationin in this paper. The final spectrum is fused by reconstructed spectrums from different scales representations, which provides more complex information and more robust features for subsequent spectral identification. Results on simulated hyperspectral images (HSIs) and real-world HSIs demonstrate that the proposed MSDAE model can effectively remove noise interference and lead to great improvements of the target detection. In addition, the proposed method shows significant potential in preserving small targets.

Deep learning models have been widely applied to extract the high-level features of hyperspectral images (HSIs) due to their strong data mining ability, and some encouraging results have been obtained. Sufficient labelled samples are the guarantee of the network performance. However, the collection of labelled hyperspectral data is time consuming and labour consuming. To get rid of the limitation of labelled samples, an unsupervised feature extractor is proposed based on transfer learning and an improved Wasserstein generative adversarial network (WGAN-GP) in this paper. In addition, considering the spectral and spatial characteristics of HSIs, the framework is designed in three-dimensional (3D) form which helps fully exploit the information in hyperspectral data. Thanks to transfer learning, the well-trained discriminator of WGAN-GP is used as a feature extractor for learning the spectral-spatial features without involving labelled samples. Moreover, in order to facilitate the optimization of WGAN-GP, we propose a novel dimensionality reduction (DR) method using $1\times 1$ convolutions to get lower-dimensional target data. Experimental results on real hyperspectral datasets demonstrate the efficiency of the proposed method combining the proposed DR and WGAN-GP(DR+WGAN-GP), which has great potential prospects in HSI classification.

Deep learning models have strong abilities in learning features and they have been successfully applied in hyperspectral images (HSIs). However, the training of most deep learning models requires labeled samples and the collection of labeled samples are labor-consuming in HSI. In addition, single-level features from a single layer are usually considered, which may result in the loss of some important information. Using multiple networks to obtain multi-level features is a solution, but at the cost of longer training time and computational complexity. To solve these problems, a novel unsupervised multi-level feature extraction framework that is based on a three dimensional convolutional autoencoder (3D-CAE) is proposed in this paper. The designed 3D-CAE is stacked by fully 3D convolutional layers and 3D deconvolutional layers, which allows for the spectral-spatial information of targets to be mined simultaneously. Besides, the 3D-CAE can be trained in an unsupervised way without involving labeled samples. Moreover, the multi-level features are directly obtained from the encoded layers with different scales and resolutions, which is more efficient than using multiple networks to get them. The effectiveness of the proposed multi-level features is verified on two hyperspectral data sets. The results demonstrate that the proposed method has great promise in unsupervised feature learning and can help us to further improve the hyperspectral classification when compared with single-level features.

The extraction of regions of interest (ROIs) plays a vital role in enhancing the precision of target analysis and identification, especially for hyperspectral small targets in wide fields of view. While the use of deep learning for ROI extraction has demonstrated significant potential, its efficacy has been hindered by a scarcity of labeled data. This study proposes a novel ROI extraction method employing a fully convolutional network and channel-spatial attention (CSFCN) for hyperspectral small targets through self-supervised learning. Firstly, a strategy for pseudo-label assignment is designed based on the feature similarity and spatial continuity of hyperspectral images (HSIs). Secondly, an unsupervised segmentation model for HSIs is established, incorporating an attention mechanism based on FCN. Finally, ROIs are extracted based on segmentation results. The experimental results on two real-world HSIs, HSI_a and HSI_b , show that the proposed method can effectively improve the accuracy of extracted ROIs. The overall accuracy (OA) and the intersection over union (IOU) values of the proposed CSFCN reached 25.62%, 44.44% ( HSI_a ) and 44.76%, 100% ( HSI_b ), far exceeding traditional unsupervised segmentation results. The superior experimental results demonstrate that the proposed method has promising application prospects.

Many deep learning-based models have been designed to extract features for hyperspectral classification. The high-dimensional target data in hyperspectral images (HSIs) can be mapped to corresponding low-dimensional features by the deep learning models. However, for most models, the input size for different targets is always same while different targets often perform differently with the same input size, especially when there are small targets. In order to solve this problem and balance different targets, a novel multi-size and multi-model framework based on three-dimensional convolutional autoencoder, called 3D-M²CAE, is proposed in this paper. The establishment and optimization of the framework rely on progressive growing and transfer learning. For each observed target, three 3D-CAEs with progressive input sizes and layers are used to obtain features, which can improve the feature robustness to spatial variation. In order to reduce the amount of calculation and save training time, three three-dimensional convolutional autoencoders (3D-CAEs) are constructed and trained in a progressive way with the help of transfer learning. The weights of the middle layers of the latter 3D-CAE are transferred from the former optimized 3D-CAE, which speeds up and facilitates network training. Benefiting from this training method, the features of the same target from different sizes are obtained in a more efficient way. Moreover, the proposed framework has great robustness to different targets and the obtained features can provide more information for subsequent processing. The effectiveness of the proposed method is verified on two hyperspectral data sets. The results demonstrate that the proposed method has great potential in improving the classification results of small targets.