Convolutional neural networks (CNN) have led to a successful breakthrough for hyperspectral image classification (HSIC). Due to the intrinsic spatial-spectral specificities of a hyperspectral cube, feature extraction with 3D convolution operation is a straightforward way for HSIC. However, the overwhelming features obtained from the original 3D CNN network suffers from the overfitting and more training cost problem. To address this issue, in this paper, a novel HSIC framework based on a simplified 2D-3D CNN is implemented by the cooperation between a 2D CNN and a 3D convolution layer. Firstly, the 2D convolution block aims to extract the spatial features abundantly involved spectral information as a training channel. Then, the 3D CNN approach primarily concentrates on exploiting band co-relation data by using a reduced kernel. The proposed architecture achieves the spatial and spectral features simultaneously based on a joint 2D-3D pattern to achieve superior fused feature for the subsequent classification. Furthermore, a deconvolution layer intends to enhance the robustness of the deep features is utilized in the proposed CNN network. The results and analysis of extensive real HSIC experiments demonstrate that the proposed light-weighted 2D-3D CNN network can effectively extract refined features and improve the classification accuracy.

Hyperspectral image classification (HSIC) methods based on convolutional neural network (CNN) continue to progress in recent years. However, high complexity, information redundancy, and inefficient description still are the main barriers to the current HSIC networks. To address the mentioned problems, we present a spatial-spectral dense CNN framework with a feedback attention mechanism called FADCNN for HSIC in this article. The proposed architecture assembles the spectral-spatial feature in a compact connection style to extract sufficient information independently with two separate dense CNN networks. Specifically, the feedback attention modules are developed for the first time to enhance the attention map with the semantic knowledge from the high-level layer of the dense model, and we strengthen the spatial attention module by considering multiscale spatial information. To further improve the computation efficiency and the discrimination of the feature representation, the band attention module is designed to emphasize the weight of the bands that participated in the classification training. Besides, the spatial-spectral features are integrated and mined intensely for better refinement in the feature mining network. The extensive experimental results on real hyperspectral images (HSI) demonstrate that the proposed FADCNN architecture has significant advantages compared with other state-of-the-art methods.

Wealthy spectral information provided by hyperspectral image (HSI) offers great benefits for many applications in hyperspectral data exploitation. However, processing such high-dimensional data volumes that may result in redundant bands due to its high interband correlation will be a challenge. For target detection and classification, this is particularly true since there may only need a relatively small number of bands that respond one particular target of interest well, while most of other bands do not. Band selection (BS) is a major dimensionality reduction technique to remove the redundant bands and selects a few bands to represent the entire image. However, how to eliminate the effect of uninteresting targets with similar spectra on detection of interesting targets is a severe issue arising in target detection for BS. This article develops a new approach called target-constrained interference-minimized BS (TCIMBS) which can be used to select band subset for specific target detection, while annihilating targets of no interest and suppressing interferers and background. Its idea is derived from target-constrained interference-minimized filter (TCIMF). By taking advantage of TCIMF, two band prioritization (BP) criteria called forward minimum variance BP (FMinV-BP) and backward maximum variance BP (BMaxV-BP) along with their three band search-based BS counterparts called sequential forward TCIMBS (SF-TCIMBS), sequential backward TCIMBS (SB-TCIMBS), and improved SB-TCIMBS (SB-TCIMBS*) are derived. The experimental results suggest that TCIMBS can improve the detection accuracy and also achieve better performance in comparison with several state-of-the-art methods.

Deep convolutional neural networks (CNN) have led to a successful breakthrough for hyperspectral image (HSI) classification. In this paper, a CNN system embedded with an extracted hashing feature is proposed for HSI classification that utilizes the semantic information of the HSI. First, a series of hash functions are constructed to enhance the presentation of the locality and discriminability of classes. Then, the sparse binary hash codes calculated by the discriminative learning algorithm are combined into the original HSI. Next, we design a CNN framework with seven hidden layers to obtain the hierarchical feature maps with both spectral and spatial information for classification. A deconvolution layer aims to improve the robustness of the proposed CNN network and is used to enhance the expression of deep features. The proposed CNN classification architecture achieves powerful distinguishing ability from different classes. The extensive experiments on real hyperspectral images results demonstrate that the proposed CNN network can effectively improve the classification accuracy after the embedding of the extracted semantic features.

When predicting data for which limited supervised information is available, hyperspectral target detection methods based on deep transfer learning expect that the network will not require considerable retraining to generalize to unfamiliar application contexts. Meta-learning is an effective and practical framework for solving this problem in deep learning. This article proposes a new meta-learning based hyperspectral target detection using Siamese network (MLSN). First, a deep residual convolution feature embedding module is designed to embed spectral vectors into the Euclidean feature space. Then, the triplet loss is used to learn the intraclass similarity and interclass dissimilarity between spectra in embedding feature space by using the known labeled source data on the designed three-channel Siamese network for meta-training. The learned meta-knowledge is updated with the prior target spectrum through a designed two-channel Siamese network to quickly adapt to the new detection task. It should be noted that the parameters and structure of the deep residual convolution embedding modules of each channel in the Siamese network are identical. Finally, the spatial information is combined, and the detection map of the two-channel Siamese network is processed by the guiding image filtering and morphological closing operation, and a final detection result is obtained. Based on the experimental analysis of six real hyperspectral image datasets, the proposed MLSN has shown its excellent comprehensive performance.

Oriented to adaptive recognition of the new land-cover categories, incremental classification (IC) that aims to complete adaptive classification with continuous learning is urgent and crucial for hyperspectral image classification (HSIC). Nevertheless, deep-learning-based HSIC models adopted the learning paradigm with fixed classes yield unsatisfactory inference in the situation of IC due to the catastrophic forgetting problem. To eliminate the recognition gap and maintain the old knowledge during IC, in this paper, we propose a novel approach called the distillation-constrained prototype representation network (DCPRN) for hyperspectral image incremental classification (HSIIC). The primary goal of DCPRN is to enhance the discriminative capability for recognizing the original classes in HSIIC, while effectively integrating both the original and incremental knowledge to facilitate adaptive learning. Specifically, the proposed framework incorporates a prototype representation mechanism, which serves as a bridge for knowledge transfer and integration between the initial and incremental learning phases of HSIIC. Additionally, we present a dual knowledge distillation module in incremental learning, which integrates discriminative information at both the feature and decision level. In this way, the proposed mechanism enables flexible and dynamic adaptation to new classes and overcomes the limitations of fixed-category feature learning. Extensive experimental analysis conducted on three popular data sets validates the superiority of the proposed DCPRN method compared with other typical HSIIC approaches.