Search

Objectives: The quality of network structure is the key to obtain excellent classification performance. Most of existing deep models are designed by the trial-and-error method. When the same model processes different hyperspectral images (HSIs), different network hyperparameters need to be set artificially to adapt to the data characteristics of different images, greatly increasing the difficulty of network design. Aiming at the problem of deep learning model design in hyperspectral image classification, this paper proposes an automatic network structure search method for HSI classification. Methods: In the proposed method, the differentiable structure search technology is first used to search the network structure on the source HSI data sets, then the deep network model is constructed by stacking cells, and finally the classification performance is evaluated on the target HSIs. It should be pointed out that, the proposed method only conducts one search on the source HSI data sets, and the obtained deep model can be applied to other target HSI classification tasks, which can effectively improve the utilization rate of the model. To improve the generalization ability and classification accuracy of the model, multi-source and multi-resolution HSIs are employed to construct the precollected data set, and the partial channel connection operation is introduced to further improve the search efficiency. Results: Three HSIs inculding Houston, Chikusei and Pavia Center are selected as source data. Four HSIs including University of Pavia, Indian Pines, Salinas and Houston 2018 are selected as the target HSIs. The source data is used to determine the network model and the target HSIs are used to evaluate the classification performance. Experimental results show that the proposed method can automatically search out a universal deep network model suitable for HSI classification task, and this model can achieve better classification performance than the conventional deep learning models. The overall classification accuracies of University of Pavia, Indian Pines, Salinas and Houston 2018 are 98.15%, 98.74%, 97.30% and 74.47%, respectively. Conclusions: Compared with other deep models, the model in this paper is automatically determined by differentiable architecture search algorithm according to the data characteristics of HSIs, effectively improving the classification accuracy of target HSIs while avoiding complex network design. Extensive experiments verify the applicability and effectiveness of the proposed method in HSI classification. [ABSTRACT FROM AUTHOR]

Objectives: The deep learning model such as convolutional neural network(CNN) has achieved satisfactory results in hyperspectral images classification. However, because traditional neurons can only perform scalar computation, the existing deep learning models cannot model the instantiation parameters of hyperspectral image features, so they cannot achieve satisfactory classification results under the condition of restricted neighborhood scope. Aiming at the problem, we design a new network model by introducing capsule network structure, to further improve the classification accuracy. Methods: The model is composed of traditional convolution layers, capsule layers and fully connected layers, which has a stronger feature representation ability. This model can further improve the classification accuracy of hyperspectral images using the vector calculation of capsule neurons and the spatial relationship between features encoded by the weight matrix. Specifically, the hyperspectral image patches are firstly processed by a convolution layer to extract local features. Next, the primary capsule layer and the digital capsule layer are used to extract the deeper abstract features at higher levels and classify the input data. In addition, the fully connected layers are used for reconstruction to further enhance the abstract modeling ability and generalization ability of the capsule network. Results: Three public hyperspectral images data sets including Pavia University, Indian Pines and Salinas are selected for experiments. The results show that the proposed method outperform the support vector machine(SVM)⁃based and the traditional CNN-based classification methods. Specifically, compared with the SVM-based methods, the proposed method improves the overall classification accuracy by about 4.7%—7.2%, 8.2%—10.9% and 2.5%—6.9% on three different datasets. Compared with the traditional CNN-based methods, the proposed method improves the overall classification accuracy by about 0.8—3.7%, 2.7—5.5% and 1.3—2.5% on three different datasets. Conclusions: In conclusion, the pro⁃ posed network model has better classification performance than that of traditional algorithms. Under the condition of sufficient training samples, the overall classification accuracy of the proposed model is higher than the traditional SVM⁃based and CNN⁃based classification models, and it has a lower time cost. In additional, the proposed model has better adaptability under the condition of further reducing training samples and pixel neighborhoods. [ABSTRACT FROM AUTHOR]

Objectives: In recent years, hyperspectral images classification based on deep learning has made important progress. In view of the scarcity of training samples for hyperspectral image classification, this pa‐ per proposes a lightweight attention depth-wise relation network (LWAD-RN) to solve the problem of small sample hyperspectral image classification. Methods: The network consists of an embedding layer and a relation layer. In the embedding layer, a lightweight convolutional neural network combining attention mechanism is used to extract pixel features, and a dense network structure is introduced. The relation value is calculated in the relation layer for classification, and the task-based mode is used to train the network. Three groups of public hyperspectral image datasets are used to implement experiments. Results and Conclusions: The results show that LWAD-RN can effectively improve the classification accuracy under the condition of small samples (5 training samples per category), and the efficiency of model training and classification is improved. The proposed LWAD-RN can obtain ideal classification accuracy under the condition of small samples, and the lightweight network structure can improve the model training and classification efficiency. However, under the condition of small samples, the quality of training samples will have an important impact on the performance of the model. Therefore, follow-up studies should be conducted on how to select high-quality training samples more accurately and efficiently to ensure the stability of the model and better meet the needs of practical application. [ABSTRACT FROM AUTHOR]