Objective Now hyperspectral images have high spatial and spectral resolution, and play an important role in the fields of land monitoring, environmental protection, earthquake prevention, and disaster reduction. However, the high dimensionality and large data volume of hyperspectral data bring many problems (e.g., strong correlation among bands, mixed redundant pixels, and data information redundancy) to hyperspectral classification. With the continuous development of deep learning technology, the convolutional neural network (CNN), as one of its representative algorithms, provides a new solution for hyperspectral image classification. There are three common hyperspectral classification methods based on the CNN network. Among them, 1D CNN extracts spectral information, and 2D CNN extracts spatial information. In contrast, 3D CNN is usually composed of three-dimensional convolution kernels, which can extract two-dimensional spatial features and one-dimensional spectral features at the same time. Although 3D CNN has a better effect in spatial-spectral information fusion, this model is more complex, which increases the cost of network calculation and the number of parameters. With the rapid expansion of data volume, the classification accuracy and speed of a complex model are not satisfactory. Here we propose a lightweight fusion CNN algorithm, 3D-2D-1D CNN, for hyperspectral image classification. This algorithm organically integrates CNNs of different dimensions, reduces the calculation amount of 3D CNN operations, and makes full use of the hyperspectral spatial-spectral joint information. We hope that our basic strategy and findings can be helpful to improve the applicability and computational efficiency of the model.

Methods A hybrid algorithm 3D-2D-1D CNN model (Fig. 1) is described as follows. Firstly, the hyperspectral data is processed by 3D CNN. In the hyperspectral cube, a three-dimensional convolution kernel is used for convolution calculation. Each feature map of the convolutional layer is connected to multiple spectrograms, so the information of the fused spatial and spectral features can be extracted at the same time. After that, 2D CNN is used to extract spatial information, and subsequently the output information is subjected to 1D CNN operations to further extract high-dimensional spectral information. Different feature information is extracted by performing two-dimensional and one-dimensional convolution operations, respectively. Finally, classification is performed according to the extracted feature maps. The algorithm proposed here retains the spatial-spectral joint information extraction of 3D CNN, and replaces part of the subsequent three-dimensional convolution layer process with a 2D convolution process and a 1D convolution process with less calculation. According to the rules of convolution operation, the proposed algorithm improves operation efficiency (Fig. 2).

Results and Discussions The 3D-2D-1D CNN proposed here has the best accuracy among all three algorithms for comparison (Table 3). The classification accuracy of machine learning algorithms is higher than that of general classification algorithms (such as SVM). The classification accuracy of the intermediate output value of the proposed 3D-2D CNN model (new) remains stable, indicating that the fusion algorithm can achieve the complementarity of multi-dimensional convolution features. The 3D-2D-1D CNN model has high classification accuracy on data with large differences in spectral characteristics (Fig. 4), which proves that the model can extract more abstract spectral information. The model has fast convergence, reaching convergence in 20 iterations (Fig. 6). Deep learning algorithms consume more computing resources than the general classification algorithm SVM. Compared with 2D CNN, 3D CNN has a more complex model, so the training time is greatly increased (Table 5). Compared with the existing basic 3D CNN model and the existing improved algorithm 3D-2D CNN, the 3D-2D-1D CNN model has the highest classification accuracy and the fastest calculation speed. As the amount of data in the dataset increases, the speed of the proposed model increases fast, indicating that the model can be applied in large-scale data analysis with a great potential. In general, the model proposed here is the optimal in terms of classification accuracy and speed.

Conclusions In view of the high computational cost of 3D CNN, a 3D-2D-1D CNN fusion model is established. The model can effectively reduce the network parameters and calculation cost, and has a high classification accuracy. Firstly, the network structure of 3D-2D-1D CNN can not only reach or exceed the classification accuracy of 3D CNN under the same parameter settings, but also it can greatly reduce the time cost and calculation cost, which is a more efficient network structure. Moreover, the model has a great data advantage in big data calculation. Second, compared with SVM, 2D CNN and other methods, the fusion model proposed here uses the complementary information of spatial-spectral combination at the same time, so it is the most excellent in overall classification accuracy, average classification accuracy, and Kappa coefficient. Finally, because the proposed model uses 3D CNN hierarchically, the fusion model can extract spatial-spectral features, spatial features, and spectral features in stages. The model improves the ability of 3D CNN to extract the features of hyperspectral data, and at the same time effectively utilizes the speed advantages of 2D CNN and 1D CNN, so it can effectively reduce the calculation cost and maintain a high classification accuracy.