Hyperspectral image classification based on optimized convolutional neural networks with 3D stacked blocks

3D convolution can fully utilize the spectral-spatial characteristics of hyperspectral image (HSI), and stacked blocks with deep layers are capable of extracting hidden features and utilizing discriminant information for classification. Naturally, a 3D convolutional neural network (CNN) based on stacked blocks named SB-3D-CNN is presented for HSI classification. Moreover, the proposed network introduces the attention mechanism before the fully connected layer, which can filter out interfering information effectively. Then we optimized the architecture to obtain optimal results on three commonly used datasets of Indian Pines, Salinas and Pavia University. Experimental results demonstrate that the optimized architecture achieves better classification rates than related recent works. Because the classification accuracies on the three datasets have reached saturation, we transferred the optimized architecture to a more complex dataset adopting the airborne hyperspectral data, which obtains from Guangxi province in south China. The results show that the optimized architecture achieves superior classification accuracies compared with other state-of-the-art methods. These results also demonstrate the optimized SB-3D-CNN has the advantages of validity and portability to more complex data.

Automated summary

In this study, a 3D CNN network based on stacked blocks was proposed for HSI classification. The proposed network includes an attention mechanism to filter out interfering information. The optimized architecture achieved higher classification rates compared to related works and demonstrated effectiveness and adaptability on a more complex dataset.