Hyperspectral image classification method based on squeeze-and-excitation networks, depthwise separable convolution and multibranch feature fusion

Hyperspectral image classification (HSIC) is a hot topic discussed by most researchers. In recent years, deep learning and especially CNN have provided very good results in HSIC. However, there is still a need to develop new deep learning-based methods for HSIC. In this study, a new CNN-based method is proposed to reduce the number of trainable parameters and increase HSIC accuracy. The proposed method consists of 3 branches. Squeeze-and-excitation network (SENet) in the first branch, a hybrid method consisting of the combination of 3D CNN and 2D DSC in the second branch, and 2D DSC in the third branch are used. The main purpose of using a multi-branch network structure is to further enrich the features extracted from HSI. SENet used in the first branch are integrated into the proposed method as they increase the classification performance while minimally increasing the total number of parameters. In the second and third branches, hybrid CNN methods consisting of 3D CNN and 2D Depthwise separable convolution were used. With the hybrid CNN, the number of trainable parameters is reduced and the classification performance is increased. In order to analyze the classification performance of the proposed method, applications were carried out on the WHU-Hi-HanChuan, WHU-Hi-LongKou and Indian pines datasets. As a result of the applications, 97.45%, 99.84% and 96.31% overall accuracy values were obtained, respectively. In addition, the proposed method was compared with nine different methods developed in recent years from the literature and it was seen that it obtained the best classification result.