Multipath feature fusion for hyperspectral image classification based on hybrid 3D/2D CNN and squeeze-excitation network

Hyperspectral Images (HSI) are commonly used for classification thanks to their rich spectral feature information along with their spatial feature information. Convolutional Neural Network (CNN) based deep learning methods are commonly used in HSI classification (HSIC) applications to process the high nonlinearity and high dimensionality of HSI. This study proposes a method consisting of a combination of multipath Hybrid CNN and a Squeeze and Excitation (SE) network for HSIC. Features extracted with different kernel sizes in the multipath method are used together to extract richer feature information from HSI in this proposed method (PM). In the Hybrid CNN used in PM, 3D CNN was used to extract the spectral-spatial features. However, computational complexity increases with 3D CNN. Computational complexity is decreased with the use of Hybrid CNN. In addition, 2D CNN used in Hybrid CNN provides more spatial feature information to be extracted. However, in this study, 2D depthwise separable convolution (DSC) was used instead of 2D CNN. By using 2D DSC instead of standard 2D CNN, computational cost and the number of trainable parameters is significantly decreased. Finally, the PM is combined with the SE network to advance the HSIC accuracies. The SE network is designed to enhance the representation quality of CNN. WHU-Hi-HongHu (WHHH), WHU-Hi-HanChuan (WHHC), and WHU-Hi-LongKou (WHLK) datasets were used to evaluate the classification accuracies of the PM. Using a 5% training sample with WHLK, WHHC and WHHH, OA values ​​of 99.86%, 97.51% and 97.64% were obtained. Furthermore, the PM was compared with the latest technology methods in the literature and outperformed all methods.