Hybrid 3D Convolution and 2D Depthwise Separable Convolution Neural Network for Hyperspectral Image Classification

Abstract

Convolutional neural networks (CNNs) are one of deep learning methods that are often used to solve the problem of hyperspectral image classification (HSIC). CNN has a strong feature learning ability that can ensure more distinctive features for higher quality HSIC. The traditional CNN-based methods mainly use the 2D CNN for HSIC. However, with 2D CNN, only spatial features are extracted in HSI. Good feature maps cannot be extracted from spectral dimensions with the use of 2D CNN alone. By using 3D CNN, spatial-spectral features are extracted simultaneously. However, 3D CNN is computationally complex. In this study, a hybrid CNN method, which is a combination of 3D CNN and 2D CNN, is improved to solve the two problems described above. Using hybrid CNN decreases the complexity of the method compared to using only 3D CNN and can perform well against a limited number of training samples. On the other hand, in Hybrid CNN, depthwise separable convolution (DSC) is used, which decreases computational cost, prevents overfitting and enables more spatial feature extraction. By adding DSC to the developed hybrid CNN, a hybrid depthwise separable convolutional neural network is obtained. Extensive applications on frequently used HSI benchmark datasets show that the classification performance of the proposed network is better than compared methods.

Multiscale Feature Fusion for Hyperspectral Image Classification Using Hybrid 3D-2D Depthwise Separable Convolution Networks

“…As a result of these applications, the obtained OA figures were 98.69%, 96.50%, and 96.84% for Pavia Center, PU, and SA, respectively. Fırat et al [28] enhanced a hybrid approach aimed at HRSIC, which fuses the potential of 3D CNN and the 2D DSC process. Within their proposed framework, evaluations were executed employing 30, 15, and 15 PCs for IP, PU, and SA, respectively, following a PCA preprocessing step.…”

Hybrid Convolutional Network Combining 3D Depthwise Separable Convolution and Receptive Field Control for Hyperspectral Image Classification

“…The authors used the proposed convolution in the lightweight network Mobilenetv1 [34]. Fırat et al introduced 2D depthwise separable convolution in HSIC tasks to decrease the computational cost [35]. Sandler et al upgraded the depthwise separable convolution and proposed the inverted residual structure [36].…”