3D residual spatial-spectral convolution network for hyperspectral remote sensing image classification

Abstract

Hyperspectral remote sensing images (HRSI) are 3D image cubes that contain hundreds of spectral bands and have two spatial dimensions and one spectral dimension. HRSI analysis are commonly used in a wide variety of applications such as object detection, precision agriculture and mining. HRSI classification purposes to assign each pixel in HRSI to a unique class. Deep learning is seen as an effective method to improve HRSI classification. In particular, convolutional neural networks (CNNs) are increasingly used in remote sensing field. In this study, a hybrid 3D residual spatial-spectral convolution network (3D-RSSCN) is proposed to extract deep spatiospectral features using 3D CNN and ResNet18 architecture. Simultaneously spatiospectral features extraction is provided using 3D CNN. In deeper CNNs, ResNet architecture is used to achieve higher classification performance as the number of layers increases. In addition, thanks to the ResNet architecture, problems such as degradation and vanishing gradient that may occur in deep networks are overcome. The high dimensionality of the HRSIs increases the computational complexity. Thus, most of studies apply dimension reduction as preprocessing. In the proposed study, principal component analysis (PCA) is used as the preprocessing step for optimum spectral band extraction. The proposed 3D-RSSCN method is tested with Indian pines, Pavia University and Salinas datasets and compared against various deep learning-based methods (SAE, RPNet, 2D CNN, 3D CNN, M3D CNN, HybridSN, FC3D CNN, SSRN, FuSENet, S3EResBoF). As a result of the applications, the best classification accuracy among these methods compared in all datasets is obtained with the proposed 3D-RSSCN. The proposed 3D-RSSCN method has the best accuracy and time performance in classifying.