A Multichannel Hybrid 2D–3D-CNN for Hyperspectral Image Classification With Small Training Sample Sizes
Shih-Yu Chen, Po-Yu Chu, Kuan-Liang Liu, Yucheng Wu
Abstract
This article presents an innovative multichannel hybrid 2D–3D-convolutional neural network (MH-2D-3D-CNN) model specifically designed for the challenging task of hyperspectral image classification (HSIC) with small labeled training sample sizes (SLTSSs). By integrating three-channel network architectures and incorporating target detection and attention mechanisms with 3D- and 2D-CNN techniques, our model emphasizes the extraction and fusion of spectral and spatial features, essential for accurate classification. In the first channel, we employ the normalized adaptive matched detector (NAMD) to enhance spectral features, enabling the extraction of intricate spectral information. Concurrently, 2D-CNN is applied to capture spatial features, enhancing the understanding of spatial structural information within the images. In the second channel, the convolutional block attention module (CBAM) is employed alongside 2D-CNN to further enhance the saliency of spatial features. The third channel adopts a concatenated approach, combining 3D-CNN and 2D-CNN to extract spectral and spatial features simultaneously, reinforcing spatial feature extraction capabilities. The merged feature maps from the three channels undergo final classification through fully connected layers, enhancing the model’s stability in feature extraction and learning performance. Experimental validation was conducted on three widely used public datasets (Indian Pines [IP], University of Pavia [PU], and Salinas [SA]) with a fixed 1% labeled training sample size. The results demonstrate remarkable overall accuracy (OA) scores of 87.49%, 98.13%, and 99.22%. These results outperform state-of-the-art methods and showcase significant advancements, particularly in SLTSS scenarios.