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

Abstract

Hyperspectral remote sensing images (HRSI) comprise three-dimensional image cubes, containing a single spectral dimension alongside two spatial dimensions. HRSI are presently among the foremost essential datasets for Earth observation. The task of HRSI classification is intricate due to the influence of spectral mixing, leading to notable variability within classes and resemblances across classes. Consequently, the field of HRSI classification has garnered significant research attention in recent times. Convolutional Neural Networks (CNNs) are harnessed to address these issues, enabling both feature extraction and classification. This study introduces a novel approach for HRSI classification called the hybrid 3D-2D depthwise separable convolution network (Hybrid DSCNet), which leverages multiscale feature integration. Within the Hybrid DSCNet, diverse kernel sizes contribute to an enriched feature extraction process from HRSI. The conventional 3D-2D CNN, while effective, comes with a computational load. Instead of using the standard 3D-2D CNN, this study adopts the 3D-2D DSC architecture. This approach partitions the conventional convolution into two components: pointwise and depthwise convolution, yielding a substantial reduction in trainable parameters and computational complexity. To evaluate the proposed method, the Indian Pines dataset along with WHU-Hi subdatasets (LongKou-LK, HanChuan-HC, and HongHu-HH) were employed. Employing a 5% training sample, impressive overall accuracy scores were achieved: 94.51%, 99.78%, 97.06%, and 97.27% for Indian Pines, WHU-LK, WHU-HC, and WHU-HH, respectively. Comparative analysis of the proposed approach with cutting-edge techniques within the literature reveals its superior performance across the four HRSI datasets. Notably, the Hybrid DSCNet attains enhanced classification accuracy while maintaining lower computational overhead.