Hyperspectral Image Classification Based on Novel Hybridization of Spatial-Spectral-Superpixelwise Principal Component Analysis and Dense 2D-3D Convolutional Neural Network Fusion Architecture

Abstract We propose a hybridized technique named Spatial-Spectral-Superpixelwise PCA-based Dense 2D-3D CNN Fusion Architecture (3SPCA-D-2D-3D-CNN), that deals with the detailed and complex study of dimensionality reduction and classification of Hyperspectal images (HSI). Our work is 2-fold: At first (1), 3SPCA is applied on raw HSI that adopts superpixels-based local reconstruction to filter the images, whereas PCA-based supplementary global features acquire the relevant and low-dimensional local features. Every HSI pixel is reconstituted by the pixels of its closest neighbors in the same superpixel block to reduce noise and improve spatial information. Next, PCA is conducted on every zone and the full HSI to get local and global features. The local-global and spatial-spectral properties are then concatenated. Secondly (2), the D-2D-3D-CNN fusion architecture is made up of three 3D convolution blocks, two 2D convolution blocks with varied kernel sizes and filters, and four fully connected (FC) dense layers, totaling nine distinguishing and information-enriched features. These features can generate precise class labels and apply them to the appropriate landcovers. The proposed method has been applied to three publicly available HSI landcover datasets, the Indian Pines, the Salinas Valley, and the Pavia University. It achieved respectively 98.33%, 99.99%, and 98.73% average accuracy scores. Due to its improved Feature Extraction capacity from a limited number of training samples and its classification performance with fewer epochs, this method outperforms other relevant state-of-the-art CNN-based methods.