Deep Feature Segmentation Model Driven by Hybrid Convolution Network for Hyper Spectral Image Classification

Abstract

Hyper-spectral image (HSI) classification can support different applications, such as agriculture, military, city planning, land utilization, and identifying distinct regions. It is treated as a crucial topic in the research community. Recent advancements in convolution neural networks (CNN) have shown the unique capability of extracting meaningful features and classification. However, CNN works with square images with fixed dimensions and cannot extract local information of images having distinct geometric variations with context and content relationships; hence, there is a scope for improvement in correctly identifying class boundaries. Encouraged by the facts, we propose an HSI feature segmentation model by the hybrid convolution network (GCNN-RESNET152) for the HSI classification. First, pre-trained CNN on ImageNet is used to obtain the multi-layer feature. Second, the 3D discrete wavelet transform image is fed into the graph convolution network GCN model to gain patch-to-patch correlations feature maps. Finally, the features are integrated using the concatenation method of the three-weighted coefficients. Finally, the linear classifier is used to predict the semantic classes of pixel HSI. The proposed model is tested on four benchmark datasets: Houston University (HU), Indian Pines (IP), Kennedy Space Station (KSS), and Pavia University (PU). The result is compared with state-of-art algorithms and is superior in terms of overall, average, and kappa accuracy.