Model stacking of Spatial Spectral Network with Capsule Networks for Classification of Hyper-Spectral Images

With the development of remote sensing technology, the application of Hyperspectral Images (HSI) is vastly utilized in various fields. In specific, the classification methods of HSI plays a major role in detecting the targets, environment management, mineral mapping and so on. The existing approaches based on classification of hyper spectral images faces problems related to poor classification results due to constrained receptive field and inability to acquire long range features. The spatial spectral network has the ability to perform pixel wise classification of HSI but training the model for large labeled datasets can be challenging and leads to overfitting issues. Caps Nets are designed to capture hierarchical relationships between features. In hyperspectral image classification, where complex spectral and spatial patterns may exist, Caps Nets can potentially model these hierarchical relationships in an effective manner and helps in effective classification. So, this research introduced the model stacking of SSN and Caps Net which enhance the overall performance of classifying the HSI with better accuracy. The data is acquisitioned from Indian pines and University of Pavia to estimate efficiency of proposed technique in classifying HIS. Experimental outcomes exhibits that the proposed approach achieved better results in overall metrics. The overall accuracy of proposed approach for Indian pines dataset is 97.82% which is comparably higher than the existing Weighted Feature Fusion of Convolutional Neural Network and Graph Attention Network (WFCG) with 90.86%.