Opt Deep CSSAN: Optimized Deep Convolutional Spectral-Spatial Attention Network for hyperspectral image...

Hyperspectral technology contains the most basic and accurate data in topography through taking hundreds of finely categorized spectral bands at the same time, which serves very useful places, like the surveillance of agriculture, geological exploration, and national security. For all efforts in or related to hyperspectral data analysis, the greatest efforts, indeed, go into image classification. Deep learning-based feature extraction frameworks thus exert their influence over several contemporary applications. This work presents a method for Hyperspectral Image Classification (HSIC) by merging deep learning models. Initially, band selection is performed by utilizing Double Exponential Smoothing-Artificial Flora Optimization (DES-AFO) algorithm by integration of Double Exponential smoothing (DES) in Artificial Flora Optimization (AFO). Then, feature engineering is done where, the feature extraction is done by Empirical wavelet transform (EWT), Convolutional Neural Network (CNN), together with the features extracted using ResNet50. Then, the dimension of the extracted features is reduced for computational efficiency and data compression using Canonical Correlation Analysis (CCA). Finally, classification is performed using Optimized Deep Convolutional Spectral-Spatial Attention Network (Opt Deep CSSAN), where Deep CSSAN is proposed by combining deep CNN and Spectral-Spatial Attention Network (SSAN). Moreover, proposed Deep CSSAN is trained using DES-AFO. Experimental evidence highlights that DES-AFO based Opt Deep CSSAN technique exhibited superior performance relative to standard methods with 96.9 % accuracy, 97.1 % of TPR, 95.8 % of Kappa, 96.9 % of TNR and 91.5 % of PPV.