Hyperspectral image classification using fractional meerkat optimization algorithm based 3D convolutional residual spectral–spatial attention network

Hyperspectral Imaging (HSI) is an important remote sensing technique for gathering detailed information from a scene. In the last decade, HSI classification became one of utmost research-intensive realms in remote sensing, involving mapping each pixel in an HSI to a particular class based on its spectral features. The growing application of Deep Learning (DL) strategies in HSI classification has attracted considerable interest, given their capacity to learn and classify features automatically from high-dimensional, complicated data. This research proposed an optimization-enabled hybrid DL approach for HSI classification. Initially, HSI is preprocessed using Adaptive Minimum Noise Fraction (Adaptive MNF). Then, feature extraction is executed by extracting spatial features using Gabor filter, Local Binary Pattern (LBP), spectral features using Spectral Angle Mapper (SAM), statistical features, like variance, mean, skewness, kurtosis of spectral bands and Shannon entropy. Lastly, for HSI classification, a 3D CNN is integrated with the Residual Spectral–Spatial Attention Network (RSSAN) to propose a novel approach, named 3D ConvRSSAN. In addition, a new training algorithm, named Fractional Meerkat Optimization Algorithm (FrMOA), is proposed by improving upon the original Meerkat Optimization Algorithm (MOA) based on fractional calculus principles to advance the optimization process further. Results from experiments demonstrated that FrMOA-based 3D ConvRSSAN reached an accuracy of 98.4%, TPR of 98.9%, TNR of 97.1%, kappa of 98.4%, F1-score of 98.8% and PPV of 97.8%.