Deep concatenated features with improved heuristic-based recurrent neural network for hyperspectral image classification

Hyperspectral remote sensing is one of the important approaches in the area of remote sensing owing to the latest enhancements in the Hyper Spectral Imaging (HSI) technology. The classification represents a direct approach in the HSI field that provides every pixel a particular semantic label based on its behavior automatically. Nowadays, deep learning-oriented techniques have gained wide attention in the area of HSI classification. Although Convolutional Neural Network (CNN)-oriented techniques are subjected to the HSI classification, their performances are not up to the expectation. This is because; the majority of the traditional techniques cannot utilize the intrinsic behavior of distinct pixels in HSI efficiently. The inherent relationships are ignored between distinct category dependencies, distinct spectral bands, and distinct spatial pixels. Moreover, deep learning methods are required to build a huge and complicated network, and hence the training process tends to be time-consuming. Hence, these work tactics to design and implement a novel HSI classification method using deep structured architectures. The major stages of the offered method are feature extraction, and classification. Initially, the HSI from RIT-18 benchmark sources is collected. Before processing it for feature extraction, the images are split into 'n' number of patches, where the length of every patch is 16 x 16. Then the feature extraction begins, in which the spatial and spectral features, as well as the Enhanced CNN, are utilized for acquiring the deep features from the entire patches. In the CNN, the architecture is enhanced by the "Modified Velocity-based Colliding Bodies Optimization (MV-CBO)". Then, the entire features acquired from all the patches are concatenated. Finally, the utilization of deep structured architectures termed modified Recurrent Neural Network (RNN) is utilized in the classification phase, which classifies the images into different categories as per the dataset. The RNN architecture is also modified by the MV-CBO to attain high classification accuracy. From the simulation results, the accuracy rate of the MV-CBO-M-RNN at a 75% learning rate is correspondingly secured at 3.16%, 4.26%, 1.03%, and 2.08% more advanced than DNN, RNN, CNN, and NN. The validation of the recommended technique on challenging public datasets, and the experimental evaluation over baseline approaches validates the efficiency and robustness of the suggested model.