A Hybrid Deep Learning-ViT Model and A Meta-Heuristic Feature Selection Algorithm for Efficient Remote Sensing Image Classification

Recent deep learning techniques driven by large datasets demonstrate the significant impact of feature learning in remote sensing for land use and cover classification, particularly exemplified by CNNs. While the pre-trained models showed good classification performance, they struggled to classify remote-sensing images with high precision accurately. In this study, we introduced XNANet, a self-attention-based CNN network for image classification. Bayesian optimization has been used to initialize the hyperparameters of the proposed model to improve training on the radiographic images. We suggested a novel network-level approach via the fusion of deep structure utilizing tiny-32 ViT and XNANet. For the first time, the tiny-32 vision transformer architecture has been utilized for RS images and combined with XNANet through network-level fusion. Following the fusion process, the model focused on RS image datasets and obtained deep features from the self-attention layer. The features that have been extracted are subject to selection, utilizing a novel meta-heuristic feature selection algorithm, RF-DE. The selected features are categorized using three popular classifiers. The proposed architecture’s experimental process was executed on the AID, RSSCN7, and SIRI-WHU datasets, resulting in accuracies of 98.9%, 99.3%, and 99.7%, respectively. Similarly, RF-DE was evaluated against six popular feature selection algorithms, yielding accuracies of 98.9%, 99.3%, and 99.7%, respectively. An in-depth statistical analysis was conducted to evaluate the suggested ensemble and RF-DE and demonstrate that the fusion model attained enhanced accuracy with RF-DE. Furthermore, recent techniques and proposed methods are compared, demonstrating enhanced precision, recall, and accuracy.