A Novel Deep Learning Framework With Meta-Heuristic Feature Selection for Enhanced Remote Sensing Image Classification

We propose a novel deep learning architecture, called XcelNet17, for image classification in remote sensing. Comprising fourteen convolutional and three fully connected layers, XcelNet17 outperforms several benchmark architectures available in the literature in terms of classification accuracy. Additionally, we present BA-ABC, a new hybrid feature selection algorithm that inherits the strengths of the Bat Algorithm (BA) and the Artificial Bee Colony (ABC) algorithm. Together these contributions significantly enhance the performance and accuracy of remote sensing image classification tasks. The proposed framework is thoroughly trained and verified using five benchmark datasets typically used for remote sensing image classification, namely AID, RSSCN7, SIRI-WHU, UC Merced, and WHU RS-19. Our simulation results suggest that in terms of classification accuracy, XcelNet17 outperforms most of the well established networks including AlexNet, VGG16, VGG19, ResNet50, and DarkNet19 by obtaining accuracy values in the range of 94.6% and 99.9%. Furthermore, the proposed features selection method, when integrated with XcelNet17, yields much improved classification accuracy in comparison to various benchmarks including WOA, GWO, BA, ABC, and ACO algorithms. For example, an 8% superior performance on WHU-RS19 dataset has been observed. The attained results are further validated by an in-depth statistical analysis.