Design of Super Resolution and Fuzzy Deep Learning Architecture for the Classification of Land Cover and Landsliding using Aerial Remote Sensing Data

Khan, Junaid Ali, Khan, Muhammad Attique, Al-Khalidi, Mohammed ORCID: https://orcid.org/0000-0002-1655-8514, AlHammadi, Dina Abdulaziz, Alasiry, Areej, Marzougui, Mehrez, Zhang, Yudong and Khan, Faheem (2025) Design of Super Resolution and Fuzzy Deep Learning Architecture for the Classification of Land Cover and Landsliding using Aerial Remote Sensing Data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 18. pp. 337-351. ISSN 1939-1404

Preview

Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (16MB) | Preview

Abstract

The diversity, noise, inter-image interference, image distortion and increase in the number of classes in aerial remotely sensed dataset cause exertion in the classification. The efficacy and stability of convolutional neural networks increase in image classification with the specified use of feature selection algorithm that causes remarkably improved decision making. To address the associated difficulties, a fuzzy deep learning architecture has been designed with a super resolution technique that consisting of 40 convolutional, four polling, four inverted bottleneck blocks, and one fully connected layer. The fuzzy optimistic formula is implemented in 4 blocks as an activation function where information is fused from the previous layers and present block while the rest are using the ReLU transfer function to handle the issue of noise and inter-image interference. Feature selection is performed based on the physics of chaotic particle swarm optimization hybrid with the active set algorithm. The accuracy of the proposed architecture is examined on three diverse datasets: Bijie Earth Landslide/Non-Landslide, EuroSAT and NWPU-RESISC45, comprised of varying classes. The results are compared with state-of-the-art models like the hybrid version of VGGNet-16, Yolov4, ResNet-50, DenseNet-121 and other reported techniques. Moreover, the stability and computational complexity of the presented architecture are computed on 50 independent runs. It has been observed that the proposed architecture is stable, accurate, and viable and exploits a smaller number of learnable parameters than the models considered in comparison.