An efficient privacy-preserving multilevel fusion-based feature engineering framework for UAV-enabled land cover classification using remote sensing images.

In recent years, unmanned aerial vehicles (UAVs) have attracted more attention. UAVs have numerous manifest benefits over traditional manned aircraft, mainly regarding operator safety, operational expense, and the possibility of complex/hazardous environments such as land cover classification and accessibility for civil applications. A land cover image classification of scenes categorizes the aerial images, captured using drones by masking some ground matters and kinds of land covers, into several semantical forms. Current technological advances have made it simpler to set up an unmanned aerial system with composite topology to reach refined missions that were formerly impossible without real human connections. Nevertheless, networked UAVs are vulnerable to malicious attacks, and therefore intrusion detection systems (IDSs) are logically derived to address the vulnerabilities and/or attacks. Deep learning (DL) methods are essential for processing security problems in UAV networks. This paper presents a Privacy-Preserving Intrusion Detection Model for UAV-Based Remote Sensing Applications in Land Cover Classification Using Multilevel Fusion Feature Engineering (IDUAVRS-LCCMFFE) technique. The main intention of the IDUAVRS-LCCMFFE technique is to provide an effective model for land cover classification using UAV images in dynamic environments. Initially, the image pre-processing stage applies a joint bilateral filter (JBF) model to enhance image quality by removing noise. Furthermore, the feature extraction process uses the fusion models comprising NASNetMobile, ResNet50, and VGG19. Moreover, the proposed IDUAVRS-LCCMFFE model employs the Elman recurrent neural network (ERNN) model for the land cover classification process. Finally, the hyperparameter selection of the ERNN model is accomplished by implementing the salp swarm algorithm (SSA) model. The experimentation of the IDUAVRS-LCCMFFE approach is examined under the ToN-IoT dataset, and the outcome is computed under different measures. The performance validation of the IDUAVRS-LCCMFFE approach portrayed a superior accuracy value of 99.66% and 96.47% under ToN-IoT and EuroSat datasets.