Non-compliant parking in open parking facilities remains a persistent challenge in urban environments, reducing parking efficiency, obstructing access, and undermining the reliability of smart parking systems. Existing vision-based monitoring approaches predominantly rely on fixed camera infrastructure, which is limited in spatial coverage and often lacks location-referencing capabilities for parking violations. This paper proposes a geo-aware system based on an Unmanned Aerial Vehicle (UAV) for detecting non-compliant parking using geo-tagged imagery. The proposed system integrates autonomous UAV patrol, frame-level UAV geolocation extraction, and deep learning-based object detection to identify vehicles parked outside designated spaces and associate each detection event with the corresponding UAV telemetry data. A processing pipeline synchronizes video frames with telemetry data, enabling the generation of geo-referenced outputs for enforcement and navigation purposes. The system is evaluated through real-world experiments conducted in an open parking facility, using a YOLOv8-based detection model trained on a custom annotated dataset. Experimental results show that the system achieves an accuracy of 99.56% with zero observed false positives under the tested conditions (evaluated dataset and post-processing pipelines), while providing reliable location-referencing information for detected parking violations. The proposed approach enables mobile, geo-aware sensing of parking facilities, addresses key limitations of static monitoring systems, and provides a practical solution for improving parking compliance, reducing operational inefficiencies, and supporting sustainable urban mobility management.
Wireless network security is a growing concern in Industrial Internet of Things (IIoT) as jamming attacks disrupt communication of IIoT entities in Private 5G networks. While machine learning (ML) algorithms detect jamming attacks effectively, centralized data collection from entities compromises sensitive information and privacy in distributed IIoT environment. This article presents a multi-faceted ensemble federated learning-based jamming detection framework, MF-EFLJD, to detect and further improve detection performance. To address the issue of non-independent and identical distribution (non-IID) and high-dimensional wireless channel data persisting across femtocells, this framework proposes sequential integration of Principal Component Analysis (PCA), Genetic Algorithm (GA), and Shapley Additive Explanation (SHAP) values to obtain optimized and structured data. Moreover, the framework trains 1D-CNN-based Federated Learning (FL) component on optimized data of entities to detect jamming attacks under non-IID conditions at each femtocell. Furthermore, to boost the performance detection of FL under non-IID, we propose Random Forest (RF)-based stacking ensemble component with an unweighted majority soft voting approach. Under non-IID setting α=0.7, the proposed system demonstrates that FL achieves accuracies of 0.9376 (±0.0134), 0.8806 (±0.0178), and 0.9576 (±0.0169) for femtocells F1, F2, and F3, respectively, in detecting jammed and non-jammed signals. Additionally, the ensemble component achieves 0.9963 (±0.0031) accuracy, depicting improvement in performance across femtocells.
The low-altitude economy (LAE) is driving rapid growth in autonomous multi-UAV services that must operate over wide, infrastructure-poor areas under intermittent connectivity and strict energy budgets. For such services, the central challenge is not only sensing but resilient networking and coordination: distributing missions, aggregating telemetry, and recovering coverage when UAVs fail or run out of battery. This paper presents a resilient, cloud-assisted ad hoc coordination architecture for multi-UAV LAE services, using precision agriculture as a demanding, representative use case. Missions, monitoring and recovery operate on compact survey-line units rather than waypoints, which keeps coordination and recovery messages at hundreds of bytes and separates them from imagery. On this substrate we propose a battery-aware reallocation mechanism with a look-ahead trigger and a feasibility test based on one empirically calibrated energy model. We also propose a decentralized recovery protocol in which the shared line plan provides failure detection, bounds each recovery to a small cohort of UAVs and makes the reassignment a deterministic function of a frozen record, so that survivors agree without a ground station or auction rounds. The operational workflow was field-validated with DJI Mavic 3M UAVs on agricultural sites in Spain. Survey kinematics were measured, and an energy model was fitted on ten flight logs and cross-validated against measured battery consumption (energy error within ±3.3%). Coordination traffic was network-emulated, and all failure-recovery experiments were software-simulated with a single co-simulator using the calibrated parameters. In simulation, recovery kept every landing above the reserve, even with a ＋30% energy-model bias. In the six-UAV vineyard experiments, recovery required approximately 1–3 KB of signalling per event. The share of orphaned work recovered within the sortie was governed by the fleet load factor: complete for lightly loaded fleets, but only 6–8 percentage points for a six-UAV vineyard mission sized to the battery limit. The decentralized protocol matched centralized recovery without a ground station, stayed consistent under cascading failures, 50% loss and reordering, and scaled to 48 UAVs at under 80 B/s per node.
Underwater acoustic networks combine limited bandwidth, long propagation delays, intermittent connectivity, and costly transmissions, making indiscriminate DTN replication particularly expensive. This work proposes NsbKnowledgeGradient (NKG), a selective DTN forwarding policy that uses locally acquired knowledge from Node State Beacon (NSB) exchanges to evaluate observed forwarding opportunities without relying on a global topology view, physical distance, future contacts, or predefined routes. NKG is evaluated in ns-3 with Aqua-Sim NG together with a staged control-plane organization that separates an initial learning period from lower-frequency knowledge maintenance during operation. The study compares NKG with classical replicative, probabilistic, and copy-limited DTN policies and further examines parameter sensitivity, the contribution of different NKG decision components, constrained-buffer behavior, communication energy, and scalability up to 50 nodes. In the main experimental matrix, staged NKG achieves 76.08% aggregate delivery compared with 70.70% for Epidemic, with a paired mean improvement of 5.37 percentage points across ten seeds. This is obtained with approximately 30.84 forwarding requests and 10.53 kB of offered DTN traffic per delivered bundle, compared with 176.06 requests and 67.10 kB for Epidemic. The additional experiments show that this advantage depends on maintaining sufficient decision-relevant knowledge: severe storage pressure favors the redundancy of aggressive replication, while larger deployments can exceed the breadth of a fixed knowledge configuration. In the 40- and 50-node experiments, broader knowledge propagation restores delivery, but at a higher control cost. Overall, the results show that local knowledge can support efficient selective forwarding, provided that the knowledge required by the decision can be maintained at a cost compatible with the acoustic channel.
