Unmanned Aerial Vehicle (UAV) remote sensing, with its advantages of rapid information acquisition and low cost, has been widely applied in scenarios such as emergency response. However, due to the long imaging distance and complex imaging mechanisms, targets in remote sensing images often face challenges such as small object size, dense distribution, and low inter-class discriminability. To address these issues, this paper proposes a multi-modal remote sensing object detection network called RemoteDet-Mamba, which is based on a patch-level four-direction selective scanning fusion strategy. This method simultaneously learns unimodal local features and fuses cross-modal patch-level global semantic information, thereby enhancing the distinguishability of small objects and improving inter-class discrimination. Furthermore, the designed lightweight fusion mechanism effectively decouples densely packed targets while reducing computational complexity. Experimental results on the DroneVehicle dataset demonstrate that RemoteDet-Mamba achieves superior detection performance compared to current mainstream methods, while maintaining low parameter count and computational overhead, showing promising potential for practical applications.

In emergency communications constrained by bandwidth limitations, unmanned aerial vehicle (UAV)-based coded caching presents a promising approach for the efficient dissemination of high-bandwidth-demanding services. This paper focuses on content download and content repair in aerial caching networks, where UAVs deliver contents to both ground users and invalid UAVs. To address potential data loss due to limited power and high mobility, fault-tolerant codes are utilized to maintain data availability and reliability. Initially, we derive the expressions of communication cost and success rate for content download and content repair. The size of coded fragments, determined by the coding design, affects both the success rate and transmission cost, while the resource allocation, which influences the cooperative relationships, also impacts these two aspects. The interplay between coding design and resource allocation is thus established to jointly optimize the overall performance. Then, we design a joint optimization problem of erasure coding schemes, coding parameters, matching relations, and UAV trajectories to maximize the overall success rate. Moreover, we propose a hierarchical multi-agent parameterized deep Q-network (H-MA-PDQN) algorithm integrating a dual-component structure for long-term coding and immediate resource allocation to solve the mixed integer nonlinear programming (MINLP), and each agent employs a PDQN with hybrid discrete-continuous action space. Simulation results demonstrate that our proposed H-MA-PDQN algorithm increases the success probability by 26.7% and 66.7% and reduces the transmission cost by 27.3% and 42.9% compared with the DQN and greedy-based strategies, respectively.

In mountainous environments, terrain occlusion causes non-line-of-sight (NLoS) transmission, significantly reducing the signal propagation range. To improve emergency rescue efficiency, a mobile uncrewed aerial vehicle (UAV)-based integrated localization and communication (ILAC) network should be deployed to achieve optimal performance through adaptive trajectory planning and resource allocation. However, irregular and unpredictable terrain occlusions, coupled with dynamic users, make traditional optimization ineffective and reinforcement learning (RL) inefficient. To address these challenges, this paper proposes a hybrid action space soft Actor-Critic with geographic information-based state aggregation (GeoAgg-HSAC) decision-making scheme. First, an RL state aggregation method based on graph contrastive learning is designed. Through a pre-trained graph neural network (GNN), the UAV network states experiencing the same occlusion are mapped to similar low-dimensional representations. This method reduces the state dimension and allows similar states to share policy experience, thereby improving sample efficiency and accelerating convergence. A hybrid action space SAC network is then designed, which simultaneously makes decisions for continuous UAV trajectories and discrete resource allocation. Finally, a simulation environment based on real mountain terrain and wireless data is built for the experiment. The experimental results show that the proposed scheme has significant advantages for optimizing communication and localization performance.

In emergency rescue missions, the perception of embodied agents (e.g., rescue robots) rely heavily on high-quality semantic information. However, deploying complex embodied semantic models within constrained Unmanned Aerial Vehicle (UAV) networks introduces severe resource conflicts. We propose a semantic-aware framework tailored for embodied intelligence networks, jointly optimizing model caching, user association, semantic compression factors, and computational resources to minimize total system delay while satisfying energy and semantic fidelity constraints. To solve this multi-timescale mixed-integer problem, we develop a Two-Timescale integrated Matching and Multi-Agent Deep Reinforcement Learning (2T-M2ADRL) algorithm. It utilizes Hierarchical Bipartite Matching (HBM) for discrete large timescale topologies and a Multi-Head Attention based MADDPG (MHA-MADDPG) for small timescale resource orchestration. Furthermore, we construct an offline surrogate model to bypass evaluating semantic similarity via Large Language Models (LLMs) during training. Extensive simulations confirm our framework outperforms baselines, ensuring optimal latency and semantic accuracy.

Multi-device collaborative computing has emerged as a promising paradigm to address the computational limitations of edge devices. However, the existing methods fail to simultaneously achieve high precision and low latency perception, while adapting to bandwidth fluctuations. Specifically, feature compression methods cause imbalanced category precision loss. We propose a category-adaptive feature compensation mechanism based on this, which analyzes gradient responses specific to each target category. It dynamically adjusts the feature reconstruction weights. Then, the feature compression method needs to consider the trade-off between high accuracy and low latency. We implement it by training a variable-sized discrete dictionary to dynamically map intermediate features to dictionary indices, adaptively reducing the amount of data transmitted. Experimental results show that our method can reduce the transmission latency by up to 98%, and limit the accuracy loss to less than 2%.

In disaster scenarios, infrastructure damage and wireless resource scarcity pose significant challenges for providing prompt and reliable communication and sensing (C&S) services. Recently, unmanned aerial vehicle (UAV) enabled integrated sensing and communication (ISAC) has emerged as a promising technique to tackle the above issues by leveraging flexibility and mobility of multiple UAVs to offer high-quality and cost-efficient C&S services. In parallel, rate-splitting multiple access (RSMA) facilitates customized transmission by partitioning messages into private and common parts with adjustable rates, thereby making it well-suited for on-demand data transmission in disaster scenarios. In this paper, we propose a framework that utilizes coordinated RSMA for ISAC (Coordinated RSMA-ISAC) in an emergency UAV system. This framework enables multiple transmit UAVs to simultaneously communicate with several communication survivors (CSs) and detect a potentially trapped survivor (TS) in a coordinated manner with imperfect channel state information (CSI). In addition, an optimization problem is formulated to jointly optimize the UAV-CS association, UAV deployment, and transmit beamforming to maximize the weighted sum rate (WSR) of the system, subject to the sensing signal-to-noise ratio (SNR) requirement. To efficiently solve such a mixed-integer non-convex programming (MINCP) problem, an iterative algorithm is proposed by applying the generalized Benders decomposition (GBD) technique. Specifically, the original problem is decoupled into a master problem for pure integer programming and a primal problem for non-convex programming. Then, we further use successive convex approximation (SCA), semi-definite relaxation (SDR), and cutting-plane techniques to solve the decoupled problems. Simulation results verify the effectiveness of the proposed algorithm, and demonstrate that the coordinated RSMA-ISAC framework outperforms conventional space division multiple access (SDMA), non-orthogonal multiple access (NOMA), and orthogonal multiple access (OMA) in terms of both C&S performance.

Streaming 3D reconstruction under a strict constant-memory budget hinges on how the recurrent state is updated as the stream evolves. We profile TTT3R-style per-token gates across five benchmarks and discover a structural bottleneck: the gate is intrinsically bounded in magnitude (median 0.31; never exceeding 0.6) and nearly frame-invariant, yielding an effective memory horizon of only ∼3 frames per state token, which serves as the structural origin of long-sequence drift. We trace this to a missing axis: existing inference-time methods modulate updates only at the per-token, intra-frame level, while the orthogonal frame-level question of how strongly each frame should contribute to the state has been treated as content-independent. We close this gap with a scalar frame-level gate α_t ∈ (0, 1] derived in closed form from frame-to-frame changes of internal features – a continuous relaxation of classical Simultaneous Localization and Mapping (SLAM) keyframe selection that requires no parameters, no training, and no extra forward pass. Across six benchmarks spanning camera pose, video depth, and 3D reconstruction at sequence lengths up to 4,541 frames, our gate cuts ATE by 51% on long TUM-RGBD pose sequences, reduces AbsRel by 12.8% on Bonn video depth, and on KITTI long-sequence pose estimation surpasses both LongStream and Keyframe-VO, while retaining strictly constant memory at zero training cost.

In this letter, we propose an airborne maneuverable bi-static integrated sensing and communication system where both the transmitter and receiver are unmanned aerial vehicles. By timely forming a dynamic bi-static range based on the motion information of the target, such a system can provide an adaptive two dimensional tracking and communication services. Towards this end, a trajectory optimization problem for both transmits and receive UAV is formulated to achieve high-accurate motion state estimation by minimizing the time-variant Cramer Rao bound, subject to the sufficient communication signal-to-noise ratio to maintain communication channel prediction error. Then we develop an efficient approach based on the successive convex approximation technique and the S-procedure to address the problem. Numerical results demonstrate that our proposed airborne maneuverable bi-static ISAC system is able to obtain higher tracking accuracy compared with the static or semi-dynamic ISAC system.

Fine-tuning large vision-language models such as CLIP on edge devices remains challenging due to severe memory and computational constraints. Although existing parameter-efficient fine-tuning methods can effectively reduce training costs, they are not specifically designed for cloud-edge collaborative scenarios. To address this limitation, we propose CE-CLIP, a cloud-edge collaborative fine-tuning framework. Specifically, we keep the pretrained CLIP backbone frozen and maintain it at the cloud, and deploy a lightweight multimodal side network at the edge enabling efficient adaptation of visual and textual features. We further propose a semantic token reduction module to compress features extracted in the cloud, significantly reducing communication overhead. In addition, we introduce a feature consistency loss to preserve the pretrained model's generalization capability during adaptation. Experimental results show that CE-CLIP achieves an average accuracy of 82.1% on few-shot adaptation tasks and outperforms most existing approaches in base-to-new generalization, demonstrating its strong adaptability. Furthermore, CE-CLIP features only 0.72M trainable parameters, 159 MB of edge memory usage, and 1.38 MB of communication overhead per image, demonstrating an effective balance among accuracy, efficiency, and adaptability in the cloud-edge collaboration paradigm.

In emergency scenarios, uncrewed aerial vehicles (UAVs) show significant potential as aerial base stations (BSs) to establish reliable communication links and provide localization services through integrated air-ground collaboration. This paper proposes a novel energy-efficient collaborative framework based on the solo-UAV-rescuer cooperative (SURC) paradigm, which synergistically enhances both communication capacity and localization accuracy. From a system optimization perspective, we formulate an optimization problem using a normalized combination of three critical metrics: achievable data rate, localization accuracy, and energy consumption. Specifically, to maximize the system's utility, we design a signal perception-based localization method that incorporates angle-of-arrival (AOA) localization information for guidance, and develop a beamforming scheme to facilitate high data rate communication. Building on these methods, we propose a deep reinforcement learning (DRL)-based synergistic communication and localization reinforcement (SYNCORE) approach that dynamically optimizes three key operational parameters: UAV trajectory planning, flight time, and transmission power control, achieving reliable services with energy-efficient operation. Based on the simulation results, we validate that the proposed scheme enhances communication and localization performance, while also improving energy efficiency, surpassing the baseline schemes.

In this paper, we propose a novel uncrewed aerial vehicle (UAV) and uncrewed ground vehicle (UGV) enabled air-ground collaborative system (UAV-UGV-AGCS) where a UAV and a UGV are utilized to realize multi-functional integration. Specifically, the UAV simultaneously provides communication, data collection services, and tracks a ground mobile target (GMT); while the UGV assists in collecting data from clusters of sensor nodes (SNs) and receiving sensing echo signals, thereby achieving high-rate communication, efficient data collection, and real-time tracking. Based on the dynamic characteristics and collaboration of the UAV and UGV, a joint trajectory optimization problem is formulated to maximize the average system throughput, subject to the UAV and UGV motion constraints as well as energy consumption limits. In particular, we derive the posterior Cram & eacute;r-Rao bound (PCRB) as a tracking performance metric. To tackle the non-convexity problem, we first design the tracking method based on the extended Kalman filtering (EKF) to estimate the motion parameters of GMT in each time slot and group the SNs via the immune optimization (IO)-based clustering algorithm. Subsequently, the problem is decomposed into two subproblems and solved via a successive convex approximation (SCA)-based iterative optimization algorithm to obtain the optimal solutions. Simulation results demonstrate that the proposed framework achieves efficient data collection and superior communication performance, thereby meeting the demands of high-precision tracking by dynamically adjusting trajectories of UGV and UAV, compared with other benchmark network designs.

Driven by the rapid proliferation of edge inference, collaborative inference among edge devices has emerged as a promising paradigm for addressing resource constraints in computationally intensive tasks. However, existing approaches often suffer from inefficient utilization of heterogeneous resources and lack dynamic adaptation to varying device capabilities and network conditions. To address these challenges, we present a novel collaborative inference framework that prioritizes energy-accuracy co-optimization through workload partitioning and resource-aware model pruning. Specifically, we first design a workload partitioning strategy aimed at maximizing resource utilization efficiency in heterogeneous environments. Dynamically distributes image slices to edge devices by closely matching tasks with their specific computational, energy, and network profiles, ensuring balanced load and minimized latency. Moreover, we propose an adaptive pruning method that is aware of energy-accuracy tradeoffs. It is equipped with dual predictors that model the accuracy and energy consumption under different pruning rates in real-time, enabling dynamic pruning rate selection to optimally balance performance with energy efficiency. The experimental results on the VisDrone2019 dataset demonstrate that our method achieves significant improvements over existing approaches, improving detection accuracy by up to 1.3% in mAP50 while increasing inference speed by 5.3× compared to slicing methods. Furthermore, our adaptive pruning strategy reduces energy consumption by up to 32% while maintaining comparable accuracy across various edge devices, such as the NVIDIA AGX Orin and Jetson TX2.

Next-generation mobile networks are expected to facilitate fast AI model deployment on end devices (EDs). By enabling collaborative model caching across EDs, mobile networks can efficiently support distributed AI inference services through device-to-device (D2D) cooperation. In this paper, we investigate a D2D-assisted model caching and collaborative computing framework that aims to balance the trade-off among inference delay, accuracy, and energy consumption by managing model caching, data offloading, and computation resources efficiently during the provisioning of diverse AI services. Specifically, considering the AI performance is constrained by multi-dimensional resources, a new metric named Service Hit Rate (SHR) is proposed to decouple the joint impacts of computation, communication, and caching resources on service success. Aiming to maximize the SHR, we propose a matching-aided multi-agent reinforcement learning (MARL) framework. First, a hierarchical bipartite matching algorithm is utilized for model deployment and helper assignment. Then, an attention-based MARL algorithm is employed to allocate computation resource for models cached on the same EDs. Simulation results demonstrate that the proposed algorithm significantly improves the system SHR for AI services.

The ultra-dense deployment of base stations (BSs) results in significant energy costs, while the increasing use of fluctuating renewable energy sources (RESs) threatens the safe operation of electric network (EN). These issues can be addressed by coordinating BSs’ active/sleep states with RES generation. However, the coordinated decision-making is challenging due to the conflicting goals of maximizing RES utilization, guaranteeing communication quality of service (QoS). In this paper, we design an electric-cellular collaborative network (ECCN) and formulate a joint optimization problem to minimize electric supply and QoS degradation costs, subjecting to EN’s safety constraints. Considering the uncertainty of RES generation and BS traffic, we propose a heterogeneous multi-agent safe reinforcement learning (HMAS-RL) algorithm to solve the problem. HMAS-RL inherits from the centralized critic and decentralized actor framework, where two heterogeneous agents learn a cooperative policy for joint decision-making through two actor networks design that share observations and a team reward. A cost critic network is designed to handle safety constraints, eliminating the need for manual penalty term design and tuning. We validate the proposed method using an IEEE 33-bus electric distribution test feeder. Results demonstrate HMAS-RL achieves superior performance in RES utilization, communication QoS, and EN safety constraints maintenance.

Electricity data plays a pivotal role in power management systems. Smart meters, as key tools for recording this data, often encounter anomalies due to meter malfunctions, operational errors, or unauthorized electricity usage, all of which jeopardize the stability of power grids. To this end, we propose the future-guidance anomaly detection network, called FG-Net, designed for real-time analysis of electricity time series data. FG-Net is designed to memorize historical data and assimilate future data, ensuring comprehensive learning of complete data information. Specifically, we leverage the comprehensive data insights gained from a complete information network to guide the predictions of the historical information network. Subsequently, we developed a self-matching feature guidance (SFG) strategy that harnesses the strengths of the complete information network to offset the limitations of the historical information network, thus providing effective guidance. The experimental results on two power grid time series datasets with different anomaly volatility, the Low Carbon London dataset and the Ausgrid Solar Home dataset, demonstrate the proposed anomaly detection method's accuracy and efficiency.

Model-based shallow machine-learning methods and data-driven deep-learning (DL) methods have been advanced to address hyperspectral and multispectral image fusion (HSMS fusion). Nonetheless, model-based approaches, which meticulously craft regularization terms within optimization models using hand-engineered priors, often struggle to pinpoint the optimal solution efficiently. DL-based methods, which train on extensive datasets to learn a non-linear mapping for generating a high spatial and spectral resolution image (HS2I), exhibit limited generalization capabilities when applied to novel and diverse test datasets. To improve the generalization ability and optimization efficiency of the existing HS-MS fusion methods, a novel deep low-rank prior (DLRP)-based self-supervised HS-MS fusion approach is devised. It incorporates a low-rank learning paradigm to produce the fused HS2I, subject to the constraints imposed by loss functions. Instead of deriving the solution via solving a low-rank approximation optimization problem, deep image prior (DIP) learned by a two-dimensional CNN and a one-dimensional CNN are integrated as the low-rank prior.

To mitigate the rising energy costs in edge computing, edge servers (ESs) can receive revenues from reducing their energy usage by contracting with virtual power plant (VPP). ESs also respond to user equipment (UE) by providing computation offloading services. However, such two-layer coordinated trading of computation offloading and energy management in a VPP-ES-UE architecture should address the information asymmetry issue and uncertain task arrivals, posing challenges to maximizing the stakeholders' utility. In this paper, we formulate the two-layer coordinated trading as a hierarchical contracting problem, which addresses the information asymmetry using two contract models. We design a VPP-ES energy contract on a large timescale and an ES-UE computation contract on a small timescale, where ESs need to coordinate the future task arrivals under the energy reduction target assigned by VPP. We construct a two-timescale virtual queue to achieve the energy reduction goal by the long-term queue stability constraint and employ Lyapunov optimization to transform the original problem into an online optimization problem without requiring future information. An online two-timescale hierarchical contract optimization algorithm is proposed to solve the transformed problem. The simulation results demonstrate that our method achieves higher social welfare compared to other benchmarks.

Unmanned aerial vehicle (UAV)-assisted integrated sensing and communication (ISAC) networks have emerged as a promising technique in next-generation wireless systems. However, due to the time-variant number and dynamics of the served users, resource allocation and trajectory optimization in UAV-assisted ISAC networks may encounter challenges in terms of the inconsistency in input dimensions, resulting in difficulties in achieving optimal performance. To this end, this paper investigates an end-to-end graph-enhanced multi-agent reinforcement learning method for resource allocation and trajectory optimization in UAV-assisted ISAC networks. Specifically, we first propose a novel heterogeneous recurrent graph neural network (HetRGNN) to capture the ISAC network topology as well as the mobility patterns of UAVs and users across different time slots, while handling the features of a dynamically varying number of users. Subsequently, a multi-agent reinforcement learning (MARL) scheme is implemented to address the nonconvex optimization problem. The simulation results show the fast convergence of the proposed HetRGNN-MARL and demonstrate its superior performance in optimizing resource allocation and UAV trajectories.

Forests are frequently impacted by climate conditions, vegetation density, and intricate terrain and geology, which contribute to natural disasters. Personnel engaged in or supporting rescue operations in such environments rely on robust communication systems to ensure their safety, highlighting the criticality of channel measurements in forest environments. However, according to current research, there is limited research on channel detection and modeling in forest areas in the existing literature. This paper describes the channel measurements campaign of air and ground in the Arxan National Forest Park of Inner Mongolia. It presents measurement results and propagation models for ground-to-ground (G2G) and air-to-ground (A2G) scenarios. The measurement campaign uses orthogonal frequency division multiplexing signals centered at 1.4 GHz for channel sounding. In the G2G measurement, in addition to using omnidirectional antennas to record data, we also use directional antennas to record the arrival angle information of the signal at the receiver. In the A2G measurement, we pre-plan the flight trajectory of the unmanned aerial vehicle so that it can fly at a fixed angle relative to the ground. We present path loss models suitable for G2G and A2G in forest environments based on the analysis of measurement results. The results indicate that the proposed model reduces error margins compared with other path loss models. Furthermore, we derive the multipath model expression specific to forest environments and conduct statistical analysis on key channel parameters e.g., shadow fading factor, root mean square delay spread, and Rician K factor. Our findings reveal that signal propagation obstruction due to tree crowns in A2G communication is more pronounced than tree trunk obstructions in G2G communication. Adjusting the elevation angle between air and ground can enhance communication quality.

Radio maps reflect the spatial distribution of signal strength and are essential for applications like smart cities, IoT, and wireless network planning. However, reconstructing accurate radio maps from sparse measurements remains challenging. Traditional interpolation and inpainting methods lack environmental awareness, while many deep learning approaches depend on detailed scene data, limiting generalization. To address this, we propose MARS, a Multi-scale Aware Radiomap Super-resolution method that combines CNNs and Transformers with multi-scale feature fusion and residual connections. MARS focuses on both global and local feature extraction, enhancing feature representation across different receptive fields and improving reconstruction accuracy. Experiments across different scenes and antenna locations show that MARS outperforms baseline models in both MSE and SSIM, while maintaining low computational cost, demonstrating strong practical potential.