This paper investigates semantic text communication in relay networks. We propose two token-level semantic forwarding schemes: semantic lossy forwarding (SLF) and semantic predict-and-forward (SPF). Inspired by their recent success in Large Language Models (LLMs), both utilize transformer archi tectures at each network node to encode and decode semantic information. Motivated by Block Markov coding, we establish a dynamic semantic state at the relay, updated upon receiving a new source signal. In SLF, this state decodes the received token; in SPF, it predicts the next source token for proactive forwarding. Extending the scope to sentence-level transmission, we introduce semantic sentence forwarding (SSF), which transmits a single embedding representing the entire message semantics. SSF integrates principles from SLF and SPF, incorporating previously decoded context to decode the received source signal autoregressively. We also leverage capabilities offered by LLMs in order to assess how well semantics are conveyed. We introduce a GPT-enabled semantic similarity score and use it in addition to commonly used BLEU and SBERT semantic metrics to evaluate performance. Extensive experiments demonstrate that proposed semantic forwarding schemes outperform conventional baselines, with SSF further improving upon SLF and SPF by effectively capturing contextual information.
Sound Event Detection (SED) and Direction of Arrival (DOA) estimation have attracted increasing attention in smart agriculture for tasks such as pest detection, machine-anomaly monitoring, and intruder localization. By integrating SED and DOA, Sound Event Localization and Detection (SELD) provides richer spatio-temporal acoustic information for advanced agricultural applications, such as targeted pest deterrence, fault-source tracing, and event-aware robotic response. However, SELD systems typically rely on large-scale neural networks to handle frame-wise multi-source detection and localization, which significantly limits their deployment on edge devices in real farming environments. To address this challenge, we propose Efficient SELDNet, a lightweight SELD architecture designed for resource-constrained edge platforms. Efficient SELDNet introduces two key modules: (1) an Efficient Multi-Scale Convolution (MSC) module that enhances spatial-spectral feature extraction while reducing computational load, and (2) a Convolutional Conformer module that replaces the standard linear layers in Transformer blocks with MSC modules, thereby substantially reducing the number of parameters. To evaluate the effectiveness and generalization ability of the proposed architecture, extensive experiments are conducted on two public SELD datasets, TNSSE21 and STARSS23, as well as a newly constructed agriculture-oriented synthetic SELD dataset. Experimental results show that Efficient SELDNet achieves comparable accuracy to the state-of-the-art (SOTA) baseline while reducing model size by up to 86% and floating-point operations by 79%. Furthermore, to evaluate cost-effective deployment feasibility in agricultural edge scenarios, we conducted CPU-only inference experiments on both a server-grade CPU and a low-power ARM CPU platform. The results demonstrate that Efficient SELDNet maintains real-time CPU-only inference with a smaller memory footprint, and the large version reduces both latency and memory usage compared with the SOTA baseline on the ARM CPU platform.
The rapid growth of low-altitude aerial services and applications, driven by uncrewed aerial vehicles (UAVs), calls for a new class of digital infrastructure beyond conventional terrestrial networks. The low-altitude wireless network (LAWN) has been proposed as dynamically reconfigurable three-dimensional architectures that integrate aerial and ground nodes to provide connectivity, sensing, and control in open, safety-critical airspace. This tutorial presents a comprehensive treatment of LAWNs from the joint perspectives of artificial intelligence (AI) and signal processing. We first review the historical evolution and architectural foundations of LAWNs, introducing altitude-based layers and functional planes, and summarizing the regulatory and standardization landscape. Building on this system view, we then discuss signal processing fundamentals for LAWNs, including 3D channel and system models, performance metrics, waveform and receiver design, localization and tracking, and multi-functionality co-design. Next, we survey AI techniques for LAWNs, covering discriminative and generative models for perception, control, resource management, and security, as well as emerging paradigms such as foundation models, large language models, and digital twins for mission planning and closed-loop optimization. To illustrate AI-signal processing integration in practice, we provide a case study of an AI-driven multi-tier LAWN with hybrid satellite, high-altitude, and ground nodes. The tutorial concludes by outlining key research challenges in architecture design, signal processing-AI co-design, safety and security, experimentation, and standardization, and by highlighting opportunities for LAWNs to evolve into dependable, AI-native infrastructure for the intelligent skies.
Reliable and energy-efficient trajectory planning is essential for ensuring stable multi-hop communication in Low-Altitude Wireless Networks (LAWNs) operating in dynamic three-dimensional environments. However, jointly optimizing Unmanned Aerial Vehicle (UAV) trajectories and routing paths is computationally intractable, since feasibility changes abruptly with obstacle mobility, link variations, and partial observations. To address the challenges, we propose Hierarchical Adaptive Reasoning and Planning for LAWN (HARP-LAWN), a hierarchical coordination framework that separates long-horizon global reasoning from short-horizon local feasibility control. The global layer uses a Large Language Model (LLM) to semantically interpret coarse network and obstacle information, generating coordination points and routing structures without solving the full optimization problem. The local execution layer refines these coarse plans through event-triggered, in-context replanning based on each UAV's partial observations using onboard LLMs. This enables real-time collision avoidance and link maintenance without the extensive policy retraining required by conventional learning-based methods. Simulation results show that HARP-LAWN provides robust zero-shot adaptability, outperforming reinforcement learning baselines in dynamic environments, and effectively scales to dense UAV deployments without exacerbating communication overhead. The analysis of local execution time and energy consumption further confirms that the hierarchical design enables efficient adaptation, with different LLM models exhibiting distinct reliability-latency-energy trade-offs.
Precise three-dimensional (3D) tracking of unmanned aerial vehicles (UAVs) trajectory is crucial for the emerging application of low-altitude economy. However, conventional tracking approaches that rely on the base station (BS) are fundamentally constrained by the limited spatial resolution of small antenna arrays. In this paper, we establish a new framework supported by extremely large-scale reconfigurable intelligent surface (XL-RIS) and exploit the virtual RIS observation, i.e., the high-dimensional signal impinging on an XL-RIS, as a far richer sensing paradigm for near-field 3D UAV tracking. Due to the inherent passive-reflection property, the high-dimensional signal $\bm {y}_{r}$ of XL-RIS is not directly observable and must be reconstructed from the low-dimensional BS measurement $\bm {y}$, which appears severely ill-posed. We provide a compact information-theoretic analysis showing that the virtual RIS observation contains no less position information than the compressed BS observation, while VOR exploits offline training data as side information and is benchmarked by a Cramér-Rao-bound-based reconstruction limit. Grounded by this theory, we next develop a DenseNet-based Virtual RIS Observation Reconstruction (VOR) algorithm that learns the low-dimensional near-field channel manifold to recover $\bm {y}_{r}$ effectively. Then, the reconstructed observation is processed by the proposed Spatial-Spectral Feature Extractor (SSFE), which comprises a CNN for spatial encoding, a time-frequency analyser, and a near-field AoA extractor leveraging element-wise Fresnel-region phase variation. Finally, the extracted features are passed to the Bi-LSTM Trajectory Predictor (BiTP) for accurate temporal trajectory prediction. Extensive simulations with random-walk, wave, and spiral UAV trajectories show consistent MSE reductions over BS-only baselines under an estimated-AoA protocol with multiple RIS probing profiles. Pilot-aware large-aperture tests show the benefit of pairing larger RIS panels with sufficient probing observations.
