Task-Oriented Semantic Sensing in Low-Altitude Networks with Model-Driven Deep Learning

Abstract

Integrated sensing and communication (ISAC) is a key enabling technology for low-altitude wireless networks (LAWNs). However, conventional sensing methods usually rely on full environmental reconstruction, which leads to considerable computational and communication overhead. To address this challenge, semantic sensing (SemS) has recently emerged as a task-oriented sensing paradigm that directly extracts taskrelevant information from the environment for downstream decision-making. Existing SemS approaches mainly rely on endto-end learning, while such purely data-driven methods often provide limited physical interpretability and may become unstable in dynamic low-altitude environments. In this paper, we propose a model-driven deep learning approach for SemS in ISAC systems by integrating neural networks with classical radar signal processing principles. Specifically, we develop a joint optimization architecture for an OFDM-based ISAC system, in which the transmitter-side encoder adaptively optimizes pilot patterns via differentiable top-K selection, while the receiver-side unfolded semantic decoder sequentially integrates neural interpolation, delay-Doppler (DD)-domain mapping based on the twodimensional fast Fourier transform (2D-FFT), and lightweight classification. Numerical results demonstrate that the proposed method achieves faster convergence and higher task accuracy than conventional schemes.